Skip to content

🙌 OpenHands: AI-Driven Development

feat(verify-openhands): map 20 more behaviors, with live recipes (#18105)

Engel Nyst208993e6c195
diff --git a/.agents/skills/verify-openhands/references/feature-map/F01-first-run-and-sign-in.md b/.agents/skills/verify-openhands/references/feature-map/F01-first-run-and-sign-in.md
index ca440a93c7..998d2c4091 100644
--- a/.agents/skills/verify-openhands/references/feature-map/F01-first-run-and-sign-in.md
+++ b/.agents/skills/verify-openhands/references/feature-map/F01-first-run-and-sign-in.md
@@ -79,7 +79,7 @@ Public run A (`launch --new --public`), backend step walked:
Public run B (`launch --new --public`), onboarding skipped:
-- **Skip to API-key screen (`F01.onboarding-skip`, `F01.api-key-entry`).** Run `control-openhands browser goto /`, `control-openhands browser click 'testid=onboarding-skip'` and `control-openhands browser wait 'testid=api-key-entry-screen'`. After `sleep 12`, `control-openhands browser toasts --history` is `[]`: the screen raises no error toast. Run `control-openhands browser reload`, `control-openhands browser wait 'testid=api-key-entry-screen'`, `sleep 12` and `control-openhands browser toasts --history` again: still `[]` when the API-key screen is the first paint. Run `control-openhands browser value 'testid=api-key-entry-host'` (the origin; the field is disabled) and `control-openhands browser enabled 'testid=api-key-entry-submit'` (`false`). Fill only the key with `control-openhands browser fill 'testid=api-key-entry-api-key' qa-wrong-key`; `enabled` is still `false` because Host Name is required. Run `control-openhands browser fill 'testid=api-key-entry-name' 'QA Public'` (`enabled` turns `true`), `control-openhands browser click 'testid=api-key-entry-submit'` and `control-openhands browser text 'testid=api-key-entry-status'`. It reads `Invalid API key. Please check the key and try again.`, the screen stays, and `control-openhands browser screenshot --feature F01.api-key-entry --name wrong-key` shows the red line above Connect. While the screen is still up, run `control-openhands browser viewport phone` and `control-openhands browser bbox 'testid=api-key-entry-form'`. It gives `insideViewport` `true` and `pageHorizontalOverflow` `false`. Return with `control-openhands browser viewport desktop`.
+- **Skip to API-key screen (`F01.onboarding-skip`, `F01.api-key-entry`).** Run `control-openhands browser goto /`, `control-openhands browser click 'testid=onboarding-skip'` and `control-openhands browser wait 'testid=api-key-entry-screen'`. After `sleep 12`, `control-openhands browser toasts --history` is `[]`: the screen raises no error toast. Run `control-openhands browser reload`, `control-openhands browser wait 'testid=api-key-entry-screen'`, `sleep 12` and `control-openhands browser toasts --history` again: still `[]` when the API-key screen is the first paint (the toast these checks guard against reads `No backend is configured.`; `control-openhands browser count 'text=No backend is configured'` is `0`). Run `control-openhands browser value 'testid=api-key-entry-host'` (the origin; the field is disabled) and `control-openhands browser enabled 'testid=api-key-entry-submit'` (`false`). Fill only the key with `control-openhands browser fill 'testid=api-key-entry-api-key' qa-wrong-key`; `enabled` is still `false` because Host Name is required. Run `control-openhands browser fill 'testid=api-key-entry-name' 'QA Public'` (`enabled` turns `true`), `control-openhands browser click 'testid=api-key-entry-submit'` and `control-openhands browser text 'testid=api-key-entry-status'`. It reads `Invalid API key. Please check the key and try again.`, the screen stays, and `control-openhands browser screenshot --feature F01.api-key-entry --name wrong-key` shows the red line above Connect. While the screen is still up, run `control-openhands browser viewport phone` and `control-openhands browser bbox 'testid=api-key-entry-form'`. It gives `insideViewport` `true` and `pageHorizontalOverflow` `false`. Return with `control-openhands browser viewport desktop`.
- **Connect (`F01.api-key-entry`).** Run `control-openhands browser fill 'testid=api-key-entry-name' 'QA Public'`, `control-openhands browser fill 'testid=api-key-entry-api-key' --value-file "$OH_VERIFY_RUN/private/session-key"`, `control-openhands browser click 'testid=api-key-entry-submit'` and `control-openhands browser wait 'testid=api-key-entry-screen' --state detached`. `testid=root-layout` is shown, the consent form names `“QA Public”`, and `control-openhands browser snapshot 'testid=backend-selector'` shows `combobox "QA Public"`. After `control-openhands browser reload`, `testid=api-key-entry-screen` counts `0` and `testid=root-layout` counts `1`.
- **Stale key re-prompt (`F01.api-key-entry`).** Answer consent if it is open (`control-openhands browser click 'testid=confirm-telemetry-preferences'`). Run `control-openhands restart --rotate-key` (`rotatedKey` `true`), then `control-openhands browser reload` and `control-openhands browser wait 'testid=api-key-entry-screen'`; `testid=root-layout` counts `0`. Connect again as above with the new `$OH_VERIFY_RUN/private/session-key`. After a reload, `testid=root-layout` counts `1`, and `control-openhands browser eval "JSON.parse(localStorage.getItem('openhands-backends')).length"` is `1`, so no duplicate backend is added.
- **Close on the last step (`F01.onboarding-hello-close`).** Run `control-openhands browser reset` and `control-openhands browser goto /`. Walk to the last step: `control-openhands browser fill 'testid=onboarding-backend-api-key' --value-file "$OH_VERIFY_RUN/private/session-key"`, `control-openhands browser click 'testid=onboarding-backend-next'`, `control-openhands browser wait 'testid=onboarding-step-choose-agent'` (consent does not return: the backend already has an answer), `control-openhands browser click 'testid=onboarding-agent-next'` and `control-openhands browser wait '[data-testid=onboarding-modal][data-current-step="1"]'`. Then fill the key with `control-openhands browser fill 'testid=onboarding-step-setup-llm >> testid=llm-api-key-input' --value-env DEEPSEEK_API_KEY`, open `testid=sdk-section-advanced-toggle` and fill `llm-custom-model-input` with `deepseek/deepseek-flash` as in the LLM step. Run `control-openhands browser click 'testid=onboarding-llm-next'` and `control-openhands browser wait '[data-testid=onboarding-modal][data-current-step="2"]'`. Check `control-openhands conversation list` (`count` `0`), then run `control-openhands browser click 'testid=onboarding-hello-close'` and `control-openhands browser wait 'testid=onboarding-modal' --state detached`. `testid=root-layout` counts `1`, the URL is `/` and `localStorage.getItem('openhands-onboarded')` is `"1"`. `conversation list` still shows `count` `0`, and after `browser reload` the modal count is `0`.
diff --git a/.agents/skills/verify-openhands/references/feature-map/F03-home.md b/.agents/skills/verify-openhands/references/feature-map/F03-home.md
index f914a2966a..9ec1b7edac 100644
--- a/.agents/skills/verify-openhands/references/feature-map/F03-home.md
+++ b/.agents/skills/verify-openhands/references/feature-map/F03-home.md
@@ -29,7 +29,7 @@ Workspaces (local backend):
- `F03.open-workspace-dialog`: **Open Workspace** opens a dialog with a workspace combobox; Confirm stays disabled until a pick; confirming swaps the pill for a preview bar that reopens the dialog; X and Escape close it.
- `F03.workspace-dropdown-search`: typing filters workspaces; with more than one group, workspaces are grouped by parent with an **Other** group; no match (or no workspaces) shows `No workspaces yet`.
-- `F03.selection-after-reload`: after a reload the preview bar is gone, but the dialog preselects the last confirmed workspace.
+- `F03.selection-after-reload`: after a reload the preview bar is gone, but the dialog preselects the last workspace picked in its dropdown, confirmed or not: a pick closed with the X survives leaving for another page and coming back, and a reload.
- `F03.workspace-mode`: the preview bar's mode menu switches **Local Repo** / **New Worktree**; the choice persists across reloads.
- `F03.folder-browser`: **+ Add Workspace** opens **Add workspaces** at the server home with Favorites/Locations, Up and the current path; **Add this directory** adds and selects the folder, **Add all subdirectories** adds a parent whose children become workspaces, **Cancel** adds nothing.
- `F03.folder-browser-locations`: the browser's **Favorites** (Home, plus `workspace` once a no-folder conversation exists) and **Locations** (`/`) jump to that folder; Up is disabled at `/`.
@@ -101,7 +101,7 @@ Workspaces:
- **Open Workspace (`F03.open-workspace-dialog`).** The dialog is still open from the previous bullets, with `workspace` selected (its disabled Confirm before a pick is checked at the start of Folder browser). Run `control-openhands browser click 'testid=workspace-dropdown'`, `control-openhands browser click 'testid=workspace-dropdown-menu >> role=option[name="qa-f03-repo"]'`, `control-openhands browser enabled 'testid=workspace-launch-button'` (`true`), `control-openhands browser click 'testid=workspace-launch-button'`. The dialog closes and `control-openhands browser text 'testid=home-git-control-bar-preview'` is `qa-f03-repo` and `Local Repo` (`browser screenshot --feature F03.open-workspace-dialog --name preview-bar`). `control-openhands browser click 'testid=home-git-control-bar-preview >> role=button[name*="qa-f03-repo"]'` reopens the dialog with `browser value 'testid=workspace-dropdown'` `qa-f03-repo`; `control-openhands browser click 'testid=close-open-workspace-dialog'` closes it, and so does `control-openhands browser press Escape`.
- **Launch in a folder (`F03.launch-in-workspace`).** Open Workspace (above) left `qa-f03-repo` confirmed in the preview bar. Run `control-openhands browser fill 'testid=chat-input' 'Run pwd and reply with only its output.'` and `control-openhands browser click 'testid=home-chat-launcher >> testid=submit-button' --observe 'text=Creating conversation' --expect-url '/conversations/[^/]+$'`. `observed` shows `Creating conversation…` then `<absent>`; the returned URL holds `<id>`. `control-openhands conversation wait <id> --timeout 180` reports `workspace` `<run>/workspace/qa-f03-repo`, and `control-openhands conversation events <id> --kinds ObservationEvent,MessageEvent` shows the terminal observation and the reply `<run>/workspace/qa-f03-repo`.
- **Search, groups and empty (`F03.workspace-dropdown-search`).** Run `control-openhands browser goto /` (Launch in a folder left its conversation), `control-openhands browser click 'testid=open-workspace-button'` and `control-openhands browser click 'testid=workspace-dropdown'`. With the dropdown open, `control-openhands browser fill 'testid=workspace-dropdown' repo` and `control-openhands browser text 'testid=workspace-dropdown-menu'` is `qa-f03-repo`. With the parent and the standalone `workspace` saved, the menu text is `workspace`, `qa-f03-folder`, `qa-f03-repo`, `Other`, `workspace` and `control-openhands browser count 'testid=workspace-group-header'` is `2` (`--name grouped`). `control-openhands browser fill 'testid=workspace-dropdown' zzz-none` shows the empty state: `control-openhands browser text 'testid=workspace-dropdown-empty'` is `No workspaces yet` (`--name no-match`) and `control-openhands browser count 'testid=add-workspaces-button'` is still `1`. Clear with `control-openhands browser fill 'testid=workspace-dropdown' ''`; `browser count 'testid=workspace-dropdown-empty'` is `0` again.
-- **Reload (`F03.selection-after-reload`).** After confirming qa-f03-repo, `control-openhands browser reload`; `control-openhands browser count 'testid=home-git-control-bar-preview'` is `0` and the pill is back. Click it; `control-openhands browser value 'testid=workspace-dropdown'` is `qa-f03-repo` (sessionStorage `oh:home-selected-workspace-path`).
+- **Reload (`F03.selection-after-reload`).** After confirming qa-f03-repo, `control-openhands browser reload`; `control-openhands browser count 'testid=home-git-control-bar-preview'` is `0` and the pill is back. Click it; `control-openhands browser value 'testid=workspace-dropdown'` is `qa-f03-repo` (sessionStorage `oh:home-selected-workspace-path`). An unconfirmed pick persists the same way: `control-openhands browser click 'testid=workspace-dropdown'`, `control-openhands browser click 'testid=workspace-dropdown-menu >> role=option[name="qa-f03-folder"]'` (value `qa-f03-folder`), then `control-openhands browser click 'testid=close-open-workspace-dialog'` (never Confirm or Launch): the preview count stays `0` and `control-openhands browser storage --session --values` shows `oh:home-selected-workspace-path` ending in `/workspace/qa-f03-folder`. Leave and come back with `control-openhands browser goto /settings` (lands on `/settings/agents`, `browser count 'testid=home-screen'` `0`) and `control-openhands browser goto /conversations`, click `open-workspace-button`: `browser value 'testid=workspace-dropdown'` is `qa-f03-folder` (`browser screenshot --feature F03.selection-after-reload --name unconfirmed-pick-restored`). Close with the X, `control-openhands browser reload`, click the pill again: the value is still `qa-f03-folder` and `browser enabled 'testid=workspace-launch-button'` is `true`. Restore the state the next bullets start from with `browser click 'testid=workspace-dropdown'`, `control-openhands browser click 'testid=workspace-dropdown-menu >> role=option[name="qa-f03-repo"]'` (value `qa-f03-repo`, the storage value ends in `/workspace/qa-f03-repo`) and `control-openhands browser click 'testid=workspace-launch-button'`: the dialog closes and `control-openhands browser text 'testid=home-git-control-bar-preview'` is `qa-f03-repo` and `Local Repo` again.
- **Workspace mode (`F03.workspace-mode`).** With the preview bar showing, `control-openhands browser click 'testid=workspace-mode-selector'`, `control-openhands browser snapshot 'testid=workspace-mode-selector-menu'` (`menuitemradio "Local Repo" [checked]`, `menuitemradio "New Worktree"`), `control-openhands browser click 'testid=workspace-mode-selector-option-new_worktree'`. `browser storage --values` shows `openhands-last-local-workspace-mode` `new_worktree`; after `browser reload` and reconfirming the workspace, `browser text 'testid=home-git-control-bar-preview'` reads `qa-f03-repo` and `New Worktree`. Restore with `testid=workspace-mode-selector-option-local_repo`.
- **Manage (`F03.manage-workspaces`).** With the parent and `workspace` saved: `control-openhands browser click 'testid=workspace-dropdown'`, `control-openhands browser click 'testid=manage-workspaces-button'`; `control-openhands browser testids 'testid=manage-workspaces-modal'` shows `manage-workspaces-row-workspace`, `manage-workspaces-parent-row-workspace`, `manage-workspaces-child-qa-f03-folder` and `manage-workspaces-child-qa-f03-repo`. `control-openhands browser click 'testid=manage-workspaces-remove-workspace'`; `control-openhands browser text 'testid=confirmation-modal'` reads `Remove "workspace" from your saved workspaces?`. `control-openhands browser click 'testid=confirmation-modal >> testid=cancel-button'` keeps the row (`browser count 'testid=manage-workspaces-row-workspace'` is `1`); repeat Remove and `control-openhands browser click 'testid=confirmation-modal >> testid=confirm-button'`: the count is `0`. Removing a workspace clears the dialog's selection only when it was the selected one: with `qa-f03-repo` selected, `control-openhands browser click 'testid=manage-workspaces-done'` leaves `browser value 'testid=workspace-dropdown'` at `qa-f03-repo`. After `browser reload` the dropdown lists only the two children, and `control-openhands api GET /api/workspaces` shows `workspaces: []` plus the parent. Cleanup proves Remove parent: `control-openhands browser click 'testid=manage-workspaces-remove-parent-workspace'` reads `Remove "workspace" and its 2 child workspaces?`; Confirm leaves `browser text 'testid=manage-workspaces-list'` at `You don't have any workspaces yet.` and `api GET /api/workspaces` empty; because the selected `qa-f03-repo` belonged to that parent, after `browser click 'testid=manage-workspaces-done'` `browser value 'testid=workspace-dropdown'` is empty and `browser enabled 'testid=workspace-launch-button'` is `false`.
@@ -131,6 +131,8 @@ Phone and cleanup:
## Gotchas
- Once a workspace is confirmed, `open-workspace-button` is not rendered (the preview bar replaces it): a click on it waits the full 30 s. Reopen through `home-git-control-bar-preview`, or reload.
+- A pick in the workspace dropdown is stored in sessionStorage at once, before Confirm or Launch, and closing the dialog with the X keeps it: a bullet that needs the earlier workspace selected must pick it again instead of relying on the last confirmation.
+- `control-openhands workspace open` and `conversation start --workspace` save that folder as a standalone workspace (`api GET /api/workspaces` lists it under `workspaces`). When another family used them first on the same stack, Add all subdirectories leaves those folders out of the `workspace` parent and the dropdown is grouped: options read `workspace, qa-f03-repo` and `Other, qa-f18-repo`. `role=option[name="qa-f03-repo"]` still matches (substring), but the menu text and the `workspace-group-header` count `2` in the Search bullet assume F03's own fixtures only.
- Folder rows navigate on a single click; there is no path field. The browser starts at the server home, `<run>/private/home`, two levels below `<run>`. After any conversation without a workspace, `<run>/private/home/workspace` exists: it shows as a `workspace` row and, after the next page load, a `workspace` favorite (`folder-browser-sidebar-workspace`), which is not `<run>/workspace`.
- A workspace and its parent can share a name (`workspace`); their test ids differ only by prefix (`manage-workspaces-row-` vs `manage-workspaces-parent-row-`).
- `chat-input` is a contenteditable `div`: it stays editable when the composer is "disabled"; assert the Send button and the absence of a new conversation instead.
diff --git a/.agents/skills/verify-openhands/references/feature-map/F05-composer.md b/.agents/skills/verify-openhands/references/feature-map/F05-composer.md
index b767fb5da6..d7a7912d9f 100644
--- a/.agents/skills/verify-openhands/references/feature-map/F05-composer.md
+++ b/.agents/skills/verify-openhands/references/feature-map/F05-composer.md
@@ -19,9 +19,11 @@ Source: `src/components/features/chat/`, `src/components/features/controls/tools
- `F05.agent-profile-switch`: before a conversation starts, `+` > Switch agent profile lists agent profiles and links to Manage agent profiles; it is absent in a started conversation.
- `F05.attach-files`: picked files show as chips and images as thumbnails (with an Upload-as-file toggle), each removable; sent files land in the workspace and are named in the message.
- `F05.attach-size-limit`: a file over 3 MB, or a selection over 3 MB in total, is refused with an error toast.
+- `F05.image-only-send`: an attached image alone, with an empty field, enables Send; the sent user message shows the thumbnail and carries the picture as an embedded data URL.
- `F05.paste`: pasted clipboard text is inserted as plain text; pasted images and dropped files attach.
- `F05.dictation`: the mic button dictates into the composer; without a microphone it toasts a failure and resets.
- `F05.llm-profile-picker`: the model pill switches the running conversation's LLM profile; the next reply uses the new model.
+- `F05.profile-identity`: the pill names the profile the conversation was started with, also after a reload and when another profile uses the same model; changing the default profile in Settings changes the home pill, not an existing conversation's.
- `F05.overflow-menu`: at very narrow widths the model pill collapses into a "More input actions" menu whose Model submenu fits the viewport and switches the profile.
- `F05.context-window-meter`: the ring shows context use; its popover opens the Usage tab and compacts the context.
- `F05.stop-resume`: Stop pauses a running agent (status `Stopped`, Play button); Play resumes it.
@@ -40,7 +42,7 @@ Source: `src/components/features/chat/`, `src/components/features/controls/tools
- Home page composer (`/`) and the composer of any conversation (`/conversations/<id>`, reached from the sidebar conversation list or by URL).
- `+` (More actions, `chat-plus-button`) inside the composer opens the tools menu; its last item **Add Files and Images** opens the native file picker (`upload-image-input`).
-- Model pill (`chat-input-llm-profile`) next to `+`; at 320 px it moves into the vertical-dots **More input actions** button.
+- Model pill (`chat-input-llm-profile`) next to `+`; at 320 px it moves into the vertical-dots **More input actions** button. On the home page it names the default profile (the one Settings → LLM marks as default); in a conversation it names the profile that conversation runs on.
- Right side of the composer: status text with Stop/Play, context ring (`context-window-meter`), mic (`chat-dictation-button`), Send (`submit-button`).
- Typing `/` at the start of a word opens the slash menu (`slash-command-menu`); `/plan` and `/code` are the only mode switches on the local backend (the Code/Plan pill is Cloud only).
- Keyboard: Enter sends, Shift+Enter newline, Ctrl+Enter / Cmd+Enter builds a plan; Escape closes the slash menu.
@@ -49,7 +51,8 @@ Source: `src/components/features/chat/`, `src/components/features/controls/tools
Preconditions:
-- Baseline state (launched, doctored, `onboard --skip` done) and `control-openhands llm preset deepseek` (profiles `deepseek-flash` active and `deepseek-pro`). Every recipe except `F05.agent-profile-switch` and the home draft needs a model.
+- Baseline state (launched, doctored, `onboard --skip` done) and `control-openhands llm preset deepseek` (profiles `deepseek-flash` active and `deepseek-pro`). Every recipe except `F05.agent-profile-switch`, the home draft, `F05.image-only-send` and `F05.profile-identity` needs a model. Without a DeepSeek key, `QA_DUMMY_KEY=qa control-openhands llm set --profile qa-zeta --model openai/gpt-4o --api-key-env QA_DUMMY_KEY --no-validate` stands in for the active profile: conversations can still be created (the agent errors at once), which is all those two bullets need, and every bullet that reads a reply is `blocked`.
+- `F05.profile-identity` needs two profiles on one model: its bullet creates `qa-alpha` and `qa-zeta` with `llm set --no-validate` right before it and deletes them afterwards.
- A conversation to type into: `control-openhands conversation start --prompt "Reply with only: ok" --wait --timeout 240` prints `id` (`<id>` below) and leaves the browser on `/conversations/<id>`.
- Attachment fixtures: `control-openhands fixture image --name qa-image` and `control-openhands fixture file --name qa-note.txt --content "QA attachment 7731"` (both land in `$OH_VERIFY_RUN/evidence/_fixtures/`).
- Desktop viewport unless a bullet says otherwise. Microphone dictation is blocked in the sandbox browser (no microphone); `+` > Git tools needs a conversation on a git workspace (not driven here).
@@ -63,9 +66,11 @@ Preconditions:
- **Agent profile before start (`F05.agent-profile-switch`).** Run `control-openhands browser goto /`, `control-openhands browser click 'testid=chat-plus-button'`, `control-openhands browser hover 'testid=switch-agent-profile-button'` and `control-openhands browser text 'testid=agent-profile-submenu'`: `AVAILABLE PROFILES`, `default` (checked; `chat-input-agent-profile-option-default`), `Manage agent profiles`. `control-openhands browser click 'testid=agent-profile-submenu >> text=Manage agent profiles' --expect-url 'settings/agents'` lands on `/settings/agents`. In a started conversation (`control-openhands browser goto /conversations/<id>`) `control-openhands browser click 'testid=chat-plus-button'` then `control-openhands browser count 'testid=switch-agent-profile-button'` is `0`.
- **Attach a file and an image (`F05.attach-files`).** On `/conversations/<id>` run `control-openhands browser upload 'testid=upload-image-input' "$OH_VERIFY_RUN/evidence/_fixtures/qa-image.png" "$OH_VERIFY_RUN/evidence/_fixtures/qa-note.txt"` and `control-openhands browser screenshot 'testid=interactive-chat-box' --feature F05.attach-files --name attached`: a `qa-note.txt TXT` chip and a thumbnail. `control-openhands browser count 'testid=interactive-chat-box >> alt=qa-image.png'` is `1`. `control-openhands browser click 'role=button[name="Upload as file"]'` turns the image toggle into `Do not upload as file` (`control-openhands browser attr 'role=button[name="Do not upload as file"]' aria-pressed` is `true`); click it again to restore. The remove buttons are named: `control-openhands browser snapshot 'testid=interactive-chat-box'` lists `button "Remove qa-note.txt"` and `button "Remove image"`. Remove the image with `control-openhands browser click 'testid=interactive-chat-box >> role=button[name="Remove image"]'`; the `alt=qa-image.png` count is `0`. Then `control-openhands browser type 'testid=chat-input' 'Reply with only the content of the attached file.'`, `control-openhands browser press Enter --selector 'testid=chat-input'`, wait as in Type and send, and `control-openhands browser text 'testid=user-message >> nth=-1'`: it ends `NEW FILES ADDED: qa-note.txt`. `control-openhands conversation events <id> --kinds MessageEvent --last 3` shows the reply `QA attachment 7731`, and `qa-note.txt` sits in the conversation `workspace` printed by `control-openhands conversation status <id>`.
- **Size limit (`F05.attach-size-limit`).** Arrange `head -c 4194304 /dev/zero > "$OH_VERIFY_RUN/evidence/_fixtures/qa-big.bin"` (and `qa-two-a.bin`, `qa-two-b.bin` with `head -c 2097152`). Run `control-openhands browser upload 'testid=upload-image-input' "$OH_VERIFY_RUN/evidence/_fixtures/qa-big.bin"` and `control-openhands browser toasts`: `Error: Files exceeding 3MB are not allowed: qa-big.bin`. Uploading `qa-two-a.bin` and `qa-two-b.bin` together toasts `Error: Total file size would be 4.0MB, exceeding the 3MB limit. Please select fewer or smaller files.` `control-openhands browser count 'testid=interactive-chat-box >> text=qa-big.bin'` is `0`.
+- **Image alone (`F05.image-only-send`).** From the home composer: `control-openhands browser goto /`, `control-openhands browser enabled 'testid=submit-button'` (`false`, empty field), `control-openhands browser click 'testid=chat-plus-button'`, `control-openhands browser upload-via 'testid=add-files-and-images-button' "$OH_VERIFY_RUN/evidence/_fixtures/qa-image.png"`, then `control-openhands browser count 'testid=home-chat-launcher >> alt=qa-image.png'` (`1`) and `control-openhands browser enabled 'testid=submit-button'`: `true` while `control-openhands browser eval "document.querySelector('[data-testid=chat-input]').innerText"` is still empty (`control-openhands browser screenshot 'testid=home-chat-launcher' --feature F05.image-only-send --name image-alone-send-enabled`). Run `control-openhands browser click 'testid=submit-button' --expect-url '/conversations/'` and read `<image-id>` from `control-openhands browser url`. After `control-openhands browser wait 'testid=user-message'` and `control-openhands browser wait 'testid=chat-message-sending' --state detached --timeout 5000` (the pending bubble carries `testid=user-message` too, so wait for its `Sending...` to go, not only for a bubble; the wait returns `"state": "detached"` at once when the indicator has already gone, the usual case on a local stack), `control-openhands browser count 'testid=user-message >> testid=image-preview'` is `1`, and `control-openhands browser count 'testid=chat-message-sending'` and `control-openhands browser count 'testid=chat-message-retry'` are `0`. Server side, `control-openhands conversation events <image-id> --kinds MessageEvent --grep 'data:image/png;base64'` matches the one user row (`count` `1`; its text is a blank and the excerpt shows the `"type":"image"` block with `image_urls`). Whether the model is handed the picture needs a key (blocked here: DEEPSEEK_API_KEY): with one, send the image together with `Which two colours are the bars in this image? Answer with the colour names only.` and expect `orange` and `blue` in the reply (`fixture image` draws those bars). Return with `control-openhands browser goto /conversations/<id>`.
- **Paste text (`F05.paste`).** Run `control-openhands browser type 'testid=chat-input' 'QA paste text'`, `control-openhands browser press Control+a`, `control-openhands browser press Control+c`, `control-openhands browser press Delete`, `control-openhands browser press Control+v`, then `control-openhands browser eval "document.querySelector('[data-testid=chat-input]').innerHTML"`: `QA paste text`, no markup. On a cleared field `control-openhands browser paste 'testid=chat-input' --file "$OH_VERIFY_RUN/evidence/_fixtures/qa-image.png"` attaches the image (`control-openhands browser count 'testid=interactive-chat-box >> alt=qa-image.png'` is `1`, with an `Upload as file` toggle) and `control-openhands browser paste 'testid=chat-input' --text 'QA pasted plain'` inserts plain text (`innerHTML` is `QA pasted plain`). `control-openhands browser drop-files 'testid=chat-input' "$OH_VERIFY_RUN/evidence/_fixtures/qa-note.txt" --stage over` makes `control-openhands browser text 'testid=interactive-chat-box'` start `Drop your files here`; the same command without `--stage` drops it: a `qa-note.txt` `TXT` chip appears. Remove the chip with `control-openhands browser click 'testid=interactive-chat-box >> role=button[name="Remove qa-note.txt"]'` and the image as in Attach; clear the field.
- **Dictation failure (`F05.dictation`).** Run `control-openhands browser click 'testid=chat-dictation-button' --observe '[role=status]' --observe-ms 3000`: the observed toasts include `Dictation failed. Check microphone access and your voice input settings.` and `control-openhands browser attr 'testid=chat-dictation-button' aria-label` is back to `Start dictation`. Real dictation is blocked without a microphone.
- **Switch LLM profile (`F05.llm-profile-picker`).** Run `control-openhands browser click 'testid=chat-input-llm-profile'` and `control-openhands browser text 'testid=chat-input-llm-profile-popover'`: `AVAILABLE PROFILES`, `deepseek-flash` / `deepseek/deepseek-flash`, `deepseek-pro` / `deepseek/deepseek-v4-pro`, `LLM Profiles`. Run `control-openhands browser click 'testid=chat-input-llm-profile-option-deepseek-pro'`, `control-openhands browser wait-text 'Switched to profile' --timeout 15000`, `control-openhands browser text 'testid=chat-input-llm-profile'` (`deepseek-pro`) and `control-openhands conversation status <id>` (`model` is `deepseek/deepseek-v4-pro`). Send `Reply with only: pro-ok` as in Type and send; the reply is `pro-ok`. After `control-openhands browser reload` the pill still reads `deepseek-pro`. Switch back with `control-openhands browser click 'testid=chat-input-llm-profile'` and `control-openhands browser click 'testid=chat-input-llm-profile-option-deepseek-flash'`.
+- **Profile identity (`F05.profile-identity`).** Arrange two profiles on one model right before this bullet: `QA_DUMMY_KEY=qa control-openhands llm set --profile qa-alpha --model openai/gpt-4o --api-key-env QA_DUMMY_KEY --no-validate --no-activate` and `QA_DUMMY_KEY=qa control-openhands llm set --profile qa-zeta --model openai/gpt-4o --api-key-env QA_DUMMY_KEY --no-validate` (`qa-zeta` becomes the default: `control-openhands api GET /api/profiles --pick active_profile` is `qa-zeta`). Run `control-openhands conversation start --prompt "Reply with only: identity-ok"` (no `--wait`; with the dummy key the agent errors at once, which does not matter here) and note its `id` as `<zeta-id>`. `control-openhands browser wait 'testid=chat-input-llm-profile'`, then `control-openhands browser text 'testid=chat-input-llm-profile'` is `qa-zeta`, not the alphabetically first profile on that model. `control-openhands browser reload`, the wait and the text again: still `qa-zeta`. The name is stamped on the conversation in this browser: `control-openhands browser eval "JSON.parse(localStorage.getItem('openhands-agent-server-conversation-metadata'))['<zeta-id>'].active_profile"` is `qa-zeta`. Change the default in Settings: `control-openhands browser goto /settings/llm`, `control-openhands browser wait 'testid=add-llm-profile'`, `control-openhands browser click '[data-testid=profile-row]:has([title="qa-alpha"]) >> testid=profile-menu-trigger'`, `control-openhands browser click 'testid=profile-actions-menu >> testid=profile-set-active'`, `control-openhands browser wait-text 'Switched to profile "qa-alpha"'` (`control-openhands api GET /api/profiles --pick active_profile` is `qa-alpha`). The home pill follows the default: `control-openhands browser goto /`, the wait, and `control-openhands browser text 'testid=chat-input-llm-profile'` is `qa-alpha`; the started conversation keeps its own: `control-openhands browser goto /conversations/<zeta-id>`, the wait, and the text is `qa-zeta` (`control-openhands browser screenshot 'testid=interactive-chat-box' --feature F05.profile-identity --name conversation-keeps-qa-zeta`). Restore: make `deepseek-flash` the default again the same way (its row's `profile-menu-trigger`, then `profile-set-active`), then `control-openhands api DELETE /api/profiles/qa-alpha --write` and `control-openhands api DELETE /api/profiles/qa-zeta --write` (arrange, not proof) so `/model` below still counts two profiles. In a model-free run make `qa-zeta` the default instead and delete only `qa-alpha`. Return with `control-openhands browser goto /conversations/<id>`.
