Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

184 KB
Raw Blame

Command box continuation — 2026-09-20

Branch: worktree-command-box. The desktop overhaul is committed as f2ff876b; integration now includes main through 60764ba9 (PR #164). The earlier sections below are a historical record; the current PR checkpoint is at the end.

The experimental JEV command bar (Cmd-Shift-J) and its review service on port 18478 were removed on 2026-10-10. Where this log still names them, it records how a build was made at the time, not what the app does now.

Direction carried forward

The terminals are the product; Harness is the frame, one keystroke away. Keep the bottom-docked search/new-agent prompt, fast launch with inherited agent/machine/project, remappable keys, and clickable alternatives. Keep the noun Tab. Do not add a status bar or permanent status segments.

The original operational brief is the session memory named command-box-handoff.md; its design history is superterminal-design-direction.md. This file records the subsequent work inside the worktree so it is discoverable without searching session memory.

Settings follow the highlighted agent — current continuation

The user could not reach a Codex profile while another engine was saved: settings were based on the saved agent, not the highlighted result. They also requested removing the repeated Agent/Machine/Project rows from both Agent and Machine selection.

  • Agent settings now follow typing and arrow selection. Permissions/Profile appear indented directly below the highlighted engine, so Down enters its settings without crossing another engine and changing the target.
  • Browsing remains a preview. Escape preserves the saved agent and profile. Choosing a permission/profile commits the highlighted engine with that setting; Enter on an engine keeps the existing choose-and-return fast path. Returning from settings restores the filter and the selected settings row.
  • Agent and Machine pickers now show a left-aligned title, choices, input, and hints. Their repeated summary rows are removed. The launch screen keeps the full summary. Scroll sizing retains room for six engine results plus inline settings, subject to window height and text size.
  • Profile/mode pickers retain their agent target when opening and returning from Machine. The saved profile is untouched by merely browsing other engines. Filtering to no matching agent cannot launch or select a setting.

Verification: 65 focused checks pass, including the Claude → highlighted Codex → Profile flow, cancellation, committed-profile preservation, and hidden summary rows in Agent/Machine (/private/tmp/agent-highlight-tests.log). Inspected normal and enlarged-text previews in agent-highlight-renders. Analysis, formatting, and git diff --check pass.

The native fake-agent creation test passed twice with packet diagnostics enabled (agent-highlight-native.log, agent-highlight-native-repeat.log). Its first run observed two unexpected terminal-input packets; this did not reproduce in either repeat, and the source was not established. Keep the diagnostics rather than weakening the empty-input assertion. These are injected Flutter keys, not physical AppKit/IME validation.

Rebuilt normal lib/main.dart with the existing JEV endpoint http://127.0.0.1:18478 (/private/tmp/agent-highlight-build.log), verified the bundle's deep/strict signature, and opened review PID 33363. Previous review PID 66703 remains open. No real agents were created by the fixtures; nothing committed or pushed. Backups are in /private/tmp/agent-highlight-before.

Agent settings without a main Options row — previous continuation

The user found Options too prominent and approved moving the settings into Agent. The launch menu now contains Agent, Machine, Project, Task, and Start/Create. Create remains selected initially; Up reaches Task directly.

  • Removed the standalone Options row. Nondefault permissions and any explicit profile are summarized beside the Agent value; defaults stay quiet.
  • Agent keeps its fast choose-and-return behavior. Permissions and Codex profile are explicit settings below its results, using the same typing, arrows, Enter, and Escape interaction. Choosing a setting returns to Agent with its search and settings-row selection restored.
  • Added a profile picker scoped to the selected machine, with Default, discovered profiles, Refresh, and Link profile folder. Discovery never picks an account. Changing engine or machine clears the profile; stale discovery replies are ignored. Older CLIs explain missing support without issuing an unsupported request. Linking blocks editing, dismissal, and creation until the response arrives.
  • An unmatched Agent query leaves no implicit selection; Enter/Cmd-Enter explains the empty result rather than using another agent or a settings row.
  • Cmd-period retains the full form for compatibility and install guidance. Updated keyboard-practice copy, review notes, fixtures, and the native creation flow to use Agent → Permissions.

Verification: 75 focused checks pass in /private/tmp/agent-settings-tests.log; all four profile-flow checks also pass after the final pending-link dismissal guard (agent-settings-fix.log). Touched-file analysis and git diff --check pass. Inspected normal and enlarged-text root, Agent, permissions, and profile layouts in /private/tmp/agent-settings-renders.

The native fake-agent creation flow also passes in /private/tmp/agent-settings-native.log (injected Flutter keys, not physical AppKit/IME input). Rebuilt normal lib/main.dart afterward, preserving JEV at http://127.0.0.1:18478; log /private/tmp/agent-settings-build.log. Deep/strict signature verification passes. Opened the review app and confirmed PID 66703. No real agents were created by these fixtures; nothing committed or pushed.

Launch menu without letter shortcuts — previous continuation

The user observed that the launch menu was the remaining creation screen with letter shortcuts and requested the same Up/Down/Enter interaction throughout.

  • Removed the a/m/p/t labels and default bindings and the footer's o option. Rows and input context now share the same left alignment without a shortcut column. Explicit custom bindings still work through the stable command IDs.
  • Options is always a selectable row above Create, including with default settings. Nondefault permissions/profile remain summarized beside Options. Up from Create opens the route to Options; Enter opens the full form.
  • Create remains selected on entry and on return from an editor. Arrows move through Agent, Machine, Project, Task, Options, and Create. Enter edits or launches the selected row. Cmd-Enter and Cmd-period remain available.
  • Updated keyboard-practice examples, design/review notes, existing test flows, and the native creation fixture to navigate the launch rows with arrows. Plain-letter tests confirm there are no default jumps or side effects.

Verification: 94 focused checks pass, including Options via Up/Enter and return to Create, plus the native fake-agent creation flow. Logs: /private/tmp/launch-arrows-tests.log, /private/tmp/launch-arrows-native.log. The native fixture uses injected Flutter keys, not physical AppKit/IME input. Touched-file analysis, formatting, and git diff --check pass. Inspected normal and enlarged-text layouts in /private/tmp/launch-arrows-renders. The native-keymap helper's launch assertions were updated; its unrelated legacy workspace assertions were not rerun.

Rebuilt normal lib/main.dart after the fixture, preserving JEV at http://127.0.0.1:18478; log /private/tmp/launch-arrows-build.log. Deep/strict signature verification passes. Opened the review app and confirmed new PID 47771 (previous review PID 66905 remains open). No real agent was created; nothing committed or pushed.

Project search without action shortcuts — previous continuation

The user requested typing to filter Project just like Agent and Machine, with recent projects above fixed New/Open/Clone actions and the input below them. They then simplified the request: remove 1–9 and n/o/g; use typing, Up/Down, Enter, and Escape throughout Project.

  • Project now focuses a search input and filters names/paths, including case and common word-separator variations. Results stay on the selected machine. The best matching recent is highlighted, or Open folder when none match.
  • New project, Open folder, and Clone GitHub repository stay visible above the input in that visual order. The recent list remains scrollable, with room for up to nine visible rows and older projects reachable by arrows. Search changes re-reveal the best result even if its cursor index is unchanged.
  • Removed action letters and numbered row labels/default shortcuts in Project. Ordinary letters/digits remain search text, including m; the header's clickable Change machine action remains, and launch m still edits Machine. Existing command IDs remain unbound for explicit custom remaps. Navigation continues to inherit picker remaps. Root launch a/m/p/t/o is unchanged.
  • Escape from New/Open/Clone restores the Project filter; reopening from launch resets it. Changing machines clears a filter belonging to the previous machine. The folder child still has no duplicate recent list.
  • Updated keyboard practice, render/native fixtures, interaction tests, and the current design and manual-review notes.

Verification: 94 focused controller/widget/keymap/practice/render checks pass, plus the native creation flow with fake agents. Logs: /private/tmp/project-filter-tests.log, /private/tmp/project-filter-native.log. The native fixture uses injected Flutter keys, not physical AppKit/IME input. Touched-file analysis has no issues; formatting and git diff --check pass. Inspected the long-history, filtered, no-match, and enlarged-text renders in /private/tmp/project-filter-renders. The older native keymap helper's Project assertion now reflects unbound action keys; its unrelated legacy workspace assertions were not rerun in this continuation.

Rebuilt normal lib/main.dart after the native fixture, preserving JEV at http://127.0.0.1:18478; log /private/tmp/project-filter-build.log. Deep/strict signature verification passes. Opened the review app and confirmed new PID 66905 (previous review PID 81012 remains open). No real agent was created; nothing committed or pushed.

Arrow-selectable launch menu and plain prompts — previous continuation

The user requested clearer GitHub cloning copy, no extra > markers, and Up/Down selection for the launch arguments as well as the child pickers.

  • GitHub now says "Paste a GitHub repository URL" and "Clone to machine:~/harnesses/". Short names such as openai/codex, HTTPS, and SSH URLs still parse as before. Selection prepares the repository; launching performs the clone. Invalid-input guidance uses a concrete example instead of the unexplained word "owner".
  • Cmd-N input labels (agent, machine, folder, repo, name, task) no longer append >. Their baselines and value columns align with the context. Result rows use the full-width highlight without an extra > selection pointer, and their names use the input/context's left inset.
  • The launch menu has an explicit Create/Start row, selected initially, so Enter retains the fast launch with inherited defaults. Up/Down cycle through the displayed arguments, any visible options summary, and Create. Enter opens the highlighted argument; a/m/p/t/o still jump directly. Returning from an editor highlights Create again. Cmd-Enter starts from any selected launch row. A missing project still opens Project instead of creating an unnamed folder. Pending creation disables argument navigation.
  • The footer describes the selected action and prints the effective remapped navigation keys. A moved pointer can select a launch row; a parked pointer does not override keyboard selection. Keyboard-practice examples and the design/review notes match the new copy and interaction.

Verification: 60 focused creation/controller/layout tests pass, followed by 21 layout/render/practice checks after the final alignment adjustment (overlapping sets). Logs: /private/tmp/launch-menu-tests.log and /private/tmp/launch-menu-final-tests.log. New coverage drives every launch argument through arrows/Enter without starting an agent, checks return to Create, and verifies live navigation remapping. The native creation workflow also passes with fake agents and arrow-selected Machine: /private/tmp/launch-menu-native.log. This uses injected Flutter keys, not a physical AppKit/IME walkthrough. Final touched-file analysis, formatting, and git diff --check pass. Inspected normal, highlighted-argument, enlarged-text, and GitHub renders in /private/tmp/launch-menu-renders.

Rebuilt normal lib/main.dart after the native fixture, retaining JEV at http://127.0.0.1:18478; log /private/tmp/launch-menu-build.log. Deep/strict signature verification passes. Opened the review app and confirmed PID 81012. No real agent was created; nothing committed or pushed.

One Project history and a focused folder prompt — previous continuation

The user requested more than four recents, removal of the duplicate history in Open folder, bottom placement for the default action, and o instead of e. Recommended and implemented one history in Project, with a separate path prompt for a known folder or browsing on the selected machine.

  • Project includes every available recent for the selected machine. The dock makes room for up to nine rows, with older entries accessible by arrows, Page Up/Down, or scrolling. Plain 1–9 pick the numbered entries. The current folder remains number 1; the list grows upward from the actions.
  • GitHub, New project, and Open folder stay fixed below the recents. Open folder is the bottom/default row, reached with Enter or o. History cannot push the actions offscreen. m still changes the shared Machine setting.
  • Open folder contains a path/completion prompt and an Open folder… browse row next to the input. It no longer repeats or searches recent projects. Empty Enter or Ctrl-O opens the selected machine's browser. Typed paths select matching folders; browse stays nearest the input. New project and GitHub remain separate intents.
  • A project keymap context inherits picker navigation/remaps and scopes o to Open folder. The launch menu retains o for options. Printed hints, shortcut settings, keyboard practice, and the native snapshot share this distinction. The macOS snapshot validator and context handler now accept it.
  • The selected recent remains visible when resizing or increasing text size. Other command docks keep their existing compact heights.

Verification: controller, completion, keyboard, and creation checks passed in focused runs. The last layout/render run passed 18 checks; the preceding layout/keymap/shortcuts run passed 44 (overlapping sets). Logs: /private/tmp/project-recents-focused.log, /private/tmp/project-recents-layout.log, and /private/tmp/project-recents-resize.log. Earlier failures were corrected and rerun in the affected files; the complete desktop suite was not repeated. Final touched-file analysis has no issues, formatting and git diff --check pass. Inspected normal, older-history, path-only, and large-text renders in /private/tmp/project-recents-renders.

The native creation workflow passes with fake agents: /private/tmp/project-recents-native-fixed.log. Its first run caught and led to fixing the macOS INVALID_KEYMAP rejection of the new context. Injected Flutter keys do not establish physical AppKit/IME behavior. The separate older tool/check_keymap_native.sh executes the new Project-context assertions but then fails at its pre-existing obsolete expectation that Cmd-R splits right; that helper has not been reconciled with the earlier shortcut redesign. Log: /private/tmp/project-recents-native-keymap.log. No production shortcut was changed to satisfy that stale expectation.

Rebuilt normal lib/main.dart after the native fixture with JEV retained at http://127.0.0.1:18478; log /private/tmp/project-recents-build.log. Deep/strict signature verification passes without a manual signing repair. Opened the review app and confirmed PID 31023. No real agent was created; nothing committed or pushed.

Quieter Quick Access symbols — previous continuation

The user approved keeping the Quick Open prefixes while removing the decorative input > / : and the selected result's > pointer. Cmd-P/T and command search now use a plain monospace input and a full-width selection highlight. Only the editable prefix identifies the mode (>, @, #, or ?); the footer keeps the mode cheat sheet. Input and result names align at the same left inset. The empty agent picker says "Find an agent…". Other field prompts are unchanged.

SwarmSearchInput.terminal separates terminal typography from the optional prompt decoration used by the advanced agent chooser. TerminalBox supplies the dock input's background, avoiding Material's extra filled-field inset.

Verification: 29 focused checks pass (Quick Access, dock geometry, picker entry, search rendering, and the existing render fixture), plus analysis and git diff --check. Logs: /private/tmp/quick-access-symbols-tests.log and /private/tmp/quick-access-symbols-analyze.log. Reviewed empty, command, and narrow/large-text renders in /private/tmp/quick-access-symbols-renders.

Rebuilt normal lib/main.dart, retaining JEV on port 18478: /private/tmp/quick-access-symbols-build.log. Opened PID 96237. The rebuilt App.framework verified independently, but the incremental build left its old signature in the enclosing app's resource seal. Refreshed the generated app's ad-hoc signature, preserving identifier, entitlements, flags, and runtime; deep/strict signature verification now passes. No source signing settings were changed. No real agent was created; nothing committed or pushed.

Quick Access prefixes and creation polish — previous continuation

The final user-approved mapping is plain text = agents, > = commands, # = projects, @ = machines, and ? = modes/help. An intermediate suggestion of @ agents and m machines was rejected; it is not in the implementation. Cmd-P/T now allow these prefixes. Cmd-Shift-P inserts > instead of opening a permanently command-only buffer. Deleting it returns to agents in the same picker, keeping the New Tab/New Pane destination. Help's prefix rows switch in place. Project/machine selection shows only that group's agents; Escape restores the parent query. Missing groups never fall back to unrelated agents. Prefix hints retain their space while typing so the dock input does not jump.

Recents previously used only the current folder and explicit creation history. They now merge known agent folders from the selected machine (including local CLI project metadata), collapse equivalent paths, and load saved history on the first opening. The menu still shows four, with the full list under Open folder. Live metadata updates refresh the menu without changing its selection.

Cmd-N's agent row no longer has >_. The editable prompt label uses an input prefix aligned to the first text baseline, including empty hints and multiline tasks, and shares the context column's left inset. Selected tabs retain their 3pt top rounding and use matching outward curves where they join the workspace, in both the native macOS tab strip and Flutter fallback.

Research sources are linked in super-terminal-interaction.md; the interaction comes from VS Code Quick Open, with object-specific meanings chosen by the user.

Verification: the full desktop run reached 2,326 passes and five skips, exposing ten failures from outdated command-prefix/menu expectations, a test-local name collision, and the hint-wrap input movement. After fixing those, all 71 checks in the affected files plus the new Quick Access/render cases pass: /private/tmp/quick-access-final.log. The whole suite was not repeated after that focused rerun. All 11 native macOS workflows pass with fake agents: /private/tmp/quick-access-native.log. This verifies injected Flutter input, not physical AppKit/IME or VoiceOver behavior. Analysis, formatting, and git diff --check pass. Final renders are in /private/tmp/quick-access-renders.

Rebuilt normal lib/main.dart after the native fixture, retaining JEV on http://127.0.0.1:18478: /private/tmp/quick-access-build.log. Deep/strict signature verification passed. Opened the review app with open -n and confirmed PID 35448. No real agent was created; nothing committed or pushed.

Shared Machine setting, immediate recents, and quieter tabs — previous continuation

The user approved removing the selected-tab underline and making Machine a first-class launch setting. Cmd-N now shows a agent, m machine, p project, and t task. The path on the launch summary omits the duplicate machine prefix; full machine:path notation remains in folder search and location previews.

Project now has three fixed actions (New project, Open folder, Clone GitHub) and up to four numbered recent projects for the selected machine. The accepted current folder is first, followed by that machine's recent history. 1–4 select immediately; full history/search/browsing stays under e. Actions stay fixed if the recent list needs scrolling. The Project header exposes the same remappable m chooser as the launch screen. Cancelling Machine returns to its source screen with its query intact; a changed machine clears a folder filter from the old machine. Accepted project choices remain separate per machine, and stale recent rows from another machine are rejected.

Both the native macOS tab strip and Flutter fallback now use the existing selected background and bold label without the 2pt underline.

Verification: 86 focused creation/practice/render checks passed, followed by 62 final keyboard/render checks including shared-Machine return, remapping, numbered recents, stale-machine rejection, and retained task/project choices. Touched-source/test analysis has no issues; formatting and whitespace checks pass. Normal and enlarged-text/narrow Project renders were inspected under /private/tmp/machine-recent-renders.

All 11 native macOS journeys pass, including the new top-level Machine chooser and direct recent selection: /private/tmp/machine-recent-native.log. These use fake agents and injected Flutter input; physical AppKit/IME and VoiceOver were not newly verified. The whole desktop suite was not rerun. The normal app was rebuilt from lib/main.dart after the native fixture, retaining JEV_COMMAND_BAR_URL=http://127.0.0.1:18478: /private/tmp/machine-recent-build.log. Deep/strict signature verification passed. Opened build/macos/Build/Products/Debug/Harness.app with open -n and confirmed PID 14979; existing app instances were left alone. Nothing was committed or pushed, and no real agent was created for verification.

Project actions and visible task availability — previous continuation

User review found that project history buried New/Open/GitHub/Machine, and Pi hid the task row. Project now opens a separate four-action menu: n New project, e Open existing project, g Clone GitHub repository, and m Change machine. All four stay visible regardless of recent history. Arrows follow the menu's top-to-bottom order, Enter chooses, and the letters follow the live keymap.

Each action opens a dedicated prompt: name >, folder >, repo >, or machine >. Existing searches only that machine's folders and never converts a missing match into a new folder. Its footer keeps Ctrl-O/clickable Browse folders visible. New previews the exact slug/path and offers Open existing on a collision. A late folder listing now refreshes that name preview too. GitHub accepts owner/repo or supported GitHub URLs and cannot silently launch in the old project on invalid input. Selection retains the repository for launch; it does not clone while merely browsing the menu. Escape backs out to Project, then the launch summary. Switching machines retains their accepted project choices separately.

Task stays visible for Pi and other unsupported engines with “Enter a task in after launch”; t explains the restriction. This reflects the current CLI first-prompt contract, not a claim that Pi itself lacks prompt support. No CLI engine-launch behavior changed. Supported agents retain the task editor and all first-message/draft/recovery protections.

Verification: 90 focused creation/practice/render checks, 92 keymap, first-task, and recovery checks, and 11 native macOS journeys pass. Logs are /private/tmp/project-menu-focused2.log, project-menu-keymap.log, and project-menu-native.log. Final touched-file analysis has no issues; formatting and whitespace checks pass. Visuals in /private/tmp/project-menu-renders include all four project actions at normal and enlarged text/narrow width, New/Existing/GitHub/Machine prompts, and Pi's visible task row. The final render fixture also passes. The entire desktop suite was not rerun for this revision. Native journeys use fake agents and injected Flutter input, with no new physical AppKit/IME or VoiceOver coverage.

The normal macOS debug app was rebuilt from lib/main.dart after the native fixture, retaining JEV_COMMAND_BAR_URL=http://127.0.0.1:18478: /private/tmp/project-menu-build.log. Deep/strict signature verification passed. Opened build/macos/Build/Products/Debug/Harness.app with open -n and confirmed the new review process (PID 30111). The earlier review instance was left alone. No real agent was created; nothing was committed or pushed.

The user also asked about the selected-tab underline. It was explained as a generic active-tab marker, not a copy of a particular terminal app. Simplifying it to the existing fill and bold label was recommended; no tab-style change was requested or made during this Project pass.

Cmd-N launch menu and focused prompts — previous continuation

The user approved trying a launch-command model after reviewing Vim, Emacs, tmux, zsh, and Magit patterns. Cmd-N reuses the current agent/project and shows a agent, p project, and t task; Enter launches with the displayed setup. Each letter opens one focused picker/editor, and Enter on an agent/project returns to the launch menu. Enter in Task launches; Escape retains it and returns to the menu. Menu letters are ordinary text inside editors. o and Cmd-period open advanced options; these keys follow the live keymap.

Tab now completes the current argument without accepting a selection or switching fields. Existing path completion/cycling and line editing remain. If there is no project, its picker opens first; compact creation no longer silently allocates an unnamed folder. New project opens name > with the machine/path preview; whitespace in a name becomes the destination slug. Machine selection remains nested in Project, with recents and cached choices scoped to the selected machine. Escape goes back through these prompts.

The user also requested a much darker background behind floating forms. New Harness and shared terminal forms (including Rename Tab) now use the existing shared 90% black dialog tint with no blur. Docks keep their workspace alignment. The updated ideal-user brief at the main checkout's docs/ideal-users.md describes technical founders building across disciplines; the task copy remains broad enough for a deck, analysis, video, or code.

Verification and review build:

  • The full desktop run completed with 2,340 passed, 5 skipped, and one outdated test failure. That test switched machines and attempted creation without choosing a project on the new machine. It now selects that project explicitly; all 26 final placement, Store, and render checks pass in /private/tmp/cmd-n-launch-final-focused.log. The full suite was not rerun after this test-only correction. An earlier default-concurrency run was interrupted under heavy load; its local-discovery timing failures passed in isolation and in the completed concurrency-2 run.
  • All 11 native macOS journeys pass, including the new launch menu, retained task, argument editing, advanced-form return, chosen destination, and terminal focus recovery: /private/tmp/cmd-n-launch-native.log. These use fake agents and Flutter-injected input. Physical AppKit keyboard/IME and VoiceOver are not newly verified. The fixture logged a foreground warning but completed all tests successfully.
  • Final touched-source/test analysis has no issues; full app/test analysis has no errors or warnings and the same three existing informational lints. Formatting and git diff --check pass.
  • Rendered and inspected the launch menu, empty and named project prompts, narrow layout at enlarged text, and dark Rename Tab/advanced-form backdrops: /private/tmp/cmd-n-launch-renders.
  • Rebuilt the normal app from lib/main.dart after the native fixture, retaining JEV_COMMAND_BAR_URL=http://127.0.0.1:18478. Build log: /private/tmp/cmd-n-launch-build.log. Deep/strict signature verification passed. Opened build/macos/Build/Products/Debug/Harness.app with open -n and confirmed the review process (PID 43696). No real agent was created for verification; nothing was committed or pushed.

PR #136 rebase and project-picker clarity — latest continuation

  • Rebased onto eb35e248 and restored all 184 working files. The only conflict was the import block in swarm_screen.dart; both sets of imports are retained. Hash comparison found no missing files and changes only in the three files that intentionally received upstream JEV updates. Nothing was committed or pushed, and the user's local main was not changed.
  • Recovery copies: branch backup/command-box-before-pr136-20260920 at d46ffaf4, stash 1b36d3c672ad8f7037b814e427f3cb3eba769056, and the file archive, patch, and hash manifest in /private/tmp/command-box-before-pr136-cn1qkthp.
  • Project editing now says on , with Search or name a project at the prompt. An empty query offers Open folder…, existing folders on that machine, and Change machine…. There is no selected-looking <name> path or empty Create row. Typing payments processing offers Create payments-processing and one machine:~/harnesses/payments-processing destination. Clone requests do not also offer an unrelated new folder.
  • Recent projects are scoped to the chosen machine. Switching to iMac cannot show M2 folders; switching back restores that machine's draft choice. Stale project rows from a different machine are rejected. Action rows are compact; real project choices keep their two-line name/path presentation.
  • Synced Score's bundled tagline to the updated Store manifest brought in by PR #134. This fixes the two catalog identity checks exposed by the rebase.

Verification artifacts are under /private/tmp/pr136-*: 73 focused desktop checks, 58 CLI checks, TypeScript checking, and the CLI build pass. The first full desktop run passed 2,335 tests (5 skipped) with only the two Score tagline mismatches; after synchronization, all 22 render/catalog/JEV checks pass. App/test/integration analysis reports only the same three informational lints; final touched-file analysis has no issues. Project empty/named states and JEV Go back renders were inspected in pr136-box-renders and pr136-jev-renders. All 11 native macOS journeys pass, including the new JEV flow, in /private/tmp/pr136-native-final.log. The fixture submits through the text-input action and waits for the completion notice's entrance animation before tapping Go back. This exercises the native workspace with injected Flutter input; physical AppKit keys, IME, and VoiceOver are not newly covered. The normal macOS debug build from lib/main.dart succeeded after the native fixture, using --dart-define=JEV_COMMAND_BAR_URL=http://127.0.0.1:18478: /private/tmp/pr136-review-build.log. Deep/strict signature verification passed. Opened build/macos/Build/Products/Debug/Harness.app with open -n and confirmed the new process (PID 47919, started 2026-09-20 01:34:14 local). An older review instance was left running; the newly opened window is the one to review.