- **Overflow menu (`F05.overflow-menu`).** Run `control-openhands browser viewport narrow` and `control-openhands browser snapshot 'testid=interactive-chat-box'`: the model pill is gone and `button "More input actions"` appears (at `phone` and `350x800` the pill stays inline). `control-openhands browser click 'role=button[name="More input actions"]' --observe '[data-testid=chat-input-overflow-menu]' --observe-ms 1000`: `observed` goes from `<absent>` to `Model` (the menu appears in the observed sequence; its delay varies from run to run, so do not assert a time) and has no later `<absent>`. `control-openhands browser count 'testid=chat-input-overflow-menu'` is `1` and `control-openhands browser attr 'role=button[name="More input actions"]' aria-expanded` is `true`. Run `control-openhands browser click 'testid=overflow-model-button'`; `control-openhands browser text 'testid=chat-input-overflow-menu'` reads `Model`, `AVAILABLE PROFILES`, `deepseek-flash` / `deepseek/deepseek-flash`, `deepseek-pro` / `deepseek/deepseek-v4-pro`, `LLM Profiles`. Run `control-openhands browser bbox 'testid=chat-input-llm-profile-option-deepseek-flash'` and `control-openhands browser bbox 'role=link[name="LLM Profiles"]'`. Expected: both `insideViewport` `true`. Known failure (#18063): the submenu opens to the right of the menu, so the profile row starts at x about 270 with width 210 and the `LLM Profiles` link sits at y about 725, below the 700 px viewport (both `insideViewport` `false`); the profile clicks below still work because the first two rows keep a strip of about 50 px on screen (from x about 270 to the 320 px edge), while a click on the `LLM Profiles` link times out (`browser screenshot --feature F05.overflow-menu --name model-submenu-narrow`). `control-openhands browser click 'testid=chat-input-llm-profile-option-deepseek-pro'` and `control-openhands browser wait-text 'Switched to profile' --timeout 15000` close the menu (count `0`), and `control-openhands conversation status <id>` shows `deepseek/deepseek-v4-pro`. Switch back the same way: the trigger, `testid=overflow-model-button`, then `testid=chat-input-llm-profile-option-deepseek-flash` (`conversation status <id>` shows `deepseek/deepseek-flash`). Open the menu once more; a second click on the trigger closes it (`aria-expanded` `false`). Return with `control-openhands browser viewport desktop`.
- **Context window (`F05.context-window-meter`).** Run `control-openhands browser attr 'testid=context-window-meter' aria-label` (`Context window usage: 1% used (99% left)`), `control-openhands browser click 'testid=context-window-meter'` and `control-openhands browser text 'testid=context-window-meter-popover'` (`Context Window`, the percentage, `Compact context`, tokens such as `11.7k / 1.0M`, `Usage`). `control-openhands browser click 'testid=context-window-meter-bar-button'` (the progress bar) closes the popover and opens the Usage tab (`control-openhands browser testids --filter usage` lists `conversation-tab-usage`, `usage-panel`). The popover's `Usage` row does the same: from another tab (`control-openhands browser click 'testid=conversation-tab-files'`), reopen the ring, `control-openhands browser click 'testid=context-window-plan-usage'` and `control-openhands browser wait 'testid=usage-panel'` (the panel mounts a moment later; an immediate `count` is `0`). Reopen the ring, then `control-openhands browser click 'testid=context-window-compact-button' --observe '[role=status]' --observe-ms 25000`: `Context compaction started`, then `Context compacted`; `control-openhands conversation events <id> --kinds CondensationRequest,Condensation --last 5` lists both kinds.
- **Stop and resume (`F05.stop-resume`) with a queued follow-up (`F05.queued-message`).** Send `Run this shell command: sleep 25 && echo slept. Then reply with only: done` (type + Enter as above), `control-openhands conversation wait <id> --until running --timeout 30` and `control-openhands browser wait 'testid=stop-button'`; `control-openhands browser attr 'testid=stop-button' aria-label` is `Stop`. While it runs, `control-openhands browser enabled 'testid=submit-button'` is `false` with an empty field; type `Also reply with: queued-ok` (enabled turns `true`) and press Enter. Run `control-openhands browser click 'testid=stop-button'` and `control-openhands browser wait 'testid=play-button' --timeout 30000`; `control-openhands browser text 'testid=interactive-chat-box'` ends `Stopped`, `control-openhands browser attr 'testid=play-button' aria-label` is `Resume the agent task` and `control-openhands conversation status <id>` is `paused`. Run `control-openhands browser click 'testid=play-button'`, wait for `running` then the terminal state: the last agent message is `done queued-ok` and `testid=user-message >> has-text=queued-ok` counts `1`.
@@ -90,6 +95,9 @@ Preconditions:
- After a client-side switch between conversations (a sidebar row, or a user-message branch in `F06.branch-from-here`) Send keeps the previous conversation's state (repro candidate, not filed yet). Switching from an empty composer to one with a saved draft restores the text next to a disabled Send; this is a race (Send came up enabled in 1 of 7 tries), so one enabled reading does not show the bug is gone. Switching from a composer with text to an empty one leaves Send enabled, and clicking it does nothing. Typing or `browser reload` fixes it, and Enter in the field sends the restored text anyway. Read `browser enabled 'testid=submit-button'` after a reload, or after typing.
- Uploaded chips and thumbnails have no test ids. Their remove buttons are named `Remove <file name>` (file chip) and `Remove image` (thumbnail), and Send, Stop and Play read `Send message`, `Stop` and `Resume the agent task` (#18033); select by those names, or by text/alt.
- `browser upload` feeds the hidden input directly. To drive the real path, `+` > Add Files and Images, use `control-openhands browser upload-via 'testid=add-files-and-images-button' <file>`, which answers the native file chooser.
+- The home composer has no `interactive-chat-box` wrapper: scope its chips and thumbnails with `testid=home-chat-launcher`.
+- `conversation events` rows show only a message's text, so an image-only message reads as a blank. Prove the attachment with `--grep 'data:image/png;base64'`, which searches whole events and quotes the `image` content block in `excerpts`.
+- The pill resolves the conversation's stamped profile before the user-level default, so profiles that share a model are told apart by name. The stamp lives in this browser's localStorage (`openhands-agent-server-conversation-metadata`), not on the server (`api GET /api/conversations/<id>` has no profile name): read it with `browser eval`, not `api GET`.
- The `Switched to profile` notes are client-side and gone after reload; assert the pill and `conversation status`.
- `/plan` messages go to a separate planner conversation; the browser stays on the parent `/conversations/<id>`, so wait on the planner id from `conversation list`.
- Only the newest goal status offers `Resume` (#18036): once a resumed goal has posted a newer status, the old `interrupted` row shows no button, also after a reload. Count `testid=goal-resume` after the resumed goal ends; while it runs the count is `0` anyway.
diff --git a/.agents/skills/verify-openhands/references/feature-map/F06-agent-activity.md b/.agents/skills/verify-openhands/references/feature-map/F06-agent-activity.md
index 3a0b913db3..cec91e7354 100644
--- a/.agents/skills/verify-openhands/references/feature-map/F06-agent-activity.md
+++ b/.agents/skills/verify-openhands/references/feature-map/F06-agent-activity.md
@@ -20,6 +20,8 @@ Source: `src/components/features/chat/chat-interface.tsx`, `src/components/conve
- `F06.code-block-copy`: hovering a fenced code block in an agent reply shows its own Copy button, which copies only the code.
- `F06.timestamps`: hovering a bubble or tool title shows the event's local date and time.
- `F06.pending-messages`: a sent message shows "Sending..." until the server echoes it; a failed send shows Failed to send with Retry and Dismiss; Retry delivers it exactly once.
+- `F06.clock-skew-send`: a message sent while the browser clock runs ahead of the server is still confirmed by the server echo: no lingering "Sending..." and no Retry.
+- `F06.failed-send-persists`: a failed send keeps its Retry row after the backend returns and after leaving and reopening the conversation from the sidebar; sending the same text again succeeds without clearing it, and Retry then delivers that attempt exactly once.
- `F06.markdown-rendering`: agent replies render markdown: tables inside a horizontal scroller with edge fades, fenced code, inline code.
- `F06.workspace-path-links`: a workspace file path in an agent reply, and the path chip of a file-editor card, open that file in the Files tab.
- `F06.thinking`: model reasoning appears as a collapsed "Thinking" row that expands.
@@ -28,7 +30,7 @@ Source: `src/components/features/chat/chat-interface.tsx`, `src/components/conve
- `F06.markdown-file-preview`: when the agent creates a Markdown file, its row stays expanded and ungrouped and shows a rendered, height-limited preview card with the file name and View, which opens the file in the Files tab.
- `F06.task-list`: task-tracker calls render a "Tasks" card with done / todo icons.
- `F06.events-match`: the stream shows one bubble per message event and one row per action, with no duplicates (compare with `conversation events`).
-- `F06.load-older-history`: only the newest 50 events load first; older pages are fetched (automatically when the content is short) so the first prompt and all actions appear.
+- `F06.load-older-history`: only the newest 50 events load first; the previous 50 are fetched when the user scrolls the transcript to its top (a "Fetching older messages…" row shows until they arrive) and automatically when the content is too short to scroll, so the first prompt and all actions appear.
- `F06.scroll-to-bottom`: scrolling up shows a scroll-to-bottom button; clicking it returns to the bottom and hides it.
- `F06.live-activity`: while the agent runs, a chip above the composer names the current action ("Thinking", then the action title) and the composer status reads Running.
- `F06.status-indicator`: while a new conversation starts, a "Starting" pill shows above the composer.
@@ -43,14 +45,14 @@ Source: `src/components/features/chat/chat-interface.tsx`, `src/components/conve
- `F06.image-attachments`: images sent with a message show as thumbnails in the bubble; expand opens a lightbox that closes with X or Escape.
- `F06.skill-install-banner`: after the agent installs a skill into the workspace, a banner offers "Start new conversation with this skill" (same workspace) and Close.
- `F06.llm-not-configured-banner`: with no usable LLM, a conversation shows "Your LLM isn't set up yet..." with Set up LLM; suggestions are hidden and sending is disabled.
-- `F06.critic-result`: with the critic enabled, agent messages carry a critic result (success likelihood, issue categories).
+- `F06.critic-result`: with the critic enabled, agent messages and the finishing action each carry a "Critic: agent success likelihood" block (stars labelled `Score: <N.N>%`, one decimal, with the percentage beside them); Expand details lists the issue categories (Potential Issues, Infrastructure, Likely Follow-up).
- `F06.hook-events`: hook executions render as their own rows in the stream.
- `F06.corrective-nudge`: when the model answers with neither a message nor a tool call, the SDK's nudge (`Your last response did not include a function call or a message. …`) shows as a muted, italic `role=note` line with an info icon (`corrective-nudge-message`), not as a user bubble (#17864).
- `F06.phone`: at 390 px the transcript fits the viewport; wide tables scroll inside their own container.
## How to get to it (user POV)
-- Open any conversation: a sidebar conversation card, or the direct URL `/conversations/<id>`.
+- Open any conversation: a sidebar conversation card, or the direct URL `/conversations/<id>`. Leaving one for the home page: the sidebar's **New Chat** link; coming back: its card again.
- Start one from the home composer (**New Chat**, `/`): type and press Send. `control-openhands conversation start` drives exactly this path.
- An empty conversation: the sidebar's create-thread button (`conversation-panel-new-thread-picker`), then **No workspace** (or a workspace).
- Composer status control (right side of the composer): Stop while running, Play while stopped.
@@ -64,7 +66,8 @@ Preconditions:
- Baseline state (launched, doctored, `onboard --skip` done) and `control-openhands llm preset deepseek` (deepseek-flash active). Every prompt below is tiny and stays in the conversation's own workspace.
- Conversation ids are printed as `"id"` by `control-openhands conversation start`, or read from `control-openhands browser url` (`/conversations/<id>`). `<id>` below is always the conversation the step just created or opened.
- `F06.llm-not-configured-banner` needs a second, fresh `control-openhands launch --new --build never` with no `llm preset` (export its run dir as `OH_VERIFY_RUN` for those commands, then `control-openhands stop` it).
-- `F06.critic-result` is blocked: it needs an OpenHands Cloud critic API key (Settings → Verification → Critic API Key). `F06.hook-events` needs a configured hook; none is set in a fresh run. `F06.corrective-nudge` is blocked: the nudge needs a model turn with neither text nor a tool call, which a normal prompt to deepseek-flash does not produce. When one happens, `control-openhands browser count 'testid=corrective-nudge-message'` is `1` and `browser count 'testid=user-message >> has-text=Your last response did not include'` is `0`.
+- `F06.critic-result` is blocked: it needs a model and an OpenHands Cloud critic API key (Settings → Verification → Critic API Key); its bullet says what a scored run shows. The automatic-fetch half of `F06.load-older-history` (its 16-command conversation) needs a model; the scroll-to-top half is driven without one, on the dummy profile below, with thirty refused user messages. `F06.hook-events` needs a configured hook; none is set in a fresh run.
+- `F06.clock-skew-send` and `F06.failed-send-persists` need no model reply: a conversation that exists is enough (without a key the agent errors at once, and the user rows still land). Without a key, create the family's conversations with a dummy profile (`QA_DUMMY_KEY=qa control-openhands llm set --profile qa-zeta --model openai/gpt-4o --api-key-env QA_DUMMY_KEY --no-validate`) and `control-openhands conversation start --prompt ...` without `--wait`. `F06.corrective-nudge` is blocked: the nudge needs a model turn with neither text nor a tool call, which a normal prompt to deepseek-flash does not produce. When one happens, `control-openhands browser count 'testid=corrective-nudge-message'` is `1` and `browser count 'testid=user-message >> has-text=Your last response did not include'` is `0`.
- The harness browser grants clipboard access: `control-openhands browser clipboard` reads what a Copy button wrote and `browser clipboard --write qa-empty` resets it first.
- A terminal call can come back as "The terminal session was reset..." after this run's agent server restarts; the model reruns it, which turns one action into a group of two (see Gotchas).
@@ -78,12 +81,14 @@ Preconditions:
- **Task list (`F06.task-list`, `F06.status-indicator`).** From `/` run `control-openhands browser goto /`, `control-openhands browser fill 'testid=chat-input' "Use your task_tracker tool to plan exactly two tasks titled 'qa step one' and 'qa step two'; mark 'qa step one' done and 'qa step two' todo. Do nothing else and reply DONE."` and `control-openhands browser click 'testid=submit-button' --expect-url '/conversations/' --observe '[data-testid=chat-status-indicator]' --observe-ms 8000`. `observed` contains `Starting` for a moment (the pill). Then `control-openhands conversation wait <id> --timeout 150` and `control-openhands browser text 'testid=chat-scroll-container'`: it contains `Tasks`, `qa step one`, `qa step two`; the screenshot shows a check icon and muted text for the done task, an empty circle for the todo one.
- **Empty state (`F06.empty-state-suggestions`).** Run `control-openhands browser click 'testid=conversation-panel-new-thread-picker'`, `control-openhands browser click 'testid=launch-no-workspace' --expect-url '/conversations/'`, then `control-openhands browser text 'testid=chat-suggestions'`: `Let's start building!` plus `Increase test coverage`, `Auto-merge PRs`, `Fix README`, `Clean dependencies`. Run `control-openhands browser click 'testid=chat-suggestions >> text=Fix README'` and `control-openhands browser text 'testid=chat-input'`: the composer holds the README-improvement prompt.
- **Pending bubble and long messages (`F06.pending-messages`, `F06.user-and-agent-messages`).** Write a 25-line message to a file (first line `Reply with only the word OK. Ignore the filler lines below.`, then `filler line 1`…`filler line 24`), run `control-openhands browser fill 'testid=chat-input' x --value-file <file>` and `control-openhands browser click 'testid=submit-button' --observe '[data-testid=chat-message-sending]' --observe-ms 3000`. `observed` shows `Sending...` then `<absent>` once the server echoes it. After `control-openhands conversation wait <id> --timeout 120`, `control-openhands browser testids 'testid=chat-scroll-container'` lists `chat-message-truncation-gradient` and `chat-message-expand` ("View More"); `control-openhands browser click 'testid=chat-message-expand'` removes both and the whole text shows. `testid=agent-message` reads `OK`.
+- **Browser clock ahead (`F06.clock-skew-send`).** On the same conversation, skew the page's clock before loading it: `control-openhands browser clock --offset-ms 300000`, `control-openhands browser goto /conversations/<id>`, `control-openhands browser wait 'testid=chat-input'`; `control-openhands browser eval "new Date().toISOString()"` is about five minutes ahead of `date -u` (read-only shell check). Run `control-openhands browser fill 'testid=chat-input' 'Reply with only: clock-ok'` and `control-openhands browser click 'testid=submit-button' --observe '[data-testid=chat-message-sending],[data-testid=chat-message-error]' --observe-ms 6000`: `observed` shows `Sending...` and then `<absent>` (the server echo confirmed it), never `Failed to send`. `control-openhands browser count 'testid=chat-message-sending'` and `control-openhands browser count 'testid=chat-message-retry'` are `0`, and `control-openhands conversation events <id> --kinds MessageEvent` gains one user row `Reply with only: clock-ok` whose `ts` is the server's time (close to `date -u`), not the browser's. Undo the skew before moving on: `control-openhands browser clock --offset-ms 0` and `control-openhands browser reload` (with a model, `control-openhands conversation wait <id> --fresh --timeout 120` first).
- **Server lost: banner, failed send, reconnect (`F06.error-banner`, `F06.pending-messages`, `F06.reconnect`).** With a conversation open run `control-openhands service stop agent-server` and `control-openhands browser wait 'testid=error-message-banner' --timeout 20000`: the banner reads `Unable to connect to server` with `error-message-banner-retry`, `-copy` and `-dismiss`; the composer shows `Disconnected`. Run `control-openhands browser fill 'testid=chat-input' 'Reply with only the word PONG.'` and `control-openhands browser click 'testid=submit-button' --observe '[data-testid=chat-message-sending],[data-testid=chat-message-error]' --observe-ms 8000`: `Sending...` turns into `Failed to send Retry Dismiss`. `control-openhands browser click 'testid=chat-message-retry'` while down fails again; `control-openhands browser click 'testid=chat-message-dismiss'` removes the bubble (`count 'testid=chat-message-error'` is `0`). Send the same message again so one failed bubble remains, `control-openhands browser click 'testid=error-message-banner-dismiss'` (banner count `0`), then `control-openhands restart`. Poll `control-openhands browser text 'testid=interactive-chat-box'` every 10 s: within about a minute after `restart` returns it no longer contains `Disconnected` (reconnected without reload). `control-openhands browser click 'testid=chat-message-retry'`, `control-openhands conversation wait <id> --timeout 120`, then `control-openhands conversation events <id> --kinds MessageEvent`: `Reply with only the word PONG.` appears once and the last `agent-message` is `PONG`.
- **Banner Retry (`F06.error-banner`).** Run `control-openhands service stop agent-server`, wait for `testid=error-message-banner` as above, `control-openhands restart`, then `control-openhands browser click 'testid=error-message-banner-retry'`. `control-openhands browser count 'testid=error-message-banner'` is `0` and the composer is no longer Disconnected. Take `browser screenshot 'testid=chat-interface' --feature F06.error-banner --name disconnected` while the banner is up.
+- **Failed send survives leaving (`F06.failed-send-persists`).** With `/conversations/<id>` open and healthy, run `control-openhands service stop agent-server`, `control-openhands browser wait 'testid=error-message-banner' --timeout 20000`, `control-openhands browser fill 'testid=chat-input' 'Reply with only: retry-ok'` and `control-openhands browser click 'testid=submit-button' --observe '[data-testid=chat-message-sending],[data-testid=chat-message-error]' --observe-ms 8000`: `Sending...` turns into `Failed to send Retry Dismiss` and `control-openhands browser count 'testid=chat-message-retry'` is `1` (`control-openhands browser screenshot 'testid=chat-interface' --feature F06.failed-send-persists --name failed-before-leaving`). Run `control-openhands restart` and `control-openhands browser wait 'testid=error-message-banner' --state hidden --timeout 90000` (the page reconnects by itself; `testid=error-message-banner-retry` hurries it while the banner is still up): the Retry row is still there (count `1`). Leave through the sidebar and come back: `control-openhands browser click 'role=link[name="New Chat"]'`, `control-openhands browser wait 'testid=home-chat-launcher'`, `control-openhands browser click 'a[href*="/conversations/<id>"] >> nth=0' --expect-url '/conversations/<id>'`, `control-openhands browser wait 'testid=user-message'`; `control-openhands browser count 'testid=chat-message-retry'` is still `1` after the history reloaded (`control-openhands browser screenshot 'testid=chat-interface' --feature F06.failed-send-persists --name retry-after-reopen`). Send the same text again (fill `Reply with only: retry-ok`, click `testid=submit-button` with the same `--observe`): the new bubble goes `Sending...` and settles while `Failed to send Retry Dismiss` stays; `control-openhands browser count 'testid=chat-message-sending'` is `0`, `control-openhands browser count 'testid=chat-message-retry'` is `1`, and the rows of `control-openhands conversation events <id> --kinds MessageEvent --grep 'retry-ok'` whose `source` is `user` number exactly one, `Reply with only: retry-ok` (in a model-free run `count` is `1`; with a model the agent's `retry-ok` reply matches the grep too, so count the `user` rows, not `count`). Then `control-openhands browser click 'testid=chat-message-retry'` and `control-openhands browser wait 'testid=chat-message-retry' --state detached --timeout 15000`: `control-openhands browser count 'testid=chat-message-sending'` is `0` and the same `--grep` now lists exactly two `user` rows `Reply with only: retry-ok` (one per attempt, none duplicated). With a model, `control-openhands conversation wait <id> --fresh --timeout 120` before the next bullet.
- **Live chip, stop and resume (`F06.live-activity`, `F06.stop-resume`).** Run `control-openhands conversation start --prompt "Run exactly one terminal command: sleep 25 && echo qa-slept. Then reply with only its output."` (no `--wait`), `control-openhands browser wait 'testid=live-activity-chip' --timeout 30000` and `control-openhands browser text 'testid=live-activity-chip'`: first `Thinking`, then the action title (e.g. `Run sleep 25 then echo qa-slept`); `browser text 'testid=interactive-chat-box'` contains `Running`. Run `control-openhands browser click 'testid=stop-button'`: `control-openhands conversation status <id>` is `paused`, the composer reads `Stopped`, `count 'testid=play-button'` is `1` and the chip is gone. Stop interrupts the running `sleep` too: `control-openhands conversation events <id>` shows an `AgentErrorEvent` (`Tool call interrupted before completion. The conversation was paused.`) followed by an `InterruptEvent`, and `browser text 'testid=chat-scroll-container'` shows an `Agent error` row under the action title. Run `control-openhands browser click 'testid=play-button'`: status `running`, then `control-openhands conversation wait <id> --timeout 120` ends `finished` with a new agent `MessageEvent` (the model reports that the command was interrupted; `qa-slept` never prints).
- **Reload mid-run (`F06.reload-mid-run`).** Run `control-openhands conversation start --prompt "Run exactly one terminal command: sleep 15 && echo qa-reloaded. Then reply with only its output."`, wait for `testid=live-activity-chip`, then `control-openhands browser reload`. The composer still reads `Running` and the chip is back. After `control-openhands conversation wait <id> --timeout 120`: `testid=agent-message` is `qa-reloaded`, `count 'testid=user-message'` is `1`, `control-openhands browser eval "document.querySelectorAll('[data-testid=chat-scroll-container] [data-testid=generic-event-message-title]').length"` is `1`.
-- **Older history (`F06.load-older-history`).** Run `control-openhands conversation start --prompt "Run these 16 terminal commands strictly one per tool call, sequentially (never combine them): echo qa1, echo qa2, echo qa3, echo qa4, echo qa5, echo qa6, echo qa7, echo qa8, echo qa9, echo qa10, echo qa11, echo qa12, echo qa13, echo qa14, echo qa15, echo qa16. Then reply DONE." --wait --timeout 400`; `control-openhands conversation events <id> --last 500` reports a count above 50. Run `control-openhands browser network --clear`, `control-openhands browser reload`, then `control-openhands browser network --last 300`: two `GET /api/conversations/<id>/events/search` requests (newest page, then the older page). `control-openhands browser text 'testid=chat-scroll-container'` starts with the prompt and shows `16 actions completed` and `DONE`.
-- **Scroll to bottom (`F06.scroll-to-bottom`).** On the 16-command conversation, `control-openhands browser click 'testid=event-group-toggle'` makes the transcript taller than the viewport while the view stays where it was. Run `control-openhands browser scroll 'testid=chat-scroll-container' --by 2000` first (the button only reacts to scroll events, so a page that is already at the top never shows it): `count 'testid=scroll-to-bottom'` is `0`. Run `control-openhands browser scroll 'testid=chat-scroll-container' --by -2000`; the count is `1`. `control-openhands browser click 'testid=scroll-to-bottom'`, then the count is `0` and `control-openhands browser eval "(()=>{const e=document.querySelector('[data-testid=chat-scroll-container]');return Math.round(e.scrollHeight-e.scrollTop-e.clientHeight)})()"` is `0`.
+- **Older history (`F06.load-older-history`).** Run `control-openhands conversation start --prompt "Run these 16 terminal commands strictly one per tool call, sequentially (never combine them): echo qa1, echo qa2, echo qa3, echo qa4, echo qa5, echo qa6, echo qa7, echo qa8, echo qa9, echo qa10, echo qa11, echo qa12, echo qa13, echo qa14, echo qa15, echo qa16. Then reply DONE." --wait --timeout 400` (note `<history-id>`); `control-openhands conversation events <history-id> --last 500` reports a count above 50. Run `control-openhands browser network --clear`, `control-openhands browser reload`, then `control-openhands browser network --last 300`: two `GET /api/conversations/<history-id>/events/search` requests (newest page, then the older page). `control-openhands browser text 'testid=chat-scroll-container'` starts with the prompt and shows `16 actions completed` and `DONE`. The other trigger is the user scrolling to the top, driven without a model on the dummy profile from Preconditions (`qa-zeta` active; in a keyed run create and activate it with that same `llm set` command for this bullet, then re-run `control-openhands llm preset deepseek` and remove it with `control-openhands api DELETE /api/profiles/qa-zeta --write`, so later families still count two profiles; arrange, not proof). Run `control-openhands conversation start --prompt "qa older 1"` (no `--wait`; note `<older-id>`), then `for N in $(seq 2 30); do control-openhands conversation send <older-id> --prompt "qa older $N"; done`: every send is accepted while the agent sits in error, about 1.3 s apiece, and each refused message adds five events (a `MessageEvent`, three `ConversationStateUpdateEvent`s, a `ConversationErrorEvent`), so the conversation ends with about 150 events (152 here). `control-openhands conversation events <older-id> --last 500` reports `"total": 100` (the CLI's one-page cap, see Gotchas) and its first row is `qa older 11`. Run `control-openhands browser network --clear`, `control-openhands browser reload`, `control-openhands browser wait 'testid=chat-input'`, then `control-openhands browser network --filter 'events/search' --last 10`: two requests, `?limit=50&sort_order=TIMESTAMP_DESC` and one with `timestamp__lt=<ts of qa older 21>`, because the newest 50 events hold only ten bubbles, shorter than the viewport, so the automatic fetch ran once and then stopped. `control-openhands browser count 'testid=user-message'` is `20`, `control-openhands browser text 'testid=chat-scroll-container'` starts `qa older 11` and holds none of `qa older 1`…`qa older 10`, and `control-openhands browser bbox 'testid=chat-scroll-container'` has `scrollHeight` `1576` against `clientHeight` `814`: taller than the viewport, so the third page waits for the user. Run `control-openhands browser scroll 'testid=chat-scroll-container' --by -100000`: `control-openhands browser network --filter 'events/search' --last 5` gains a request with `timestamp__lt=<ts of qa older 11>`, `control-openhands browser count 'testid=user-message'` is `29` and the text starts `qa older 2` (that page held messages 2–10 plus the tail of the first message's events; not seen in two runs: if the first refusal ever logs six events instead of seven, the arithmetic gives count `30`, text `qa older 1` and a last page of only the `SystemPromptEvent`), and the view stays on `qa older 11`: `control-openhands browser eval "document.querySelector('[data-testid=chat-scroll-container]').scrollTop"` is `684`, the height of the prepended rows. Scroll to the top again the same way: a fourth request (`timestamp__lt=` the oldest loaded event), the count is `30`, the text starts `qa older 1` and `scrollTop` is `76` (`control-openhands browser screenshot 'testid=chat-interface' --feature F06.load-older-history --name scroll-top-loaded-qa-older-1`). A third scroll adds no request: the short page ended the paging. `control-openhands browser count 'testid=loading-older-events'` is `0` after each scroll; the `Fetching older messages…` row lasts about 10 ms on a local stack and `browser scroll` has no `--observe`, so the `timestamp__lt=` network rows are the proof. Return to the 16-command conversation for the next bullet: `control-openhands browser goto /conversations/<history-id>` and `control-openhands browser wait 'testid=chat-input'`.
+- **Scroll to bottom (`F06.scroll-to-bottom`).** On the 16-command conversation (`<history-id>`; in a model-free run this bullet is blocked with the automatic-fetch half, prerequisite `DEEPSEEK_API_KEY`), `control-openhands browser click 'testid=event-group-toggle'` makes the transcript taller than the viewport while the view stays where it was. Run `control-openhands browser scroll 'testid=chat-scroll-container' --by 2000` first (the button only reacts to scroll events, so a page that is already at the top never shows it): `count 'testid=scroll-to-bottom'` is `0`. Run `control-openhands browser scroll 'testid=chat-scroll-container' --by -2000`; the count is `1`. `control-openhands browser click 'testid=scroll-to-bottom'`, then the count is `0` and `control-openhands browser eval "(()=>{const e=document.querySelector('[data-testid=chat-scroll-container]');return Math.round(e.scrollHeight-e.scrollTop-e.clientHeight)})()"` is `0`.
- **Confirmation mode on (`F06.confirmation-mode`).** Arrange through the settings UI, in this order (the Security Analyzer field appears only once Confirmation Mode is on): `control-openhands browser goto /settings/verification`, `control-openhands browser click 'testid=verification-settings-screen >> text=Confirmation Mode'`, `control-openhands browser click 'testid=sdk-section-all-toggle'`, `control-openhands browser click 'testid=verification-settings-screen >> role=button[name="Show suggestions"]'`, `control-openhands browser click 'role=listbox >> role=option[name="None"]'`, `control-openhands browser value 'testid=verification-settings-screen >> role=combobox[name="Security Analyzer"]'` (`None`), `control-openhands browser click 'testid=verification-settings-screen >> testid=save-button'`; `control-openhands api GET /api/settings` shows `"confirmation_mode": true` and `"security_analyzer": "none"` (None means every action is confirmed; the default LLM analyzer only stops HIGH-risk actions, so `echo` would run unprompted). Run `control-openhands conversation start --prompt "Run the terminal command 'echo qa-confirm-ok' and reply with its output." --wait --timeout 180`: it returns `"status": "waiting_for_confirmation"`. `control-openhands browser text 'testid=chat-scroll-container'` ends `Do you want to continue with this action? Cancel ⇧⌘⌫ Continue ⌘↩`, `control-openhands browser tooltip 'testid=action-confirm-button'` is `Confirm the requested action`, and a lock icon sits above the composer (screenshot). Run `control-openhands browser click 'testid=action-reject-button'`: status `idle`, the buttons are gone and `conversation events <id> --last 6` ends with `UserRejectObservation`. Send `Run the terminal command 'echo qa-confirm-two' and reply with its output.` through `testid=chat-input` / `testid=submit-button`, `control-openhands browser wait 'testid=action-confirm-button' --timeout 20000`, `control-openhands browser click 'testid=action-confirm-button'`; about 15 s later the status is `finished` and the events contain the observation `qa-confirm-two`. Expected after a Cancel: the rejected row is shown as resolved and the group stops spinning (currently fails, see Gotchas).
- **Confirmation shortcuts (`F06.confirmation-shortcuts`).** The previous bullet resolved its action, so arrange a new one in the same conversation: send `Run the terminal command 'echo qa-confirm-three' and reply with its output.` through `testid=chat-input` / `testid=submit-button` and `control-openhands browser wait 'testid=action-confirm-button' --timeout 20000`. While it awaits confirmation, `control-openhands browser press Control+Enter` does nothing (`control-openhands conversation status <id>` stays `waiting_for_confirmation`); see Gotchas. `control-openhands browser press Meta+Enter` continues it: `control-openhands conversation wait <id> --fresh --timeout 120` ends `finished` and `control-openhands conversation events <id> --kinds ObservationEvent` contains `qa-confirm-three`. Restore afterwards: `browser goto /settings/verification`, click `testid=sdk-section-all-toggle`, `Show suggestions`, `role=listbox >> role=option[name="LLM"]`, then `testid=verification-settings-screen >> text=Confirmation Mode`, then `save-button`; `api GET /api/settings` shows `"confirmation_mode": false`, `"security_analyzer": "llm"`.
- **Branch from here (`F06.branch-from-here`).** Open a conversation with at least two exchanges (the OK/PONG one above): `control-openhands browser goto /conversations/<id>` with its id (the previous bullet left `/settings/verification`). Agent message: `control-openhands browser hover 'testid=agent-message >> nth=0'`, `control-openhands browser click 'testid=agent-message >> nth=0 >> role=button[name="Branch from here"]' --expect-url '/conversations/(?!<id>)'`; after `control-openhands browser wait 'testid=user-message'` (the history mounts a moment after the URL changes, so an immediate count reads `0`) the new conversation has `count 'testid=user-message'` `1`, the agent message `OK`, and `browser eval "document.title"` contains `(branch)`. User message (edit): back on the OK/PONG conversation (`control-openhands browser goto /conversations/<id>`; the branch above has one user message), `control-openhands browser hover 'testid=user-message >> nth=1'` and `control-openhands browser click 'testid=user-message >> nth=1 >> role=button[name="Branch from here"]' --observe '[data-testid=chat-input]' --observe-ms 4000`: `observed` shows the composer change from empty to `Reply with only the word PONG.` and stay. The new conversation (`<branch-id>` from `browser url`) omits that message, `control-openhands browser text 'testid=chat-input'` is `Reply with only the word PONG.`, and nothing was sent: `control-openhands conversation events <branch-id> --kinds MessageEvent` lists only the copied first exchange (count `2`: the long message that starts `Reply with only the word OK.`, then `OK`; no `PONG`). `control-openhands browser enabled 'testid=submit-button'` should be `true` with the prefilled text; today it is `false` until the text is edited or the page is reloaded (see Gotchas). Send it with `control-openhands browser press Enter --selector 'testid=chat-input'` (Enter sends whatever the Send button's state is) and `control-openhands conversation wait <branch-id> --fresh --timeout 120`: expected `finished` with a reply; today it ends `error` (Agent Server bug, see Gotchas).
@@ -91,13 +96,13 @@ Preconditions:
- **Error events (`F06.error-events`).** Any run that ends in a `ConversationErrorEvent` (today: any message sent in a branched conversation) leaves `conversation wait <id>` at `"status": "error"`; `control-openhands browser testids 'testid=chat-interface'` lists `error-message-banner` with `warning-message-banner-icon` and the error's text, and the composer status reads `Error`. Read the code with `control-openhands api GET "/api/conversations/<id>/events/search?limit=3&sort_order=TIMESTAMP_DESC"`.
- **Skill installed banner (`F06.skill-install-banner`).** The banner keys on the installer's success line in terminal output. Put this one line in a prompt file: `Run exactly this one terminal command, verbatim, then reply DONE: mkdir -p .agents/skills/qa-skill && printf '# qa-skill\n' > .agents/skills/qa-skill/SKILL.md && echo "✅ Successfully installed 'qa-skill' to $PWD/.agents/skills/qa-skill"`, then run `control-openhands conversation start --prompt "$(cat <prompt file>)" --wait --timeout 150`. `control-openhands browser text 'testid=skill-install-restart-banner'` reads `Installed to this workspace: qa-skill. Skills load when a conversation starts, so this conversation can't use them yet.` Run `control-openhands browser click 'testid=skill-install-restart-action' --expect-url '/conversations/(?!<id>)'`; `control-openhands api GET /api/conversations/<new id>` has the same `workspace.working_dir` as the source. Back on the source (`control-openhands browser goto /conversations/<id>`), `control-openhands browser click 'testid=skill-install-restart-dismiss'` hides it (count `0`); after `browser reload` it is back (dismissal is session-only by design).
- **No LLM (`F06.llm-not-configured-banner`).** On the fresh no-LLM run: `control-openhands onboard --skip`, `control-openhands browser click 'testid=conversation-panel-new-thread-picker'`, `control-openhands browser click 'testid=launch-no-workspace' --expect-url '/conversations/'`. `control-openhands browser text 'testid=home-llm-not-configured-banner'` reads `Your LLM isn't set up yet, so conversations won't run. Finish setup to get started. Set up LLM`; `count 'testid=chat-suggestions'` is `0`; the composer cannot take text (`control-openhands browser attr 'testid=chat-input' contenteditable` is `false`, so `browser fill` fails) and `control-openhands browser enabled 'testid=submit-button'` is `false`. `control-openhands browser click 'testid=home-llm-not-configured-action'` then `browser url` ends in `/settings/llm`.
-- **Critic (`F06.critic-result`, blocked).** Toggle the hidden switch through its label on `/settings/verification`: `control-openhands browser goto /settings/verification`, `control-openhands browser click 'label:has([data-testid="sdk-settings-verification.critic_enabled"])'`, save, `api GET /api/settings` shows `"critic_enabled": true`. Without a Critic API key a `Reply with only the word OK.` conversation renders no critic block. Turn it off again the same way.
+- **Critic (`F06.critic-result`, blocked).** Toggle the hidden switch through its label on `/settings/verification`: `control-openhands browser goto /settings/verification`, `control-openhands browser click 'label:has([data-testid="sdk-settings-verification.critic_enabled"])'`, save, `api GET /api/settings` shows `"critic_enabled": true`. Without a Critic API key a `Reply with only the word OK.` conversation renders no critic block. With a key and a model, run `control-openhands conversation start --prompt "Create qa_critic.txt containing hi, then finish." --wait --timeout 180`: every scored event carries a block, so `control-openhands browser count 'text=Critic: agent success likelihood'` is at least `1` and a finishing action (`finish` tool row) shows its own block next to the agent's message; the stars carry the score as their label, which an ARIA snapshot does not list: `control-openhands browser count '[aria-label^="Score: "]'` is at least `2` and `control-openhands browser attr '[aria-label^="Score: "] >> nth=0' aria-label` reads `Score: <N.N>%` with one decimal (`Score: 82.0%`), the same number as the `(82.0%)` text beside the stars. `control-openhands browser click 'role=button[name="Expand details"] >> nth=0'` opens the categories `Potential Issues:`, `Infrastructure:` and `Likely Follow-up:` with their named issues in `control-openhands browser text 'testid=chat-scroll-container'`, and the button's label flips to `Collapse details`. Blocked here: DEEPSEEK_API_KEY and a Cloud critic key. Turn it off again the same way.
- **Phone (`F06.phone`).** Run `control-openhands browser viewport phone`, open the table conversation (`control-openhands browser goto /conversations/<stream id>`), `control-openhands browser bbox 'testid=chat-interface'` (`insideViewport` `true`, `pageHorizontalOverflow` `false`) and `control-openhands browser bbox 'testid=markdown-table-scroll'` (`scrollWidth` larger than `clientWidth`: the table scrolls inside itself). `control-openhands browser screenshot --feature F06.phone --name conversation`, then `control-openhands browser viewport desktop`.
- After each group run `control-openhands browser errors --app-only`. Outages driven with `service stop` log connection-refused console errors by design; `pageErrors` must stay `0`.
## Gotchas
-- `control-openhands conversation events <id>` returns only the last 25 events unless you pass `--last N`; event texts are cut at about 160 characters.
+- `control-openhands conversation events <id>` returns only the last 25 events unless you pass `--last N`, and never more than the newest 100 whatever `N` is (one API page; the Agent Server rejects `limit` above 100), so on a long conversation `"total"` reads `100` and the oldest rows are missing: `--from-start` reads the oldest 100 instead, and `control-openhands api GET "/api/conversations/<id>/events/search?limit=100&sort_order=TIMESTAMP"` pages through all of them with `page_id=<next_page_id>`. Event texts are cut at about 160 characters.
- `--observe` takes a CSS selector (`'[data-testid=chat-message-sending]'`), not the ` >> ` syntax; a `testid=` value there fails with "not a valid selector".
- `role=button[name="Expand"]` also matches "Expand thinking" and "Expand actions"; use `[exact]` to target tool rows. Each click flips the row to Collapse, so `nth=0` walks to the next collapsed row.
- Tool rows show no success icon; the only indicators are the `exit N` badge for non-zero exits and a clock icon (`status-icon`) for timeouts. A single action is not grouped: it renders as a bare row without `event-group`.
@@ -113,6 +118,9 @@ Preconditions:
- The `ConversationErrorEvent` banner is not restored after navigating away and back (or a reload); read the error from the events API instead.
- After `restart`, reconnection takes about 20 s without a click; click `error-message-banner-retry` to reconnect at once. The error banner does not come back after a page reload once the stack is healthy.
- "Sending..." lasts about 100 to 650 ms on a local stack, too short to hover and click `chat-message-stop`; that stop path is not driven here. The `F06.pending-messages` failure path needs `service stop agent-server`.
+- `browser clock` installs a fake clock that stays with the page across reloads and client-side navigation. Install it before the `goto` whose page should see it, and put it back with `browser clock --offset-ms 0` plus `browser reload`; compare `browser eval "new Date().toISOString()"` with `date -u` when in doubt.
+- The pending-message queue lives in page memory, so a failed bubble survives sidebar navigation and a backend `restart` but not a `browser reload` or `browser goto` of the conversation (`src/stores/optimistic-user-message-store.ts`). Leave and return through the sidebar when proving `F06.failed-send-persists`. The sidebar card's title is the first message or model-written: click its link by id, `'a[href*="/conversations/<id>"] >> nth=0'`.
+- Both the pending bubble and the echoed message carry `testid=user-message`, so `count 'testid=user-message >> has-text=...'` is one too high while a failed bubble is still shown; count attempts with `conversation events <id> --kinds MessageEvent` (user rows: `"source": "user"`) and the bubbles with `chat-message-retry`.
- Local `Stop` calls the Agent Server's interrupt (Cloud pauses the sandbox instead, see `pauseConversation`): a command already running in the terminal is cut off, recorded as `AgentErrorEvent` + `InterruptEvent`, and after Play the model only reports the interruption.
- The no-LLM banner reuses the home banner's test ids (`home-llm-not-configured-banner`). It exists only in a run where no LLM was ever configured: use a separate `launch --new`, never remove the profile from the shared run.
- Read the clipboard with `control-openhands browser clipboard`, never `browser eval "navigator.clipboard.readText()"`. An agent bubble that contains a code block holds two `copy-to-clipboard` buttons (the code block's is hidden until hovered): scope the bubble's with `>> nth=0`.
diff --git a/.agents/skills/verify-openhands/references/feature-map/F09-settings-shell.md b/.agents/skills/verify-openhands/references/feature-map/F09-settings-shell.md
index 51ed1d22a8..9ac723e419 100644
--- a/.agents/skills/verify-openhands/references/feature-map/F09-settings-shell.md
+++ b/.agents/skills/verify-openhands/references/feature-map/F09-settings-shell.md
@@ -28,6 +28,7 @@ Source: `src/routes.ts`, `src/routes/settings.tsx`, `src/routes/settings-index.t
- `F09.landmarks`: the settings screen exposes a single `main` landmark, on desktop pages and on the phone hub.
- `F09.phone-hub`: below 1024 px (phones and tablets), `/settings` stays put and shows a "Settings" heading, the eight links, the version card and the synced note, with no page header and no horizontal overflow.
- `F09.hub-resize-redirect`: widening the window to ≥1024 px while the phone hub is open replaces `/settings` with `/settings/agents`; narrowing on a sub-page keeps its URL and adds the Back chevron.
+- `F09.breakpoint-sweep`: at 767, 768, 820, 1023, 1024 and 1440 px, `/settings` is the hub below 1024 px and redirects to the desktop nav from 1024 px up, and the Application page's `main` is at least 340 px wide with no horizontal overflow.
- `F09.phone-back`: at phone width each settings page shows a Back chevron (label "Settings") in the top bar that returns to the hub.
- `F09.synced-badge`: the nav footer reads "These settings are synced from Local backend (<origin>)" for the active backend.
- `F09.cloud-links`: on a Cloud backend only, the nav adds "Integrations" and "All Cloud Settings" external links; on a local backend neither exists.
@@ -69,6 +70,7 @@ Preconditions:
- **Landmarks (`F09.landmarks`).** Run `control-openhands browser goto /settings/agents`, `control-openhands browser count 'role=main'` (`1`) and `control-openhands browser eval "document.querySelector('[data-testid=settings-screen]').tagName"` (`DIV`: the shell wrapper is not a landmark; the layout's own `<main>` is). `control-openhands browser goto /settings/app` and the same `count` is `1`. At phone width (`control-openhands browser viewport phone`, `control-openhands browser goto /settings`) the hub also counts `1`; return with `control-openhands browser viewport desktop`.
- **Phone hub and Back (`F09.phone-hub`, `F09.phone-back`).** Run `control-openhands browser viewport phone`, `control-openhands browser goto /settings` (URL stays `/settings`), `control-openhands browser snapshot 'testid=settings-mobile-hub'` (heading "Settings", the eight links, the version button, the synced text), `control-openhands browser count 'testid=settings-page-subtitle'` (`0`), `control-openhands browser bbox 'testid=settings-mobile-hub'` (`pageHorizontalOverflow` `false`) and `control-openhands browser screenshot --feature F09.phone-hub --name hub`. Then `control-openhands browser click 'testid=settings-mobile-hub >> testid=sidebar-settings-/settings/condenser'`, `control-openhands browser wait-url '/settings/condenser$'`, `control-openhands browser attr 'testid=sidebar-mobile-back-button' aria-label` (`Settings`), `control-openhands browser click 'testid=sidebar-mobile-back-button'`, `control-openhands browser wait-url '/settings$'` and `control-openhands browser visible 'testid=settings-mobile-hub'` (`true`). Drawer entry: `control-openhands browser goto /`, `control-openhands browser click 'testid=sidebar-mobile-menu-toggle'`, `control-openhands browser click 'testid=sidebar-mobile-drawer >> testid=backend-selector-settings-link'`, `control-openhands browser wait-url '/settings$'`; `control-openhands browser visible 'testid=sidebar-mobile-drawer'` is `false` (the drawer closed). Return with `control-openhands browser viewport desktop`.
- **Resize from the hub (`F09.hub-resize-redirect`).** Run `control-openhands browser goto /`, `control-openhands browser viewport phone`, `control-openhands browser goto /settings` and `control-openhands browser visible 'testid=settings-mobile-hub'` (`true`). Then `control-openhands browser viewport desktop` and `control-openhands browser wait-url '/settings/agents$'`: the hub is gone (`count 'testid=settings-mobile-hub'` is `0`) and `control-openhands browser back` lands on `/` (the redirect replaced `/settings`). Reverse: `control-openhands browser goto /settings/condenser`, `control-openhands browser viewport phone`, `control-openhands browser url` (still `/settings/condenser`) and `control-openhands browser visible 'testid=sidebar-mobile-back-button'` (`true`). Tablets get the hub too, and the boundary is 1024 px: run `control-openhands browser viewport tablet` (820 px), `control-openhands browser goto /settings` (URL stays `/settings`) and `control-openhands browser visible 'testid=settings-mobile-hub'` (`true`); then `control-openhands browser viewport 1023x900`, `control-openhands browser goto /settings` and the same `visible` (`true`); then `control-openhands browser viewport 1024x900` and `control-openhands browser wait-url '/settings/agents$'` (`browser visible 'testid=settings-navbar-desktop'` is `true`). Return with `control-openhands browser viewport desktop`.
+- **Widths around the breakpoint (`F09.breakpoint-sweep`).** For each width `W` in `767`, `768`, `820`, `1023`, `1024`, `1440` run `control-openhands browser viewport Wx1024` and `control-openhands browser goto /settings`. Below 1024 px the URL stays `/settings`, `control-openhands browser visible 'testid=settings-mobile-hub'` is `true` and `control-openhands browser bbox 'testid=settings-mobile-hub'` has `pageHorizontalOverflow` `false` (725, 426, 478 and 681 px wide); from 1024 px up `control-openhands browser wait-url '/settings/agents$'` passes and `control-openhands browser visible 'testid=settings-navbar-desktop'` is `true`. Then `control-openhands browser goto /settings/app` and `control-openhands browser bbox 'role=main'`: `width` is at least 340 (767, 468, 520, 723, 424 and 840 today: full width at 767 px, `W − 300` beside the rail up to 1023 px, `W − 600` beside the rail and the settings nav from 1024 px), `scrollWidth` equals `clientWidth` and `pageHorizontalOverflow` is `false` (`insideViewport` is `false` only because the page is taller than 1024 px). Take `control-openhands browser screenshot --feature F09.breakpoint-sweep --name app-767` and `--name app-1024`. Finish with `control-openhands browser viewport desktop`.
- **Synced badge (`F09.synced-badge`).** Run `control-openhands browser text 'testid=backend-synced-settings-badge'` and `control-openhands browser eval 'location.origin'`. The text is `These settings are synced from Local backend (<origin>)` with the same origin.
- **Cloud links (`F09.cloud-links`).** On a local backend `control-openhands browser count 'testid=settings-integrations-link'` and `control-openhands browser count 'testid=settings-cloud-link'` are both `0`. Positive case (blocked without a Cloud account): with a Cloud backend active both exist; "All Cloud Settings" points at `<cloud host>/settings?org=<org id>` and opens in a new tab unless the canvas is locked to Cloud.
- **Update card and dialog (`F09.update-card`, `F09.update-modal`, `F09.update-up-to-date`).** On a settings page run `control-openhands browser text 'testid=agent-canvas-update-toggle'` (`Agent Canvas`, `Version`, the client version, plus the badge text when present), `control-openhands browser count 'testid=agent-canvas-update-badge'` (`0` while the npm check fails; `1` with "Up to date" or "Update" when it succeeds), then `control-openhands browser click 'testid=agent-canvas-update-toggle'`, `control-openhands browser wait 'testid=agent-canvas-update-modal'` and `control-openhands browser text 'testid=agent-canvas-update-status'`. With registry access and a current client the status is `You're running the latest version. What's new`, and `control-openhands browser attr 'testid=agent-canvas-update-release-notes' href` is `https://github.com/OpenHands/OpenHands/releases` (`target` `_blank`). Without registry access the status is `Couldn't check for updates. Try again later.` Re-check: run `control-openhands browser network --clear`, then `control-openhands browser click 'testid=agent-canvas-update-check-button' --observe 'testid=agent-canvas-update-modal' --observe-ms 2000` (the observed states include `Checking for updates…` before the final status), then `control-openhands browser network`: exactly one new `GET https://registry.npmjs.org/@openhands/agent-canvas/latest` proves the button re-checked (`browser errors` does not list successful requests). `control-openhands browser enabled 'testid=agent-canvas-update-check-button'` is `true` again. Close with `control-openhands browser click 'testid=close-agent-canvas-update-modal'` (modal count `0`); reopen and `control-openhands browser press Escape` also closes it.
diff --git a/.agents/skills/verify-openhands/references/feature-map/F10-llm-profiles.md b/.agents/skills/verify-openhands/references/feature-map/F10-llm-profiles.md
index 563a3eb47d..b695b4581d 100644
--- a/.agents/skills/verify-openhands/references/feature-map/F10-llm-profiles.md
+++ b/.agents/skills/verify-openhands/references/feature-map/F10-llm-profiles.md
@@ -33,7 +33,10 @@ Source: `src/routes/llm-settings.tsx`, `src/components/features/settings/llm-pro
- `F10.create-validation`: a Save whose validation call fails toasts the provider error and keeps the editor open with nothing saved.
- `F10.custom-model`: the Advanced tab takes a free-text model and Base URL; an empty model is refused with `Model is required`.
- `F10.model-picker`: the Basic provider and model comboboxes group options as Verified Models and Other Models; changing provider clears the model; the OpenHands provider shows account and key help.
+- `F10.provider-list-complete`: the Basic provider combobox offers every provider the backend reports or has verified models for (over a hundred, `xai` included); picking a late-sorting one fills the field and loads its models.
+- `F10.cloud-provider-pagination`: on a Cloud organization backend the provider list is fetched from `providers/search` page by page, so providers past the first page (`xai`, `openrouter`) are offered too (blocked here).
- `F10.edit`: Edit loads the profile (`Editing profile "<name>" - save to apply changes`) with the stored key masked; Save stays disabled until something changes; a changed name renames before saving; the stored key is kept.
+- `F10.basic-save-keeps-base-url`: re-saving a profile from the Basic view without changing its model keeps the Base URL that view hides.
- `F10.edit-default-reapply`: saving the Default profile re-applies it to the agent's LLM settings.
- `F10.editor-discard`: Back and Cancel return to the list without saving and without an unsaved-changes prompt.
- `F10.schema-validation`: bad values in schema fields (Temperature over 2, Top P over 1, invalid JSON) are refused with a toast naming the field.
@@ -62,12 +65,14 @@ Preconditions:
- No profile or connection named `QA_*`, `deepseek-chat` or `deepseek-v3` exists (`control-openhands api GET /api/profiles`, `control-openhands api GET /api/llm/provider-connections`).
- Select a row by exact name with the CSS form `'[data-testid=profile-row]:has([title="<name>"])'`; `has-text=QA_pro` also matches `QA_pro-copy`.
- Read toasts with `control-openhands browser wait-text '<text>'` (or `browser toasts`) right after the click; they disappear after a few seconds.
-- Blocked here: completing `F10.subscription` sign-in (needs a ChatGPT Plus/Pro account), `F10.cloud-readonly` (needs a Cloud org member account), `F10.schema-unavailable` (needs an agent server without the schema endpoints, for example an older `launch --sdk-ref`), and the `F10.list-states` error (needs `/api/profiles` to fail while the rest of the agent server answers).
+- Blocked here: completing `F10.subscription` sign-in (needs a ChatGPT Plus/Pro account), `F10.cloud-readonly` (needs a Cloud org member account), `F10.cloud-provider-pagination` (needs a Cloud organization backend where the user can manage profiles), `F10.schema-unavailable` (needs an agent server without the schema endpoints, for example an older `launch --sdk-ref`), and the `F10.list-states` error (needs `/api/profiles` to fail while the rest of the agent server answers).
- **Empty list (`F10.list-empty`).** On a fresh run, navigate from the sidebar: `control-openhands browser click 'testid=backend-selector-settings-link' --expect-url '/settings'`, `control-openhands browser click 'testid=sidebar-settings-/settings/llm' --expect-url '/settings/llm'`, then `control-openhands browser text 'testid=profiles-empty'`. The text is `No profiles saved yet. Add a profile to save your LLM configurations.` and `control-openhands browser testids main` lists `add-llm-profile`, `add-models-from-provider` and `profiles-empty`.
- **Open the editor (`F10.create`).** Run `control-openhands browser click 'testid=add-llm-profile'`. `control-openhands browser text 'testid=profile-editor-title'` is `Add LLM Profile`, `control-openhands browser text 'testid=profile-editor-description'` is `Configure your LLM settings below, then click Save to create this profile.`, and the form opens on the Basic tab with provider `OpenAI` and model `gpt-5.6-sol` (`browser value 'testid=llm-provider-input'`, `browser value 'testid=llm-model-input'`).
- **API key help link (`F10.api-key-help`).** In that editor run `control-openhands browser click 'testid=sdk-section-basic-toggle'` (a new editor opens on All once the agent's LLM settings carry a non-default option; see `F10.create-prefill`), then `control-openhands browser text 'testid=llm-api-key-help-anchor'` is `Don't know your API key? Click here for instructions`; its link has `target` `_blank` (`browser eval "document.querySelector('[data-testid=llm-api-key-help-anchor] a, a[data-testid=llm-api-key-help-anchor]').target"`). Run `control-openhands browser click 'testid=llm-api-key-help-anchor >> role=link'`, then `control-openhands browser tabs` (rerun once if it still lists one page): page 1 is `https://docs.openhands.dev/.../local-setup#getting-an-api-key`. Return with `control-openhands browser close-tab 1` and `control-openhands browser tab 0`. The Advanced tab shows the same link as `testid=llm-api-key-help-anchor-advanced`.
- **Provider and model lists (`F10.model-picker`).** Run `control-openhands browser click 'testid=llm-provider-input'` and `control-openhands browser snapshot 'role=listbox'`: groups `Verified Models` (OpenHands, Anthropic, OpenAI, Mistral AI, Gemini, deepseek, Moonshot, minimax, glm, nvidia, qwen) and `Other Models`. Pick DeepSeek with `control-openhands browser fill 'testid=llm-provider-input' deepseek` and `control-openhands browser click 'testid=provider-item-deepseek'`; `browser value 'testid=llm-model-input'` is now empty. `control-openhands browser click 'testid=llm-model-input'` and `browser snapshot 'role=listbox'` list `deepseek-chat`, `deepseek-v4-pro`, … as Verified and `deepseek-coder`, `deepseek-flash`, … under Other. For the OpenHands provider (`browser fill 'testid=llm-provider-input' openhands`, `browser click 'testid=provider-item-openhands'`), `control-openhands browser text 'testid=openhands-account-help'` is `Need an OpenHands Account? Click here`; after `control-openhands browser click 'role=option[name="claude-sonnet-5"][exact]'` in the model list, `control-openhands browser text 'testid=openhands-api-key-help'` starts `You can find your OpenHands Provider LLM Key in the OpenHands LLM Key section of OpenHands Cloud`. Screenshot: `browser screenshot --feature F10.model-picker --name openhands-provider`.
+- **Every provider is offered (`F10.provider-list-complete`).** Run `control-openhands browser click 'testid=cancel-profile-btn'` (the Provider and model lists bullet leaves its editor open; skip this when the list is showing), `control-openhands browser click 'testid=add-llm-profile'`, `control-openhands browser click 'testid=sdk-section-basic-toggle'` and `control-openhands browser click 'testid=llm-provider-input'`. `control-openhands browser count 'role=listbox >> role=option'` is the size of the union of the two backend lists the page merges: the names in `control-openhands api GET /api/llm/providers --pick providers` (under `value`; `control-openhands api GET /api/llm/providers --pick providers | jq '.value | length'` is `160` today) and the provider keys of `control-openhands api GET /api/llm/models/verified --max-bytes 200000` (its `body` is `{"models": {<provider>: [<model>, …]}}`; `control-openhands api GET /api/llm/models/verified --max-bytes 200000 | jq '.body.models | keys | length'` is `12` today, four of which, `glm`, `nvidia`, `openhands` and `qwen`, are missing from the first list). The union is `jq -s '([.[0].value[]] + (.[1].body.models | keys)) | unique | length' <(control-openhands api GET /api/llm/providers --pick providers) <(control-openhands api GET /api/llm/models/verified --max-bytes 200000)` (bash process substitution): `164` today, and the listbox count is the same `164`, with every backend name present under its display label (`openhands` as `OpenHands`, `mistral` as `Mistral AI`, `azure_ai` as `Azure AI Studio`). `control-openhands browser count 'role=option[name="xai"][exact]'` is `1` (it sorts last but three, before `xiaomi_mimo`, `xinference` and `zai`). Pick it with `control-openhands browser click 'role=option[name="xai"][exact]'`: `control-openhands browser value 'testid=llm-provider-input'` is `xai`, `control-openhands browser value 'testid=llm-model-input'` is empty, and after `control-openhands browser click 'testid=llm-model-input'` the listbox offers `grok-4.20`, `grok-4.20-0309`, … under Other Models (57 today; `control-openhands browser screenshot --feature F10.provider-list-complete --name xai-picked`). Close with `control-openhands browser press Escape` (the open listbox would swallow the next click) and `control-openhands browser click 'testid=cancel-profile-btn'`; the list is back.
+- **Cloud provider pagination (`F10.cloud-provider-pagination`).** Blocked: needs a Cloud organization backend where the user can manage profiles (local backends serve the whole list in one `GET /api/llm/providers`). There, the Basic provider list comes from `providers/search` one page at a time (`control-openhands browser network --filter providers/search` lists one request per page, the later ones with a `page_id`), and `xai` and `openrouter`, which sort past the first page, are offered and pick like any other provider. Record `control-openhands evidence add --feature F10.cloud-provider-pagination --result blocked ...` naming the Cloud account.
- **Create the first profile (`F10.create`, `F10.name-autofill`).** In a fresh editor pick deepseek as above, then `control-openhands browser click 'testid=llm-model-input'` and `control-openhands browser click 'testid=model-item-deepseek-flash'`. `control-openhands browser value 'testid=profile-name-input'` now follows the model: `deepseek-flash` (right after the provider pick, while the model is still empty, it still reads `gpt-5.6-sol`). Run `control-openhands browser fill 'testid=profile-name-input' QA_flash`, `control-openhands browser fill 'testid=llm-api-key-input' --value-file <key file>` (or `--value-env DEEPSEEK_API_KEY`), `control-openhands browser click 'testid=save-profile-btn' --observe 'testid=save-profile-btn' --observe-ms 4000` and `control-openhands browser wait-text 'Profile "QA_flash" created' --timeout 60000`. The observed button states are `Save`, `Validating...`, `Saving...`, then the editor closes. After `control-openhands browser reload`, `control-openhands browser text '[data-testid=profile-row]:has([title="QA_flash"])'` reads `QA_flash`, `deepseek/deepseek-flash`, `Default`, and `control-openhands api GET /api/profiles` has `active_profile` `QA_flash`.
- **A typed name sticks; a cleared one follows again (`F10.name-autofill`).** In a fresh editor (`control-openhands browser click 'testid=add-llm-profile'`) run `control-openhands browser fill 'testid=profile-name-input' QA_typed`, then pick deepseek as above, `control-openhands browser click 'testid=llm-model-input'` and `control-openhands browser click 'role=option[name="deepseek-v4-pro"][exact]'`; `browser value 'testid=profile-name-input'` stays `QA_typed`. Run `control-openhands browser fill 'testid=profile-name-input' ''`, then `browser click 'testid=llm-model-input'` and `browser click 'testid=model-item-deepseek-flash'`: the name is `deepseek-flash`; pick `role=option[name="deepseek-v4-pro"][exact]` again and it is `deepseek-v4-pro`.
- **Name rules (`F10.name-validation`).** In the editor run `control-openhands browser fill 'testid=profile-name-input' 'QA bad'`; `control-openhands browser enabled 'testid=save-profile-btn'` is `false` and `control-openhands browser attr 'testid=profile-name-input' aria-invalid` is `true` (the same for `-QA` and a 65-character name), and the rule text under the field turns red. An existing name (`QA_flash`) also gives `enabled` `false`, but `aria-invalid` stays `false` and nothing says why (see Gotchas). `QA_flash2` gives `true`.
@@ -80,6 +85,7 @@ Preconditions:
- **Rename (`F10.rename`).** Open the `QA_pro-copy-copy` row menu and click `testid=profile-actions-menu >> testid=profile-rename`. `control-openhands browser value 'testid=rename-profile-input'` is `QA_pro-copy-copy` and `browser eval "document.activeElement.dataset.testid"` is `rename-profile-input`. `control-openhands browser fill 'testid=rename-profile-input' '-bad'` makes `control-openhands browser enabled 'role=dialog[name="Rename Profile"] >> role=button[name="Rename"]'` `false` and `control-openhands browser attr 'testid=rename-profile-rule' class` contains `text-red-400`. Filling the old name and clicking Rename closes the modal with no toast (`browser count 'testid=rename-profile-modal'` is `0`); `control-openhands browser click 'role=dialog[name="Rename Profile"] >> role=button[name="Cancel"]'` also closes it. To rename, reopen, fill `QA_renamed`, `control-openhands browser press Enter`, `control-openhands browser wait-text 'Profile renamed to "QA_renamed"'`, reload, and the count of `'[data-testid=profile-row]:has([title="QA_renamed"])'` is `1` while the old name's is `0`. Renaming the Default row (`QA_pro` → `QA_pro2`, submit with `testid=rename-profile-submit`) keeps its badge (`browser text '[data-testid=profile-row]:has([data-testid=profile-active-badge])'` reads `QA_pro2 … Default`) and `active_profile` follows.