Creation polish and keyboard learning — previous continuation

The user chose to leave JEV experimental and improve the everyday workflow. No new JEV integration or promotion was added in this continuation.

  • Project searches match words with spaces to hyphenated/underscored folder names. Selecting a recent project still chooses its machine and path together. Punctuation-only names cannot accidentally become an unnamed new project.
  • Escape retains edited agent/project defaults even before a task is typed. Drafts remain scoped to their source and bounded to 32 inactive contexts; unresolved creations are retained. Cancelling Browse restores prompt focus.
  • Quick start is an optional four-step guide alongside real work: open an agent, add a second pane in the tab, zoom, then open command search. Opening and cancelling an empty chooser does not count. It never takes typing focus. Progress and dismissal persist locally; the guide can be paused or restarted.
  • Keyboard practice is a separate, repeatable terminal-styled scratch surface. Search by action/group, press the actual shortcut, see a simulated result, then Enter to continue or Escape to return. It covers the stable command catalog and prompt editing, follows live remaps/sequences/unbindings, and tracks local progress. Actions without a binding can be practiced by typing their command-search name. It does not execute workspace/agent actions. Page Up/Down reads longer explanations at enlarged text sizes.
  • Both are discoverable on the start page, through command search, and from native Help. The shortcut reference links to practice too. Pause quick start is available through command search while the guide is active.
  • Nested keyboard hosts now defer to the host nearest the focused control. This prevents one Escape from firing twice in the scratch surface and keeps real workspace shortcuts out of practice.

Primary files: lib/state/workspace_learning.dart, lib/widgets/workspace_quick_start.dart, lib/shortcuts/keyboard_practice.dart, and their integration in lib/screens/swarm_screen.dart.

Verification:

  • 2,323 desktop tests passed, 5 skipped: /private/tmp/keyboard-learning-full.log.
  • After the final practice scrolling/typography refinements, 34 focused regressions and 14 final keyboard/layout checks passed, including the added narrow Quick start render: /private/tmp/keyboard-learning-focused-final.log, /private/tmp/keyboard-learning-layout-final.log.
  • All 10 native macOS journeys passed, including the new Help-command dispatch, guide progression, simulated practice, single-step Escape, and restored terminal input: /private/tmp/keyboard-learning-native.log. These use fake agents and Flutter-injected keys; physical AppKit keyboard, IME, VoiceOver, and Linux runtime were not newly exercised.
  • App/test/integration static analysis has no errors or warnings and the same three pre-existing informational lints: /private/tmp/keyboard-learning-analyze-final.log. git diff --check and formatting checks pass.
  • Captured and inspected the practice list, exercise, result, keyboard-scrolled explanation, and narrow guide: /private/tmp/keyboard-learning-renders.
  • Normal macOS debug build succeeded from lib/main.dart after the native fixture: /private/tmp/keyboard-learning-build.log. Signature verification passes. Review artifact: build/macos/Build/Products/Debug/Harness.app; opened with open -n for review. Nothing was committed or pushed.

Agent → Project → Task, main rebase, and JEV — previous continuation

The accepted creation layout is now:

agent     >_ Claude Code
project   iMac–Home:~/work/payments

task >    What should it do?
  • Agent and Project are vertically aligned; Task initially owns input. Tab and Shift-Tab follow Agent → Project → Task. Choosing an agent or project with Enter returns to Task. The inherited project still makes repeat creation fast.
  • Machine lives inside Project. Recent projects from all available machines show a two-line name and machine:path; selecting one chooses both together. Browse folders and Change machine are inside the project picker. Machine choice returns to Project; Escape backs out without dropping the task. Each machine remembers its project for this draft, including a More options round trip.
  • The project prompt says Type a project name. payments processing offers Create “payments processing” with machine:~/harnesses/payments-processing. If that folder already exists under the selected machine's harnesses directory, the choice becomes Open. Existing paths, path completion, and repository cloning remain supported.
  • Remote ~ resolves against the selected machine's home. Old directory/home replies cannot retarget the project after a machine switch. Browse results also check machine identity before applying.
  • Auto-approve is hidden under More options (Cmd-period). A changed permission mode or chosen Codex profile appears in one compact options summary.

Main was behind upstream and carried two unpushed local commits. Those were rebased on an isolated integration branch before applying them to this worktree; the user's local main remains at d626604d. Two additional upstream commits arrived during validation, including the JEV Sheets desktop chooser fix; this branch includes them through 252fdc18. Nothing was pushed. Existing desktop and CLI WIP was restored unstaged. The final restoration compared all 175 saved working files byte-for-byte with its stash and found no differences.

Recovery copies retained:

  • Original branch: backup/command-box-before-jev-20260919 at 3cd4014a.
  • Original WIP stash: 7a06eebf4db3805519430d90adae21e0a5c522a0.
  • Original file archive/patch/hashes: /private/tmp/command-box-before-main-mthmiuvc.
  • Validated WIP before the final two upstream commits: d6d70f2f1c9918002ba1373434520d7da7479847.

Verification:

  • 2,309 desktop tests passed, 5 skipped: /private/tmp/project-context-full-final.log. The last two upstream commits touched Store packages/docs only; all desktop/CLI working files were restored identically afterward. Updated old field expectations and synchronized FreeCAD's bundled tagline with the changed upstream Store description.
  • All 9 native macOS journeys passed: /private/tmp/project-context-native.log.
  • 75 targeted CLI tests and TypeScript checking passed: /private/tmp/project-context-cli.log, /private/tmp/project-context-cli-typecheck.log.
  • All 42 JEV Sheets tests passed, including the new in-pane chooser: /private/tmp/project-context-jev-sheets.log.
  • 5 new project-context tests cover Tab order, spaces in names, machine ownership, duplicate folder names, offline choices, remote home completion, existing destinations, draft retention, and delayed replies. JEV-to-creation integration and focus ownership also have a regression journey.
  • Rendered 107 box states and inspected ordinary, narrow and enlarged-text layouts: /private/tmp/project-context-renders. Static analysis has no errors or warnings and the same three existing infos: /private/tmp/project-context-analyze-final.log.
  • Normal macOS debug build succeeded after the native fixture, with --target lib/main.dart: /private/tmp/project-context-build.log. Artifact: build/macos/Build/Products/Debug/Harness.app. Opened with open -n; its process remains running and codesign --verify --deep --strict passes. Live UI automation could not reliably distinguish it from the installed copy with the same bundle ID. The observed launch crash reports name /Applications/Harness.app and mismatched framework Team IDs, not this worktree artifact. That installed copy was not modified. Visual verification of this change is from the rendered fixtures and native integration journeys.

Compact dock, tab height and Harness Store — current continuation

The user accepted the simulated panel's compact bottom-dock direction, asked to check the unusually tall tab bar, and approved three prominent Store entrances. Implemented:

  • Bottom-up dock results: best match immediately above the input, about six visible rows, Up/P/K away from the input and Down/N/J back toward it. Docked lists clamp at their ends; inline start-page search keeps its existing order.
  • A smaller dock height budget, default preview beside results or above them when narrow, no redundant preview for New agent, and stable footer widths.
  • New Pane shows its captured destination tab. New-agent choices follow the same spatial ordering; the inherited engine is first, while task arrows remain text-editing keys.
  • AppKit .unifiedCompact: measured 40pt versus the previous 52pt with a scratch AppKit probe. The Flutter fallback is also 40pt, with matching tab top insets and centered controls.
  • Persistent Harness Store text control at the far right of both tab bars. app.store is a remappable command, searchable as store/install/browse harnesses/packages/extensions. Browse more harnesses… in agent choices opens the Store while retaining the source's task/default draft. All paths reuse the existing singleton Store-tab behavior.
  • Utility tabs now hide the retained canvas rather than disposing its terminal renderers. A native regression caught the previous remount after Store browsing; the original view, selection and scroll can now survive the trip.
  • The taller offline help card scrolls when needed, including the breakpoint newly reached by reclaiming 12pt from the tab bar.

Verification:

  • 2,277 desktop tests passed, 5 skipped, including dock geometry/spatial keys, Store aliases/remapping, singleton tabs, native/Flutter dispatch, retained terminal views, focus and draft ownership. /private/tmp/terminal-compact-full-final.log.
  • All 9 native macOS journeys passed, including Store round trips retaining the original terminal renderer. /private/tmp/terminal-compact-native-retained.log.
  • 69 focused regressions passed while generating 107 box renders and wide/narrow/2x-text picker captures. Inspected input adjacency, preview order, footer wrapping and the compact top bar. /private/tmp/terminal-compact-regressions.log, /private/tmp/terminal-compact-final-render, /private/tmp/terminal-compact-final-picker.
  • Static analysis: no errors or warnings, the same three informational lints in app_state.dart, grid_model_picker.dart and pane_grid.dart. /private/tmp/terminal-compact-analyze-final.log.
  • The normal macOS debug build succeeded with --target lib/main.dart after native fixtures and was opened for review. Native UI inspection confirmed the compact bar and permanent Harness Store control beside the user's tabs. /private/tmp/terminal-compact-build.log.

Dock placement discussion

The user is reconsidering top versus bottom and explicitly wants AI reviewers roleplaying expert terminal developers. No human outreach is requested. The roleplayed panel splits: two favor compact bottom docks with different internal ordering; one favors top. All prioritize query/result adjacency, explicit scope and reliable focus restoration. The bottom-dock direction is now implemented as described above. The opinions, design tradeoffs and primary-source references are in command-dock-placement-review.md.

Default previews and harness marker — latest review feedback

New Tab and New Pane now open with preview visible, matching the start-page picker. Ctrl-/ still hides or restores it without changing the query, selection or input focus. The default was previously off to keep the picker compact; the user prefers immediate context. The Enter hint is shortened to open when the dock title already names its destination, avoiding a footer wrap that moved the input while filtering in narrow windows.

>_ is the ASCII harness marker before Codex, Claude and other harness names in all prompt presets, agent previews and new-agent defaults. Machine symbols use a monitor so the two kinds of context are visually distinct. Catalog data now supplies a separate harness field; offline and fork annotations do not acquire a harness marker. The customization preview includes a Codex segment.

Verification: 147 focused checks passed (128 picker/layout/keyboard and 19 placement checks), 107 box states and 21 customization states rendered, and wide/narrow prompt examples inspected. Static analysis has no errors or warnings and the same three existing infos. The normal macOS debug build succeeded with FLUTTER_TARGET=lib/main.dart. Logs: /private/tmp/terminal-preview-marker-checks.log, /private/tmp/terminal-preview-marker-placement.log, /private/tmp/terminal-preview-marker-analyze.log, and /private/tmp/terminal-preview-marker-build.log.

Workspace dock and prompt customization

The requested workspace styling is implemented and rebuilt for review. Search, New Tab, New Pane and New Agent use a bottom command dock, aligned with the pane grid's 6px outer inset. Results sit above the input. The dock does not resize terminal renderers. The editor retains its input connection when switching between task and default fields; ordered traversal takes Tab from the search input to the first result, including virtualized lists.

The New Pane pill has been removed from both native macOS and Flutter titlebars. Cmd-P, File → New Pane and command search remain. Tabs use monospace labels, flat 3px corners and a thin selection rule; pane corners share that radius. Native tab labels follow the selected terminal font.

Customize OpenHarness now opens on a Prompt tab, also reachable by typing customize in Cmd-Shift-P command search. Plain, Symbols and Powerline presets have a live preview, optional color, and machine/project/branch visibility. Choices persist together without replacing wallpaper, color palette or fonts. Search and pane headers share this presentation. Catalog identity and ranking remain based on the actual agent metadata. No repository status is fabricated. See research and design decisions.

Verification for this pass:

  • Full desktop suite: 2,270 passed, 5 skipped (/private/tmp/terminal-dock-full-final.log).
  • Final focused field/creation/customization/startup/identity regressions: 40 passed; dock and configured-key regressions: 30 passed.
  • Native macOS workspace fixture: 9 journeys passed (/private/tmp/terminal-dock-native.log).
  • Static analysis: no errors or warnings; the same three pre-existing infos.
  • Rendered 107 box states and 21 customization states. Inspected the three context styles, narrow Powerline, wide/large-text creation, search preview, and customization at ordinary and enlarged text sizes.
  • Normal macOS debug build succeeded; FLUTTER_TARGET=lib/main.dart confirmed. Artifact: build/macos/Build/Products/Debug/Harness.app. Build log: /private/tmp/terminal-dock-macos-build.log.

The user's installed or already running applications were not replaced or restarted. Relaunch the worktree artifact to see the new styles.

The unfinished Settings navigation experiment was parked before this review: its three source files were restored to their previous state, with its exact patch retained at /private/tmp/terminal-settings-wip.patch. It needs live keymap propagation, section focus, responsive layout and tests before resuming.

At that stage onboarding remained a design proposal. It is now implemented as Quick start and optional Keyboard practice; see the latest continuation above.

Terminal restart status and lifecycle — latest continuation

Restart Agent / Restart Terminal now use a compact terminal status prompt. The action still starts immediately; Escape returns to the workspace while the request continues. Reopening joins that request. Confirmed refusals stay inline with Enter/retry. Lost replies switch to Enter/check status on the original receipt. “Restart again…” is a separate action, with Cancel focused initially and an explicit reminder that the earlier restart may have completed. The confirmation stays fully visible at 480×360 with 1.7× text. Tab goes directly through controls; Page Up/Down scroll the explanation without making it an extra focus stop. Picker bindings, including live changes, apply throughout.

Command search and custom bindings now expose Restart Agent, or Restart Terminal for shells. Pane controls and the machine list use the same names and prompt. A resumed agent closes the prompt. A fresh conversation retains a clear notice with Enter/close; a fresh shell does not show a conversation-resume warning. The existing terminal views, tab membership, and focused pane remain intact.

Restart attempts belong to the model, with one pending request per source. Completed explicit intents cannot replay, and uncertain attempts only query status until the user deliberately starts another. Stale source/machine/auth identity and any intervening Stop request invalidate late results. Replies cannot restore a removed or recreated agent, undo a newer name/session, or let inventory begun before the restart restore the old session.

The CLI now persists restart receipts, including resumed/fresh outcomes. An in-memory coordinator also joins overlapping restarts of one canonical agent across clients and receipt IDs. Stop cancels that operation; target checks around asynchronous steps prevent subsequent process replacement or fresh fallback after cancellation. Already-issued system calls still finish. No real agent, shell, or running daemon was restarted during this work.

Verification:

  • Focused restart, stop, fork, rename, and pane regressions: 89 passed. New coverage includes 18 restart model tests and 6 keyboard tests. Log: /private/tmp/terminal-restart-focused-final.log.

  • A final keyboard check found that Enter could not activate a focused Close control while Restart was pending. Focused controls now own Enter before the pending-submit guard. All 27 restart model, keyboard, and pane checks pass, including Tab/Enter dismissal followed by reopening the same request. Log: /private/tmp/terminal-restart-review-focused.log.

  • Full desktop suite before that final focused-control fix: 2,266 passed, 5 skipped, no failures. Log: /private/tmp/terminal-restart-suite-final.log.

  • CLI socket, receipt, and restart tests: 117 passed across three files. Cancellation is exercised before dispatch and during hold, termination, resume preparation, respawn, and process discovery; no fresh fallback runs after cancellation. Log: /private/tmp/terminal-restart-cli-tests.log.

  • TypeScript typecheck passed with --incremental false, and the local CLI compiled successfully. Logs: /private/tmp/terminal-restart-cli-typecheck.log and /private/tmp/terminal-restart-cli-build-final.log.

  • Analyzer: no errors or warnings; the same three preexisting infos. Log: /private/tmp/terminal-restart-analyze-review.log (after the final keyboard and native fixture changes).

  • 103 rendered states in /private/tmp/terminal-restart-final. All ten new restart states were inspected, including pending reopen, refusals, uncertainty, explicit repeat confirmation, small-window/enlarged-text views, keyboard-scrolled details, fresh conversation, and shell restart. Log: /private/tmp/terminal-restart-render-final.log.

  • All thirteen touched Dart files are formatted; git diff --check passes.

  • Native macOS workspace fixture: 9 passed. Log: /private/tmp/terminal-restart-native-review.log. Earlier runs passed eight journeys, including Restart, but Find correctly released its index while the fixture window was inactive. The fixture now explicitly shows and focuses its own window for each journey, and waits for the completed query snapshot before checking all four matches. Product background indexing is unchanged.

  • Normal macOS debug rebuild succeeded after native verification, with explicit --target lib/main.dart; the generated target was checked. Artifact: build/macos/Build/Products/Debug/Harness.app. Log: /private/tmp/terminal-restart-macos-build-final.log.

The 30-item review checklist records exact keys, expected behavior, and useful feedback. Existing app instances and the installed application were not replaced or restarted. Status recovery requires the updated CLI; legacy successful replies remain supported. Desktop receipt references last for the current window, not across app restarts. Native tests use simulated traffic and injected Flutter keys. Real process replacement, physical keyboard/IME, VoiceOver, Linux runtime, settings, and remaining onboarding need further work. The broader goal remains active.

Terminal fork prompt and receipt recovery — preceding continuation

Fork Agent now uses the same thin terminal frame, font, transparent backdrop, and local key guide as search and creation. The source, machine, and actual project folder are visible above name > and task >. Task input owns focus immediately; the draft survives Escape. Native Next/Done, multiline tasks, Alt-Enter, Readline editing, composition, custom submit/accept/cancel/traversal, and live keymap reload are covered. Up/Down and Ctrl-N/P keep their editing behavior inside the task. Fork Agent is available in command search and as a remappable action without reserving a default terminal chord.

Pending forks now belong to the model. Reopening joins the existing request; repeated submission cannot start a second fork. Refusals keep the edited draft and report errors inline. A lost response locks the submitted choices and offers “check status” using the same creation receipt. Starting another fork is a separate explicit action followed by confirmation. A short uncertainty message, recovery controls, and Escape stay visible in short windows; longer details scroll with Page Up/Down.

The CLI now gives agent_fork the durable receipt support already used by creation. It reserves the intent before launch, coalesces repeated requests, and retains native/handoff information through status lookup and daemon restart. Spawn/registration failures remain unconfirmed rather than inviting an automatic retry. The desktop accepts legacy successful fork responses; lost-response status recovery requires this updated CLI. No running daemon was restarted or updated as part of verification. Desktop drafts and receipt references remain in memory for the current window.

The original destination and focus are captured when the fork starts. A late result opens there without moving the user's newer focus; closed or full background destinations leave the created agent available through New Pane or New Tab. A full tab still in focus opens a new tab named for the fork. Stale machine/auth/source identities are rejected before dispatch, and late receipts cannot undo a newer observed name or restore a fork stopped in the meantime. Restart behavior is unchanged in this continuation and remains a separate audit.

Verification:

  • Focused fork, creation recovery, placement, stop/rename, and menu regressions: 115 passed, including 19 new fork model tests and 8 keyboard tests. Log: /private/tmp/terminal-fork-focused-final.log.
  • Full desktop suite: 2,242 passed, 5 skipped, no failures. Log: /private/tmp/terminal-fork-suite-final.log.
  • Native macOS workspace fixture: 8 passed. The new journey verifies immediate task focus, multiline draft restoration, pending reopen, lost-reply status lookup without another fork, successful placement, and returning input to the original terminal. Log: /private/tmp/terminal-fork-native-final.log.
  • CLI receipt/socket checks: 89 passed; existing fork planning checks: 8 passed. Logs: /private/tmp/terminal-fork-cli-tests.log and /private/tmp/terminal-fork-cli-plan-tests.log. TypeScript typecheck passed; local CLI compilation passed (/private/tmp/terminal-fork-cli-build-final.log).
  • Analyzer: no errors or warnings; the same three preexisting infos. Log: /private/tmp/terminal-fork-analyze-final.log.
  • 93 rendered states in /private/tmp/terminal-fork-final. All nine new fork states were inspected, including pending reopen, refusals, uncertainty, 480×360 at 1.7× text, keyboard-scrolled details, and OpenCode handoff. Finder and creation previews were also inspected. Render log: /private/tmp/terminal-fork-render-final.log.
  • All fourteen touched Dart files are formatted; git diff --check passes.

The normal macOS review app was rebuilt after native testing with explicit --target lib/main.dart, also confirmed in generated configuration. Artifact: build/macos/Build/Products/Debug/Harness.app. Build log: /private/tmp/terminal-fork-macos-build-final.log. Existing user app instances remain untouched. Tests use simulated traffic; no live fork or engine-history copy has been exercised. Physical keyboard/IME, VoiceOver, Linux runtime, restart lifecycle, settings, and remaining onboarding still need further work. The broader goal remains active.

Terminal stop prompts and agent lifecycle — preceding continuation

Stop Agent / Stop Terminal now use the same compact terminal frame, font, and key guide as search and creation. The prompt identifies the agent and machine, explains what stopping does, and starts on Cancel. Controls and Escape remain visible in short windows; Page Up/Down scroll the explanation at enlarged text sizes. Stop Agent is available through command search and custom bindings, with Stop Terminal used for shell agents. No default terminal chord is taken. Pane menus use the same names and prompt.

Stopping now retains progress and errors inside the prompt. Its persistent focus node fixes Escape/custom cancellation after the focused Stop button disappears. Reopening joins the existing model-owned request instead of sending another. Failures return focus to Cancel and allow retry. Closing a pane continues to leave the agent running.

Confirmation captures the machine, agent session, auth revision, and observed deletion generation before the user accepts. A replaced target cannot receive the old confirmation's stop request; ordinary inventory/name updates are still accepted. Pending stop receipts also check identity/session, and an agent_deleted event can complete the request before its RPC receipt arrives. An old receipt cannot affect a recreated ID. All views of the stopped agent are removed across tabs before awaiting local detach cleanup, retaining other terminals. Older inventory cannot restore a stopped row or overwrite an ID recreated after the stop. Rename and new restart requests are blocked while a stop is pending.

Verification:

  • Focused stop, rename, machine-manager, pane, and handoff regressions: 103 passed, including 19 new stop model/widget checks. Log: /private/tmp/terminal-stop-focused-final.log.
  • Final full desktop suite: 2,215 passed, 5 skipped, no failures. Log: /private/tmp/terminal-stop-suite-final.log.
  • Native macOS workspace fixture: 7 passed. The new journey confirms the default Cancel, pending dismissal/reopen, inline refusal/retry, closing only the target's views, and returning input to the retained terminal. Traffic is entirely simulated. Log: /private/tmp/terminal-stop-native-final.log.
  • Analyzer: no errors or warnings; the same three preexisting infos. Log: /private/tmp/terminal-stop-analyze-final.log.
  • 84 rendered states in /private/tmp/terminal-stop-final; all eight new stop states were inspected, including pending reopen, errors, 480×360 at 1.7× text, and keyboard-scrolled details. Finder and creation previews were also inspected. Log: /private/tmp/terminal-stop-render-final.log.
  • All fifteen touched Dart files are formatted; git diff --check passes.

The normal macOS review app was rebuilt after the native fixture with explicit --target lib/main.dart, also confirmed in generated configuration. Artifact: build/macos/Build/Products/Debug/Harness.app (binary timestamp 18:53:41). Build log: /private/tmp/terminal-stop-macos-build-final.log. Existing user app instances were not restarted, and no real agent or shell was stopped.

The CLI acknowledges deletion before its asynchronous process cleanup ends; the native fixture does not prove live process termination. A CLI restart may retain the same agent/session identity; the new guard protects changed identities and observed deletion/recreation, not every same-agent restart. Requests for restart that began before Stop still need a separate lifecycle audit. Fork dialogs, settings, remaining onboarding, physical keyboard/IME, VoiceOver, and Linux runtime are further work. The broader goal remains active.

Terminal renaming and agent name consistency — preceding continuation

Tab, agent, and machine rename now share TerminalNamePrompt: the same thin terminal frame, name > editor, monospace font, and local key guide. The whole name is selected once on opening. Immediate typing, native Done, readline editing, composition, live picker remapping, inline errors, and short-window layouts share one implementation. Tab names keep the 80-character limit, and each editor has a specific accessible name.

Rename Agent is now a command-search action for the focused owned agent. It can be assigned a key without taking a default chord away from the terminal. The existing double-click path uses the same editor. Its delayed focus timer was removed: that timer could reselect a name after the user had begun typing. The direct title path is tested with background terminal output while editing. Native Rename Tab acknowledges the visible input before awaiting its value. Custom tab names remain independent of agent names.

Agent rename requests now belong to the model. Reopening joins a pending request with read-only input, repeated Enter cannot duplicate it, competing names are rejected, and different agents remain independent. Failures release busy state and retain the draft for retry. The CLI's returned name is used when present; timeout text distinguishes an unconfirmed result from a refusal. Inspection of cli/src/backendSocket.ts confirms that agent_update replies with the updated agent and broadcasts agent_renamed.

Late replies cannot rename a replacement machine, a different agent session, or a deleted/recreated ID. Auth changes and disposal invalidate outstanding requests. A newer observed name cannot be rolled back by an earlier request receipt or inventory read. Fresh inventory and sync events now update every open pane title, which previously required a separate rename event. Background agent inventory also checks the machine/auth identity before applying a reply.

Verification:

  • Focused rename, manager, commands, and focus regressions: 89 passed. Log: /private/tmp/terminal-renaming-focused-final.log.
  • New model and keyboard regressions: 16 passed independently. Log: /private/tmp/terminal-renaming-new-tests.log.
  • Final full desktop suite: 2,196 passed, 5 skipped, no failures. Log: /private/tmp/terminal-renaming-suite-final.log. The first full run found two focus tests still locating the removed AlertDialog; their editor finders now use the stable tab-name key, preserving their focus assertions.
  • Native macOS workspace fixture: 6 passed. The new journey renames a tab and agent through command search, closes/reopens a pending save, retries a refused save, uses Ctrl-W, and returns input to the original terminal. Traffic is entirely simulated. Log: /private/tmp/terminal-renaming-native-final.log.
  • Analyzer: no errors or warnings; the same three preexisting infos. Log: /private/tmp/terminal-renaming-analyze-final.log.
  • 76 rendered states in /private/tmp/terminal-renaming-final; all seven new rename states were inspected, including errors, pending work, and 480×360 with 1.7× text. The refactored machine rename was also inspected at ordinary and enlarged text sizes. Log: /private/tmp/terminal-renaming-render-final.log.
  • All fifteen touched Dart files are formatted; git diff --check passes.