- **Rename onto an existing name (`F10.rename-conflict`).** The Rename step above already renamed `QA_pro-copy-copy` and `QA_pro`, so use the remaining fixtures: open the `QA_pro-copy` row menu, click `testid=profile-actions-menu >> testid=profile-rename`, fill `testid=rename-profile-input` with `QA_pro2` and click `testid=rename-profile-submit`. The modal stays open (`count 'testid=rename-profile-modal'` is `1`), `control-openhands browser toasts` has the server's message `Profile 'QA_pro2' already exists` (no `HTTP request failed` prefix, no JSON), and `browser screenshot --feature F10.rename-conflict --name toast` shows that toast above the open modal. Close it with `control-openhands browser click 'role=dialog[name="Rename Profile"] >> role=button[name="Cancel"]'`.
- **Edit (`F10.edit`).** Open the `QA_renamed` row menu and click `testid=profile-actions-menu >> testid=profile-edit`. `browser text 'testid=profile-editor-title'` is `Edit LLM Profile`, `browser text 'testid=profile-editor-description'` is `Editing profile "QA_renamed" - save to apply changes`, `browser attr 'testid=llm-api-key-input' placeholder` is `<hidden>`, `browser count 'testid=set-indicator'` is `1` and `browser enabled 'testid=save-profile-btn'` is `false` (screenshot `--feature F10.edit --name loaded`). Run `control-openhands browser click 'testid=llm-model-input'`, `control-openhands browser click 'testid=model-item-deepseek-flash'` (Save turns `true`), `control-openhands browser fill 'testid=profile-name-input' QA_edited`, Save, and `control-openhands browser wait-text 'Profile "QA_edited" updated' --timeout 60000`. After a reload `browser text '[data-testid=profile-row]:has([title="QA_edited"])'` reads `QA_edited deepseek/deepseek-flash` and `QA_renamed` is gone. The save passed validation without retyping the key, which proves the stored key was kept.
+- **Basic re-save keeps a hidden Base URL (`F10.basic-save-keeps-base-url`).** `QA_adv` carries `base_url` `https://api.deepseek.com` (Custom model above). Open its row menu, click `testid=profile-actions-menu >> testid=profile-edit`, then `control-openhands browser click 'testid=sdk-section-advanced-toggle'`: `control-openhands browser value 'testid=base-url-input'` is `https://api.deepseek.com`. Run `control-openhands browser click 'testid=sdk-section-basic-toggle'`: `control-openhands browser count 'testid=base-url-input'` is `0` (the view hides it), `control-openhands browser value 'testid=llm-model-input'` is `deepseek-flash` and `control-openhands browser enabled 'testid=save-profile-btn'` is `false` (`control-openhands browser screenshot --feature F10.basic-save-keeps-base-url --name basic-view`). Without touching the model, run `control-openhands browser fill 'testid=llm-api-key-input' --value-env DEEPSEEK_API_KEY` (Save turns `true`), `control-openhands browser click 'testid=save-profile-btn' --observe 'testid=save-profile-btn' --observe-ms 4000` (`Save`, `Validating...`, `Saving...`) and `control-openhands browser wait-text 'Profile "QA_adv" updated' --timeout 60000`. Then `control-openhands api GET /api/profiles/QA_adv --pick config.base_url` is still `https://api.deepseek.com` and `control-openhands api GET /api/profiles/QA_adv --pick config.model` is `deepseek/deepseek-flash`; after `control-openhands browser reload` the same GETs agree. Without a key, arrange `QA_DUMMY_KEY=qa control-openhands llm set --profile QA_base --model deepseek/deepseek-chat --api-key-env QA_DUMMY_KEY --base-url https://api.deepseek.com --no-validate --no-activate` (it prints `activated` `false` and `control-openhands api GET /api/profiles --pick active_profile` is unchanged: the Default stays where it is and the editors below inherit it, as long as some profile is Default; on a run where `QA_base` is the only profile the page makes it Default at the next load, see Gotchas) and `control-openhands browser reload`, then drive the same steps on `QA_base`'s row with its own readbacks: its editor opens on Advanced (the Base URL is a non-default value) with `control-openhands browser value 'testid=base-url-input'` `https://api.deepseek.com`, Basic hides it (`count` `0`), `control-openhands browser value 'testid=llm-model-input'` is `deepseek-chat` and `control-openhands browser enabled 'testid=save-profile-btn'` is `false`; fill the key with `QA_DUMMY_KEY=qa-dummy-key control-openhands browser fill 'testid=llm-api-key-input' --value-env QA_DUMMY_KEY` (Save turns `true`) and Save: the observed states are `Save`, `Validating...` (`[disabled] [busy]`), `Save` within a second, `control-openhands browser toasts` has the provider's message `litellm.BadRequestError: DeepseekException - {"error":{"message":"Authentication Fails, Your api key: ****-key is invalid …`, `browser text 'testid=profile-editor-title'` is still `Edit LLM Profile`, `control-openhands api GET /api/profiles/QA_base --pick config.base_url` is still `https://api.deepseek.com` and `--pick config.model` is `deepseek/deepseek-chat`; record `blocked` with `DEEPSEEK_API_KEY` as the prerequisite, leave with `control-openhands browser click 'testid=cancel-profile-btn'` and remove the arranged profile with `control-openhands api DELETE /api/profiles/QA_base --write` (arrange, not proof).
- **Leave without saving (`F10.editor-discard`).** Edit `QA_adv`, `control-openhands browser click 'testid=sdk-section-all-toggle'`, `control-openhands browser fill 'testid=sdk-settings-llm.top_p' 0.5` (Save turns enabled), then `control-openhands browser click 'testid=back-to-profiles'`. The list returns at once (`browser count 'testid=add-llm-profile'` is `1`) with no prompt (`control-openhands browser events --kinds dialog` is empty), and reopening Edit → All shows `top_p` empty. `testid=cancel-profile-btn` behaves the same.
- **Schema field validation and re-apply (`F10.schema-validation`, `F10.edit-default-reapply`).** Edit the Default profile and run `control-openhands browser click 'testid=sdk-section-all-toggle'`. `control-openhands browser fill 'testid=sdk-settings-llm.temperature' 5`, Save, then `control-openhands browser wait-text 'must be at most 2'` (toast `Temperature must be at most 2`). Clear it, fill `testid=sdk-settings-llm.top_p` with `3`: `Top P must be at most 1`. Clear it, fill `testid=sdk-settings-llm.extra_headers` with `{bad`: `Invalid JSON for Extra Headers`. Clear that, fill temperature `0.3`, Save, and `wait-text 'Profile "<name>" updated'`. `control-openhands api GET /api/settings` then has `agent_settings.llm.temperature` `0.3`, and reopening Edit → All shows `0.3`.
- **New profile prefill (`F10.create-prefill`).** With that Default (temperature `0.3`) run `control-openhands browser click 'testid=add-llm-profile'`. The editor opens on All (`browser eval "[...document.querySelectorAll('[role=tab]')].map(t=>t.innerText+':'+t.getAttribute('aria-selected')).join(',')"` is `Basic:false,Advanced:false,All:true`), `browser value 'testid=llm-custom-model-input'` is the backend default `openai/gpt-5.6-sol`, and `browser value 'testid=sdk-settings-llm.temperature'` is `0.3`, inherited from the Default. Run `control-openhands browser fill 'testid=llm-custom-model-input' deepseek/deepseek-flash`, `browser fill 'testid=profile-name-input' QA_inherit`, fill the key as in Create, Save and `wait-text 'Profile "QA_inherit" created' --timeout 60000`. Expected: the saved profile has the temperature the form showed. Actual today: `control-openhands api GET /api/profiles/QA_inherit` has `config.temperature` `null`, and after `browser reload` Edit `QA_inherit` → All shows the temperature empty (`fail`; the create path saves dirty fields only, so the shown inherited values are dropped). Screenshot `--feature F10.create-prefill --name edit-after-create`.
@@ -101,11 +107,13 @@ Preconditions:
- The editor's starting tab is inferred from the values it loads. Edit opens on whichever tab the profile's non-default values need (`QA_renamed` and `QA_adv` opened on Basic, `QA_broken` on All). A new editor opens on Basic only while the agent's LLM settings are all defaults: once the Default profile has a temperature (the schema step) it opens on All, and this survives deleting every profile, because the agent settings keep the last applied values. Click the tab you need before touching its fields, or `llm-provider-input` / `llm-custom-model-input` / `llm-api-key-help-anchor` will not exist.
- `control-openhands service stop agent-server` cannot show this page's error states: the whole app switches to the disconnected-backend screen.
- In Edit, `llm-api-key-input` holds the stored *encrypted* key token (about 140 masked characters), not an empty field. Never run `browser value` on it: it prints the token. Leave it untouched to keep the key.
+- An open provider or model listbox intercepts clicks on the editor's buttons (`cancel-profile-btn`, `save-profile-btn` time out with "subtree intercepts pointer events"): `browser press Escape` first.
- Provider and model options render in a portal outside the editor and modals: select `testid=provider-item-<name>` unscoped. Only verified providers have `provider-item-*` IDs and only Other models have `model-item-*` IDs; use `role=option[name="..."][exact]` for the rest. `browser testids --filter model-item` lists only items scrolled into view; `browser count` finds the rest.
- The auto-filled Profile Name never reuses an existing profile's name: picking a model whose name is already a profile's (for example `deepseek-flash` after `llm preset deepseek`) empties the field, which disables Save. Recipes fill `profile-name-input` with a `QA_*` name anyway.
- A duplicate profile name disables Save silently: no message, and `aria-invalid` stays `false` (#17938). Check `browser enabled 'testid=save-profile-btn'`, not the field styling.
- The Rename modal does not check duplicates client-side (Rename stays enabled); the server's 409 comes back as the toast `Profile '<name>' already exists`, and `browser errors --app-only` counts one `http-error` and one `console.error` for it.
- Set as default (and deleting the Default) rewrites the agent's active LLM settings (`control-openhands llm show`). Deleting every profile leaves the last applied model and key in agent settings.
+- The app never leaves profiles without a Default: `src/routes/root-layout.tsx` runs `useEnsureActiveProfile` (`src/hooks/use-ensure-active-profile.ts`), which activates the first profile with a key (else the first profile) whenever profiles exist but none is Default, whether never activated or because the Default was deleted (this is what promotes a profile in `F10.delete-default`). So `llm set --no-activate` keeps the Default only while another profile is Default: on a run whose only profile is the arranged one, `control-openhands api GET /api/profiles --pick active_profile` flips from `null` to it at the next page load (`browser reload`) and the agent's LLM settings follow (`llm show`).
- Profiles created by Add models are named from the model (`deepseek-chat`, …), not `QA_*`. The chooser hides a model when a profile with that *name* exists, not when a profile uses that model. Clean them up before the connection: the server refuses to delete a connection while profiles reference it.
- A profile linked to a missing connection is grouped under a header showing the raw connection id (`00000000…`) instead of a name.
- The All tab logs many `[i18n] Missing translation for key "SCHEMA$LLM$..."` console warnings; they are not page errors.
diff --git a/.agents/skills/verify-openhands/references/feature-map/F17-mcp-servers.md b/.agents/skills/verify-openhands/references/feature-map/F17-mcp-servers.md
index e9c3756b8b..60a9e6dff6 100644
--- a/.agents/skills/verify-openhands/references/feature-map/F17-mcp-servers.md
+++ b/.agents/skills/verify-openhands/references/feature-map/F17-mcp-servers.md
@@ -16,6 +16,7 @@ Source: `src/routes/extensions-hub.tsx`, `src/routes/mcp.tsx`, `src/components/f
- `F17.customize-redirect`: on desktop, the sidebar **Customize** link, the URL `/customize` and the command-menu items **Customize** and **MCP servers** all land on `/mcp`.
- `F17.desktop-subnav`: from 1024 px up, Customize pages show a left aside headed **Customize** with MCP Servers, Skills, Plugins and Apps, the active row marked, and the "synced from" backend badge.
- `F17.mobile-hub`: below 1024 px (phones and tablets), `/customize` is a hub with the same four rows and badge; at phone width detail pages show a Back chevron to the hub; widening the window to 1024 px or more on `/customize` redirects to `/mcp`.
+- `F17.breakpoint-sweep`: at 767, 768, 820, 1023, 1024 and 1440 px, `/customize` is the hub below 1024 px and the desktop aside from 1024 px up, and the MCP page keeps **Add custom server**, a search box wider than 150 px and no horizontal overflow.
- `F17.cloud-nav`: with a Cloud backend active, Plugins and Apps are hidden and Skills becomes an external link to `<cloud host>/settings/skills` (blocked here).
- `F17.page-states`: `/mcp` shows the **Model Context Protocol (MCP)** header, **Add custom server**, the toolbar, an **Installed** section (empty-state copy when nothing is installed) and the **Library** grid, without page errors.
- `F17.search`: one search box filters the library and the installed list (catalog name, description and args count); no match shows `No matches for your search.` in each section; the X clears it.
@@ -35,6 +36,9 @@ Source: `src/routes/extensions-hub.tsx`, `src/routes/mcp.tsx`, `src/components/f
- `F17.server-health`: each card's health row runs **Test connection** (`Checking connection…`), then shows a verdict with a coloured dot; a failure is red, worded for the transport (a STDIO command that cannot start, a remote server that cannot be reached) with the Agent Server's detail, and offers **Retry**.
- `F17.enable-disable`: the card toggle disables a server without removing its configuration; the state persists and a disabled server is withheld from new conversations.
- `F17.custom-edit`: clicking a card opens **Edit MCP server** prefilled (secrets shown as `**********`); Save tests and persists the change.
+- `F17.stored-secret-reuse`: a saved server's env values and tokens come back from the settings API and in the editor as `**********`; the editor's Test connection and Save both send the stored values behind those placeholders, never the placeholder text, so a Save from the edit form that keeps them changes only the edited fields.
+- `F17.single-request-mutations`: adding, editing or deleting one server sends exactly one `POST`, `PATCH` or `DELETE` to `/api/settings/mcp/<name>` and leaves sibling servers and their stored credentials untouched.
+- `F17.credential-probe`: a catalog entry whose server lists its tools with any credentials (Slack) also gets a read-only credential probe during Install and Test connection: a failed probe reads `Credential check failed: <provider error>` and saves nothing. A hosted server that rejects the token at connect (GitHub) never reaches a probe and reads `Connection failed: … 401 Unauthorized …`, also saving nothing; a server the catalog does not know gets no probe.
- `F17.delete-server`: **Delete** in the editor asks for confirmation; Cancel keeps the server, Confirm removes it with `MCP server removed.`
- `F17.native-git-tabs`: on a Cloud backend, GitHub/GitLab/Bitbucket offer a recommended native-integration tab next to the MCP tab (blocked here).
- `F17.agent-uses-server`: a conversation started after the save can call the server's tools.
@@ -58,11 +62,14 @@ Preconditions:
- TLS-intercepting proxy (this sandbox): MCP stdio servers start with a minimal environment (`HOME`, `PATH`, no proxy or CA variables), so without trust settings `uvx`/`npx` cannot download packages and catalog STDIO installs fail. `launch` now seeds the run's private `HOME` when `SSL_CERT_FILE` or `NODE_EXTRA_CA_CERTS` points at a CA bundle: check with `cat "$OH_VERIFY_RUN/private/home/.config/uv/uv.toml"` (`native-tls = true`) and `cat "$OH_VERIFY_RUN/private/home/.npmrc"` (`cafile=<bundle>`). Only if they are missing, write them once: `mkdir -p "$OH_VERIFY_RUN/private/home/.config/uv"`, `printf 'native-tls = true\n' > "$OH_VERIFY_RUN/private/home/.config/uv/uv.toml"` and `printf 'cafile=/root/.ccr/ca-bundle.crt\n' > "$OH_VERIFY_RUN/private/home/.npmrc"` (use your proxy's CA bundle path). `F17.install-error` moves them aside for one step. Custom servers can also pass `UV_NATIVE_TLS=true` in their env. Skip all of this on a machine with direct internet.
- `F17.custom-add-remote` needs a local Streamable HTTP MCP server (arrange, not proof; no `fixture` verb exists for it yet): from a shell, `cd "$OH_VERIFY_RUN/workspace" && (PORT=38417 nohup npx -y @modelcontextprotocol/server-everything streamableHttp > "$OH_VERIFY_RUN/private/qa-shttp.log" 2>&1 &)`, pick any free port, and wait until the log says `MCP Streamable HTTP Server listening on port 38417`. Stop it in Cleanup by its own PID (`pgrep -f 'server-everything streamableHttp'` lists it; never `pkill -f`).
- `F17.cloud-nav` and `F17.native-git-tabs` need an OpenHands Cloud backend (with git providers configured); not available here.
+- `F17.stored-secret-reuse` and `F17.single-request-mutations` use `control-openhands fixture mcp-server --name qa-vault`; it prints the `<node>` binary and `<server.mjs>` path the recipes substitute.
+- `F17.credential-probe` sends a dummy token to GitHub's hosted MCP endpoint and, through the catalog's npx stdio server, to Slack; both need the network. Behind the TLS-intercepting proxy the stdio server's own HTTPS call fails unless it gets the proxy CA (`NODE_EXTRA_CA_CERTS`), which only the custom editor can pass.
- Never type real credentials: every key below is a dummy.
- **Desktop entry points (`F17.customize-redirect`).** Run `control-openhands browser goto /conversations`, `control-openhands browser click 'testid=sidebar-skills-link' --expect-url '/mcp(\?|$)'` and `control-openhands browser count 'testid=mcp-page'` (`1`). Run `control-openhands browser goto /customize` and `control-openhands browser wait-url '/mcp(\?|$)'`. Command menu: run `control-openhands browser goto /conversations` (the browser is on `/mcp`, where the URL wait would pass at once), `control-openhands browser press Control+k`, `control-openhands browser type 'testid=command-menu >> role=combobox' MCP`, `control-openhands browser press Enter` and `control-openhands browser wait-url '/mcp(\?|$)'`; repeat with the term `Customize` (its only option reads `Customize Browse skills, plugins, and integrations.`). Every path ends on `/mcp`.
- **Desktop sub-navigation (`F17.desktop-subnav`).** On `/mcp` run `control-openhands browser text 'testid=extensions-navbar-desktop'`: `Customize`, `MCP Servers`, `Skills`, `Plugins`, `Apps`, then `These settings are synced from Local backend (<base url>)`. `control-openhands browser attr 'testid=sidebar-extensions-/mcp' aria-current` is `page`. Run `control-openhands browser click 'testid=sidebar-extensions-/skills' --expect-url '/skills(\?|$)'`; that row's `aria-current` becomes `page`. Do the same for `/plugins` and `/apps`, then return with `control-openhands browser click 'testid=sidebar-extensions-/mcp' --expect-url '/mcp(\?|$)'`. `control-openhands browser screenshot 'testid=extensions-navbar-desktop' --feature F17.desktop-subnav --name subnav` shows the MCP Servers row highlighted.
- **Phone hub (`F17.mobile-hub`, `F17.phone`).** Run `control-openhands browser viewport phone`, `control-openhands browser goto /customize` and `control-openhands browser text 'testid=extensions-mobile-hub'` (the same five lines as the desktop aside); `control-openhands browser screenshot --feature F17.mobile-hub --name hub`. Run `control-openhands browser click 'testid=extensions-mobile-hub >> testid=sidebar-extensions-/mcp' --expect-url '/mcp(\?|$)'`; `control-openhands browser attr 'testid=sidebar-mobile-back-button' aria-label` is `Customize`, and `control-openhands browser count 'testid=extensions-navbar-desktop >> visible'` is `0`. `control-openhands browser bbox 'testid=mcp-page'` has `insideViewport` `true` and `pageHorizontalOverflow` `false`; `control-openhands browser screenshot --feature F17.phone --name mcp-page` shows the header, toolbar and one column of cards. Run `control-openhands browser click 'testid=sidebar-mobile-back-button' --expect-url '/customize(\?|$)'`; `control-openhands browser count 'testid=extensions-mobile-hub'` is `1`. Drawer entry: `control-openhands browser goto /conversations`, `control-openhands browser click 'testid=sidebar-mobile-menu-toggle'` and `control-openhands browser click 'testid=sidebar-mobile-drawer >> testid=sidebar-skills-link' --expect-url '/customize(\?|$)'`; the hub is shown. Tablets get the hub too: `control-openhands browser viewport tablet` (820 px) keeps it on `/customize`; run `control-openhands browser click 'testid=extensions-mobile-hub >> testid=sidebar-extensions-/mcp' --expect-url '/mcp(\?|$)'`, then `control-openhands browser visible 'testid=extensions-navbar-desktop'` and `control-openhands browser visible 'testid=sidebar-mobile-back-button'` (both `false`), `control-openhands browser bbox 'role=main'` (520 px wide, `pageHorizontalOverflow` `false`) and `control-openhands browser screenshot --feature F17.mobile-hub --name tablet-mcp`; the rail's `control-openhands browser click 'testid=sidebar-skills-link' --expect-url '/customize(\?|$)'` brings the hub back. Then `control-openhands browser viewport desktop` and `control-openhands browser wait-url '/mcp(\?|$)'`: widening on `/customize` redirects to `/mcp`.
+- **Widths around the breakpoint (`F17.breakpoint-sweep`).** For each width `W` in `767`, `768`, `820`, `1023`, `1024`, `1440` run `control-openhands browser viewport Wx1024` and `control-openhands browser goto /customize`. Below 1024 px `control-openhands browser visible 'testid=extensions-mobile-hub'` is `true` and `control-openhands browser click 'testid=extensions-mobile-hub >> testid=sidebar-extensions-/mcp' --expect-url '/mcp(\?|$)'` opens the MCP page; from 1024 px up `control-openhands browser wait-url '/mcp(\?|$)'` passes at once and `control-openhands browser count 'testid=extensions-mobile-hub'` is `0`. On `/mcp`, `control-openhands browser visible 'testid=extensions-navbar-desktop'` is `false` below 1024 px and `true` from 1024 px (the switch is exactly 1023 → 1024), `control-openhands browser visible 'testid=mcp-add-custom-server'` is `true`, `control-openhands browser bbox 'testid=mcp-search-input'` is wider than 150 px (622, 323, 375, 578, 257 and 673 today) and `control-openhands browser bbox 'role=main'` has `pageHorizontalOverflow` `false` with equal `scrollWidth` and `clientWidth` (`main` is full width at 767 px, `W − 300` beside the rail from 768 to 1023 px, `W − 640` beside the rail and the aside from 1024 px). `control-openhands browser fill 'testid=mcp-search-input' GitHub` keeps `control-openhands browser count 'testid=mcp-marketplace-card-github'` at `1`; `control-openhands browser click 'testid=mcp-search-clear'`. At the boundary take `control-openhands browser screenshot --feature F17.breakpoint-sweep --name mcp-1023` and `--name mcp-1024`. Finish with `control-openhands browser viewport desktop`.
- **Page and empty state (`F17.page-states`).** On a fresh run run `control-openhands browser goto /mcp`, `control-openhands browser text 'testid=mcp-installed-empty'` (`No MCP servers installed yet.` and `Pick one from the marketplace below to get started.`) and `control-openhands browser count '[data-testid^="mcp-marketplace-card-"]'` (`55` today; the catalog ships with `@openhands/extensions`). `control-openhands browser eval "[...document.querySelectorAll('main h2')].map(h=>h.textContent)"` is `["Model Context Protocol (MCP)","Installed","Library"]`. Take `control-openhands browser screenshot --feature F17.page-states --name empty-desktop`. After the family, `control-openhands browser errors --app-only` shows `pageErrors` `0`.
- **Search (`F17.search`).** Run `control-openhands browser fill 'testid=mcp-search-input' time` and `control-openhands browser count '[data-testid^="mcp-marketplace-card-"]'` (`1`, `mcp-marketplace-card-time`). Fill `qa-zzz-nothing`; `control-openhands browser text 'testid=mcp-marketplace-empty'` is `No matches for your search.` (the installed list's search needs installed cards: it is driven at the end of `F17.install-stdio`). Run `control-openhands browser click 'testid=mcp-search-clear'`; `control-openhands browser value 'testid=mcp-search-input'` is `""` and the card count is back to `55`.
- **Section filter (`F17.section-filter`).** Run `control-openhands browser click 'testid=mcp-section-filter >> testid=dropdown-trigger'`, `control-openhands browser snapshot 'testid=mcp-section-filter'` (`button "Filter MCP servers" [expanded]: All` and `menu "Filter MCP servers"` with `All` [checked], `Installed`, `Library`) and `control-openhands browser click 'testid=mcp-section-filter-library'`; the `main h2` list above loses `Installed`. Open the dropdown again and click `testid=mcp-section-filter-installed`; `control-openhands browser count 'testid=mcp-marketplace-section'` is `0` and `control-openhands browser text 'testid=mcp-section-filter >> testid=dropdown-trigger'` is `Installed`. Restore with `testid=mcp-section-filter-all`.
@@ -81,11 +88,14 @@ Preconditions:
- **Card health (`F17.server-health`).** Run `control-openhands browser click 'testid=mcp-health-probe-qa_time' --observe 'testid=mcp-health-label-qa_time' --observe-ms 20000`: `Not checked yet`, `Checking connection…`, `Reachable — credentials not verified`. `control-openhands browser text 'testid=mcp-server-health-qa_time'` adds the hint `This check only proves the server is reachable and lists its tools; it does not verify your credentials.`, `control-openhands browser attr 'testid=mcp-server-health-qa_time >> testid=mcp-health-dot' data-status` is `healthy-connectivity`, and the click did not open the editor (`browser count 'testid=mcp-custom-editor'` is `0`). Failure: the editor refuses to save a broken server, so arrange one with `control-openhands api POST /api/settings/mcp/qa_broken --data '{"transport":"stdio","command":"qa-no-such-command","enabled":false}' --write`, `control-openhands browser reload`, then click `testid=mcp-health-probe-qa_broken`: the label is `Could not start the server command: Client failed to connect: [Errno 2] No such file or directory: 'qa-no-such-command'` in red (STDIO wording, no URL hint; the card clamps it to two lines, `browser text 'testid=mcp-server-health-qa_broken'` reads it whole), the dot's `data-status` is `failed` and the button reads `Retry`; `control-openhands browser screenshot 'testid=mcp-server-item >> has-text=qa_broken' --feature F17.server-health --name failed`.
- **Delete (`F17.delete-server`).** Run `control-openhands browser click 'testid=mcp-server-detail-qa_broken'` (a click on the card's centre lands on its health row and opens nothing: see Gotchas), `control-openhands browser click 'testid=mcp-custom-editor-delete'` and `control-openhands browser text 'testid=confirmation-modal'` (`Are you sure you want to delete this server?`, `Cancel`, `Confirm`). Run `control-openhands browser click 'testid=confirmation-modal >> testid=cancel-button'`; the editor is still open. Click Delete again, then `control-openhands browser click 'testid=confirmation-modal >> testid=confirm-button'` and `control-openhands browser toasts` (`MCP server removed.`). After `control-openhands browser reload`, `control-openhands browser count 'testid=mcp-server-item >> has-text=qa_broken'` is `0`, and `control-openhands api GET /api/settings` has no `qa_broken`.
- **Edit (`F17.custom-edit`).** Run `control-openhands browser click 'testid=mcp-server-detail-qa_time'` (the detail line opens the editor on every card; a plain click on the card's centre happens to open it on `qa_time`, whose Time description fills the centre, but not on a card without a description; keyboard: `browser focus '[data-server-id=qa_time]'` then `browser press Enter`) and `control-openhands browser snapshot 'testid=mcp-custom-editor'`: heading `Edit MCP server`, no Server Type field, a `Delete` button; `control-openhands browser value 'testid=edit-mcp-server-form >> testid=env-input'` is `UV_NATIVE_TLS=**********`. Run `control-openhands browser fill 'testid=edit-mcp-server-form >> testid=args-input' $'mcp-server-time\n--local-timezone\nUTC'` (bash ANSI-C quoting for the newlines), click `testid=edit-mcp-server-form >> testid=mcp-test-connection` (still `Connected — 2 tool(s) available`: the stored env value is reused), then `control-openhands browser click 'testid=edit-mcp-server-form >> testid=submit-button'` and wait for `testid=mcp-custom-editor` to detach. The card shows a verdict at once (seeded from the save's test). After `control-openhands browser reload`, `control-openhands browser text 'testid=mcp-server-detail-qa_time'` is `uvx mcp-server-time --local-timezone UTC`.
+- **Stored secrets are reused, not their placeholder (`F17.stored-secret-reuse`).** Make a STDIO server that only starts with the right secret: `control-openhands fixture mcp-server --name qa-vault` prints `<node>` and `<server.mjs>`. Run `control-openhands browser click 'testid=mcp-add-custom-server'`, pick `STDIO`, fill `testid=add-mcp-server-form >> testid=name-input` `qa_vault` and `testid=add-mcp-server-form >> testid=command-input` `/bin/sh`, then `control-openhands browser fill 'testid=add-mcp-server-form >> testid=args-input' $'-c\ntest "$QA_VAULT_TOKEN" = qa-vault-secret-000 && exec "$QA_NODE" "$QA_SERVER"'` and `control-openhands browser fill 'testid=add-mcp-server-form >> testid=env-input' $'QA_VAULT_TOKEN=qa-vault-secret-000\nQA_NODE=<node>\nQA_SERVER=<server.mjs>'`. Click `testid=add-mcp-server-form >> testid=mcp-test-connection`; `control-openhands browser text 'testid=add-mcp-server-form >> testid=mcp-test-message' --timeout 60000` is `Connected — 1 tool(s) available`. Click `testid=add-mcp-server-form >> testid=submit-button`, `control-openhands browser wait 'testid=mcp-custom-editor' --state detached --timeout 60000` and `control-openhands browser reload`. `control-openhands api GET /api/settings --pick agent_settings.mcp_config.qa_vault` has `env` `{"QA_VAULT_TOKEN":"**********","QA_NODE":"**********","QA_SERVER":"**********"}` (every env value is redacted; the page reads this same endpoint, there is no `GET /api/settings/mcp`). Run `control-openhands browser click 'testid=mcp-server-detail-qa_vault'`; `control-openhands browser value 'testid=edit-mcp-server-form >> testid=env-input'` is the three lines `QA_VAULT_TOKEN=**********`, `QA_NODE=**********`, `QA_SERVER=**********`. Click `testid=edit-mcp-server-form >> testid=mcp-test-connection`: `control-openhands browser text 'testid=edit-mcp-server-form >> testid=mcp-test-message' --timeout 60000` is `Connected — 1 tool(s) available`, which the shell gate only allows when the stored token, binary and path were sent (`control-openhands browser screenshot 'testid=mcp-custom-editor' --feature F17.stored-secret-reuse --name masked-test-ok`). Counter-check that the gate bites: `control-openhands browser fill 'testid=edit-mcp-server-form >> testid=env-input' $'QA_VAULT_TOKEN=qa-wrong\nQA_NODE=**********\nQA_SERVER=**********'`, click Test connection again and `control-openhands browser wait 'testid=edit-mcp-server-form >> testid=mcp-test-message >> text=Connection' --timeout 90000`; the text is red and reads `Connection failed: McpError: Connection closed` or, when the client only notices at its deadline, `Connection timed out. Check the URL and try again.` (the gate exits before the MCP handshake; see Gotchas; `--name wrong-token`). `control-openhands browser click 'testid=mcp-custom-editor-close'` discards it (`count 'testid=mcp-custom-editor'` `0`); reopen with `testid=mcp-server-detail-qa_vault`: the env is the three placeholders again and Test connection is `Connected — 1 tool(s) available`. `control-openhands browser press Escape`. Save with the placeholders in place: reopen with `control-openhands browser click 'testid=mcp-server-detail-qa_vault'`, append one harmless argument (it only becomes the shell script's `$0`) with `control-openhands browser fill 'testid=edit-mcp-server-form >> testid=args-input' $'-c\ntest "$QA_VAULT_TOKEN" = qa-vault-secret-000 && exec "$QA_NODE" "$QA_SERVER"\nqa-v2'`, leave `env-input` as the three placeholders, run `control-openhands browser network --clear`, then `control-openhands browser click 'testid=edit-mcp-server-form >> testid=submit-button' --observe 'testid=edit-mcp-server-form >> testid=submit-button' --observe-ms 6000` (`Save Server`, `Save Server [disabled]`, `<absent>`) and `control-openhands browser wait 'testid=mcp-custom-editor' --state detached --timeout 60000`; `control-openhands browser network --filter '/api/settings/mcp'` lists one `PATCH /api/settings/mcp/qa_vault` (`200`) and the card's health label reads `Reachable — credentials not verified`. After `control-openhands browser reload`, `control-openhands api GET /api/settings --pick agent_settings.mcp_config.qa_vault.env` is still `{"QA_VAULT_TOKEN":"**********","QA_NODE":"**********","QA_SERVER":"**********"}`, `control-openhands api GET /api/settings --pick agent_settings.mcp_config.qa_vault.args` ends with `qa-v2`, and `control-openhands browser text 'testid=mcp-server-detail-qa_vault'` ends with `"$QA_SERVER" qa-v2`. Reopen with `control-openhands browser click 'testid=mcp-server-detail-qa_vault'` (the env is the three placeholders) and click `testid=edit-mcp-server-form >> testid=mcp-test-connection`: `control-openhands browser text 'testid=edit-mcp-server-form >> testid=mcp-test-message' --timeout 60000` is `Connected — 1 tool(s) available` again, which the gate only allows if the Save kept the real token, binary and path rather than storing the placeholder text (`--name saved-with-placeholders`). `control-openhands browser press Escape`.