The normal macOS app is rebuilt after native testing with explicit --target lib/main.dart, also confirmed in generated configuration. Artifact: build/macos/Build/Products/Debug/Harness.app. Build log: /private/tmp/terminal-renaming-macos-build-final.log. Existing user app instances remain untouched. The broader product goal remains active; stop/fork dialogs, settings, onboarding, physical keyboard/IME, VoiceOver, and Linux runtime remain useful next audits.

Machines Manager and account edits — preceding continuation

Machines Manager now uses the terminal frame and font, a machine > filter, compact naturally ordered rows, and a contextual key guide. This computer comes first; presence, linking, shared access, and agent counts remain distinct. Search includes names, hostnames, IDs, and status. Direct matches outrank loose fuzzy matches, and background updates retain the selected row.

Enter opens an action > list. Escape restores the machine query, then returns to the terminal. Nested rename, delete, password, and link prompts hide the parent panel and restore its action query and editing connection afterward. The live picker keymap follows every nested route. Refresh keeps the input focused and reports errors in place. Small windows scroll the body while the input and footer stay visible; prompt labels and text share a baseline.

Rename uses a name > prompt with the current name selected. Native Done and keyboard submission retain focus for retries. Delete names its account-level effect and starts on Cancel. Shared machines explain their view-only access, and this computer cannot be deleted here. Successful linking reports the result even when the retained filter no longer matches an available action.

Machine rename/delete requests now belong to the model. Identical pending requests join, competing edits are rejected, and reopening a prompt joins the existing operation. Failures retain the edited input or confirmation and offer a retry. Late replies cannot mutate a replacement machine or a different auth session. Successful deletion closes that machine's panes across this window's tabs; failed local stream cleanup does not repeat the committed account action. Terminal disposal runs even when stream close fails.

An inventory response begun before an edit, or an offline cache response after it, can no longer restore an old name or deleted machine. Newer live inventory can still reflect another client's changes or a re-registered machine. This protection belongs to the current window/auth session and clears on auth reset.

Verification:

  • Focused manager/menu/link/inventory set: 74 passed before the final prompt alignment and linked-status refinement; the final full suite includes those changes and the linked-status assertion. Log: /private/tmp/terminal-machines-manager-focused.log.
  • Final full desktop suite: 2,180 passed, 5 skipped, no failures. Log: /private/tmp/terminal-machines-manager-suite-final.log.
  • Native macOS workspace fixture: 5 passed. The added journey covers commands → machine/action search → rename → pending reopen → delete/cancel → pending reopen → retained terminal input. API responses and terminal traffic are fake. Log: /private/tmp/terminal-machines-manager-native-final.log.
  • Analyzer: no errors or warnings; the same three preexisting infos. Log: /private/tmp/terminal-machines-manager-analyze-final.log.
  • 69 rendered states in /private/tmp/terminal-machines-manager-final. All twelve new manager states were inspected, including pending/error views, shared access, missing machines, and 480×360 at 1.7× text. Search and creation previews were also checked. Log: /private/tmp/terminal-machines-manager-render-final.log.
  • All twelve touched Dart files are formatted; git diff --check passes.

The normal macOS review app is rebuilt after the native fixture with explicit --target lib/main.dart: build/macos/Build/Products/Debug/Harness.app. Generated configuration confirms the normal target. Build log: /private/tmp/terminal-machines-manager-macos-build-final.log. Existing user app instances remain untouched.

The broader product goal remains active. Other secondary dialogs, settings, onboarding, and platform parity remain useful next audits. Native fixture keys are injected through Flutter; physical AppKit keyboard/IME, VoiceOver, Linux runtime, and live daemon behavior remain outside this evidence.

Local password and outgoing links — preceding continuation

“This computer’s password” now uses the same thin terminal frame, transparent backdrop, monospace font, and local key guide as search and creation. The two masked prompts read password > and again >. Initial typing works immediately; Enter advances and then submits, native Next/Done keep focus, and Escape cancels an edit or returns to the chooser. Composition owns its keys. Readline editing, visibility controls, retry focus, and live remapping work across the nested route. The summary's focused control owns Enter after the input is removed.

The machine chooser hides while a nested password prompt is open, leaving one visible panel. Returning restores its query, selection, and focus. Outgoing trust is a separate “Links from this computer” view, keeping it distinct from this computer's incoming password. Fingerprints and dates remain available. Clear and unlink confirmations focus Cancel first; errors remain visible and retryable. Short windows scroll the body while the key guide stays visible.

Password mutations now belong to the model. Reopening joins a pending operation without retaining the closed password buffer; concurrent mutations are rejected. A late status read cannot overwrite a completed change, unexpected exceptions release busy state, and disposed/auth-stale replies cannot update current state. Linked-machine refreshes coalesce, retain known rows on failure, and show errors. Repeated unlink requests join; a failed refresh cannot restore a removed row.

The old clear confirmation claimed existing remote access would be revoked. Inspection of cli/src/lib/e2ee/manager.ts, store.ts, and the CLI clear command shows that clearing removes the password verifier; established trust/sessions are separate. The UI and Dart API comment now accurately say that clearing prevents new password-based links and leaves existing links/sessions connected. No actual password, trust pin, or live agent was changed during verification.

Verification:

  • Focused setup/password/keymap/boot set: 81 passed before the final summary ownership assertion; that assertion also passed in the final full suite. Log: /private/tmp/terminal-local-password-focused.log.
  • Final full desktop suite: 2,161 passed, 5 skipped, no failures. Log: /private/tmp/terminal-local-password-suite-final.log.
  • Native macOS workspace fixture: 4 passed, including a new commands → chooser → set password → pending reopen → change/cancel → retained terminal journey. All CLI responses and terminal traffic are fake. Log: /private/tmp/terminal-local-password-native-final.log.
  • Analyzer: no errors or warnings; the same three preexisting infos. Log: /private/tmp/terminal-local-password-analyze-final.log.
  • 57 rendered states in /private/tmp/terminal-local-password-final. All eight new states were visually inspected: initial, pending, failed, set, clear, outgoing links, unlink, and 480×360 with 1.7× text. Log: /private/tmp/terminal-local-password-render-final.log.
  • All nine touched Dart files are formatted; git diff --check passes.

The normal macOS review artifact is rebuilt after the native fixture with --target lib/main.dart: build/macos/Build/Products/Debug/Harness.app. Build log: /private/tmp/terminal-local-password-macos-build-final.log. Existing user app instances remain untouched. Physical AppKit input/IME, VoiceOver, Linux runtime, and live daemon behavior remain outside the fake-fixture evidence. The broader product goal remains active; Machines Manager, other settings and onboarding surfaces, and platform parity remain useful next audits.

Machine chooser and terminal setup guides — preceding continuation

“Link machine” now starts with a compact machine > chooser instead of two large installation cards. It searches machine names and setup actions, shows presence separately from trust, and puts direct matches before loose fuzzy matches. Existing machines can go straight to password entry. Desktop and SSH setup are explicit choices; already-linked machines explain how to open their agents. The local computer is excluded from the remote list.

Both setup guides use the terminal frame and monospace text. Commands and the download URL are selectable, all actions are keyboard focusable, and Tab reveals controls even in short windows at enlarged text sizes. Clipboard errors offer retry; commands are only copied, never executed. The desktop guide points to the current command-palette path, and the local password dialog was accurately titled “This computer’s password”. Its subsequent visual and interaction audit is recorded above.

Machine arrivals preserve the selected row and do not fold away an open guide. Escape returns from a guide to the original query, then closes the chooser. Nested password dialogs return to that query as well. Refresh remains available in every stage, coalesces repeated requests, and preserves focus and filtering on failure. Active composition retains ownership of its keys.

The chooser carries the live keymap across both direct workspace and Machines Manager routes. Navigation, accept, cancel, traversal, and refresh use picker bindings. A new picker.refresh command defaults to Cmd-R/Ctrl-R; only the machine chooser supplied its action at this stage. Custom bindings take precedence in hints. The configured accept key activates focused guide controls. An unbound Enter or Escape cannot fall through and open/close the search anyway. Live binding updates, disabled defaults, both route paths, and Linux refresh are tested.

Verification:

  • Focused setup/keymap/boot/manager set: 64 passed. Log: /private/tmp/terminal-machine-setup-keymap-checks.log.
  • Final full desktop suite: 2,147 passed, 5 skipped, no failures. Log: /private/tmp/terminal-machine-setup-suite-final.log.
  • Native macOS workspace fixture: 3 passed. Its linking journey now uses the actual command palette and machine chooser to reopen a pending request, verifies the retained query after linking, and returns input to the original terminal. Traffic remains entirely simulated. Log: /private/tmp/terminal-machine-setup-native-final.log.
  • Analyzer: no errors or warnings; the same three preexisting infos. Log: /private/tmp/terminal-machine-setup-analyze-final.log.
  • Forty-nine rendered states in /private/tmp/terminal-machine-setup-final. All five new machine chooser/guide states were inspected, including an empty result and the SSH guide at 600×420 with 1.7× text. Log: /private/tmp/terminal-machine-setup-render-final.log.
  • Formatting of all eleven touched Dart files and git diff --check passed.

The normal macOS debug app rebuilt with --target lib/main.dart after the native fixture: build/macos/Build/Products/Debug/Harness.app. Log: /private/tmp/terminal-machine-setup-macos-build-final.log. Existing app instances were not restarted. Work remains uncommitted and unpublished. Widget platform variants establish the Linux key behavior; Linux runtime, physical AppKit keyboard/IME, VoiceOver, and real daemon setup remain unverified. The broader product goal remains active.

Terminal machine linking and connection ownership — preceding continuation

The remote-password entry now shares search and creation's thin, top-anchored terminal frame, monospace text, transparent backdrop, and compact key guide. The target machine appears above a password > input. The body scrolls in short windows while the recovery instructions remain visible. Troubleshooting details stay optional, and the prompt follows terminal-font changes.

The keyboard audit reproduced two focus failures. Native Done unfocused the input, leaving Escape outside the prompt's handler while connecting. The dialog's fallback autofocus also took focus before its password editor, which the native fixture and a separate widget reproduction both caught. Done now retains focus; the prompt explicitly claims initial input after mounting. Enter submits from the input, Tab reaches controls, Escape returns to the original terminal, and active composition owns Enter/Escape. Shared readline editing works in the password field and is disabled while connecting. Errors are announced and return focus for a keyboard retry.

Connection attempts now belong to AppNotifier, keyed by machine and auth revision. Reopening the prompt joins the same pending future without copying the old prompt's password or launching another CLI request. Progress remains visible, repeated Enter is ignored, and different machines proceed independently. Unexpected link-process exceptions release the busy state and offer retry. Disposed or stale-auth replies cannot update the current workspace. The CLI remains responsible for the actual password exchange.

Verification:

  • Focused linking/boot/reconnect/offline set: 41 passed. Log: /private/tmp/terminal-link-final-checks.log.
  • Full desktop suite: 2,137 passed, 5 skipped, no failures. Log: /private/tmp/terminal-link-suite.log.
  • Native macOS workspace fixture: 3 passed, including retry, close/reopen while pending, automatic close on success, and input returning to the same retained terminal. All traffic is fake. Log: /private/tmp/terminal-link-native-workspace.log.
  • The dialog-focus regression failed before its fix; reproduction log: /private/tmp/terminal-link-dialog-focus-before.log.
  • Analyzer: no errors or warnings; the same three preexisting infos. Includes both native fixtures. Log: /private/tmp/terminal-link-analyze-final.log.
  • Forty-four rendered states in /private/tmp/terminal-link-final. The four new link states (ready, pending, error, and 600×420 at 1.7× text) were inspected, along with the current chooser and creation prompt. Log: /private/tmp/terminal-link-render-final.log.
  • The normal macOS app rebuilt successfully with --target lib/main.dart after native testing. Log: /private/tmp/terminal-link-macos-build.log.
  • Formatting of all six touched Dart files and git diff --check passed.

Artifact: build/macos/Build/Products/Debug/Harness.app. Existing app instances were not restarted. Physical AppKit keyboard/IME, VoiceOver, Linux runtime, and live daemon behavior remain outside this fixture evidence. The separate “Link another machine” setup guide and this computer's password-management dialog still need an interaction audit. Work remains uncommitted and unpublished; the broader product goal remains active.

Native workspace journey and prompt editing — preceding continuation

Output Find still had a rounded, proportional-font toolbar. It now uses the same thin terminal frame and monospace family as search and creation, with a / prompt and match count. It occupies the existing header, so opening Find does not resize or reflow terminal output. Narrow panes and larger text sizes put match controls in a compact keyboard menu, keeping the query readable. Enter activates a focused control; Enter/Shift-Enter step matches from the query. The menu closes with Find or its pane, and Escape restores typing.

The shared prompt editing audit reproduced six failures before the fix: Ctrl-H/D could split emoji or combining characters on macOS and Linux, Ctrl-W could erase a word on the preceding task line, and Ctrl-U erased all preceding lines. Character deletion now uses grapheme boundaries, word kill respects whitespace/path separators, and line kill stops at the current line's start. Ctrl-Y restores the killed text. These keys also work in output Find, with query updates isolated from terminal input. Composition and locked prompts remain protected.

The new native workspace fixture runs a real macOS window/titlebar around fake sessions and a fake creation connection. Its two keyboard journeys cover:

  • Staying at the current scrollback location while output streams; searching output and returning to the original location and terminal input.
  • Opening a shared agent in a named tab, closing/reopening it, switching panes, moving a pane, zooming, resizing, and retaining the original terminal view.
  • Editing a multiline task and inherited agent/project/mode, dismissing and resuming the draft through New Pane, checking the exact simulated creation request, and immediately finding that agent to name a new tab.

Both native fixtures now require FLUTTER_TEST=1. They never bootstrap the real CLI or load real Harness state. CLAUDE.md documents their invocation and the required normal-app rebuild afterward. Injected Flutter keys do not establish physical AppKit keyboard/IME or VoiceOver behavior.

Verification:

  • Focused prompt/Find/creation set: 50 passed. Log: /private/tmp/terminal-prompt-find-checks.log.
  • Full desktop suite: 2,131 passed, 5 skipped, no failures. Log: /private/tmp/terminal-workspace-suite.log.
  • Analyzer: no errors or warnings; the same three preexisting infos. Includes both native fixtures. Log: /private/tmp/terminal-workspace-analyze-final.log.
  • Native terminal fixture: 2 passed. A combined run then failed to launch its second test app before any workspace assertion ran. Log: /private/tmp/terminal-workspace-native.log.
  • Running the native workspace fixture independently: 2 passed. Log: /private/tmp/terminal-workspace-native-workspace.log. Use separate native invocations, as documented in CLAUDE.md.
  • Forty rendered states in /private/tmp/terminal-workspace-final; the three new Find states at normal/enlarged text and with keyboard options open were inspected. These use fake terminal output, not a live PTY. Log: /private/tmp/terminal-workspace-render.log.
  • Formatting of all eight touched Dart files and git diff --check passed.

The normal macOS debug app rebuilt successfully with --target lib/main.dart after the native fixtures. Artifact: build/macos/Build/Products/Debug/Harness.app. Build log: /private/tmp/terminal-workspace-macos-build.log. Existing app instances were not restarted. Work remains uncommitted and unpublished. The broader product audit remains active.

Clipboard ownership and Linux shell input — preceding continuation

The terminal audit reproduced three input failures: Linux Ctrl-A/Ctrl-V were intercepted for selection/paste, a delayed clipboard read could paste into a replacement agent, and an empty shell paste emitted Ctrl-V (quoted-insert). A follow-up reproduction showed Linux Alt word-motion keys emitted no input.

  • Linux clipboard keys are Ctrl-Shift-C/V/A. Ctrl-A/B/C/D/R/V/Z reach the program unchanged. The shortcut guide now names Linux's actual copy/paste/ select-all chords. Other existing platform clipboard shortcuts are unchanged.
  • Linux sessions use xterm's Linux input target. Left-Alt printable keys send an escape prefix with the actual character, preserving case, punctuation, and numeric arguments. The fallback letter handler now respects Shift. macOS Option and Linux AltGr remain native text composition.
  • Paste captures its target session and stream before reading the clipboard. It revalidates ownership after text and image reads, cancelling if the pane changes agents, reconnects, closes, or becomes read only. A pending paste cannot follow a reused pane to a different agent.
  • Empty clipboard content sends no quoted-insert to a shell. Clipboard read errors show a local retry message without freezing the stream. Paste clears the old selection. Read-only output can still be copied, older daemons retain bracketed paste, and Linux's paste chord reaches remote image upload.
  • The vendored xterm patch is documented in third_party/xterm/README.autonomous.md. Protocol behavior remains covered by the existing session tests; the new UI tests record actual outgoing input, paste, and image-upload frames using fake sessions and an in-memory clipboard.

Verification:

  • Focused input/clipboard/IME/shortcut set: 60 passed. Log: /private/tmp/terminal-input-final-checks.log.
  • Full desktop suite: 2,122 passed, 5 skipped, no failures. Log: /private/tmp/terminal-input-suite.log.
  • Analyzer: no errors or warnings; the same three preexisting infos. Includes the native terminal fixture. Log: /private/tmp/terminal-input-analyze-final.log.
  • Native macOS fixture: 2 passed. Log: /private/tmp/terminal-input-native.log. It exercised streamed keyframe replacement, alternate-buffer scroll routing, Unicode input injection, and switching across eleven agents while retaining exactly one terminal view. It ran in its own process with FLUTTER_TEST=1, fake terminal senders, no AppNotifier bootstrap, and no real agent traffic. The SDK desktop runner was inspected: it tracks and stops only processes it launches. Existing app sessions were not restarted.
  • This native fixture does not establish physical AppKit keyboard/IME behavior, VoiceOver behavior, a complete workspace walkthrough, or Linux runtime parity. Those remain outstanding. Earlier cgWindowNotFound capture failures still apply to the attempted walkthrough of the existing app window.

The normal macOS debug app was rebuilt successfully with --target lib/main.dart after the native fixture. Artifact: build/macos/Build/Products/Debug/Harness.app. Build log: /private/tmp/terminal-input-macos-build.log. The final clipboard checks passed all 14 tests after analyzer cleanup (/private/tmp/terminal-input-final-clipboard.log); formatting of all seven touched Dart files and git diff --check passed. Work remains uncommitted and unpublished.

The broad product goal remains active; terminal selection/scrollback and the other product areas still need the end-to-end review tracked in the audit.

Workspace arrangement and compact pane headers — preceding continuation

Directional pane moves previously removed and reinserted a pane, shifting unrelated neighbors and bypassing pinned-slot updates. They now use the same swap operation as direct reordering, keep pins with their panes, and request visibility for the moved terminal. Shift-Cmd-arrows stop at the layout edge; the existing wrap behavior for directional focus is unchanged.

  • The Flutter tab strip now reveals the selected tab after keyboard selection, closing, reordering, and window resizing. Background updates preserve a deliberately scrolled strip. The native strip already had reveal behavior.
  • Keyboard resize uses a compact monospace guide. Arrow keys resize, Shift takes larger steps, Tab changes divider, and Escape returns to terminal input.
  • Narrow pane headers prioritize a monospace agent name and machine identity. Extra project/branch detail remains in the tooltip. Model selection has a compact trigger; other actions move into an arrow/Enter/Escape menu. Wider headers retain their direct controls. Hover does not shift the title.
  • Reconnect/take-control remains directly accessible at narrow widths. Enlarged text no longer lets the action row reduce the title to zero width. Header typography follows changes to the terminal font preference.
  • The model menu now uses the thin terminal prompt frame, clamps to the actual window bounds, and scrolls when necessary. It owns focus, supports arrows and Enter, restores focus after Escape/selection, and dismisses on native chrome actions. Its Open Grid action is keyboard focusable. Outside clicks continue to reach the control beneath the menu.

Focused verification: 50 passed for headers, model menus, arrangement, and resize; 71 passed for viewer and terminal-session regressions. Logs: /private/tmp/terminal-arrange-final-checks.log and /private/tmp/terminal-arrange-regression-checks.log.

The first full suite exposed a viewer assertion that expected the now-collapsed icons, plus a resync timeout test that raced a 45 ms wall-clock delay. The viewer check now opens the real compact menu. The resync test uses the widget test's controlled clock and retains its exact three-retry/one-reopen/failure assertions; no production timeout changed.

Render verification: 37 states in /private/tmp/terminal-arrange-final, including four-pane resize, many tabs, compact actions, and model menus at 600 px / 1.7× text. The new states were inspected. The fixture exercises workspace chrome with fake sessions and matching agent metadata; no real PTY or agent is started. Log: /private/tmp/terminal-arrange-render.log.

Final verification:

  • Full desktop suite: 2,108 passed, 5 skipped, no failures. Log: /private/tmp/terminal-arrange-suite-final.log.
  • Analyzer: no errors or warnings; the same three preexisting infos. Log: /private/tmp/terminal-arrange-analyze-final.log.
  • macOS debug build succeeded with Swift Package Manager. Artifact: build/macos/Build/Products/Debug/Harness.app. Log: /private/tmp/terminal-arrange-macos-build-final.log.
  • Formatting (13 touched Dart files) and git diff --check passed. Existing app instances were not restarted; native runtime and VoiceOver checks remain outstanding. No changes were committed or published.

The broad product goal remains active. Native runtime, terminal correctness, and remaining product areas are still tracked in the product audit.

Draft ownership and visible recovery keys — preceding continuation

The canceled compact prompt previously saved only a global task string. A reproduction showed that reopening from another project put that task in the wrong context, and canceling after an advanced-form round trip lost the chosen Codex profile. It also reset the edited machine, project, and permissions.

  • Drafts now retain the complete setup in memory for their source machine, agent, and project. Separate agents working in the same project have separate buffers, as do machines with identical folder paths. The buffers last for the current window. They are not persisted across app restarts.
  • An empty chooser or Cmd-N resumes that source's draft. Explicit search text starts a fresh task with inherited defaults. Clearing the task removes it; successful creation consumes only that source's draft. Advanced options keeps the original source context through edits and returns.
  • An unsubmitted draft honors the newly chosen New Tab/New Pane destination. A dismissed unresolved creation retains its original receipt, even with no task. Resuming sends agent_create_status with that same creation ID and completes at the original destination. It does not launch another agent.
  • Errors previously replaced all key hints. They now retain the available keys underneath. A pending prompt says pending agent and exposes check status and close; editing and advanced-options hints are omitted. Busy requests show progress. All message text uses the shared monospace styling.
  • The dismissed receipt carries its warning acknowledgement so an existing “Escape again” message still means what it says after reopening.

Focused tests: 60 passed including the render fixture, and 30 recovery checks passed after preserving the warning acknowledgement. Exact outgoing requests cover clone URLs, profiles, permissions, first task, placement, and receipt identity. Logs: /private/tmp/terminal-context-final-checks.log and /private/tmp/terminal-context-recovery-checks.log.

Render verification: 31 states in /private/tmp/terminal-context-final, adding pending creation and reopened pending creation at 600 px / 1.7× text. Both new states were inspected. The fixture simulates a timeout through its overridden connection; no real agent creation or daemon traffic is involved.

Final verification:

  • Full desktop suite: 2,102 passed, 5 skipped, no failures. Log: /private/tmp/terminal-context-suite.log.
  • Analyzer: no errors or warnings; the same three preexisting infos. Log: /private/tmp/terminal-context-analyze-final.log.
  • macOS debug build succeeded with Swift Package Manager. Artifact: build/macos/Build/Products/Debug/Harness.app. Log: /private/tmp/terminal-context-macos-build.log.
  • Formatting and git diff --check passed. Existing app instances were not restarted; native runtime and VoiceOver verification remain outstanding.

The broad product goal remains active. This completes the scoped draft audit; the remaining product areas are listed below and in the product audit.

Mixed catalogs and workspace navigation — preceding continuation

Compact search now names the agent and machine before the project and branch. Removing provider logos had left otherwise identical Codex and Claude tasks indistinguishable without a preview. The plain-text identity also uses discovered agent names such as Robot Studio. Existing domain metadata remains searchable.

  • Wide results stay on one line. Narrow results, enlarged text, and the list beside a preview put identity on a second line. Row height follows the actual text size, and the selected row stays visible when preview or geometry changes.
  • Commands and shortcut help now say Next Tab, Previous Tab, and Select tabs. History says Reopen Closed Tab or Pane in both the palette and native menu. Existing IDs and bindings are unchanged.
  • A complete keyboard journey verifies finding a specific agent among duplicate task names, opening a shared tab, closing and reopening it, tab and pane switching, zoom, and closing/reopening a pane. The same running sessions survive every transition, and picker keys do not reach terminal input.

Verification:

  • Full desktop suite: 2,098 passed, 5 skipped, no failures. Log: /private/tmp/terminal-catalog-suite.log.
  • Analyzer: no errors or warnings; the same three existing infos. Log: /private/tmp/terminal-catalog-analyze.log.
  • Render fixture: 29 states in /private/tmp/terminal-catalog-final, including mixed local/remote/offline agents, duplicate tasks, preview, and 600 px with 1.7× text. These states were visually inspected. Log: /private/tmp/terminal-catalog-final-render.log.
  • macOS debug build succeeded. Artifact: build/macos/Build/Products/Debug/Harness.app. Log: /private/tmp/terminal-catalog-macos-build.log.

The narrow-preview test waits for the scheduled post-layout scroll before inspecting the selected row's viewport bounds. The keyboard journey lets the terminal's existing 50 ms resize debounce finish before teardown. No production timer or scrolling change was required for those two test timing corrections. Native app interaction remains unverified; existing app instances were not restarted and no real agents were created.