+- **One request per mutation (`F17.single-request-mutations`).** With `qa_vault` as the sibling, run `control-openhands browser network --clear`, `control-openhands browser click 'testid=mcp-add-custom-server'`, pick `STDIO`, fill `testid=add-mcp-server-form >> testid=name-input` `qa_docs`, `testid=add-mcp-server-form >> testid=command-input` `<node>` and `testid=add-mcp-server-form >> testid=args-input` `<server.mjs>`, click `testid=add-mcp-server-form >> testid=submit-button` and `control-openhands browser wait 'testid=mcp-custom-editor' --state detached --timeout 60000`. `control-openhands browser network --filter '/api/settings/mcp'` has `total` `1`, a `POST /api/settings/mcp/qa_docs` with status `201`, and `control-openhands browser network --filter qa_vault` has `total` `0`. Edit: `control-openhands browser network --clear`, `control-openhands browser click 'testid=mcp-server-detail-qa_docs'`, `control-openhands browser fill 'testid=edit-mcp-server-form >> testid=args-input' $'<server.mjs>\n--qa-v2'`, click `testid=edit-mcp-server-form >> testid=submit-button` and the same `wait`; the filter now lists one `PATCH /api/settings/mcp/qa_docs` (`200`) and `control-openhands api GET /api/settings --pick agent_settings.mcp_config.qa_docs.args` ends with `--qa-v2`. Delete: `control-openhands browser network --clear`, `control-openhands browser click 'testid=mcp-server-detail-qa_docs'`, `control-openhands browser click 'testid=mcp-custom-editor-delete'`, `control-openhands browser click 'testid=confirmation-modal >> testid=confirm-button'` and `control-openhands browser wait '[data-server-id=qa_docs]' --state detached --timeout 10000`; the filter lists one `DELETE /api/settings/mcp/qa_docs` (`200`). Sibling: `control-openhands api GET /api/settings --pick agent_settings.mcp_config.qa_vault` is unchanged from the end of the previous bullet (command `/bin/sh`, args still ending with `qa-v2`, the three `**********` env values, `enabled` `true`), and after `control-openhands browser click 'testid=mcp-server-detail-qa_vault'` its Test connection still reads `Connected — 1 tool(s) available`: the stored secrets survived the sibling's writes. `control-openhands browser press Escape`.
+- **Credential probe (`F17.credential-probe`).** GitHub (hosted): run `control-openhands browser click 'testid=mcp-marketplace-card-github'`, `control-openhands browser fill 'testid=mcp-install-modal >> testid=mcp-install-field-api_key' ghp_qaInvalid000`, `control-openhands browser network --clear`, then `control-openhands browser click 'testid=mcp-install-modal >> testid=mcp-install-submit' --observe 'testid=mcp-install-modal >> testid=mcp-install-submit' --observe-ms 20000` (`Install`, `Verifying…`, `Install`) and `control-openhands browser text 'testid=mcp-install-modal-error'`. The server refuses the token before any tool is listed, so the probe never runs and the text is `Connection failed: HTTPStatusError: Client error '401 Unauthorized' for url 'https://api.githubcopilot.com/mcp/' …`; the modal stays open (`control-openhands browser count 'testid=mcp-install-modal'` is `1`), `control-openhands browser network --filter 'api/(mcp|settings)'` shows one `POST /api/mcp/test` and no `/api/settings/mcp` write, `control-openhands api GET /api/settings --pick agent_settings.mcp_config.github` has no value and `control-openhands api GET /api/settings/secrets` lists no `GITHUB_PERSONAL_ACCESS_TOKEN` (`control-openhands browser screenshot 'testid=mcp-install-modal' --feature F17.credential-probe --name github-invalid`); `control-openhands browser click 'testid=mcp-install-cancel'`. Slack (stdio; it lists 8 tools with any token, so only the probe, a channel listing, can tell a bad token): `control-openhands browser click 'testid=mcp-marketplace-card-slack'`, `control-openhands browser fill 'testid=mcp-install-modal >> testid=mcp-install-field-SLACK_TEAM_ID' T_QA_INVALID`, `control-openhands browser fill 'testid=mcp-install-modal >> testid=mcp-install-field-SLACK_BOT_TOKEN' xoxb-qa-invalid-000` (`type` `password`), `control-openhands browser click 'testid=mcp-install-modal >> testid=mcp-install-submit'` and `control-openhands browser wait 'testid=mcp-install-modal-error' --timeout 150000` (the first run downloads the package through npx, about a minute: a plain click, then one long `wait`, rather than `--observe-ms`, which records a few transient states). `control-openhands browser text 'testid=mcp-install-modal-error'` is `Credential check failed: invalid_auth` with direct internet; behind the TLS-intercepting proxy the probe's own HTTPS call fails first and it reads `Credential check failed: fetch failed`. Either way the modal stays open and `control-openhands api GET /api/settings --pick agent_settings.mcp_config.slack` has no value (`--name slack-invalid`); `control-openhands browser click 'testid=mcp-install-cancel'`. For Slack's own verdict behind the proxy, run the same command as a custom server, which keeps the entry's probe (#18034): `control-openhands browser click 'testid=mcp-add-custom-server'`, pick `STDIO`, fill `testid=add-mcp-server-form >> testid=name-input` `qa_slack`, `testid=add-mcp-server-form >> testid=command-input` `npx`, `control-openhands browser fill 'testid=add-mcp-server-form >> testid=args-input' $'-y\n@zencoderai/slack-mcp-server'` and `control-openhands browser fill 'testid=add-mcp-server-form >> testid=env-input' $'SLACK_TEAM_ID=T_QA_INVALID\nSLACK_BOT_TOKEN=xoxb-qa-invalid-000\nNODE_EXTRA_CA_CERTS=/root/.ccr/ca-bundle.crt'` (your proxy's CA bundle), then `control-openhands browser click 'testid=add-mcp-server-form >> testid=mcp-test-connection'` and `control-openhands browser text 'testid=add-mcp-server-form >> testid=mcp-test-message' --timeout 90000`: `Credential check failed: invalid_auth`, class `text-red-500` (`control-openhands browser screenshot 'testid=mcp-custom-editor' --feature F17.credential-probe --name slack-custom-invalid-auth`). `control-openhands browser click 'testid=add-mcp-server-form >> testid=submit-button' --observe 'testid=add-mcp-server-form >> testid=submit-button' --observe-ms 20000` (`Add Server`, `Add Server [disabled]`, `Add Server`: Save tests first) keeps the editor open (`control-openhands browser count 'testid=mcp-custom-editor'` is `1`) with the same red message, and `control-openhands api GET /api/settings --pick agent_settings.mcp_config.qa_slack` has no value; close it with `control-openhands browser click 'testid=mcp-custom-editor-close'` (an Escape pressed while the Save's test still runs is ignored). A server the catalog does not know gets no probe: `control-openhands browser click 'testid=mcp-health-probe-qa_vault' --observe 'testid=mcp-health-label-qa_vault' --observe-ms 20000` ends on `Reachable — credentials not verified` (the Card health wording).
- **The agent uses it (`F17.agent-uses-server`).** With `qa_time` enabled run `control-openhands conversation start --prompt "Use the qa_time MCP server's get_current_time tool with timezone UTC. Do not use the terminal. Reply with only the tool's datetime value." --wait --timeout 240`, then `control-openhands conversation events <id> --kinds ActionEvent,ObservationEvent,MessageEvent` with the `id` it printed. There is an `ActionEvent` with tool `qa_time_get_current_time` (MCP tools reach the agent as `<server id>_<tool>`), its observation reads `[Tool 'qa_time_get_current_time' executed.] { "timezone": "UTC", "datetime": … }`, and the agent's reply is that `datetime`. `control-openhands conversation events <id> --grep qa_time_get_current_time --from-start` also matches the `SystemPromptEvent` (the tool was offered). `control-openhands browser screenshot --feature F17.agent-uses-server --name conversation` shows the tool step (its title is model-written, e.g. `Get current time in UTC via qa_time MCP`) and the datetime reply. `conversation start` leaves the browser on the conversation: `control-openhands browser goto /mcp` next.
- **Enable and disable (`F17.enable-disable`).** On `/mcp` run `control-openhands browser attr 'testid=mcp-installed-toggle-qa_time' aria-checked` (`true`; `aria-label` `Disable server`), `control-openhands browser click 'testid=mcp-installed-toggle-qa_time'`, `control-openhands browser reload` and the same `attr` (`false`; label `Enable server`). `control-openhands api GET /api/settings` shows `agent_settings.mcp_config.qa_time.enabled` `false` with command, args and env intact; `control-openhands browser screenshot 'testid=mcp-server-item >> has-text=qa_time' --feature F17.enable-disable --name disabled` shows the + icon. With no other server exposing that tool, run `control-openhands conversation start --prompt "Without running any command: is a tool named get_current_time available to you right now? Reply with only YES or NO." --wait --timeout 240`; the reply is `NO`, `control-openhands api GET /api/conversations/<id> --pick agent.mcp_config.qa_time.enabled` is `false`, and `control-openhands conversation events <id> --grep qa_time_get_current_time --from-start` has `count` `0` (no `SystemPromptEvent` lists it). Then `control-openhands browser goto /mcp`. Click the toggle again and reload; it is `true`.
- **Phone modals (`F17.phone`).** At `browser viewport phone` on `/mcp` run `control-openhands browser click 'testid=mcp-add-custom-server'`, `control-openhands browser bbox 'testid=mcp-custom-editor'` (`insideViewport` `true`, `pageHorizontalOverflow` `false`) and `control-openhands browser screenshot --feature F17.phone --name custom-editor`; close with `testid=mcp-custom-editor-close`. Do the same for `testid=mcp-marketplace-card-stripe` with `testid=mcp-install-modal` (`--name install-modal`), then `control-openhands browser viewport desktop`.
- **Cloud backend (`F17.cloud-nav`, `F17.native-git-tabs`).** Blocked without an OpenHands Cloud backend. On the local backend `control-openhands browser click 'testid=mcp-marketplace-card-github'` shows only the MCP form (`browser count 'testid=mcp-install-tab-native'` is `0`). With Cloud active, `browser text 'testid=extensions-navbar-desktop'` should list only MCP Servers and the cloud Skills link (`browser attr 'testid=sidebar-extensions-/skills' href` ends in `/settings/skills`), and the GitHub modal should show `mcp-install-tab-native` (Recommended) and `mcp-install-tab-mcp`.
-- **Cleanup.** On `/mcp` (not the conversation page `conversation start` left open), delete `qa_time`, `qa_shttp` and any `time`, `time_1`, `tavily`, `filesystem` cards through the editor's Delete: for each id, `browser focus '[data-server-id=<id>]'`, `browser press Enter`, `browser click 'testid=mcp-custom-editor-delete'`, `browser click 'testid=confirmation-modal >> testid=confirm-button'`, `browser wait '[data-server-id=<id>]' --state detached --timeout 10000`. After `browser reload`, `browser count 'testid=mcp-installed-empty'` is `1` and `api GET /api/settings --pick agent_settings.mcp_config` is `{}`. Delete the `TAVILY_API_KEY` secret on `/settings/secrets` (F14: `testid=secret-item >> has-text=TAVILY_API_KEY >> testid=delete-secret-button`, then `testid=confirmation-modal >> testid=confirm-button`), stop the Streamable HTTP fixture by its PID, and restore `deepseek-flash` as the active profile if you changed it.
+- **Cleanup.** On `/mcp` (not the conversation page `conversation start` left open), delete `qa_time`, `qa_shttp`, `qa_vault` and any `time`, `time_1`, `tavily`, `filesystem` cards through the editor's Delete: for each id, `browser focus '[data-server-id=<id>]'`, `browser press Enter`, `browser click 'testid=mcp-custom-editor-delete'`, `browser click 'testid=confirmation-modal >> testid=confirm-button'`, `browser wait '[data-server-id=<id>]' --state detached --timeout 10000`. After `browser reload`, `browser count 'testid=mcp-installed-empty'` is `1` and `api GET /api/settings --pick agent_settings.mcp_config` is `{}`. Delete the `TAVILY_API_KEY` secret on `/settings/secrets` (F14: `testid=secret-item >> has-text=TAVILY_API_KEY >> testid=delete-secret-button`, then `testid=confirmation-modal >> testid=confirm-button`), stop the Streamable HTTP fixture by its PID, and restore `deepseek-flash` as the active profile if you changed it.
## Gotchas
@@ -106,4 +116,8 @@ Preconditions:
- Catalog installs are add-only: a second Time install becomes `time_1`. Since #18034 (closed #17930) STDIO servers match the catalog by command plus leading args instead of by name, so `time_1` and a custom server that runs a catalog command under its own name (`qa_time`) keep that entry's description, badge and search keywords, while a server that only borrows a catalog name does not (a STDIO `fetch` running `qa-no-such-command` shows only its command) (`F17.install-stdio`, `F17.custom-add`). Twelve library entries (provider-OAuth-only, such as Sentry, HubSpot or the Cloudflare OAuth servers) open an empty modal whose Install does nothing (`F17.install-unsupported`) (#17922).
- The section filter's trigger and menu are named `Filter MCP servers` since #18035 (closed #17939); a checkout without it reuses the conversation filter's `Filter conversations`. The test ids work on both.
- A malformed header line in the custom editor shows `Environment variables must follow KEY=value format` (`mcp-server-form.tsx` runs headers through `validateEnvFormat`) (#17958); see `F17.custom-validation`.
+- A catalog stdio server runs with the Agent Server's minimal environment: behind a TLS-intercepting proxy its own HTTPS calls fail (`fetch failed`) even when npx could download it, and the install modal has no field for `NODE_EXTRA_CA_CERTS`; use the custom editor with the same command to see the provider's real answer (`F17.credential-probe`).
+- GitHub's hosted MCP endpoint answers an invalid token with HTTP 401 at connect, before tools are listed, so its `get_me` probe never runs and the modal shows the connection wording (`Connection failed: HTTPStatusError … 401 Unauthorized`), not `Credential check failed`.
+- The shell-gated `qa_vault` fixture exits before the MCP handshake, which the Agent Server reports either at once as `Connection failed: McpError: Connection closed` or only at its deadline as `Connection timed out. Check the URL and try again.` (both seen on one stack); neither is the `Could not start the server command` wording, which needs a command that cannot be spawned at all. Wait on `text=Connection`, which matches both and not `Connected — …`.
+- A browser verb's client call waits 120 s, or 10 s longer than the verb's own `--timeout` or `--observe-ms` when that is longer, so `browser wait 'testid=mcp-install-modal-error' --timeout 150000` runs its full deadline and a miss ends with Playwright's timeout and a failure screenshot (before this, every call was aborted at 120 s with `TimeoutError: The operation was aborted due to timeout` while the page kept going). For a first Slack install use a plain click, then that `wait`; `--observe-ms` is for a few transient states, not a minute-long outcome.
- Known issue #17941: turning off an enabled card toggle can drop the click when the pointer moves in and presses at once (`F17.enable-disable`).
diff --git a/.agents/skills/verify-openhands/references/feature-map/F18-skills.md b/.agents/skills/verify-openhands/references/feature-map/F18-skills.md
index 157a4d4eb5..effe59b4c5 100644
--- a/.agents/skills/verify-openhands/references/feature-map/F18-skills.md
+++ b/.agents/skills/verify-openhands/references/feature-map/F18-skills.md
@@ -24,6 +24,8 @@ Source: `src/routes/skills-settings.tsx`, `src/components/features/skills/`, `sr
- `F18.toggle-local`: switching a personal/project skill off writes `disabled_skills`; new conversations then do not load it, and do again once it is back on.
- `F18.toggle-error`: a failed save shows an error toast and nothing is persisted.
- `F18.skill-in-chat`: an enabled skill is loaded into new conversations and fires on its trigger; a disabled one is not.
+- `F18.project-skill`: a skill committed under `<workspace>/.agents/skills/<name>/SKILL.md` is loaded into every conversation started in that workspace, in Local Repo and New Worktree mode alike, and its trigger word activates it (the chat shows a **Skill Ready** row naming it); workspace skills are not listed on the Skills page.
+- `F18.skill-deleted`: a personal skill whose folder was deleted disappears from the Skills page after a reload and is not loaded into new conversations: its trigger word activates nothing.
- `F18.detail-modal`: the detail modal shows source, an Enabled/Disabled switch, description, pills and the read-only content; switching off disables **Use skill**.
- `F18.detail-close`: the X, **Close**, Escape and a backdrop click close the detail modal.
- `F18.detail-pills`: the modal shows the full pill set (type, category, recommended, license, compatibility, version, triggers); pills that do not fit collapse into a `+N` button whose popover lists the rest without closing the modal.
@@ -40,6 +42,7 @@ Source: `src/routes/skills-settings.tsx`, `src/components/features/skills/`, `sr
- Command menu (`Control+k`/`Meta+k`): searching "Skills" lists only **Customize** ("Browse skills, plugins, and integrations."), which lands on `/mcp`; then click Skills. There is no direct Skills command.
- Phone: `/customize` shows the Customize hub (`extensions-mobile-hub`) with a Skills row; on `/skills` the header back button (`sidebar-mobile-back-button`) returns to the hub.
- Install banner: inside a conversation, after the agent ran `/add-skill <GitHub URL>` (local backend only).
+- Project skills: a `SKILL.md` committed under `<workspace>/.agents/skills/<name>/` in the folder picked with Open Workspace on Home (Local Repo or New Worktree mode); the conversation's chat shows **Skill Ready** when a message triggers one.
- Cloud: Customize → **Skills** on a Cloud backend (external link).
## Driving it with control-openhands
@@ -50,6 +53,8 @@ Preconditions:
- A fresh run lists only the bundled catalog, `<N>` skills, all with source `public`. `<N>` is the catalog length the checkout bundles: run `node --input-type=module -e "import {SKILLS_CATALOG} from '@openhands/extensions/skills'; console.log(SKILLS_CATALOG.length)"` in the checkout (68 with `@openhands/extensions` 0.29.0, #17998). The Source and Type facet groups stay hidden until a personal/project skill exists. `add-javadoc` and `flarglebargle` are off; `docker` is on (recommended).
- `F18.skill-in-chat`, `F18.toggle-local`, `F18.copy-source` and `F18.install-banner` need an active LLM profile (`control-openhands llm preset deepseek`). `F18.install-banner` also needs the agent to reach github.com.
- Personal skill fixture (arranged through the agent, not proof): run `control-openhands conversation start --prompt 'Use one shell command to create the file ~/.agents/skills/qa-hello/SKILL.md (create the folders) with exactly these 7 lines: "---", "name: qa-hello", "description: QA test skill that answers qa-ping with QA-PONG-7731.", "triggers:", "- qa-ping", "---", "When the user says qa-ping, reply with exactly QA-PONG-7731 and nothing else." Then reply done.' --wait --timeout 240`. Afterwards `/skills` shows `<N+1> result(s)` and a `qa-hello` card (type pill `Auto-discovery`), and the Enabled facet count goes from 12 to 13.
+- Project skill fixture for `F18.project-skill` (arranged, not proof; no model needed), in this family's own repo so the commit never lands in F08's `qa-repo`: `control-openhands fixture git-repo --name qa-f18-repo`, `control-openhands fixture skill --repo qa-f18-repo --name qa-f18-skill --trigger qa-f18-ping`, then commit it with an arrange step outside the UI, `git -C "$OH_VERIFY_RUN/workspace/qa-f18-repo" add -A && git -C "$OH_VERIFY_RUN/workspace/qa-f18-repo" -c user.name=qa -c [email protected] commit -q -m "Add qa-f18-skill"`: `fixture skill` writes the file but does not commit it, and a New Worktree checkout holds only committed files. A fixture skill's name must not be a substring of its trigger (see Gotchas).
+- `F18.skill-deleted` creates its own personal fixture right before its bullet (`control-openhands fixture skill --name qa-gone --trigger qa-farewell`, written into the run's private HOME) and deletes it inside the bullet, so the counts in the other bullets are unaffected.
- `F18.cloud-link` needs an OpenHands Cloud backend and account (blocked in a sandbox).
- **Open from the sidebar (`F18.page`).** From `/` run `control-openhands browser click 'testid=sidebar-skills-link' --expect-url '/mcp(\?|$)'`, `control-openhands browser click 'testid=sidebar-extensions-/skills' --expect-url '/skills(\?|$)'`, then `control-openhands browser text 'testid=skills-settings-description'` (`Discover skills to add to your workspace. Search above the cards or filter by state, recommendation, and category, then open a card to see its details and use the skill. Enable or disable default skills. Disabled skills will not be loaded into agent context.`, describing this layout since #18035), `control-openhands browser text 'testid=skills-new-conversation-notice'` (`Skill changes apply to new conversations only.`) and `control-openhands browser text 'testid=skills-result-summary'` (`<N> result(s)` on a fresh run). `control-openhands browser screenshot --feature F18.page --name desktop` shows the Customize nav, the facet rail and the card grid.
@@ -65,6 +70,8 @@ Preconditions:
- **Toggle a built-in skill (`F18.toggle`).** Run `control-openhands browser attr 'testid=skill-toggle-add-javadoc' aria-checked` (`false`; `aria-label` is `Enable skill`), `control-openhands browser click 'testid=skill-toggle-add-javadoc'` (no modal opens; `aria-checked` turns `true` and the Enabled facet count goes up by one), wait a second, `control-openhands browser reload` and read `aria-checked` again: `true`. `control-openhands api GET /api/settings --pick misc_settings.app_preferences.enabled_skills` lists `add-javadoc` (`--pick` paths start inside `body`). To switch an enabled skill off from its card, `control-openhands browser hover 'testid=skill-toggle-add-javadoc'` first (`attr ... data-showing-remove` is `true`), then click (see Gotchas).
- **Toggle a personal skill (`F18.toggle-local`).** Run `control-openhands browser hover 'testid=skill-toggle-qa-hello'`, `control-openhands browser click 'testid=skill-toggle-qa-hello'`, wait a second, `control-openhands browser reload` and `control-openhands browser attr 'testid=skill-toggle-qa-hello' aria-checked` (`false`). `control-openhands api GET /api/settings` shows `disabled_skills: ["qa-hello"]` under `body.misc_settings.app_preferences` and `enabled_skills` without it. `control-openhands conversation start --prompt "qa-ping" --wait --timeout 240`, then `control-openhands api GET '/api/conversations/<id>/events/search?limit=5' --max-bytes 2000000` (id from `conversation start`): the user `MessageEvent` has `"activated_skills": []`. Back on `/skills` (`control-openhands browser goto /skills`; `conversation start` left the browser on the conversation), click the toggle on again (no hover needed), start another `qa-ping` conversation: `"activated_skills": ["qa-hello"]` and `control-openhands browser text 'testid=agent-message >> nth=-1'` is `QA-PONG-7731`; `disabled_skills` is `[]` again.
- **Skill reaches chat (`F18.skill-in-chat`).** Run `control-openhands browser goto /skills` (the previous bullet ends on a conversation), `control-openhands browser click 'testid=skill-toggle-flarglebargle'` (`aria-checked` `true`), then `control-openhands conversation start --prompt "flarglebargle" --wait --timeout 240`, `control-openhands api GET '/api/conversations/<id>/events/search?limit=5' --max-bytes 2000000` (`"activated_skills": ["flarglebargle"]`) and `control-openhands browser text 'testid=agent-message >> nth=-1'` (a reply praising how smart the user is). `control-openhands browser screenshot --feature F18.skill-in-chat --name enabled`. Switch it off again on `/skills` (`control-openhands browser goto /skills`, hover, click; `enabled_skills` no longer lists it) and start another `flarglebargle` conversation: `"activated_skills": []` and the agent says it does not recognise the word.
+- **Project skill in a workspace (`F18.project-skill`).** With the committed `qa-f18-skill` fixture, run `control-openhands browser goto /skills`, `control-openhands browser reload` and `control-openhands browser count 'testid=skill-card-qa-f18-skill'`: `0` (the catalog lists the Agent Server's own skills, not a workspace's). Then `control-openhands conversation start --workspace qa-f18-repo --prompt "qa-f18-ping" --wait --timeout 240` (note `<repo-skill-id>`; its `workspace` is the fixture's path) and, in the chat it leaves open, `control-openhands browser text 'testid=generic-event-message-title'` (`Skill Ready`), `control-openhands browser click 'role=button[name="Expand"]'` and `control-openhands browser count 'role=button[name="qa-f18-skill"]'` (`1`, under `Triggered Skill Knowledge:`); `control-openhands browser screenshot --feature F18.project-skill --name skill-ready`. With an LLM the agent also obeys the skill, and this is the moment to read it, while the chat is still open: `control-openhands browser text 'testid=agent-message >> nth=-1'` ends with `QA_SKILL_OK`; without a key the conversation ends in `error` right after the events below and that reply check is blocked (prerequisite: `DEEPSEEK_API_KEY`). Agent side, `control-openhands conversation events <repo-skill-id> --grep qa-f18-skill --from-start` matches the `SystemPromptEvent` (its `<available_skills>` lists `<name>qa-f18-skill</name>`) and the user `MessageEvent` (`skills: ["qa-f18-skill"]`, with excerpts showing `"activated_skills":["qa-f18-skill"]` and the `…/qa-f18-repo/.agents/skills/qa-f18-skill/SKILL.md` path). New Worktree: run `control-openhands workspace open qa-f18-repo`, `control-openhands browser click 'testid=workspace-mode-selector'`, `control-openhands browser click 'testid=workspace-mode-selector-option-new_worktree'` (`control-openhands browser text 'testid=home-git-control-bar-preview'` reads `qa-f18-repo` and `New Worktree`), then `control-openhands conversation start --stay --prompt "qa-f18-ping" --wait --timeout 240` (note `<worktree-id>`): its `workspace` is `/tmp/conversation-worktrees/<worktree-id>/qa-f18-repo`, a worktree of the fixture on branch `openhands/<worktree-id>` (read-only check: `git -C /tmp/conversation-worktrees/<worktree-id>/qa-f18-repo log --oneline` lists `Add qa-f18-skill`), `control-openhands browser text 'testid=generic-event-message-title'` is `Skill Ready` again (and the same reply check applies here, with a key), and `control-openhands conversation events <worktree-id> --grep qa-f18-skill --from-start` matches the same events with the worktree's `SKILL.md` path (a `ConversationStateUpdateEvent` carrying the agent context matches in both conversations too). Restore the mode: `control-openhands workspace open qa-f18-repo`, `control-openhands browser click 'testid=workspace-mode-selector'`, `control-openhands browser click 'testid=workspace-mode-selector-option-local_repo'` (the preview reads `Local Repo`).
+- **Deleted personal skill (`F18.skill-deleted`).** Run `control-openhands fixture skill --name qa-gone --trigger qa-farewell`, `control-openhands browser goto /skills`, `control-openhands browser reload`, `control-openhands browser count 'testid=skill-card-qa-gone'` (`1`; `control-openhands browser attr 'testid=skill-source-qa-gone' title` is the run's `…/private/home/.agents/skills/qa-gone/SKILL.md` path and `control-openhands browser text 'testid=skills-result-summary'` counts one result more than before) and `control-openhands browser screenshot --feature F18.skill-deleted --name present`. Prove it loads: `control-openhands conversation start --prompt "qa-farewell" --wait --timeout 240` (note `<gone-a>`), `control-openhands browser text 'testid=generic-event-message-title'` (`Skill Ready`) and `control-openhands conversation events <gone-a> --grep qa-gone --from-start`: the `SystemPromptEvent` and the user `MessageEvent` match, the latter with `skills: ["qa-gone"]`. Delete the folder (outside the UI; it is inside the run): `rm -rf "$OH_VERIFY_RUN/private/home/.agents/skills/qa-gone"`. Then `control-openhands browser goto /skills`, `control-openhands browser reload`, `control-openhands browser count 'testid=skill-card-qa-gone'` (`0`; the summary is back to its earlier count), `control-openhands conversation start --prompt "qa-farewell" --wait --timeout 240` (note `<gone-b>`), `control-openhands browser count 'text=Skill Ready'` (`0`; `control-openhands browser screenshot --feature F18.skill-deleted --name absent`) and `control-openhands conversation events <gone-b> --grep qa-gone --from-start`: `count` is `0`, and `control-openhands conversation events <gone-b> --kinds MessageEvent --from-start` shows the user message without a `skills` row. No model is needed: the skill set is fixed before the first model call.
- **Save failure (`F18.toggle-error`).** Run `control-openhands browser goto /skills`, wait for `testid=skill-toggle-add-javadoc` and note its `aria-checked` (`true` when this follows the toggle bullet, which left `add-javadoc` on). Then `control-openhands service stop agent-server`, `control-openhands browser click 'testid=skill-toggle-add-javadoc' --timeout 5000` (the card flips at once), wait about 2 s (the `PATCH /api/settings` is tried three times before the error shows; `browser network` lists three `502`s) and `control-openhands browser toasts --history`: `HTTP request failed (502 Bad Gateway): "Bad Gateway: connect ECONNREFUSED 127.0.0.1:<agent-server port>"`. `control-openhands browser screenshot --feature F18.toggle-error --name toast`. Run `control-openhands restart --timeout 240`, `control-openhands browser goto /skills`: `aria-checked` on `testid=skill-toggle-add-javadoc` is back to the value noted before the click and `enabled_skills` is unchanged. Clear the expected 502 noise with `control-openhands browser errors --clear`.
- **Detail modal and switch (`F18.detail-modal`).** With `add-javadoc` on, run `control-openhands browser click 'testid=skill-card-add-javadoc'`, `control-openhands browser testids 'testid=skill-detail-modal'` (name, source, `skill-modal-enable-row-add-javadoc`, description, pills incl. license `MIT`, `skill-modal-field-content-add-javadoc`, `skill-detail-close`, `skill-detail-use-skill-add-javadoc`) and `control-openhands browser enabled 'testid=skill-detail-use-skill-add-javadoc'` (`true`). Switch off with `control-openhands browser click 'testid=skill-modal-enable-row-add-javadoc >> text=Enabled'`; `control-openhands browser eval "document.querySelector('[data-testid=skill-modal-toggle-add-javadoc]').checked"` is `false`, the row reads `Disabled`, `enabled` on Use skill is `false` and the card toggle behind reads `aria-checked` `false`. `control-openhands browser screenshot --feature F18.detail-modal --name disabled`. Close, reload: the card stays off and `enabled_skills` no longer lists `add-javadoc`.
- **Close the modal (`F18.detail-close`).** With `testid=skill-card-docker` opened each time, each of `control-openhands browser click 'testid=skill-detail-modal-close'`, `control-openhands browser click 'testid=skill-detail-close'`, `control-openhands browser press Escape` and `control-openhands browser mouse-click 100 500` (backdrop) leaves `control-openhands browser count 'testid=skill-detail-modal'` at `0`.
@@ -94,3 +101,6 @@ Preconditions:
- With the modal's `+N` pill popover open (focus on its trigger), Escape closes neither the popover nor the detail modal; on a card the same Escape closes the popover (#17957). Cause not confirmed: the `+N` trigger's `onKeyDown` in `src/components/features/skills/skill-card-pill-row.tsx` stops propagation, which may keep the key from the popover's `document` and `ModalBackdrop`'s `window` listeners. Click elsewhere in the modal first. Recorded as `fail` on `F18.detail-pills`.
- The Add skill example, its Copy button and the add-skill **Use skill** message share one constant. Since #18035 it is `.../skills/code-review`; before, it pointed at `skills/codereview`, which is 404 upstream (the skill was renamed).
- Skill changes apply to new conversations only; an open conversation keeps the skills it started with.
+- `conversation events --grep` searches whole events, the user's message text included: a trigger that contains the skill name (`qa-gone-ping` for `qa-gone`) matches the user `MessageEvent` even when the skill is not loaded. Give fixture skills a trigger that does not contain their name, or grep for `<name>qa-gone</name>`.
+- `fixture skill --repo` writes the project `SKILL.md` but does not commit it; a New Worktree conversation checks out the repo's commits, so commit the skill first. Workspace skills are loaded per conversation and never appear on `/skills`.
+- Without an LLM key the skill set is still observable: the `SystemPromptEvent` (available skills), the user `MessageEvent` (`activated_skills`) and the chat's **Skill Ready** row (`generic-event-message-title`; Expand lists the skills under `Triggered Skill Knowledge:`) all exist before the first model call, after which the conversation ends in `error`.
diff --git a/.agents/skills/verify-openhands/references/feature-map/F20-canvas-apps.md b/.agents/skills/verify-openhands/references/feature-map/F20-canvas-apps.md
index 05a71fcf6c..6be72840fc 100644
--- a/.agents/skills/verify-openhands/references/feature-map/F20-canvas-apps.md
+++ b/.agents/skills/verify-openhands/references/feature-map/F20-canvas-apps.md
@@ -24,6 +24,7 @@ Source: `src/routes/canvas-extensions.tsx`, `src/routes/canvas-extension-page.ts
- `F20.rail-active`: on the app's page its rail entry is marked current (`aria-current="page"`, highlighted); with the sidebar collapsed the entry is an icon whose tooltip and `aria-label` carry the label. App pages are not listed in the command menu.
- `F20.rail-label`: the rail entry is labelled with the page's `nav_label` (`Extension demo` for the demo fixture), falling back to its title.
- `F20.page-render`: the rail entry and the direct URL mount the app into `main` named after the page title (demo: h1 `Hello from a Canvas Extension`); a deeper URL passes the rest of the path to the app (`Nested extension path: nested`).
+- `F20.app-backend-view`: an app page learns from the host whether it has a backend view; the host offers one only when the Agent Server advertises the `canvas_app_backend_bridge_v1` capability with an app-backend ingress URL, so on the pinned server the demo's status line reads `App backend view unavailable` on every path of its page.
- `F20.page-unavailable`: an unknown app, an unknown page of a known app, or a disabled or uninstalled app shows **Not available** / `This app is disabled, missing, or does not provide this page.`
- `F20.phone`: at 390 px the Customize hub lists Apps, the Apps page, the Add form and an app page fit without horizontal overflow, and the app's entry is in the sidebar drawer.
- `F20.refresh`: **Update** re-installs from the recorded source, toasts `App updated.` and keeps the enabled state; a failed update toasts the error and leaves the card unchanged.
@@ -63,6 +64,7 @@ Preconditions:
- **Trust confirmation (`F20.enable-confirm`).** Run `control-openhands browser click 'testid=canvas-extension-card-demo-page >> role=switch'` and `control-openhands browser snapshot 'testid=confirmation-modal'`: the paragraph `This app runs trusted JavaScript inside Agent Canvas and can make authenticated requests to the active Agent Server. Review its source and revision before enabling it.`, buttons `Cancel` and `Enable trusted app`. `control-openhands browser click 'testid=confirmation-modal >> testid=cancel-button'` closes it and `aria-checked` stays `false`. Keyboard: `control-openhands browser focus 'testid=canvas-extension-card-demo-page >> role=switch'` and `control-openhands browser press Space` open it again (`browser count 'testid=confirmation-modal'` is `1`); `control-openhands browser press Escape` closes it and `api GET /api/canvas-extensions/installed` still says `"enabled": false`. Backdrop: click the switch again, `control-openhands browser mouse-click 20 500`, then `browser count 'testid=confirmation-modal'` is `0` and the API still says `"enabled": false`. Reopen with the click, then `control-openhands browser click 'testid=confirmation-modal >> testid=confirm-button'`, `control-openhands browser reload` and read `aria-checked`: `true`; `control-openhands browser text 'testid=canvas-extension-card-demo-page'` contains `Enabled`.
- **Rail entry (`F20.rail-entry`, `F20.rail-label`).** After enabling, run `control-openhands browser wait 'testid=sidebar-canvas-extension-demo-page-hello' --timeout 10000`, `control-openhands browser attr 'testid=sidebar-canvas-extension-demo-page-hello' href` (`/extensions/demo-page/hello`) and `control-openhands browser text 'testid=sidebar-canvas-extension-demo-page-hello'`. Expected `Extension demo` (the fixture's `nav_label`); today it reads `Hello from an extension` (see Gotchas).
- **Render the page (`F20.page-render`).** Run `control-openhands browser click 'testid=sidebar-canvas-extension-demo-page-hello' --expect-url '/extensions/demo-page/hello(\?|$)'` and `control-openhands browser snapshot 'role=main[name="Hello from an extension"]'`: heading `Hello from a Canvas Extension` and paragraph `Host API 1 on backend default-local`. Direct URL: `control-openhands browser goto /extensions/demo-page/hello/nested` and the same snapshot show the paragraph `Nested extension path: nested`. `control-openhands browser screenshot --feature F20.page-render --name rail`.
+- **App backend view (`F20.app-backend-view`).** Still on `/extensions/demo-page/hello/nested` (where Render the page ends) run `control-openhands browser text 'testid=demo-extension-app-backend-status'`, `control-openhands api GET /server_info --pick capabilities` and `control-openhands api GET /server_info --pick app_backend_ingress_url`. The host hands a page `appBackendView` only when the capabilities list `canvas_app_backend_bridge_v1` and the ingress URL is not `null`: with the pinned server (1.53.0: neither) the line reads `App backend view unavailable`; a server advertising both would make the demo read `App backend view available` (not driven here). `control-openhands browser goto /extensions/demo-page/hello` shows the same line, `control-openhands browser snapshot 'role=main[name="Hello from an extension"]'` lists it as the last paragraph, and `control-openhands browser screenshot --feature F20.app-backend-view --name unavailable`.
- **Active and collapsed rail entry (`F20.rail-active`).** With the demo enabled, run `control-openhands browser goto /apps`; `control-openhands browser attr 'testid=sidebar-canvas-extension-demo-page-hello' aria-current` is `null`; after `control-openhands browser click 'testid=sidebar-canvas-extension-demo-page-hello' --expect-url '/extensions/demo-page/hello(\?|$)'` it is `page`. Run `control-openhands browser click 'testid=sidebar-collapse-toggle'`: `browser visible 'testid=sidebar-canvas-extension-demo-page-hello'` stays `true` (about 40 px wide), `browser attr ... aria-label` and `control-openhands browser tooltip 'testid=sidebar-canvas-extension-demo-page-hello'` read `Hello from an extension`; `control-openhands browser screenshot --feature F20.rail-active --name collapsed` shows the highlighted icon. Expand again with `browser click 'testid=sidebar-collapse-toggle'` (its `aria-label` is `Collapse sidebar` again). Command menu: `browser goto /`, `browser press Control+k`, `browser type 'testid=command-menu >> role=combobox' Hello` and `browser snapshot 'testid=command-menu'` read `No commands found`; `browser press Escape`.
- **Unavailable page (`F20.page-unavailable`).** Run `control-openhands browser goto /extensions/qa-missing/x` and `control-openhands browser snapshot 'role=main'`: heading `Not available` and `This app is disabled, missing, or does not provide this page.` `control-openhands browser goto /extensions/demo-page/nope` shows the same card while the demo is enabled.
- **Phone (`F20.phone`).** Run `control-openhands browser viewport phone`, `control-openhands browser goto /customize` (the URL stays `/customize`), `control-openhands browser click 'testid=extensions-mobile-hub >> testid=sidebar-extensions-/apps' --expect-url '/apps(\?|$)'`, `control-openhands browser bbox 'testid=canvas-extensions-screen'` (`insideViewport` `true`, `pageHorizontalOverflow` `false`) and `control-openhands browser screenshot --feature F20.phone --name apps`. Open the form with `canvas-extensions-add-button`; `control-openhands browser bbox 'testid=add-canvas-extension-modal'` is inside the viewport with no page overflow; close it with `add-canvas-extension-dismiss`. Then `control-openhands browser click 'testid=sidebar-mobile-menu-toggle'`, `control-openhands browser visible 'testid=sidebar-mobile-drawer >> testid=sidebar-canvas-extension-demo-page-hello'` (`true`), `control-openhands browser click 'testid=sidebar-mobile-drawer >> testid=sidebar-canvas-extension-demo-page-hello' --expect-url '/extensions/demo-page/hello(\?|$)'` and `control-openhands browser bbox 'role=main[name="Hello from an extension"]'` (inside, no overflow). Return with `control-openhands browser viewport desktop`.
@@ -93,3 +95,4 @@ Preconditions:
- `control-openhands launch` passes the launcher only a fixed list of environment variables (path, locale, proxy) plus its own flags: `OH_AGENT_SERVER_VERSION=<old>` in the environment is ignored (the run still gets the pinned server); use `--sdk-ref v<version>`. Any server older than 1.47.0 is gated by Canvas before `/apps` renders.
- Stopping the Agent Server makes the next full page load show the **Manage backends** gate; always `restart` and `doctor` afterwards.
- A trailing `browser errors --app-only` sweep shows a `400` per failed install by design; count only other errors.
+- `App backend view unavailable` on the demo page is the expected text while `GET /server_info` lists no `canvas_app_backend_bridge_v1` capability (Agent Server 1.53.0); it does not mean the app failed to load. Cloud backends and the no-backend state never get a backend view either.
diff --git a/.agents/skills/verify-openhands/references/feature-map/F25-backends-and-cloud.md b/.agents/skills/verify-openhands/references/feature-map/F25-backends-and-cloud.md
index 1f78169d05..c0dc6da989 100644
--- a/.agents/skills/verify-openhands/references/feature-map/F25-backends-and-cloud.md
+++ b/.agents/skills/verify-openhands/references/feature-map/F25-backends-and-cloud.md
@@ -21,9 +21,10 @@ Source: `src/components/features/backends/`, `src/contexts/active-backend-contex
- `F25.cloud-advanced-host`: on the Cloud tab, **Advanced** reveals a **Host** field (placeholder `https://app.all-hands.dev`) for a self-hosted Cloud.
- `F25.cloud-device-flow`: **Connect to OpenHands** opens a popup and shows `Starting authentication...`, then `Waiting for authorization...` with the verification link and **Cancel**; Cancel returns to the idle button and closes the popup. An unreachable host shows `Failed to start device flow: ...` with **Try again**. Approving the code (adds a backend named OpenHands Cloud) needs a Cloud account.
- `F25.agent-server-guidance`: the Agent-server tab has a **Local / Remote** toggle. Local shows a collapsible **Before you connect** with `agent-canvas --backend-only --port 8001` and a local-setup docs link; Remote shows **Recommended setup**, **Connection details** and a remote-setup docs link.
-- `F25.add-agent-server`: **Connect** stays disabled until Host Name and a valid Host are filled (and, on Remote, an API key). A dead host or a wrong key shows `Could not connect to <host>` with the reason inline; a reachable host with the right key adds the backend, makes it active (BM-001) and redirects a `/conversations/<id>` page to `/conversations` without an error toast, also on the backend's first activation, without asking the new backend for the old conversation id.
+- `F25.add-agent-server`: **Connect** stays disabled until Host Name and a valid Host are filled; the API Key field is optional for Connect on Local and required on Remote, but a key-protected host still rejects an empty key. A dead host, an empty key or a wrong key shows `Could not connect to <host>` with the reason inline (`Disconnected (check URL or network)` for a dead host, `Invalid API key` for an empty or wrong key); a reachable host with the right key, entered on the default Local location, adds the backend, makes it active (BM-001) and redirects a `/conversations/<id>` page to `/conversations` without an error toast, also on the backend's first activation, without asking the new backend for the old conversation id.
- `F25.switch-backend`: choosing another option shows a full-screen `Switching to <name>` overlay for about a second, then the sidebar and pages show that backend's data; on `/conversations/<id>` the app moves to `/conversations` (BM-002).
- `F25.backend-pinned-url`: in-app links carry `?backend=<id>`; opening such a URL pins the tab to that backend; an unknown id is ignored.
+- `F25.per-tab-backend`: the active backend is per browser tab: switching it in one tab leaves the other open tabs where they were, a new tab opened from a plain link starts from the last choice made in any tab, and a sidebar conversation row opened in a new tab (Control-click) lands on the backend that owns the conversation because its link carries `?backend=<owner>`.
- `F25.manage-backends`: Manage backends lists every backend with status dot, name, version (`v1.50.1`), host, status text and a LOCAL/CLOUD pill, plus Edit and Remove; Escape, a backdrop click, X and **Done** close it.
- `F25.manage-add`: **Add Backend** in Manage backends opens the same Add form stacked over Manage; its X returns to Manage, and a successful add keeps Manage open with the new row while the new backend becomes active.
- `F25.manage-select`: clicking a healthy row makes that backend active and closes the modal; on a detail page it leaves it for that section's list, like the selector does (BM-002).
@@ -72,10 +73,11 @@ Preconditions:
- **Device flow error (`F25.cloud-device-flow`).** Run `control-openhands browser fill 'testid=add-backend-cloud-host' http://127.0.0.1:9`, `control-openhands browser click 'testid=add-backend-login-button' --observe 'testid=add-backend-device-flow'` and `control-openhands browser wait 'testid=add-backend-auth-error' --timeout 15000`. The observation goes `Starting authentication...` → `Failed to start device flow: Failed to fetch` / `Try again`, and `control-openhands browser network --last 5` shows `POST /oauth/device/authorize` to `http://127.0.0.1:9`. `control-openhands browser click 'testid=add-backend-auth-retry'` starts again and fails the same way. Each attempt leaves an `about:blank` popup: `control-openhands browser tabs`, then `control-openhands browser close-tab <i>` for each extra page and `control-openhands browser tab 0`. Close the modal with `testid=add-backend-close`.
- **Device flow awaiting and Cancel (`F25.cloud-device-flow`).** Reopen Add Backend (Cloud tab, default host), run `control-openhands browser click 'testid=add-backend-login-button' --observe 'testid=add-backend-device-flow' --observe-ms 6000`, then `control-openhands browser snapshot 'testid=add-backend-device-flow'`. If `https://app.all-hands.dev` is reachable (`control-openhands browser network --external --last 5` lists it), the snapshot shows `Waiting for authorization...`, `Browser opened. Complete sign-in to continue. If browser didn't open, visit:`, a link `https://app.all-hands.dev/oauth/device/verify?user_code=<code>` and button `Cancel`; `browser tabs` lists a second page on that URL. Screenshot with `--feature F25.cloud-device-flow --name awaiting`, then `control-openhands browser click 'testid=add-backend-auth-cancel'`: the panel is back to `Connect to OpenHands` / `Advanced` and `browser tabs` lists one page. Never approve the code. If Cloud is unreachable you get the error state of the previous bullet instead.
- **Agent-server guidance (`F25.agent-server-guidance`).** Run `control-openhands browser click 'testid=add-backend-option-agent-server'` and `control-openhands browser snapshot 'testid=add-backend-agent-server-panel'`: radiogroup `Agent-server location` with `Local` checked, heading `Before you connect`, textboxes `Host Name`, `Host`, `API Key` and `button "Connect" [disabled]`. `control-openhands browser click 'testid=add-backend-local-guidance-toggle'` and `browser snapshot 'testid=add-backend-local-guidance'` show the code `agent-canvas --backend-only --port 8001` and link `Read the local backend setup guide`. `control-openhands browser click 'testid=add-backend-location-option-remote'` and `control-openhands browser click 'testid=add-backend-remote-guidance-toggle'`; `browser snapshot 'testid=add-backend-remote-guidance'` shows `Recommended setup`, `Connection details` and `Read the remote backend setup guide`.
-- **Connect rules and errors (`F25.add-agent-server`).** Still on Remote: `control-openhands browser fill 'testid=add-backend-name' QA_Second`, `control-openhands browser fill 'testid=add-backend-host' 127.0.0.1:9`; `control-openhands browser enabled 'testid=add-backend-submit'` is `false` (Remote needs a key). `control-openhands browser fill 'testid=add-backend-api-key' wrong-key` makes it `true`. `control-openhands browser click 'testid=add-backend-location-option-local'` keeps the fields. `control-openhands browser click 'testid=add-backend-submit' --observe 'testid=add-backend-submit'` shows `Checking…`, then `control-openhands browser text 'testid=add-backend-error'` starts `Could not connect to http://127.0.0.1:9` / `Disconnected (check URL or network). ...`. Fill `add-backend-host` with `<second-url>` and submit again: the error reads `Could not connect to <second-url>` / `Invalid API key`. Close with `testid=add-backend-close`.
+- **Connect rules and errors (`F25.add-agent-server`).** Still on Remote: `control-openhands browser fill 'testid=add-backend-name' QA_Second`, `control-openhands browser fill 'testid=add-backend-host' 127.0.0.1:9`; `control-openhands browser enabled 'testid=add-backend-submit'` is `false` (Remote needs a key). `control-openhands browser fill 'testid=add-backend-api-key' wrong-key` makes it `true`. `control-openhands browser click 'testid=add-backend-location-option-local'` keeps the fields. `control-openhands browser click 'testid=add-backend-submit' --observe 'testid=add-backend-submit'` shows `Checking…`, then `control-openhands browser text 'testid=add-backend-error'` starts `Could not connect to http://127.0.0.1:9` / `Disconnected (check URL or network). ...`. Fill `add-backend-host` with `<second-url>` and submit again: the error reads `Could not connect to <second-url>` / `Invalid API key`. On Local the key is optional for Connect but not for a key-protected host: `control-openhands browser fill 'testid=add-backend-api-key' ''` leaves `control-openhands browser enabled 'testid=add-backend-submit'` at `true` (`browser value 'testid=add-backend-api-key'` reports `length` `0`), and `control-openhands browser click 'testid=add-backend-submit'` then `control-openhands browser wait 'testid=add-backend-error' --timeout 15000` shows the same `Could not connect to <second-url>` / `Invalid API key` (`browser screenshot --feature F25.add-agent-server --name local-no-key-error`). Close with `testid=add-backend-close`.
- **Add the second backend (`F25.add-agent-server`).** Run `control-openhands browser goto /conversations/<id>`, `browser click 'testid=backend-selector'`, `browser click 'testid=add-backend-menu-item'`, `browser click 'testid=add-backend-option-agent-server'`, fill `add-backend-name` with `QA_Second` and `add-backend-host` with `<second-url>`, then `control-openhands browser fill 'testid=add-backend-api-key' --value-file "$OH_VERIFY_RUN_2/private/session-key"` and `control-openhands browser click 'testid=add-backend-submit' --expect-url '/conversations(\?|$)'`. The URL is `/conversations`, `browser count 'testid=add-backend-modal'` is `0`, `browser snapshot 'testid=backend-selector'` shows `combobox "QA_Second"`, and `browser count 'testid=conversation-card'` is `0`. Before anything reloads the page, `control-openhands browser toasts --history` lists only `Loading...` and the new backend's `Your LLM isn't set up yet, so conversations won't run. ...` (the second stack has no LLM profile; no error toast), and `control-openhands browser network --external --filter 'ids='` lists no request: the new backend is never asked for the first stack's conversation id. (`--external` leaves out this stack's own `GET /api/conversations?ids=<id>` from the `goto /conversations/<id>` above, which the request log keeps across page loads; `<second-url>` is another origin, so a request to it is listed.) The new backend asks its own telemetry consent ("This preference is saved for the local backend “QA_Second”"): run `control-openhands onboard --skip` (it leaves the browser on `/`). After `control-openhands browser reload` the selector still reads `QA_Second`.
- **Switch with the overlay (`F25.switch-backend`).** From `/` run `browser click 'testid=backend-selector'` and `control-openhands browser click 'testid=backend-selector >> role=option[name="Connected Local"]' --observe 'testid=environment-switch-overlay' --observe-ms 2500`. The observation shows `Switching to Local`, then `<absent>` about one second later; the selector reads `Local` and `browser count 'testid=conversation-card'` is `1`. Then `control-openhands browser click 'testid=conversation-card' --expect-url '/conversations/[0-9a-f-]+'`, open the selector and `control-openhands browser click 'testid=backend-selector >> role=option[name="Connected QA_Second"]' --expect-url '/conversations(\?|$)'`: the URL is `/conversations`, the selector reads `QA_Second` and the card count is `0`.
- **Pinned URL (`F25.backend-pinned-url`).** With `QA_Second` active run `control-openhands browser goto '/conversations/<id>?backend=default-local'` and `control-openhands browser wait 'testid=chat-interface'`: the selector reads `Local` and the conversation opens. `control-openhands browser goto '/conversations?backend=qa-unknown-id'` keeps `Local` (card count `1`). The sidebar card links themselves end in `?backend=default-local` (`browser snapshot`).
+- **Per-tab backend (`F25.per-tab-backend`).** With `Local` active and `QA_Second` registered, run `control-openhands browser goto /conversations` and `control-openhands browser eval "JSON.parse(sessionStorage.getItem('openhands-active-backend')).backendId"` (`default-local`); note QA_Second's id with `control-openhands browser eval "JSON.parse(localStorage.getItem('openhands-backends')).find(b=>b.name==='QA_Second').id"` (`<second-id>`, a generated UUID). Open a second tab from the sidebar: `control-openhands browser click 'testid=sidebar-conversations-link' --modifiers Control`, then `control-openhands browser tabs` lists two pages at `/conversations` (the **New Chat** link carries no `?backend=`). `control-openhands browser tab 1`, `control-openhands browser wait 'testid=backend-selector'`: `browser snapshot 'testid=backend-selector'` reads `combobox "Local"` and `control-openhands browser eval "sessionStorage.getItem('openhands-active-backend')"` is `null` (a new tab inherits nothing and starts from the localStorage fallback). Switch this tab: `browser click 'testid=backend-selector'`, `control-openhands browser click 'testid=backend-selector >> role=option[name="Connected QA_Second"]'`, `control-openhands browser wait 'testid=environment-switch-overlay' --state detached --timeout 10000`; the selector reads `QA_Second`, `browser count 'testid=conversation-card'` is `0` and `control-openhands browser eval "JSON.parse(localStorage.getItem('openhands-active-backend')).backendId"` is `<second-id>`. Back in `control-openhands browser tab 0` nothing moved: the selector still reads `Local`, the card count is `1`, the sessionStorage eval is still `default-local` while the localStorage eval is `<second-id>` (`browser screenshot --feature F25.per-tab-backend --name first-tab-unchanged`). A plain new tab starts from the last choice made in any tab: from tab 0 run `control-openhands browser click 'testid=sidebar-conversations-link' --modifiers Control`, `control-openhands browser tab 2`, `browser wait 'testid=backend-selector'`: the selector reads `QA_Second`, the card count is `0` and its sessionStorage item is `null` (`--name plain-new-tab-last-choice`). A conversation row opens on its owner instead: `control-openhands browser tab 0`, `control-openhands browser click 'testid=conversation-card' --modifiers Control`, `control-openhands browser wait-tab '/conversations/<id>' --timeout 20000` prints `index` `3` and the URL `/conversations/<id>?backend=default-local`; `control-openhands browser tab 3`, `control-openhands browser wait 'testid=chat-interface' --timeout 30000`, `browser url` still ends in `/conversations/<id>?backend=default-local` (no bounce to `/conversations`), the selector reads `Local` and `browser toasts --history` lists only `Loading...` (no `This conversation does not exist…`); `browser screenshot --feature F25.per-tab-backend --name new-tab-on-owner`. Close the extra tabs highest first: `control-openhands browser close-tab 3`, `control-openhands browser close-tab 2`, `control-openhands browser close-tab 1`, `control-openhands browser tab 0`; `browser tabs` lists one page and the selector reads `Local`. Opening the pinned link also recorded `default-local` as the last choice (the localStorage eval in tab 3 reads `default-local`), so run the plain-new-tab check before the row check.
- **Manage list and closing (`F25.manage-backends`).** Run `browser click 'testid=backend-selector'`, `control-openhands browser click 'testid=manage-backends-menu-item'` and `control-openhands browser snapshot 'testid=manage-backends-modal'`: heading `Manage backends`, two rows `Connected Local v1.50.1 <this run's URL> Connected Local` and `Connected QA_Second v1.50.1 <second-url> Connected Local`, each with `Edit` and `Remove`, then `Add Backend` and `Done`. `browser text 'testid=manage-backends-status-QA_Second'` is `Connected` and `browser text 'testid=manage-backends-version-QA_Second'` is `v1.50.1` (whatever `doctor` reports). Screenshot with `--feature F25.manage-backends --name list`. Close it four ways, reopening between: `browser press Escape`, `control-openhands browser mouse-click 20 500` (backdrop), `browser click 'testid=close-manage-backends-modal'`, `browser click 'testid=manage-backends-done'`; each time `browser count 'testid=manage-backends-modal'` is `0`.
- **Select from Manage (`F25.manage-select`).** With `Local` active, `control-openhands browser click 'testid=conversation-card' --expect-url '/conversations/[0-9a-f-]+'`, open Manage Backends, then `control-openhands browser click 'testid=manage-backends-row-QA_Second >> role=button >> nth=0' --expect-url '/conversations(\?|$)'`. Like the selector, the app leaves the detail page for `/conversations` (BM-002): `browser count 'testid=manage-backends-modal'` is `0`, the selector reads `QA_Second`, `browser count 'testid=chat-interface'` and `browser count 'testid=conversation-card'` are `0`, and `browser screenshot --feature F25.manage-select --name after-select` shows QA_Second's Home with its empty conversation list.
- **Edit validation and errors (`F25.edit-backend`).** Open Manage Backends, `control-openhands browser click 'testid=manage-backends-edit-QA_Second'`, then `browser value 'testid=edit-backend-host'` (`<second-url>`), `browser attr 'testid=edit-backend-api-key' type` (`password`) and `browser text 'testid=edit-backend-status'` (`Connected`, `·`, `Local`, `v1.50.1` on separate lines). Fill `edit-backend-name` with an empty string and `edit-backend-host` with `not a host`, then `browser focus 'testid=edit-backend-api-key'` and `browser snapshot 'testid=edit-backend-form'`: alerts `Name is required` and `Enter a valid URL (e.g. http://localhost:8080)`; `browser enabled 'testid=edit-backend-submit'` is `false`.
@@ -101,6 +103,8 @@ Preconditions:
- Manage-row test ids use the backend name (`manage-backends-row-QA_Second`), so they change after a rename and are not unique when two backends share a name; the Cloud **Log back in** id uses the backend id instead (`manage-backends-login-<id>-login-button`).
- Selector options are named `<status> <name>` (`Connected Local`, `Disconnected QA_Renamed`): select them with `role=option[name="Connected <name>"]`.
- After a navigating option click, `--expect-url` must allow `?backend=...`: end regexes with `(\?|$)`.
+- `browser wait-tab <regex>` matches any open page, including the current one: for a second tab at the same URL (`/conversations` from **New Chat**) read `browser tabs` and switch with `browser tab <i>`; `wait-tab` is only reliable when the new tab's URL is unique (`/conversations/<id>`). Close extra tabs highest index first, or the indexes shift under you.
+- A tab opened with `click --modifiers Control` starts with an empty sessionStorage, so it boots from the localStorage fallback (the backend last chosen in any tab) unless its URL carries `?backend=`. Opening a `?backend=` link writes that backend to localStorage too: a per-tab check that relies on "last choice" must run before opening pinned links.
- The Add Backend Agent-server form has no inline "required" messages; it only disables **Connect**. The messages exist in the Edit form.
- Footer clicks from the collapsed popover or the phone drawer open the modals a moment later: `browser wait` for them instead of `count`.
- The device-flow popup survives errors: every failed attempt leaves an `about:blank` tab (#17953). Cancel closes it; after an error close it with `browser close-tab`.
@@ -110,7 +114,7 @@ Preconditions:
- `/oauth/device/verify` is a Cloud page that also ships in Canvas: on a local stack its POST target does not exist, so Authorize always ends in `Error`. The "Authentication Required" state is unreachable (`useIsAuthed` always resolves `true`).
- `/shared/conversations/<id>` on a local backend also raises a raw toast `HTTP request failed (404 Not Found): {"detail":"Not Found"}` next to `Conversation not found` (#17953).
- Known failure (repro candidate, not an exemption): `F25.edit-escape` closes the Manage modal underneath and drops the Edit draft (Cancel and X close only Edit, `F25.edit-cancel`) (#17953).
-- The add bullet's `network --external --filter 'ids='` check guards an earlier race: on a backend's first activation from `/conversations/<id>`, about 1 attempt in 5 asked the new backend for the old conversation id. It did not happen in 15 first activations at 4eff80237: each time `QA_Second` was removed, its consent unset with `control-openhands api PATCH /api/settings --run "$OH_VERIFY_RUN_2" --write --data '{"misc_settings_diff":{"app_preferences":{"user_consents_to_analytics":null}}}'` (arrange only) and the backend added again from `/conversations/<id>`. A backend that lacks an id answers `[null]`, which now reads as "not found" (only the `This conversation does not exist…` toast), not as incompatible data. No merged change was shown to fix the race itself (#18009 only changed how a `[null]` answer reads; #18024 only moved the redirect into a shared hook, still called after `addBackend`), so a hit (an `ids=` request to `<second-url>`, now followed by the `This conversation does not exist…` toast) is a new bug to file.
+- The add bullet's `network --external --filter 'ids='` check guards an earlier race: on a backend's first activation from `/conversations/<id>`, about 1 attempt in 5 asked the new backend for the old conversation id. It did not happen in 15 first activations at 4eff80237: each time `QA_Second` was removed, its consent unset with `control-openhands api PATCH /api/settings --run "$OH_VERIFY_RUN_2" --write --data '{"misc_settings_diff":{"app_preferences":{"user_consents_to_analytics":null}}}'` (arrange only) and the backend added again from `/conversations/<id>`. A backend that lacks an id answers `[null]`, which now reads as "not found" (only the `This conversation does not exist…` toast), not as incompatible data. No merged change was shown to fix the race itself (#18009 only changed how a `[null]` answer reads; #18024 only moved the redirect into a shared hook, still called after `addBackend`), so a hit (an `ids=` request to `<second-url>`) is a bug to file. At 95115e84f6dd it hit on 1 of 2 first activations from `/conversations/<id>` (one `GET <second-url>/api/conversations?ids=<id>` answered `200`, with no `This conversation does not exist…` toast after it). The hit followed a failed Connect in the same modal (an empty key, `Invalid API key`) before the successful one; the clean retry, which connected once, had none, so a repro should try both sequences. Record the Add bullet as `fail` when the request appears, even though the redirect and the selector look right.
- Toasts are not page or HTTP errors, so `browser errors --app-only` misses them: read `browser toasts --history` after every add or switch, before anything reloads the page (the history starts with each page load).
- `F25.locked-cloud` needs `scripts/static-server.mjs --lock-to-cloud <url>`, which neither `bin/agent-canvas.mjs` nor `control-openhands launch` can pass today; even with it, reconnecting needs a Cloud account.