Draft recovery and picker ownership — preceding continuation

Advanced options now returns to the compact task prompt with the complete edited draft. Escape, the back hint, and an outside click retain the task, machine, agent, project or clone intent, permission mode, and Codex profile. An explicit Default profile remains default even when discovery finds one named profile. The compact prompt displays the selected profile; changing machine or agent clears it before the next launch.

  • An unresolved creation keeps its original receipt across the return. Enter checks that request instead of sending another create. Pending input is read-only, including the custom readline editing keys. Busy requests still prevent dismissal.
  • A dialog pop could complete before the screen's cached route flag updated, suppressing the returning prompt. The open guard now reads the live route.
  • The advanced agent picker previously let Cmd-Enter reach the parent form and launch its old selection. It now accepts the highlighted agent; an unbound modified Enter cannot submit the parent form. IME composition does not choose an agent, dismiss the picker, or launch a request.
  • The form now receives the workspace's live keymap. Picker remappings, unbindings, aliases, and visible hints remain effective after opening the separate dialog route, including config changes while it is open.
  • The agent picker measures its anchor and available space during layout, replacing the snapshot taken when it opened. Resizing or changing text size keeps the panel, input, preview, and key guide inside the window. The result list gives up space before the input or key guide does.

Focused verification: 48 tests passed, including exact creation payloads, receipt recovery, profile round trips, real workspace key dispatch, IME, remapped keys, and narrow/short/enlarged-text layouts. Log: /private/tmp/advanced-agent-final-checks.log.

Render verification: 25 isolated macOS fixture states passed and were written to /private/tmp/terminal-draft-preview; the returned draft, enlarged draft, and resized agent searches were inspected. The fixtures reject real agent creation. Native app interaction remains unverified for the reason below.

Final verification:

  • Full desktop suite: 2,094 passed, 5 skipped, no failures. Log: /private/tmp/terminal-draft-suite.log.
  • Analysis of lib and test: no errors or warnings; the same three preexisting informational lints. Log: /private/tmp/terminal-draft-analyze-final.log.
  • Formatting and git diff --check passed.
  • macOS debug build succeeded with Swift Package Manager. Artifact: build/macos/Build/Products/Debug/Harness.app. Log: /private/tmp/terminal-draft-macos-build.log. Existing native app instances were not restarted. No real agent was created for verification.

The broad product goal remains active. This pass completes the draft/back and nested-picker audit; the remaining areas are listed below and in the product audit.

Creation integrity and terminal forms — preceding continuation

Advanced options now shares the compact terminal appearance: an 840 px top-anchored thin frame, monospace text, full-width machine/project choice rows, small controls, and a visible keyboard guide. The backdrop stays clear. The agent selector uses a > prompt and Ctrl-/ preview toggle, preserves input focus, and supports readline editing and Page Up/Down. Provider descriptions and installation status remain available in the preview.

  • The compact prompt's key guide wraps instead of hiding later hints offscreen. Creation sizes itself from the actual fields, editor, and guide. Long task text scrolls in its editor at the available height, keeping actions visible.
  • Advanced handoff preserves named new projects, clone requests, existing folders, permission modes, and the first task. A successful full-form launch consumes the saved task. Tests inspect the resulting creation request.
  • Explicit project choices survive switching to another machine and back in advanced options. The dialog's per-machine choice bucket includes inherited choices and new-project names.
  • Outside-click and keyboard dismissal now share the compact prompt's pending creation guard; a lost receipt cannot be bypassed by clicking outside.
  • Numbered agents, projects, and folder completions use natural ordering after search score and recency (1, 2, 10). The already-resolved collection package is now an explicit dependency; no package version changed.
  • The form's agent picker changes selection only when the pointer moves; rebuilding or scrolling beneath a parked pointer leaves keyboard selection alone.

Verification:

  • Full desktop suite: 2,083 passed, 5 skipped, no failures, recorded in /private/tmp/terminal-refinement-suite-final.log.
  • Earlier in this continuation, the handoff/order changes alone passed 2,082 tests. Focused checks exercise narrow and enlarged-text layouts, multiline tasks, actual creation payloads, machine round trips, and preview toggling.
  • After visual review corrected the two-line agent row height, all 27 focused picker/form tests passed. The responsive tests verify that both text lines remain inside each row's selection bounds. Log: /private/tmp/terminal-final-picker-tests.log.
  • Final analysis of lib and test: no errors or warnings; the same three preexisting informational lints. Log: /private/tmp/terminal-refinement-analyze-final.log.
  • macOS render fixtures passed and produced the final screenshots in /private/tmp/terminal-final-preview, including advanced options, agent search and preview, and narrow/large-text forms. The fixture uses an isolated connection and rejects agent creation. Inspected representative search, creation, advanced, and large-text states, and corrected the row-height issue found during that review.
  • Final macOS debug build succeeded with Swift Package Manager. Artifact: build/macos/Build/Products/Debug/Harness.app. Log: /private/tmp/terminal-refinement-macos-build.log. Existing app instances were not restarted during this continuation.
  • Native UI inspection remains unavailable: the selected app returned cgWindowNotFound, and inventory reports multiple OpenHarness entries. Render fixtures and keyboard tests do not claim a live native walkthrough.

The ongoing product audit is in terminal-product-audit.md.

Terminal appearance and interaction — preceding continuation

The search and creation overlays now use the terminal font family, a thin border, compact rows, text prompts, and an ASCII key guide. Large provider marks, rounded cards, repeated row-action buttons, and the creation explanatory paragraph/button were removed from these overlays. The workspace stays visible.

  • New Tab/New Pane search starts with a compact list and a > prompt. The New agent row stays above matches; its detail carries any proposed first task. A match count is always visible.
  • Ctrl-/ toggles the preview. Its text and status are also compact and monospaced. Once opened, its geometry stays stable while navigating, including the create row, so Tab traversal does not lose its place.
  • Page Up/Down page results when preview is hidden, or scroll the preview when shown. Escape/Ctrl-C dismiss the picker; Ctrl-C still reaches the terminal after dismissal. All bindings remain scoped and remappable.
  • Creation exposes labeled agent, machine, mode, and project defaults above task >. Tab edits defaults through the same prompt; Enter chooses an option and returns to the task. Existing path completion, drafts, receipt recovery, and cancellation guards remain. Defaults stack with narrow/large-text layouts.
  • The command palette has a : prompt and an empty input instead of a second > inside the editor.
  • Design references and interaction rules record the studied fzf, zsh, Vim, tmux, and Lazygit patterns.

Verification:

  • Full desktop suite: 2,075 passed, 5 skipped, no failures.
  • After the final hidden-preview paging addition: 88 targeted tests passed, including creation, placement, preview paging, input ownership, remapping, cancellation, and terminal key routing.
  • Analysis of lib and test: no errors/warnings; the same three preexisting informational lints in app_state, grid_model_picker, and pane_grid.
  • Rendered and inspected search, creation, defaults, preview, command palette, and large-text/narrow layouts. The render fixture uses Menlo for the terminal font aliases; runtime uses the user's terminal font. PNGs are in /private/tmp/terminal-box-preview.
  • macOS debug build succeeded with Swift Package Manager. Artifact: build/macos/Build/Products/Debug/Harness.app. Opened a fresh review instance with open -n; existing instances were not stopped. No real agent was created during verification.

New Tab / New Pane — built after the design discussion

The primary choice is now where to put the work:

  • ⌘T / New Tab opens the harness picker with a pending new-tab destination.
  • ⌘P / New Pane opens the same picker for the current tab, automatically tiling after an addition. From the Store or Orchestrator it offers a new tab.
  • Both show a pinned New agent row, selected with an empty query. Typing selects the best existing match and carries the text into the first task if New Harness is chosen. Projects, machines, and tab names are searchable metadata; the results are individual harnesses, with no bulk group actions.
  • New Tab allocates only after choosing an existing harness or receiving a successful creation receipt. Cancellation, failures, and pending recovery leave the layout alone. A recovered receipt allocates exactly once.
  • A tab opened through New Tab takes the chosen or newly created agent's name. Adding/focusing panes preserves an existing tab name, including custom names.
  • A harness can appear in multiple tabs. Within a tab it appears once: choosing a harness already there focuses its pane. Views share the session, and closing one view leaves the others intact.
  • The picker and creation actions show the destination. Existing creation inheritance (agent, machine, project) remains, including multiple feature harnesses on the same project and panes from different machines.
  • Native and Flutter tab strips now have New Tab and New Pane controls. The File menu follows the same model. Directional split edges were removed; directional splits and Orchestrator remain in the ⇧⌘P command palette, without default directional/orchestrator bindings. ⌘N remains a secondary direct-creation shortcut. Existing command IDs remain remappable.
  • A pending creation cannot be replaced by another picker or the full form. Cancellation after transitioning from search restores terminal focus rather than focusing the disposed search editor. The full form receives the pending destination and task.

The placement intent is lib/state/harness_placement.dart. Regression coverage is in test/harness_placement_test.dart, with existing entry, keymap, shared session, layout, and recovery tests updated for the new behavior. Keyboard documentation is in ../docs/keyboard.md.

Verification of this implementation

  • Desktop suite: 2,074 passed, 5 skipped, no failures. The old pane-control count assertion now includes the existing Fork button.
  • Final analysis of lib and test: no errors or warnings; the same three existing informational lints. The final cleanup subset also passed (68 tests).
  • Focused tests cover successful, failed, and recovered creation; retained context and task; cancellation; shortcut/focus recovery; cross-machine panes; shared views; duplicate selection; full tabs; and stale results.
  • Rendered and inspected the New Tab and New Pane pickers, matching/no-match results, creation destination, permission selection, and separate command palette. Fixture glyph substitutions remain as recorded below.
  • macOS debug build succeeded with Swift Package Manager. Final artifact: build/macos/Build/Products/Debug/Harness.app.
  • Live creation against the user's running harnesses was not performed.

Earlier continuation work

  • Verified the final held-Tab fix through real key-down/repeat/up events in the screen's keymap dispatcher. The chosen agent is committed once; subsequent repeats preserve the machine and project. Existing controller tests also cover a current machine absent from the choices and Escape restoring a path-completion stem.
  • Search now puts openable matches before unavailable matches, preserving ranking within each group and any explicitly selected row. Location-tree ordering and the create row's position keep their existing contracts.
  • The search overlay measures its editor and hints before sizing the results to whole rows. This works with side-by-side and stacked previews and larger text. A create-only result uses exactly one row plus list padding.
  • Risky permission modes use amber text in the list and selected mode label. Codex's choices read Read only, Ask first, Auto-approve, Full access. The default remains Auto-approve by ID, verified in an actual form submission.
  • New-harness errors and creation status use explicit accessibility announcements. Updates coalesce per frame, messages take precedence over row announcements, and clearing an error allows it to be announced again.
  • Corrected the keyboard documentation: text carried from search becomes the first task, not a new project name.

Relevant additions: test/box_polish_test.dart. The implementation changes are in state/swarm_search.dart, widgets/swarm_switcher.dart, screens/swarm_screen.dart, state/new_harness.dart, widgets/new_harness_box.dart, widgets/box_chrome.dart, and core/permission_modes.dart, all under lib/.

Earlier verification

Using Flutter 3.47.2 / Dart 3.13.2:

  • Before changes: 2,056 desktop tests passed, 5 skipped, 1 failed.
  • After changes: 2,064 desktop tests passed, 5 skipped, the same 1 failed. The unchanged failure is swarm_interactions_test.dart, "pane controls zoom, confirm stopping and close only this view": it expects six icon buttons after a seventh Fork button was added on the base branch.
  • Static analysis of lib and the changed tests: no errors or warnings; three existing informational lints in app_state.dart, grid_model_picker.dart, and pane_grid.dart.
  • CLI: all 19 new-command/project-folder tests and TypeScript checking passed.
  • All 12 rendered box states were inspected. Known render-fixture artifacts remain: some shortcut glyphs render as boxes and some logos load late.
  • The normal macOS debug app built successfully with Swift Package Manager. Artifact: build/macos/Build/Products/Debug/Harness.app.
  • Live interaction with the newly built app and native VoiceOver have not been verified. Accessibility tests verify emitted announcements.

Next work

The broad product goal remains active. Next audits include pane moves and resizing, many-tab workflows, secondary dialogs, terminal correctness, reconnect/offline states, settings, and platform parity. Mixed-catalog identity, the keyboard close/reopen journey, and scoped draft ownership are verified above. Renamed/moved-project relinking still needs a design decision. Physical AppKit keyboard/IME, VoiceOver, Linux runtime, and live daemon verification remain outstanding; separate fake-traffic native fixtures are recorded above.

The original handoff's independent terminal-reviewer re-review has not been run in this continuation; verification above was performed locally. Details and unresolved findings are maintained in terminal-product-audit.md.

First workspace guide and predictable harness entry — 2026-09-20

Built the approved pared-down keyboard wallpaper as native Flutter line art: Cmd-T new tab, Cmd-P new pane, Cmd-S Harness Store, and the command dock. The callouts invoke the actual actions and read the active keymap. A small All keyboard shortcuts link opens the existing help sheet. The production empty workspace uses this guide; the old full-form compatibility path is retained.