- The Cloud tab title truncates to `OpenHan…` at 390 px when the Agent-server tab is selected; the layout still fits.
diff --git a/.agents/skills/verify-openhands/references/feature-map/F26-runtime-variants.md b/.agents/skills/verify-openhands/references/feature-map/F26-runtime-variants.md
index 67b1e2757c..2c45ab69f4 100644
--- a/.agents/skills/verify-openhands/references/feature-map/F26-runtime-variants.md
+++ b/.agents/skills/verify-openhands/references/feature-map/F26-runtime-variants.md
@@ -21,14 +21,17 @@ Source: `bin/agent-canvas.mjs`, `scripts/dev-with-automation.mjs`, `scripts/dev-
- `F26.cli-public-needs-key`: `--public` without `LOCAL_BACKEND_API_KEY` exits 1 with guidance.
- `F26.cli-missing-build`: a package without `build/` exits 1 with `No build found` and build instructions.
- `F26.cli-port-in-use`: an occupied ingress (or service) port stops the launch with `Cannot start: the following ports are already in use` naming each busy port.
+- `F26.loopback-bind-default`: without `--host` the ingress listens on `127.0.0.1` only: the machine's own non-loopback address refuses the connection while loopback serves the app.
+- `F26.session-key-rotated`: in local mode a stale session key left in the browser (the launcher's key changed since the last visit) is replaced by the injected key on load: the app opens on Home without the API-key screen, and the stored `Local` backend holds the new key (public mode asks for it instead: F01's api-key-entry row).
- `F26.session-key-persist`: in local mode without `LOCAL_BACKEND_API_KEY` the launcher generates a session key, saves it under `$HOME/.openhands/agent-canvas/api-key.txt` and injects the same key again on the next start.
- `F26.session-key-pinned`: `LOCAL_BACKEND_API_KEY=<key>` in local mode injects exactly that key into the page and the API accepts only it (the saved generated key gets 401).
-- `F26.frontend-only`: `--frontend-only` serves the SPA; `/server_info`, `/api/*`, `/sockets/*` and `/api/automation/*` answer 503, and the UI opens on the **Add a backend** onboarding step without an error toast.
+- `F26.frontend-only`: `--frontend-only` serves the SPA; `/server_info`, `/api/*`, `/sockets/*` and `/api/automation/*` answer 503, and a browser new to that origin opens on the **Add a backend** onboarding step without an error toast.
+- `F26.frontend-only-returning`: on a frontend-only origin, a browser that already stores a backend for that origin (the `default-local` entry a full launcher on the same port seeded, or one added there earlier) gets the recovery gate instead of onboarding: `agent-server-onboarding-screen` with Manage backends in recovery mode and the `Local` row `Disconnected` with the 503 detail (the gate itself is F25's recovery-gate row); the failing probes raise generic `An error occurred` toasts there.
- `F26.backend-only`: `--backend-only` serves the APIs (key-protected) and answers 503 `No backend configured for this route` for `/` and static assets.
- `F26.cross-connect`: a frontend-only UI connects to a separate backend-only instance through **Add a backend**; the shell then shows that backend as Connected and all API traffic goes to it.
- `F26.remote-backend`: from a normal Canvas, backend selector → **Add Backend** → **Agent-server** → **Remote** shows the self-hosting guidance and connects to another instance (stand-in for a self-hosted VM, `docs/SELF_HOSTING.md`).
- `F26.seeded-local-backend`: the first load of a launcher-served UI seeds one backend `Local` (id `default-local`) with the page origin and the injected key.
-- `F26.host-bind-lan`: `--host 0.0.0.0` warns, does not inject the session key, and the UI behaves like `--public` (Add a backend with Next disabled; Skip leads to the API-key screen).
+- `F26.host-bind-lan`: `--host 0.0.0.0` warns, listens on the machine's LAN address, and serves an index.html that carries `__AGENT_CANVAS_AUTH_REQUIRED__` and neither `__AGENT_CANVAS_SESSION_API_KEY__` nor the key itself (the loopback launcher's page carries the key marker); the UI behaves like `--public` (Add a backend with Next disabled; Skip leads to the API-key screen).
- `F26.host-bind-lan-optin`: the opt-in the launcher's warning recommends (`--allow-lan-session-key`) should restore key injection on a LAN bind. Fails today (flag ignored).
- `F26.host-bind-env`: `OH_BIND_HOST=0.0.0.0` (the environment form of `--host`) prints the same two "not loopback" warnings and does not inject the session key, even with `LOCAL_BACKEND_API_KEY` set.
- `F26.runtime-services`: `/server_info.runtime_services` lists the agent-server, ingress, frontend and automation URLs as the agent sees them, and a new conversation's system prompt carries a matching `<RUNTIME_SERVICES>` block.
@@ -70,17 +73,20 @@ Preconditions:
- **Public mode needs a key (`F26.cli-public-needs-key`).** Run `T=$(mktemp -d); env -u LOCAL_BACKEND_API_KEY HOME=$T OH_CANVAS_SAFE_STATE_DIR=$T/state timeout 60 node bin/agent-canvas.mjs --public --port 18990; echo "exit=$?"; rm -rf $T`. After `✓ uvx found` it prints `✗ PUBLIC MODE requires LOCAL_BACKEND_API_KEY environment variable.` and `exit=1`. (The successful `--public` launch is `control-openhands launch --new --public`, mapped in F01.)
- **Missing build (`F26.cli-missing-build`).** Never move this checkout's `build/` (other runs serve it); copy the launcher instead: `T=$(mktemp -d); mkdir -p $T/bin $T/config; cp bin/agent-canvas.mjs $T/bin/; cp package.json $T/; cp config/defaults.json $T/config/; node $T/bin/agent-canvas.mjs --port 18990; echo "exit=$?"; rm -rf $T`. It prints `Error: No build found at <T>/build`, the `npm install` / `npm run build` hint and `exit=1`.
- **Port in use (`F26.cli-port-in-use`).** Point a second launcher at your run's ingress port: `P=$(control-openhands status | jq -r .ports.ingress); T=$(mktemp -d); HOME=$T OH_CANVAS_SAFE_STATE_DIR=$T/state timeout 60 node bin/agent-canvas.mjs --port $P; echo "exit=$?"; rm -rf $T`. It prints `Cannot start: the following ports are already in use:` with `• ingress: port <P>` and `Another agent-canvas instance may already be running.`, then exits 1. Your run is untouched (`control-openhands doctor` stays ok).
+- **Loopback-only ingress (`F26.loopback-bind-default`).** Your run was started without `--host`. Run `P=$(control-openhands status | jq -r .ports.ingress); curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:$P/` (`200`), then find the machine's own address portably, `LAN=$(node -e 'for (const a of Object.values(require("os").networkInterfaces()).flat()) if (a.family === "IPv4" && !a.internal) { console.log(a.address); break }'); echo "LAN=$LAN"`, and `curl -s -o /dev/null -w '%{http_code}\n' --connect-timeout 3 http://$LAN:$P/; echo "exit=$?"`. At that address curl prints `000` and `exit=7` (connection refused). If `LAN` is empty the machine has no non-loopback address: record the loopback half and leave the refused-connection half `not-run` with that reason (it is the behavior this ID asserts, so an empty `LAN` is never a pass). Read-only second view, Linux form: `awk -v p=$(printf '%04X' $P) 'NR>1 && $4=="0A" && $2 ~ ":"p"$" {print $2}' /proc/net/tcp` prints only `0100007F:<hex port>` (`127.0.0.1` in `/proc/net/tcp`'s byte order; elsewhere list the listener with the OS's socket tool, for example `ss -ltn 'sport = :'$P` or `lsof -nP -iTCP:$P -sTCP:LISTEN`, and expect `127.0.0.1:<P>` only), and `grep -a 'Listening on' $OH_VERIFY_RUN/private/stack.log` shows the ingress banner `Listening on: http://localhost:<P>/`. The LAN half of `F26.host-bind-lan` below is the contrast: there the same curl answers `200`.
- **Backend-only (`F26.backend-only`).** Start it: `Q=$OH_VERIFY_RUN/private/f26-backend; mkdir -p $Q; HOME=$Q/home OH_CANVAS_SAFE_STATE_DIR=$Q/state OH_CANVAS_SAFE_BACKEND_PORT=18961 OH_CANVAS_SAFE_AUTOMATION_PORT=18962 OH_CANVAS_SAFE_VITE_PORT=18963 LOCAL_BACKEND_API_KEY=qa-f26-backend-key OH_SECRET_KEY=qa-f26-secret DO_NOT_TRACK=1 nohup node bin/agent-canvas.mjs --backend-only --port 18960 > $Q/launcher.log 2>&1 & echo $! > $Q/pid`. Wait until `curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:18960/health` is `200` (30–90 s; the first start downloads the agent-server into the private `HOME`). Then `curl -s http://127.0.0.1:18960/` prints `No backend configured for this route` (status 503, also for `/index.html` and `/assets/x.js`), `/server_info` is 200, `/api/settings` is 401 without a key and 200 with `-H 'X-Session-API-Key: qa-f26-backend-key'`. Keep it running for the next two bullets.
- **Frontend-only (`F26.frontend-only`).** Start it: `Q=$OH_VERIFY_RUN/private/f26-frontend; mkdir -p $Q; HOME=$Q/home OH_CANVAS_SAFE_STATE_DIR=$Q/state OH_CANVAS_SAFE_BACKEND_PORT=18971 OH_CANVAS_SAFE_AUTOMATION_PORT=18972 OH_CANVAS_SAFE_VITE_PORT=18973 DO_NOT_TRACK=1 nohup node bin/agent-canvas.mjs --frontend-only --port 18970 > $Q/launcher.log 2>&1 & echo $! > $Q/pid` (ready in a few seconds). `for u in / /server_info /api/settings /sockets/events/x /api/automation/v1 /vscode/; do curl -s -o /dev/null -w "%{http_code} $u\n" http://127.0.0.1:18970$u; done` prints `200 /` and `503` for the four backend paths and for `/vscode/`: the launcher serves the VS Code editor and reserves its path, and `grep -a 'vscode -> 503' $Q/launcher.log` prints `[static] /vscode -> 503 (rejected)` (see Gotchas). In the browser: `control-openhands browser goto http://127.0.0.1:18970/ --allow-external`, `control-openhands browser value 'testid=onboarding-backend-name'` (`Local`), `control-openhands browser value 'testid=onboarding-backend-host'` (`http://127.0.0.1:18970`, the page origin), `sleep 12`, `control-openhands browser toasts --history` (`[]`: no error toast) and `control-openhands browser screenshot --feature F26.frontend-only --name add-backend-step` (the **Add a backend** card).
- **Cross-connect (`F26.cross-connect`).** On that page run `control-openhands browser fill 'testid=onboarding-backend-name' 'QA Backend Only'`, `control-openhands browser fill 'testid=onboarding-backend-host' 'http://127.0.0.1:18960'`, `control-openhands browser fill 'testid=onboarding-backend-api-key' qa-f26-backend-key`, `control-openhands browser click 'testid=onboarding-backend-next'` and `control-openhands browser wait 'testid=onboarding-step-choose-agent' --timeout 20000`. On a backend state that has never answered telemetry consent, the consent dialog opens on top (`control-openhands browser wait 'testid=telemetry-consent-form' --timeout 10000` succeeds; it appears a second or two after the step, so an immediate `browser count` can still read `0`; consent is stored on the backend, so a second pass with the same `$Q` skips it): `control-openhands browser uncheck 'testid=telemetry-consent-form >> role=checkbox'`, `control-openhands browser click 'testid=confirm-telemetry-preferences'`, then `control-openhands browser click 'testid=onboarding-skip'` and `control-openhands browser wait 'testid=root-layout' --timeout 20000`. After `control-openhands browser reload`, `control-openhands browser snapshot 'testid=backend-selector'` shows `status "Connected"` and `combobox "QA Backend Only"`. `control-openhands browser goto http://127.0.0.1:18970/settings/secrets --allow-external` and `control-openhands browser count 'testid=secret-item >> has-text=OPENHANDS_AUTOMATION_API_KEY'` is `1` (the backend-only instance's seeded secret), and `control-openhands browser network --filter 18960 --last 5` shows `byOrigin` with only `http://127.0.0.1:18960`. `control-openhands browser errors --app-only` has `pageErrors` `0`.
- **Remote backend from a normal Canvas (`F26.remote-backend`).** Needs the backend-only instance. Run `control-openhands browser goto /`, `control-openhands browser click 'testid=backend-selector'`, `control-openhands browser click 'testid=add-backend-menu-item'`, `control-openhands browser click 'testid=add-backend-option-agent-server'` and `control-openhands browser click 'testid=add-backend-location-option-remote'`. `control-openhands browser text 'testid=add-backend-agent-server-panel'` includes `Run the remote backend with --public, set a strong LOCAL_BACKEND_API_KEY, and expose it through an SSH tunnel, ngrok, or a TLS reverse proxy.` Fill `testid=add-backend-name` with `QA_Remote`, `testid=add-backend-host` with `http://127.0.0.1:18960` and `testid=add-backend-api-key` with `qa-f26-backend-key`, click `testid=add-backend-submit`, then `control-openhands browser wait 'testid=add-backend-modal' --state detached --timeout 15000`, `control-openhands browser reload` and `control-openhands browser snapshot 'testid=backend-selector'`: `status "Connected"`, `combobox "QA_Remote"` (the new backend becomes active). If you run this bullet without Cross-connect first, the backend-only state has not answered telemetry consent yet: the consent dialog (`This preference is saved for the local backend “QA_Remote” at http://127.0.0.1:18960.`) opens over the page and blocks the selector; answer it as in Cross-connect (`control-openhands browser wait 'testid=telemetry-consent-form' --timeout 10000`, `control-openhands browser uncheck 'testid=telemetry-consent-form >> role=checkbox'`, `control-openhands browser click 'testid=confirm-telemetry-preferences'`). Switch back with `control-openhands browser click 'testid=backend-selector'` and `control-openhands browser click 'role=option[name="Local"]'`, then `control-openhands browser wait 'testid=backend-selector >> role=combobox[name="Local"]' --timeout 5000` (the switch applies after a short delay: a snapshot taken right after the click still reads `combobox "QA_Remote": Local`; after the wait it shows `combobox "Local"`). Clean up: `backend-selector` → `testid=manage-backends-menu-item`; `control-openhands browser text 'testid=manage-backends-row-QA_Remote'` reads `QA_Remote`, `v<version>`, `http://127.0.0.1:18960`, `Connected`, `LOCAL`, where `<version>` is the backend-only instance's own report, `curl -s http://127.0.0.1:18960/server_info | jq -r .version` (the `versions.agentServer` pin in `config/defaults.json` unless your shell exports an `OH_AGENT_SERVER_*` override, which a hand-started launcher inherits); it passes the UI's gate: `printf '%s\n' "$(jq -r .compatibility.minimumAgentServer config/defaults.json)" "$(curl -s http://127.0.0.1:18960/server_info | jq -r .version)" | sort -VC && echo meets-minimum || echo below-minimum` prints `meets-minimum`. Then click `testid=manage-backends-remove-QA_Remote`, `testid=confirmation-modal >> testid=confirm-button`, `testid=manage-backends-done`, `browser reload`.
- **Seeded Local backend (`F26.seeded-local-backend`).** On your run, `control-openhands browser eval "JSON.parse(localStorage.getItem('openhands-backends')).map(b=>({id:b.id,name:b.name,host:b.host,kind:b.kind,hasKey:!!b.apiKey}))"` returns exactly one entry `{id: "default-local", name: "Local", host: "http://127.0.0.1:<ingress>", kind: "local", hasKey: true}` (after the Remote cleanup above), and `control-openhands browser eval "({key: Boolean(window.__AGENT_CANVAS_SESSION_API_KEY__), authRequired: window.__AGENT_CANVAS_AUTH_REQUIRED__ === true})"` is `{key: true, authRequired: false}`.
+- **Rotated session key (`F26.session-key-rotated`).** Still on your run's `/`, `control-openhands browser eval "JSON.parse(localStorage.getItem('openhands-backends'))[0].apiKey === window.__AGENT_CANVAS_SESSION_API_KEY__"` is `true` (compare keys inside the page; never print one). Run `md5sum $OH_VERIFY_RUN/private/session-key | cut -c1-12`, then `control-openhands restart --rotate-key` (`rotatedKey` `true`: the launcher comes back injecting a new key) and the `md5sum` line again; the hash changed. The open page still holds the old key, so its polls get `401` until it reloads: run `control-openhands browser errors --clear`, then `control-openhands browser reload` and `control-openhands browser wait 'testid=home-screen' --timeout 20000`. `control-openhands browser count` is `1` for `testid=root-layout` and `0` for `testid=api-key-entry-screen`, `testid=agent-server-onboarding-screen` and `testid=onboarding-modal`: Home, not the key prompt public mode shows in the same state (`F01.api-key-entry`). Second view: `control-openhands browser eval "JSON.parse(localStorage.getItem('openhands-backends')).map(b=>({id:b.id,host:b.host,keyMatchesPage:b.apiKey===window.__AGENT_CANVAS_SESSION_API_KEY__}))"` is `[{id: "default-local", host: "http://127.0.0.1:<ingress>", keyMatchesPage: true}]` against the new page key, `control-openhands browser eval "JSON.parse(localStorage.getItem('openhands-agent-server-config')).sessionApiKey === window.__AGENT_CANVAS_SESSION_API_KEY__"` is `true` (the legacy key is overwritten too), `control-openhands browser network --filter 'api/settings' --last 2` shows `status` `200` under `recent`, and `control-openhands browser errors --app-only` has `pageErrors` `0` and `appErrors` `0`. `control-openhands api GET /api/settings` works with the rotated key, and `control-openhands browser screenshot --feature F26.session-key-rotated --name home-after-rotation` shows Home.
- **Stop the partial stacks.** `kill -TERM $(cat $OH_VERIFY_RUN/private/f26-frontend/pid) $(cat $OH_VERIFY_RUN/private/f26-backend/pid)`; after about 5 s the ports 18960–18963 and 18970–18973 are free again (check as in Preconditions). Kill by the saved pid only, never by pattern.
-- **LAN bind (`F26.host-bind-lan`).** Start a full stack on all interfaces: `Q=$OH_VERIFY_RUN/private/f26-lan; mkdir -p $Q; HOME=$Q/home OH_CANVAS_SAFE_STATE_DIR=$Q/state OH_CANVAS_SAFE_BACKEND_PORT=18961 OH_CANVAS_SAFE_AUTOMATION_PORT=18962 OH_CANVAS_SAFE_VITE_PORT=18963 LOCAL_BACKEND_API_KEY=qa-f26-lan-key OH_SECRET_KEY=qa-f26-secret DO_NOT_TRACK=1 nohup node bin/agent-canvas.mjs --host 0.0.0.0 --port 18960 > $Q/launcher.log 2>&1 & echo $! > $Q/pid`, and wait until `curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:18960/` is `200`. `grep -a 'not loopback' $Q/launcher.log` shows `[auth] Bind host 0.0.0.0 is not loopback — session key will not be injected into HTML` and `[static] WARNING: bind host 0.0.0.0 is not loopback; refusing to inject the session API key into HTML.` Then `control-openhands browser goto http://127.0.0.1:18960/ --allow-external` and `control-openhands browser eval "({key: Boolean(window.__AGENT_CANVAS_SESSION_API_KEY__), authRequired: window.__AGENT_CANVAS_AUTH_REQUIRED__ === true})"` → `{key: false, authRequired: true}`. The onboarding shows `testid=onboarding-step-check-backend` with `control-openhands browser enabled 'testid=onboarding-backend-next'` `false` (no key typed). `control-openhands browser click 'testid=onboarding-skip'` and `control-openhands browser wait 'testid=api-key-entry-screen' --timeout 15000` succeed; after `sleep 12`, `control-openhands browser toasts --history` is `[]` (no error toast on either screen); `control-openhands browser screenshot --feature F26.host-bind-lan --name api-key-screen`. Fill `testid=api-key-entry-name` with `QA_LAN` and `testid=api-key-entry-api-key` with `qa-f26-lan-key`, click `testid=api-key-entry-submit`; `control-openhands browser wait 'testid=root-layout' --timeout 10000` succeeds. Stop it with `kill -TERM $(cat $Q/pid)`.
+- **LAN bind (`F26.host-bind-lan`).** Start a full stack on all interfaces: `Q=$OH_VERIFY_RUN/private/f26-lan; mkdir -p $Q; HOME=$Q/home OH_CANVAS_SAFE_STATE_DIR=$Q/state OH_CANVAS_SAFE_BACKEND_PORT=18961 OH_CANVAS_SAFE_AUTOMATION_PORT=18962 OH_CANVAS_SAFE_VITE_PORT=18963 LOCAL_BACKEND_API_KEY=qa-f26-lan-key OH_SECRET_KEY=qa-f26-secret DO_NOT_TRACK=1 nohup node bin/agent-canvas.mjs --host 0.0.0.0 --port 18960 > $Q/launcher.log 2>&1 & echo $! > $Q/pid`, and wait until `curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:18960/` is `200`. `grep -a 'not loopback' $Q/launcher.log` shows `[auth] Bind host 0.0.0.0 is not loopback — session key will not be injected into HTML` and `[static] WARNING: bind host 0.0.0.0 is not loopback; refusing to inject the session API key into HTML.` The served page proves it: `curl -s http://127.0.0.1:18960/ | grep -o '__AGENT_CANVAS_\(SESSION_API_KEY\|AUTH_REQUIRED\)__' | sort -u` prints only `__AGENT_CANVAS_AUTH_REQUIRED__` and `curl -s http://127.0.0.1:18960/ | grep -c qa-f26-lan-key` prints `0`, while your loopback run's page, `curl -s "$(control-openhands status | jq -r .baseUrl)/" | grep -o '__AGENT_CANVAS_\(SESSION_API_KEY\|AUTH_REQUIRED\)__' | sort -u`, prints only `__AGENT_CANVAS_SESSION_API_KEY__`. The bind is real: with `LAN` from the loopback bullet, `curl -s -o /dev/null -w '%{http_code}\n' --connect-timeout 3 http://$LAN:18960/` is `200` here (refused on your run, `F26.loopback-bind-default`; with an empty `LAN` this check is `not-run`). The browser part below expects a profile that has never visited `127.0.0.1:18960` (true in this file's order; on a repeat pass run `control-openhands browser reset` first and restore your run afterwards, see Gotchas). Then `control-openhands browser goto http://127.0.0.1:18960/ --allow-external` and `control-openhands browser eval "({key: Boolean(window.__AGENT_CANVAS_SESSION_API_KEY__), authRequired: window.__AGENT_CANVAS_AUTH_REQUIRED__ === true})"` → `{key: false, authRequired: true}`. The onboarding shows `testid=onboarding-step-check-backend` with `control-openhands browser enabled 'testid=onboarding-backend-next'` `false` (no key typed). `control-openhands browser click 'testid=onboarding-skip'` and `control-openhands browser wait 'testid=api-key-entry-screen' --timeout 15000` succeed; after `sleep 12`, `control-openhands browser toasts --history` is `[]` (no error toast on either screen); `control-openhands browser screenshot --feature F26.host-bind-lan --name api-key-screen`. Fill `testid=api-key-entry-name` with `QA_LAN` and `testid=api-key-entry-api-key` with `qa-f26-lan-key`, click `testid=api-key-entry-submit`; `control-openhands browser wait 'testid=root-layout' --timeout 10000` succeeds. Stop it with `kill -TERM $(cat $Q/pid)`.
- **LAN opt-in (`F26.host-bind-lan-optin`).** Expected: the opt-in the warning names (`Pass --host 127.0.0.1 (default) for local mode, or --allow-lan-session-key only if you accept LAN exposure.`) makes the npm launcher inject the key again. Start the LAN stack as above with `--host 0.0.0.0 --allow-lan-session-key --port 18960` (log to `$Q/launcher2.log`), wait for `200`, then `curl -s http://127.0.0.1:18960/ | grep -o '__AGENT_CANVAS_\(SESSION_API_KEY\|AUTH_REQUIRED\)__' | sort -u`. Today it prints only `__AGENT_CANVAS_AUTH_REQUIRED__` and the log repeats both "not loopback" warnings: the launcher ignores the flag (fail; see Gotchas). Stop it with `kill -TERM $(cat $Q/pid)`.
- **Generated session key persists (`F26.session-key-persist`).** With 18960–18963 free, define a helper and start a loopback stack without a key: `Q=$OH_VERIFY_RUN/private/f26-keys; mkdir -p $Q; start(){ HOME=$Q/home OH_CANVAS_SAFE_STATE_DIR=$Q/state OH_CANVAS_SAFE_BACKEND_PORT=18961 OH_CANVAS_SAFE_AUTOMATION_PORT=18962 OH_CANVAS_SAFE_VITE_PORT=18963 OH_SECRET_KEY=qa-f26-secret DO_NOT_TRACK=1 "$@" > $Q/launcher.log 2>&1 & echo $! > $Q/pid; for i in $(seq 1 60); do c=$(curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:18960/); [ "$c" = 200 ] && break; sleep 3; done; echo "code=$c"; }; start env -u LOCAL_BACKEND_API_KEY nohup node bin/agent-canvas.mjs --port 18960` (`code=200`). `curl -s http://127.0.0.1:18960/ | grep -o '__AGENT_CANVAS_SESSION_API_KEY__[^;<]*' | md5sum` prints a hash (never print the key itself); `$Q/home/.openhands/agent-canvas/api-key.txt` exists (64 hex characters) and `curl -s -o /dev/null -w '%{http_code}' -H "X-Session-API-Key: $(cat $Q/home/.openhands/agent-canvas/api-key.txt)" http://127.0.0.1:18960/api/settings` is `200`. `kill -TERM $(cat $Q/pid); sleep 6`, run the same `start ...` line again in the same shell and repeat the `md5sum` line: the hash is identical.
- **Pinned session key (`F26.session-key-pinned`).** Same shell, after `kill -TERM $(cat $Q/pid); sleep 6`: `start env LOCAL_BACKEND_API_KEY=qa-f26-pin-key nohup node bin/agent-canvas.mjs --port 18960`. `curl -s http://127.0.0.1:18960/ | grep -o '__AGENT_CANVAS_SESSION_API_KEY__[^;<]*'` prints `__AGENT_CANVAS_SESSION_API_KEY__="qa-f26-pin-key"`; `/api/settings` is `200` with `-H 'X-Session-API-Key: qa-f26-pin-key'` and `401` with the saved generated key. In the browser, `control-openhands browser goto http://127.0.0.1:18960/ --allow-external` and `control-openhands browser eval "({key: window.__AGENT_CANVAS_SESSION_API_KEY__ === 'qa-f26-pin-key', authRequired: window.__AGENT_CANVAS_AUTH_REQUIRED__ === true})"` → `{key: true, authRequired: false}`.
- **Bind host from the environment (`F26.host-bind-env`).** Same shell, after `kill -TERM $(cat $Q/pid); sleep 6`: `start env LOCAL_BACKEND_API_KEY=qa-f26-pin-key OH_BIND_HOST=0.0.0.0 nohup node bin/agent-canvas.mjs --port 18960`. `grep -a 'not loopback' $Q/launcher.log` shows the same `[auth]` and `[static]` warnings as `--host 0.0.0.0`, and `curl -s http://127.0.0.1:18960/ | grep -o '__AGENT_CANVAS_\(SESSION_API_KEY\|AUTH_REQUIRED\)__' | sort -u` prints only `__AGENT_CANVAS_AUTH_REQUIRED__`. Stop it with `kill -TERM $(cat $Q/pid)`; after about 6 s ports 18960–18963 are free.
+- **Frontend-only, returning user (`F26.frontend-only-returning`).** Same shell, after `kill -TERM $(cat $Q/pid); sleep 6`. Arrange a browser that onboarded on this origin against the full launcher: `start env LOCAL_BACKEND_API_KEY=qa-f26-pin-key nohup node bin/agent-canvas.mjs --port 18960` (`code=200`), `control-openhands browser reset` (a fresh profile, which also forgets your run's onboarding: restore it at the end), `control-openhands browser goto http://127.0.0.1:18960/ --allow-external`, `control-openhands browser wait 'testid=telemetry-consent-form' --timeout 15000` (this backend state has never answered consent; a second pass on the same `$Q` skips the form), `control-openhands browser uncheck 'testid=telemetry-consent-form >> role=checkbox'`, `control-openhands browser click 'testid=confirm-telemetry-preferences'`, `control-openhands browser wait 'testid=onboarding-step-choose-agent' --timeout 15000`, `control-openhands browser click 'testid=onboarding-skip'` and `control-openhands browser wait 'testid=root-layout' --timeout 20000`. `control-openhands browser eval "JSON.parse(localStorage.getItem('openhands-backends')).map(b=>({id:b.id,name:b.name,host:b.host}))"` is `[{id: "default-local", name: "Local", host: "http://127.0.0.1:18960"}]`. Now serve the same port without backends: `kill -TERM $(cat $Q/pid); sleep 6; start nohup node bin/agent-canvas.mjs --frontend-only --port 18960` (`code=200`, and `curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:18960/server_info` is `503`). `control-openhands browser goto http://127.0.0.1:18960/ --allow-external` and `control-openhands browser wait 'testid=agent-server-onboarding-screen' --timeout 20000` succeed: the gate of `F25.recovery-gate`, not onboarding. `control-openhands browser count` is `1` for `testid=manage-backends-modal` and `0` for `testid=onboarding-step-check-backend`, `testid=first-run-onboarding-screen`, `testid=api-key-entry-screen`, `testid=root-layout`, `testid=close-manage-backends-modal` and `testid=manage-backends-done`; `control-openhands browser press Escape` leaves the modal (`count` `1`). `control-openhands browser text 'testid=manage-backends-row-Local'` reads `Local`, `http://127.0.0.1:18960`, `Disconnected`, `HTTP request failed (503 Service Unavailable): "Service Unavailable (no backend configured for this route)"` and `LOCAL`, and `control-openhands browser testids` lists `manage-backends-add` (**Add Backend**) as the way on. `control-openhands browser screenshot --feature F26.frontend-only-returning --name recovery-gate`. The 503 probes are expected errors in the sweep: `control-openhands browser errors` (without `--app-only`, which keeps only your run's origin) lists `external:http-error` rows with `503` for `http://127.0.0.1:18960/server_info`, `/api/settings` and `/api/llm/models/verified`, and `pageErrors` is `0`. Stop it with `kill -TERM $(cat $Q/pid)`; after about 6 s ports 18960–18963 are free. Restore your run: `control-openhands browser goto /` and `control-openhands onboard --skip`.
- **Runtime services (`F26.runtime-services`).** Run `control-openhands api GET /server_info --pick runtime_services.mode` (`agent-canvas`) and `control-openhands api GET /server_info --pick runtime_services.services.automation.url_from_agent` (`http://localhost:<automation port>`; `control-openhands status` prints the ports). Then `control-openhands conversation start --prompt "Reply with the single word: ok" --wait --timeout 180` (prints `<id>`) and `control-openhands conversation events <id> --grep RUNTIME_SERVICES --from-start`: the `SystemPromptEvent` matches, with an excerpt `<RUNTIME_SERVICES>\nYou are running inside an agent-canvas dev stack started ...`. `control-openhands conversation events <id> --grep "localhost:<automation port>" --from-start` shows `* Automation backend: http://localhost:<port>` plus its `Docs:` and `OpenAPI:` URLs.
- **Library build (`F26.lib-build`).** Never build inside the shared checkout (it regenerates i18n and typegen files the other runs hash); build in a copy: `L=$OH_VERIFY_RUN/private/f26-lib; mkdir -p $L; tar --exclude=./node_modules --exclude=./build --exclude=./.git --exclude=./dist -cf - . | tar -xf - -C $L; ln -s "$PWD/node_modules" $L/node_modules; (cd $L && npm run build:lib); echo "exit=$?"`. It ends with `✓ built in` and `exit=0` (about 30 s). Then `(cd $L && node -e 'const p=require("./package.json"),fs=require("fs");let ok=0,miss=[];for(const[k,v]of Object.entries(p.exports))for(const f of(typeof v==="string"?[v]:Object.values(v)))fs.existsSync(f)?ok++:miss.push(k+" -> "+f);console.log(JSON.stringify({entries:Object.keys(p.exports),ok,missing:miss}))')` prints 9 entries (`.`, `./browser`, `./conversation`, `./files`, `./settings`, `./sidebar`, `./terminal`, `./i18n`, `./package.json`), `ok` 25 and `missing` `[]`. `grep -ohE 'AgentServerUIProviders|AgentServerUIRoot|CloudOrganizationBoundary|configureTelemetry|setTelemetryConsent' $L/dist/lib/index.d.ts | sort -u` lists all five. Remove the copy with `rm -rf $L`.
@@ -96,6 +102,8 @@ Preconditions:
- The browser daemon refuses other origins unless you pass `--allow-external` to `browser goto`. Each port is its own origin, so a partial-stack page starts at first run again (onboarding and a consent dialog), independent of your run's onboarding.
- In a frontend-only UI the **Add a backend** step's Next is enabled with an empty key, while on a LAN or `--public` stack it is disabled until a key is typed: the host the step probes differs (503 vs. auth-required).
- A partial-stack origin keeps its stored backend in the browser profile across launches, so a second pass lands straight in the shell. `control-openhands browser reset` brings back the first run on every origin; then `control-openhands browser goto /` and `control-openhands onboard --skip` restore your own run's baseline.
+- Which screen a frontend-only origin shows depends on the browser, not the launcher: a profile new to the origin gets **Add a backend** (first run owns the initial backend), while a profile that stores a backend for that origin (the `default-local` entry a full launcher on the same port seeded, or any backend added there earlier) gets the recovery gate of `F25.recovery-gate` as soon as that backend answers 503 (`F26.frontend-only-returning`). The no-toast check belongs to the first-run state only: on the gate, `browser toasts --history` lists generic `An error occurred` status toasts raised by the failing 503 probes (two within 12 s here), with no hint of which request failed.
+- Stopping a backend-only instance while the frontend-only page that uses it stays open is the same backend-down state as `F25.recovery-gate` after a reload; keep the returning-user bullet above as the frontend-only-specific check (it reproduces the `default-local` case).
- The launcher serves the bundled VS Code editor. #17660 made that opt-in (`OH_CANVAS_ENABLE_VSCODE=true`) and #18048 reverted it, so only a checkout between the two needs the variable (`launch --vscode` sets it). A launcher that serves it reserves `/vscode/`: a frontend-only stack answers `503` there and logs `[static] /vscode -> 503 (rejected)`; a stack with an agent-server lists `/vscode → http://127.0.0.1:<agent-server port + 1000>` in its ingress banner and hands the agent-server that editor port (19961 for the hand-started launchers here, 19000 for the desktop app), which the launcher's own port check does not cover. On a checkout between the two without the variable, `/vscode/` is an ordinary SPA URL (`200`). Here the agent-server has no editor binary (`VSCode server binary not found, VSCode will be disabled` in the launcher log, `/api/vscode/status` `{"running":false,"enabled":true}`), so the editor port stays free even then.
- The port-in-use error is printed twice, the second time with a Node stack trace, and suggests `PORT=<other>` rather than `--port` (#17949). The `--public` key error's example says `npm run dev -- --public` instead of the npx command.
- A backend-only instance's `/server_info.runtime_services` still describes the ingress as routing `/* to the frontend`, which it does not run (#17949).
diff --git a/.agents/skills/verify-openhands/references/feature-map/F27-workspace-tools.md b/.agents/skills/verify-openhands/references/feature-map/F27-workspace-tools.md
index 60cb3c1263..8e7da025a2 100644
--- a/.agents/skills/verify-openhands/references/feature-map/F27-workspace-tools.md
+++ b/.agents/skills/verify-openhands/references/feature-map/F27-workspace-tools.md
@@ -35,6 +35,7 @@ Source: `src/components/features/terminal/`, `src/hooks/use-terminal.ts`, `src/c
- `F27.overview-pin`: the Overview "..." menu pins and unpins rows (Workspace, Git parts); an unpinned row stays hidden after a reload.
- `F27.overview-changes-link`: the Overview Changes row closes the Overview and opens the drawer on Commits.
- `F27.overview-menu-open`: the Overview "..." menu's Git > Commits item opens the drawer on Commits with Uncommitted collapsed; Git > Changes opens Commits with Uncommitted expanded; both close the Overview.
+- `F27.overview-identity`: the Overview's Workspace row shows the attached folder's basename (`None` without a workspace), and once the workspace's `origin` is a GitHub remote the Git block adds an `owner/repo` link to the repository and a branch link.
- `F27.overview-drawer`: Overview menu > Pull Requests opens the secondary drawer with a close button; without a git repository it reads "Connect a git repository to view pull requests and issues".
- `F27.canvas-ui-open-tab`: the agent's `canvas_ui_control` `open_tab` opens the drawer on the requested tab (files, browser, terminal, planner, tasklist), and the choice persists like a user tab click.
- `F27.canvas-ui-navigate-file`: `canvas_ui_control` `navigate_to_file` (and clicking a file path in chat) opens Files with that file selected.
@@ -64,6 +65,7 @@ Preconditions:
- `F27.plugins-modal` needs a conversation started with at least one plugin (the menu item is hidden otherwise).
- `F27.terminal-waiting` needs a conversation in the error state: arrange one with a profile whose key is invalid (no model cost; see the bullet).
- `F27.usage-provider-balance` (card shown) needs an OpenRouter key; with DeepSeek only the hidden case is checkable.
+- `F27.overview-identity` needs a repo with a GitHub origin, its own so the shared `qa-repo` of F08 keeps none: `control-openhands fixture git-repo --name qa-f27-repo --remote https://github.com/qa-example/qa-f27-repo.git` (`--remote` only sets `origin`, also when the repo already exists, which the output reports as `remoteChanged`; no model is needed for the rows).
- **Empty tools (`F27.terminal-empty`, `F27.browser-empty`, `F27.planner-empty`).** Run `control-openhands conversation start --prompt "Reply with only the word OK. Do not run any tools." --wait --timeout 240` (note the `id` as `<quiet-id>`), then `control-openhands browser click 'testid=right-panel-toggle'`, `control-openhands browser click 'testid=conversation-tab-terminal'` and `control-openhands browser wait-text 'No terminal output yet' --timeout 10000`. Then `control-openhands browser click 'testid=conversation-tab-browser'`, `control-openhands browser text 'testid=browser-chrome-url'` (`No URL loaded`), `control-openhands browser count 'testid=browser-chrome-open-external'` (`0`), `control-openhands browser eval "[...document.querySelectorAll('[data-testid=browser-chrome-bar] button')].map(b=>[b.disabled,b.getAttribute('aria-label')])"` (`[[true,"Open in New Tab"]]`) and `control-openhands browser wait-text 'No page loaded yet' --timeout 5000`. Then `control-openhands browser click 'testid=conversation-tab-planner'`, `control-openhands browser wait-text 'There is currently no plan for this repo' --timeout 10000`, `control-openhands browser enabled 'role=button[name="Create a plan"]'` (`true`) and `control-openhands browser enabled 'testid=planner-tab-build-button'` (`false`). The tab bar has no `conversation-tab-tasklist`. Screenshot with `control-openhands browser screenshot --feature F27.browser-empty --name empty`.
- **Usage totals (`F27.usage-metrics`, `F27.usage-provider-balance`).** In `<quiet-id>` run `control-openhands browser click 'testid=conversation-tab-usage'`, `control-openhands browser text 'testid=usage-panel'` and `control-openhands browser attr 'testid=context-meter-bar' style`. The text reads `Context Window`, `1% used (99% left)`, `<tokens> / 1,000,000`, `Compact context`, `Token Usage`, Input, Cache Hit, Cache Write, Output, Total and `Total Cost $0.00…`; the bar style is `width: <percent>%`. `control-openhands browser count 'testid=provider-balance-card'` is `0` with DeepSeek (the `GET /api/llm/balance` 404 is expected). Screenshot with `control-openhands browser screenshot --feature F27.usage-metrics --name usage`.
@@ -79,6 +81,7 @@ Preconditions:
- **Launch a child (`F27.launch-child`).** On `/conversations/<tools-id>` (code mode, not a plan-mode conversation) run `control-openhands conversation start --stay --prompt "Call the launch_child_conversation tool with target local, isolation shared and task 'Reply with only the word PONG. Do not run any tools.' Then reply OK."` and poll `control-openhands browser toasts` every few seconds: one toast reads `Launched a local child conversation Open`. Take the child id from `control-openhands conversation events <tools-id> --kinds MessageEvent` (the `[child-conversation] {"status":"launched",…,"conversation_id":"<child-id>"…}` message), then `control-openhands conversation wait <child-id> --timeout 120` (`finished`) and `control-openhands conversation events <child-id> --kinds MessageEvent` (agent `PONG`). `control-openhands browser count 'testid=conversation-card >> has-text=PONG'` is `1`.
- **Overview (`F27.overview-peek`, `F27.overview-pin`, `F27.overview-changes-link`, `F27.overview-drawer`).** On `/conversations/<quiet-id>` (`control-openhands browser goto /conversations/<quiet-id>`; Launch a child left `<tools-id>`) run `control-openhands browser click 'testid=right-panel-toggle'` (the peek exists only while the drawer is open; a fresh page load has it closed), `control-openhands browser hover 'testid=conversation-overview-toggle'` and `control-openhands browser text 'testid=conversation-overview-peek'` (`Overview`, `Changes`, `+0`, `-0`, `Workspace`, `None`). `control-openhands browser click 'testid=conversation-overview-toggle'`; `control-openhands browser attr 'testid=conversation-overview-toggle' aria-pressed` is `true` and `right-panel-toggle` is `false`. Pin: `control-openhands browser click 'testid=conversation-overview-ellipsis'`, `control-openhands browser click 'testid=conversation-overview-menu-pin-workspace'`; `control-openhands browser count 'testid=conversation-overview-workspace'` is `0` and the pin button's `aria-pressed` is `false`; after `control-openhands browser reload` and `browser click 'testid=conversation-overview-toggle'` the count is still `0`. Re-pin the same way (count `1`). Git parts work the same: `conversation-overview-menu-pin-git-changes` hides `testid=conversation-overview-diffs` (count `0`, also after a reload) and a second click restores it. Close the menu with `control-openhands browser press Escape`: `control-openhands browser count 'testid=conversation-overview-context-menu'` is `0`, `control-openhands browser eval "document.activeElement.dataset.testid"` is `conversation-overview-ellipsis` and `testid=conversation-overview-panel` still counts `1` (a second click on the ellipsis closes the menu too). Changes: `control-openhands browser click 'testid=conversation-overview-diffs'`; the overview panel count is `0`, `right-panel-toggle` is pressed and `tabs-pane-header` starts with `Commits`. Drawer: reopen the Overview, `control-openhands browser click 'testid=conversation-overview-ellipsis'`, `control-openhands browser click 'testid=conversation-overview-menu-open-git-pull_requests'`, `control-openhands browser text 'testid=conversation-overview-drawer'` (`Pull Requests` / `Connect a git repository to view pull requests and issues`), then `control-openhands browser click 'testid=conversation-overview-drawer-close'` and `control-openhands browser wait 'testid=conversation-overview-drawer' --state hidden --timeout 5000` (an immediate `visible` still reads `true` during the slide-out animation).
- **Overview menu opens Commits (`F27.overview-menu-open`).** On `/conversations/<tools-id>` (it has an uncommitted `.agents_tmp/PLAN.md`; `control-openhands browser goto /conversations/<tools-id>`, since the Overview bullet is on `<quiet-id>`) run `control-openhands browser reload`, `control-openhands browser click 'testid=conversation-overview-toggle'`, `control-openhands browser click 'testid=conversation-overview-ellipsis'` and `control-openhands browser click 'testid=conversation-overview-menu-open-git-commits'`: `control-openhands browser count 'testid=conversation-overview-panel'` is `0`, `tabs-pane-header` starts with `Commits` and `control-openhands browser attr 'testid=uncommitted-changes-row-toggle' aria-expanded` is `false`. Repeat with `conversation-overview-menu-open-git-changes`: same tab, `aria-expanded` `true` and `control-openhands browser count 'testid=uncommitted-changes-row-content'` `1` (the `.agents_tmp/PLAN.md` row). Screenshot with `control-openhands browser screenshot --feature F27.overview-menu-open --name changes`.
+- **Workspace and repository identity (`F27.overview-identity`).** Run `control-openhands conversation start --workspace qa-f27-repo --prompt "List the files" --wait --timeout 240` (note `<repo-id>`) and, staying on the page it leaves open, `control-openhands browser click 'testid=conversation-overview-toggle'`, `control-openhands browser wait 'testid=conversation-overview-git-repo' --timeout 20000`, `control-openhands browser text 'testid=conversation-overview-workspace'` (`Workspace`, `qa-f27-repo`: the folder's basename, not its path), `control-openhands browser text 'testid=conversation-overview-git-repo'` (`qa-example/qa-f27-repo`), `control-openhands browser attr 'testid=conversation-overview-git-repo' href` (`https://github.com/qa-example/qa-f27-repo`), `control-openhands browser text 'testid=conversation-overview-git-branch'` (`main`; `control-openhands browser attr 'testid=conversation-overview-git-branch' href` is `https://github.com/qa-example/qa-f27-repo/tree/main`) and `control-openhands browser snapshot 'testid=conversation-overview-panel'`: Changes `+0 -0`, the repo and branch links, the `Commits` and `Pull Requests` items, then `Workspace qa-f27-repo`. `control-openhands browser screenshot --feature F27.overview-identity --name repo-rows`. The repo and branch rows come from a git probe the page runs while the conversation is active: without an LLM key the conversation ends in `error`, and after `control-openhands browser reload` and `control-openhands browser click 'testid=conversation-overview-toggle'` the panel lists only `Workspace qa-f27-repo` (`control-openhands browser count 'testid=conversation-overview-git-repo'` is `0`; see Gotchas), so read the rows before any reload.
- **No browser tool here (`F27.browser-screenshot`, blocked).** `control-openhands conversation start --prompt "Use your browser tool to navigate to https://example.com and then reply with the page title only." --wait --timeout 240` made the agent fall back to `curl`. Confirm with `control-openhands browser click 'testid=chat-plus-button'`, `control-openhands browser click 'testid=show-agent-tools-button'`, `control-openhands browser click 'testid=system-message-modal >> role=tab[name="Available Tools"]'` and `control-openhands browser snapshot 'testid=system-message-modal'`: the tools are terminal, file_editor, task_tracker, canvas_ui_control, launch_child_conversation, finish, think, switch_llm, invoke_skill (no browser tools). Close with `control-openhands browser click 'testid=close-system-message-modal'`. With a working browser tool, expect `browser-chrome-url` to hold the URL and `browser-chrome-open-external` count `1`.
- **Info modals at phone width (`F27.info-modals-phone`).** Run `control-openhands browser viewport phone`; for each of `skills`/`skills-modal`, `hooks`/`hooks-modal`, `agent-tools`/`system-message-modal`: `control-openhands browser click 'testid=chat-plus-button'`, `control-openhands browser click 'testid=show-skills-button'`, `control-openhands browser wait 'testid=skills-modal' --timeout 10000`, `control-openhands browser bbox 'testid=skills-modal'`, `control-openhands browser screenshot --feature F27.info-modals-phone --name skills`, `control-openhands browser click 'testid=close-skills-modal'` (substitute the names). Each box is 351 px wide at x 19.5 with `insideViewport` `true` and `pageHorizontalOverflow` `false`.
- **Agent tab at phone width (`F27.canvas-ui-phone`).** Still at phone width on `/conversations/<id>`, run `control-openhands conversation start --stay --prompt "Call the canvas_ui_control tool with command open_tab and tab terminal, then reply OK." --wait --timeout 200`, then `control-openhands browser url` (it ends `/conversations/<id>/panel`), `control-openhands browser count 'testid=tabs-pane-header'` (`1`) and `control-openhands browser attr 'role=tab[name="Terminal"]' aria-selected` (`true`). The panel page has no `right-panel-toggle`. Screenshot with `control-openhands browser screenshot --feature F27.canvas-ui-phone --name panel-terminal`.
@@ -98,6 +101,7 @@ Preconditions:
- The Overview column is 28 px narrower than its content (`scrollWidth` 296, `clientWidth` 268). After some clicks inside it, Playwright's scroll-into-view shifts it and the card looks cut off on the left; `browser reload` restores it before taking screenshots.
- The agent may summarize a client-tool action with its own title (for example "Launch local child conversation to reply PONG") instead of "Launching a child conversation"; assert the toast and events, not the chat title.
- The Overview peek appears only while the drawer is open (`canPeekOnHover` needs `isRightPanelShown`); a `browser goto` or reload starts with the drawer closed.
+- The Overview's repo and branch rows depend on a bash probe (`useLocalGitInfo`) that runs only while the conversation's status is idle, running, waiting for confirmation or finished and the agent is not in error. Right after `conversation start` the probe has run and its result stays cached on the page, but on an errored conversation a reload hides the rows again, and `conversation pause` does not apply to an errored conversation. Read the rows before any `goto` or `reload`.
- At phone width (and in a narrow drawer) the tab row shows only the tabs that fit; the rest are reachable only through the tabs menu (`conversation-tabs-menu-open-<tab>`), and which tabs fit depends on the active tab's label width.
- Plain `agent-canvas` instances on one machine share the tmux socket `/tmp/tmux-<uid>/openhands` (OpenHands/OpenHands#17946), so another instance stopping can reset this one's terminal. `control-openhands launch` gives each run its own `TMUX_TMPDIR`, so in verification runs a reset means this run's agent server restarted.
- Agent Tools & Metadata remembers the last tab (it reopens on Available Tools).
diff --git a/.agents/skills/verify-openhands/references/feature-map/README.md b/.agents/skills/verify-openhands/references/feature-map/README.md
index 7105876fdd..47a3c717af 100644
--- a/.agents/skills/verify-openhands/references/feature-map/README.md
+++ b/.agents/skills/verify-openhands/references/feature-map/README.md
@@ -102,7 +102,10 @@ Maintenance baseline: main@ed815e141409c7b52991ab8d13c553b595da9165 (2026-10-06)
`control-openhands conversation start --workspace qa-repo --prompt ...`. It
drives Open Workspace and the folder browser, which has no path field
(`workspace open qa-repo` does only the picking). `--remote` only sets
- `origin`, which is enough for the repo/branch links and Pull/Push chips.
+ `origin`, which is enough for the repo/branch links and Pull/Push chips;
+ on a repo that already exists it still sets or updates `origin` and reports
+ `remoteChanged`. A family that needs a remote on the shared repo uses its
+ own name instead (`qa-f27-repo`), so F08's `qa-repo` keeps none.
- Sidebar and row navigation can append `?backend=<id>`: end `--expect-url`
regexes with `(\?|$)`, not `$`.
- Some confirmation buttons render without the `data-testid` their source
@@ -230,31 +233,31 @@ How to write and prove new entries: [../mapping.md](../mapping.md).
| F02 | [App shell, sidebar and command menu](F02-app-shell.md) | rail and phone drawer, collapse, home pinning, getting-started checklist, command menu, toasts | every page; `Control+k` / `Meta+k` | baseline; LLM for checklist progress | 25 |
| F03 | [Home and starting work](F03-home.md) | home composer, workspace picker, plugin picker, recommended automations rail, no-LLM banner | `/`, **New Chat**, command menu | runs before and after `llm preset`; fixture repo | 39 |
| F04 | [Conversation list and folders](F04-conversation-list.md) | sidebar list, View presets, workspace folders, tags, rename/pin/archive/delete, load more | sidebar on every page | several conversations; fixture repos | 40 |
-| F05 | [Composer, slash commands and plan mode](F05-composer.md) | send, drafts, attachments, model pill, slash commands, `/goal`, plan mode, dictation | home and conversation composer | LLM | 23 |
-| F06 | [Agent activity](F06-agent-activity.md) | messages, tool events, thinking, empty state, confirmation mode, stop/resume, errors, branching | `/conversations/<id>` | LLM | 33 |
+| F05 | [Composer, slash commands and plan mode](F05-composer.md) | send, drafts, attachments, model pill, slash commands, `/goal`, plan mode, dictation | home and conversation composer | LLM | 25 |
+| F06 | [Agent activity](F06-agent-activity.md) | messages, tool events, thinking, empty state, confirmation mode, stop/resume, errors, branching | `/conversations/<id>` | LLM | 35 |
| F07 | [Conversation page, header and menu](F07-conversation-page.md) | title rename, status menu, ⋯ menu (skills, hooks, tools, export, download, cost, stop, delete), git bar, overview | sidebar row, URL, `/panel` | LLM; fixture repo; Cloud items blocked | 25 |
| F08 | [Workspace drawer: files and changes](F08-workspace-files-and-changes.md) | Files tab, Commits and uncommitted changes, diff viewer, file links in chat | panel toggle, file links | LLM; fixture repo | 33 |
-| F09 | [Settings shell and navigation](F09-settings-shell.md) | settings navigation, phone hub, deep links, update card, backend note | gear, command menu, URLs | baseline; deep links before `llm preset` | 24 |
-| F10 | [LLM profiles](F10-llm-profiles.md) | list, add, edit, rename, duplicate, delete, default, advanced fields, validation | `/settings/llm` | fresh run; DeepSeek key | 31 |
+| F09 | [Settings shell and navigation](F09-settings-shell.md) | settings navigation, phone hub, deep links, update card, backend note | gear, command menu, URLs | baseline; deep links before `llm preset` | 25 |
+| F10 | [LLM profiles](F10-llm-profiles.md) | list, add, edit, rename, duplicate, delete, default, advanced fields, validation | `/settings/llm` | fresh run; DeepSeek key | 34 |
| F11 | [Provider connections](F11-provider-connections.md) | add, edit, rotate key, delete connections; profiles from a connection | `/settings/llm` | fresh run; DeepSeek key | 21 |
| F12 | [Model router](F12-model-router.md) | meta-profiles, templates, run on first message, routing effect | `/settings/meta-llm` | LLM profiles | 25 |
| F13 | [Agent profiles](F13-agent-profiles.md) | OpenHands and ACP profiles, editor, name rules, active profile, MCP and secret scoping | `/settings/agents` | fresh run; LLM | 20 |
| F14 | [Secrets](F14-secrets.md) | list, add, edit, rename, delete, agent access | `/settings/secrets` | LLM for agent access | 11 |
| F15 | [Condenser, agent context and verification](F15-agent-behavior-settings.md) | condenser fields, agent context, confirmation mode and critic, their effect on runs | `/settings/condenser`, `/agent-context`, `/verification` | LLM for the effects | 23 |
| F16 | [Application settings](F16-application-settings.md) | language, theme, analytics, sound, checklist, title model, voice input, git identity | `/settings/app` | LLM for title and git checks | 19 |
-| F17 | [Customize hub and MCP servers](F17-mcp-servers.md) | catalog, install, test, edit, delete, custom servers, agent use | **Customize**, `/mcp` | `uvx`/`npx`; LLM | 26 |
-| F18 | [Skills catalog](F18-skills.md) | facets, search, enable/disable, add skill, Use skill, personal and project skills | `/skills` | `fixture skill`; LLM | 19 |
+| F17 | [Customize hub and MCP servers](F17-mcp-servers.md) | catalog, install, test, edit, delete, custom servers, agent use | **Customize**, `/mcp` | `uvx`/`npx`; LLM | 30 |
+| F18 | [Skills catalog](F18-skills.md) | facets, search, enable/disable, add skill, Use skill, personal and project skills | `/skills` | `fixture skill`; LLM | 21 |
| F19 | [Plugins and plugin launch](F19-plugins.md) | catalog, install, update, uninstall, enable, launch deep links | `/plugins`, `/launch` | LLM | 25 |
-| F20 | [Canvas apps](F20-canvas-apps.md) | install from path or git, enable, update, uninstall, extension pages | `/apps` | none | 23 |
+| F20 | [Canvas apps](F20-canvas-apps.md) | install from path or git, enable, update, uninstall, extension pages | `/apps` | none | 24 |
| F21 | [Automations dashboard and actions](F21-automations-dashboard.md) | cards and list, filters, sort, Run now, enable, export, import, delete, pin | `/automations` | LLM; automation service | 34 |
| F22 | [Creating automations](F22-automation-creation.md) | templates, setup dialog, custom automations, import | `/automations/templates`, `/automations/new/<id>` | LLM | 28 |
| F23 | [Automation detail, runs and editing](F23-automation-detail.md) | detail page, runs, logs, edit dialog, triggers, debug | `/automations/<id>` | LLM; runs | 39 |
| F24 | [Automation Git Sync](F24-git-sync.md) | configure, sync cycles, encryption, status | `/automations/git-sync` | `fixture git-repo` | 26 |
-| F25 | [Backends, Cloud and sharing](F25-backends-and-cloud.md) | add, edit, remove and switch backends, per-backend consent, Cloud login, shared pages | backend selector | a second stack; Cloud account (blocked) | 27 |
-| F26 | [Launcher modes, Docker, desktop and library](F26-runtime-variants.md) | launcher flags, partial stacks, LAN bind, Docker, Electron, embeddable library | a terminal | Docker/Electron where available | 27 |
-| F27 | [Workspace tools](F27-workspace-tools.md) | terminal, browser, planner, task list, usage, `canvas_ui_control` | drawer tabs | LLM | 29 |
+| F25 | [Backends, Cloud and sharing](F25-backends-and-cloud.md) | add, edit, remove and switch backends, per-backend consent, Cloud login, shared pages | backend selector | a second stack; Cloud account (blocked) | 28 |
+| F26 | [Launcher modes, Docker, desktop and library](F26-runtime-variants.md) | launcher flags, partial stacks, LAN bind, Docker, Electron, embeddable library | a terminal | Docker/Electron where available | 30 |
+| F27 | [Workspace tools](F27-workspace-tools.md) | terminal, browser, planner, task list, usage, `canvas_ui_control` | drawer tabs | LLM | 30 |
-27 families, 716 sub-features. `control-openhands map ids` lists every ID with its file.
+27 families, 736 sub-features. `control-openhands map ids` lists every ID with its file.
### Neighbouring families
diff --git a/.agents/skills/verify-openhands/scripts/control-openhands.mjs b/.agents/skills/verify-openhands/scripts/control-openhands.mjs
index 35d855e460..6fda00aa16 100755
--- a/.agents/skills/verify-openhands/scripts/control-openhands.mjs
+++ b/.agents/skills/verify-openhands/scripts/control-openhands.mjs
@@ -41,6 +41,7 @@ import {
sourceLiterals,
} from "./lib/testids.mjs";
import { tmuxPathFor } from "./lib/tmux-path.mjs";
+import { browserCallLimit } from "./lib/call-limit.mjs";
const here = dirname(fileURLToPath(import.meta.url));
const skillDir = resolve(here, "..");
@@ -383,7 +384,7 @@ async function browserCall(run, cmd, args = {}, { timeout = 120_000 } = {}) {
"x-control-token": info.token,
},
body: JSON.stringify({ cmd, args }),
- signal: AbortSignal.timeout(timeout),
+ signal: AbortSignal.timeout(browserCallLimit(args, timeout)),
});
const result = await response.json();
if (!result.ok) {
@@ -2191,8 +2192,35 @@ async function cmdFixture({ positional, flags }) {
: undefined;
if (kind === "git-repo") {
const dir = join(workspace, name ?? "qa-repo");
+ const remote =
+ flags.remote && flags.remote !== true ? String(flags.remote) : undefined;
if (existsSync(join(dir, ".git"))) {
- out({ ok: true, path: dir, existed: true });
+ // The repo is kept as it is, except that --remote still means "origin
+ // is this URL": a family that needs repo links on a repo another family
+ // made gets them, and says so in its output.
+ let remoteChanged = false;
+ if (remote) {
+ const current = spawnSync("git", ["remote", "get-url", "origin"], {
+ cwd: dir,
+ encoding: "utf8",
+ });
+ const existing =
+ current.status === 0 ? current.stdout.trim() : undefined;
+ if (existing !== remote) {
+ execFileSync(
+ "git",
+ ["remote", existing ? "set-url" : "add", "origin", remote],
+ { cwd: dir },
+ );
+ remoteChanged = true;
+ }
+ }
+ out({
+ ok: true,
+ path: dir,
+ existed: true,
+ ...(remote ? { remote, remoteChanged } : {}),
+ });
return;
}
mkdirSync(join(dir, "src"), { recursive: true });
@@ -2222,8 +2250,6 @@ async function cmdFixture({ positional, flags }) {
]) {
execFileSync("git", args, { cwd: dir, env: gitEnv });
}
- const remote =
- flags.remote && flags.remote !== true ? String(flags.remote) : undefined;
// A remote URL lets the UI show repo/branch links and Pull/Push chips;
// nothing is fetched or pushed.
if (remote)
@@ -3904,7 +3930,8 @@ Examples:
control-openhands conversation events <id> --last 10
`,
fixture: `control-openhands fixture git-repo [--name qa-repo] [--remote https://github.com/qa-example/qa-repo.git]
- # git repo in <run>/workspace (README, src/calc.py, test); --remote only sets origin
+ # git repo in <run>/workspace (README, src/calc.py, test); --remote only sets origin,
+ # also on a repo that already exists (the output then carries existed and remoteChanged)
control-openhands fixture git-remote [--name qa-remote] # bare repo <run>/workspace/qa-remote.git to push to
control-openhands fixture mcp-server [--name qa-mcp] # stdio MCP server (tool qa_echo); prints command and args
control-openhands fixture folder [--name qa-folder]
diff --git a/.agents/skills/verify-openhands/scripts/control-openhands.test.mjs b/.agents/skills/verify-openhands/scripts/control-openhands.test.mjs
index 0bf02f7f91..d6c2bee6c3 100755
--- a/.agents/skills/verify-openhands/scripts/control-openhands.test.mjs
+++ b/.agents/skills/verify-openhands/scripts/control-openhands.test.mjs
@@ -35,6 +35,7 @@ import {
import { routePattern } from "./lib/route-pattern.mjs";
import { resolveTestids } from "./lib/testids.mjs";
import { tmuxPathFor } from "./lib/tmux-path.mjs";
+import { browserCallLimit } from "./lib/call-limit.mjs";
import { buildLocator, parseRole, toCss } from "./lib/selectors.mjs";
const here = dirname(fileURLToPath(import.meta.url));
@@ -910,6 +911,18 @@ test("evidence report shows notes, family counts, and changes against a baseline
);
});
+test("a browser call waits for the verb's own deadline plus a margin", () => {
+ // The daemon runs one Playwright waitFor with the verb's --timeout; the
+ // client used to abort every call at 120 s, so `wait --timeout 150000`
+ // failed with an aborted request while the page kept waiting.
+ assert.equal(browserCallLimit({}), 120_000);
+ assert.equal(browserCallLimit({ timeout: 30_000 }), 120_000);
+ assert.equal(browserCallLimit({ timeout: 150_000 }), 160_000);
+ assert.equal(browserCallLimit({ observeMs: 130_000 }), 140_000);
+ assert.equal(browserCallLimit({ timeout: "125000" }), 135_000);
+ assert.equal(browserCallLimit({ timeout: undefined }), 120_000);
+});
+
test("the tmux directory of a run comes from its whole path", () => {
// Two runs whose ids end alike must not share a tmux server (the earlier
// scheme used the last six characters of the run id).
diff --git a/.agents/skills/verify-openhands/scripts/lib/call-limit.mjs b/.agents/skills/verify-openhands/scripts/lib/call-limit.mjs
new file mode 100644
index 0000000000..0ebfe02315
--- /dev/null
+++ b/.agents/skills/verify-openhands/scripts/lib/call-limit.mjs
@@ -0,0 +1,18 @@
+// How long the CLI waits for one browser daemon call.
+//
+// Every verb answers within a default limit, but a verb's own deadline
+// (`--timeout` on wait/text/attr…, `--observe-ms` on click) may run past it.
+// The client then waits that long plus a margin, so a long wait ends with
+// Playwright's TimeoutError and a failure screenshot, not an aborted request
+// while the page action keeps running.
+
+export const DEFAULT_CALL_LIMIT_MS = 120_000;
+export const CALL_LIMIT_MARGIN_MS = 10_000;
+
+export function browserCallLimit(args = {}, base = DEFAULT_CALL_LIMIT_MS) {
+ const asked = Math.max(
+ Number(args?.timeout) || 0,
+ Number(args?.observeMs) || 0,
+ );
+ return Math.max(base, asked + CALL_LIMIT_MARGIN_MS);
+}