The first empty workspace automatically opens the real New Harness draft, with an editable ~/harnesses/first-project destination. Machine-scoped engine detection chooses an installed agent, preferring a saved choice when available. Discovery does not overwrite an agent or machine the person has begun editing. No harness starts until Enter/click. A persisted marker keeps Escape dismissed; restored tabs or discovered existing harnesses bypass automatic entry.

The creation/search vocabulary now calls the whole session a harness: New harness, Start harness, New harness…, and Find a harness…. The inner engine field remains agent. New Tab and New Pane still name placement.

The New harness… action is pinned outside the scrolling search matches, across the dock width directly above the input. Typing, clearing, scrolling, and preview changes leave it there. Typed queries still select their best match; Down reaches creation. Preserved result state across preview changes fixes keyboard focus being lost during Tab traversal.

Cmd-S opens Harness Store. Layout is Cmd-Shift-L, including the native menu; Cmd-L remains focus right. Shortcut help and the guide follow remaps.

Validation: 95 focused unit/widget/render checks passed; two native macOS journeys passed (first workspace and edited creation/placement). Static analysis of changed production code and new fixtures is clean. Dart-to-AppKit export and 130 native keymap checks passed. The subsequent legacy titlebar harness cannot compile: stale openButton, createButton, and managerCaption references in tool/swarm_titlebar_checks.swift. It was not changed for this task.

The first native fixture exposed that local creation prepares its project folder before sending agent_create. Its empty first-project test folder was removed, and the fixture now overrides folder preparation with an in-memory path. No real agent was created. The native fixtures' injected Flutter keys do not prove physical AppKit/IME event delivery.

Rebuilt the normal lib/main.dart debug app with the existing JEV endpoint http://127.0.0.1:18478, verified its code signature, and opened it. No commits or pushes. Logs: /private/tmp/harness-guide-{focused,native,native-keys,analyze,build}.log. First-run render: /private/tmp/harness-guide-render/first-harness.png.

Entry polish, focus, titles, and Store — 2026-09-20 follow-up

This supersedes the earlier guide and entry behavior above. Agent settings now live behind a small gear shown on hover/focus; permissions and Codex profiles no longer insert rows into the agent list. New Harness stays pinned above the search input. Labels use title case, the action reads Start, the optional task reads Add a task (optional), and typed tasks have no task: prefix.

The welcome illustration is larger, with one tall pane on the left and two on the right, using Codex, Claude Code, and Grok examples. Tab, Pane, and Store callouts explain the concepts and are static. Only All Keyboard Shortcuts is clickable; it shows the current Cmd-/ binding. Removed the redundant command dock callout. Each visit to the empty workspace opens New Harness when there are no harnesses, or search when harnesses already exist. Escape stays dismissed for that visit. Empty first projects default to ~/harnesses/first-project.

Installed-agent detection now has a separate status beside Agent. Start stays labeled Start and cannot activate during detection. The visible snapshot tracks detection completion so a stale loading label cannot conceal an enabled action.

Creation hands keyboard focus to the newly mounted terminal after closing the dock, with guards against stealing focus after navigation. Four native macOS fixture journeys pass, including New Tab and New Pane accepting input without a click. The fixture needed a valid protocol UUID to record encoded input. These injected Flutter keys do not establish physical AppKit/IME behavior.

Automatic CLI labels display as Untitled Pane, and new tabs as Untitled Tab. Engine session titles take over until the user renames; explicit tab naming ownership and its source harness persist across layout restore and reopening. Placeholder labels are never sent to the CLI as explicit names.

Store Open and Try this prompt invoke the workspace dock with the chosen harness, machine, and example task. The Store button is now a compact filled pill with an original six-color polymath mark, shared by its tab and History. The native button preserves keyboard focus, enabled state, and remapped shortcut tooltip; its normal, hover, and disabled states were rendered and inspected. Asset vector source: tool/render_store_mark.swift.

Validation: 56 Store/navigation/title checks pass; agent settings, welcome, project context, command dock, and placement checks passed earlier in this follow-up. All four native creation/onboarding fixture journeys pass. Logs: /private/tmp/harness-store-name-regression.log and /private/tmp/harness-settings-native.log. Button preview: /private/tmp/harness-store-preview/button.png. No real harnesses were created by these fixtures, and no commits or pushes were made.

The normal lib/main.dart review build was rebuilt with the existing JEV URL, passed codesign --verify --deep --strict, and was opened. Analysis has no errors or warnings; two pre-existing style infos remain in app_state.dart and pane_grid.dart. Build log: /private/tmp/harness-final-build.log.

Editable generated project defaults — 2026-09-20

This supersedes the fixed first-project default above. The first-workspace draft shows a concrete name using the existing harness-and-date convention, for example ~/harnesses/codex-2026-09-20-17-22. The timestamp is captured when the suggestion is made; Start uses the reviewed name even after a delay. Project → New Project prefills and selects the proposed name for replacement. Edited names remain explicit and survive agent changes. Machine changes keep separate project choices, and an untouched suggestion follows its agent.

A read-only directory listing offers an unused candidate. Start reserves the folder exclusively, so a concurrent collision advances only generated names to seconds and then a suffix. Existing folders and their files are preserved. Long labels retain the timestamp and suffix within the remote naming limit. Remote confirmed collisions get a fresh receipt; uncertain replies and replies belonging to another receipt cannot trigger another start. User-chosen names are never silently renumbered. A known name conflict leaves the form editable. The action remains Start and its pending label is Starting harness….

Validation: 103 focused unit/widget tests pass (one optional render skipped), plus four native macOS creation/onboarding journeys and the separate first-run render. Inspected /private/tmp/harness-generated-project-render/first-harness.png. Analysis has no errors or warnings and one pre-existing informational lint in app_state.dart. git diff --check and strict bundle signature verification pass. Normal lib/main.dart was rebuilt with the existing JEV endpoint and opened. Logs: /private/tmp/harness-generated-project-{test,native,analyze,render,build}.log. These fixtures do not create live agents. No commits or pushes were made.

Live computer-use inspection was attempted again, but the native tool returned Sky Computer Use native pipe startup failed. Native fixtures use injected Flutter events; physical keyboard/IME and live daemon behavior remain separate from the evidence above. The broader product audit goal remains active.

Consistent creation entry points — 2026-09-20

Fixed Store product switches restoring the previous product's draft. Draft ownership now includes the entry source and explicit product, in addition to machine, source harness, and project context. An already-open dock processes a new product request instead of just focusing its old Start row. Store Open/Try do not inherit a workspace search task, project, or split. Explicit product and machine choices win over incompatible saved defaults; user edits remain in their own compatible drafts. Uncertain starts retain their exact receipt and cannot be silently replaced by another task.

Cmd-T and Cmd-P retain their common focused-pane defaults and draft recovery; the current shortcut decides placement. Found and removed a separate legacy popup route in the native Models menu. It now shares the product entry helper with Store and pane-menu requests, including the native keyboard handoff. The current rules and coverage matrix are in new-harness-entry-rules.md.

Validation: 2,398 unit/widget tests passed; 6 optional tests skipped. All 7 native macOS creation journeys passed, including immediate input after Cmd-T/Cmd-P and Store Open/Try/Models starts. Focused product-route checks also passed (66). Updated stale test expectations for the already-approved Start wording, layout shortcut, pinned creation row, and Store discovery action. Full analysis has no errors or warnings (18 existing informational lints, mostly vendored xterm); targeted analysis reports one existing style info. git diff --check passes. Rebuilt normal lib/main.dart with the existing JEV endpoint, verified the bundle signature, and opened it for review.

Logs: /private/tmp/harness-entry-full-final.log, /private/tmp/harness-entry-native-final.log, /private/tmp/harness-entry-analyze-final.log, and /private/tmp/harness-entry-build.log. Fixtures use fake transport; no live harnesses were started. No commits or pushes were made. Live computer-use inspection remains unavailable as noted above; this is automated fixture coverage, not a claim of physical keyboard/IME validation.

Startup and keyboard entry — 2026-09-20

Removed redundant chmod subprocesses from persisted settings reads. Each operation still stats the current path and repairs permissions when needed; there is no cached permission assumption. Tests cover external changes to the directory and lock, a second store instance, and replacement of the directory at the same path. Locking, fresh reads, atomic writes, and private file modes remain covered by the persistence suite.

The existing startup benchmark uses fresh store objects, temporary files, 20 warmups, and 100 measured samples per operation in the headless debug runner. Settings initialization median/p95 changed from 21.656/41.626 ms to 4.205/7.378 ms. Settings plus keymap changed from 22.604/25.238 ms to 6.052/9.858 ms. These are warm-filesystem initialization measurements, not cold-launch, first-frame, or native input latency results. Raw logs: /private/tmp/harness-startup-baseline.log and /private/tmp/harness-startup-optimized.log.

Settings now focuses Search settings immediately. Enter selects the first matching section, Down reaches its row, and Escape returns to the workspace. Filtering preserves the current section until selection; empty/unmatched queries do not switch sections. Composition retains its keys. A native journey verifies Cmd-comma → search → Enter → Escape → terminal input with no mouse click or leaked input. Pending sign-in now accepts Escape to cancel and restores Sign in focus for Enter to retry. Late completion of a cancelled attempt cannot sign in or disturb a new attempt.

Validation: 2,406 unit/widget tests passed, 6 optional tests skipped; all 18 native workspace journeys passed. The focused settings/sign-in suite has 16 passing checks; the earlier startup, persistence and sign-in suite has 64. Full analysis has no errors or warnings and the same 18 informational lints. Formatting is unchanged across the nine touched Dart files. Logs: /private/tmp/harness-startup-settings-full.log, /private/tmp/harness-startup-settings-native-full.log, /private/tmp/harness-settings-signin-final-test.log, and /private/tmp/harness-startup-settings-analyze.log.

Live computer-use inventory was retried and again failed with Sky Computer Use native pipe startup failed; no live UI inspection is claimed. Native tests use simulated sessions and injected Flutter keys. No real harnesses, sign-ins, or user settings were changed by the tests.

Rebuilt the normal application from explicit lib/main.dart with the existing JEV endpoint after the native fixtures. The build and strict deep signature verification pass; git diff --check is clean. Build log: /private/tmp/harness-startup-settings-build.log. No commit or push was made.

Keyboard help at enlarged text and reliable config reload — 2026-09-20

Reviewed all five production Settings sections at the minimum 880×560 window, in light and dark themes at 1× and 1.8× text. The render fixture navigates via the Settings search and Enter, uses synthetic account data, and never invokes a mutating settings action. It also renders the keyboard-shortcuts sheet in each theme/scale. All 24 final images were inspected in /private/tmp/harness-settings-review/; the fixture passes 20 checks.

Found and fixed clipped keycaps and a truncated shortcut-context selector at large text. Keycaps now have a minimum height and grow to fit glyphs or wrapped multi-stroke bindings; Tab and Return symbols scale too. The deck reduces its column count as text grows, and the context selector gives its label enough width. Three new layout tests failed before the changes and pass afterward. The real-font context-width check also reproduced the truncation before its fix. Settings keyboard coverage now includes Tab into the content, Page Down/ Page Up, changing shortcut context, Escape closing only its open menu, and then Escape returning to the terminal. Both native Settings and keyboard- practice journeys pass with fake traffic and injected Flutter keys.

The full regression run exposed an intermittent dotfile-symlink reload failure. An isolated ordinary run passed, but a controlled filesystem-resolution delay reproduced a missed save after switching targets: the new map was published before the new target's watch was installed. KeymapStore.reload now establishes the watches before reading and publishing the map. The delayed test observes the real filesystem replacement event and passes with this ordering; it does not hide the issue with a longer timeout. The focused watcher, remapping, and Settings suite passes 38 checks. The generated keyboard-config template now correctly describes moving New Tab from Cmd-T and lists the project context. Existing user configuration files are preserved.

Logs: /private/tmp/harness-shortcuts-scale-before.log, /private/tmp/harness-settings-context-before.log, /private/tmp/harness-keymap-watch-before.log, /private/tmp/harness-settings-help-watch.log, /private/tmp/harness-settings-help-review.log, and /private/tmp/harness-settings-help-native.log.

Further Settings review remains: the Usage range selector still abbreviates its current range at enlarged text, and the About card could use more room for its action descriptions. This pass checks the visible initial states and keyboard navigation; provider data, pairing, account changes, and update flows still require their own full interaction audit. Live computer-use remains unavailable as recorded above.

Final validation for this pass: 2,431 unit/widget tests passed; 6 optional tests skipped. Full analysis has no errors or warnings and the same 18 informational lints. The normal lib/main.dart macOS build completed after the native fixtures and passes strict deep signature verification. Formatting and git diff --check pass. Logs: /private/tmp/harness-settings-help-full-final.log, /private/tmp/harness-settings-help-analyze-final.log, and /private/tmp/harness-settings-help-build.log. No commit, push, real account change, device pairing, or production configuration edit was performed.

Honest update results and readable Usage/About — 2026-09-20

The existing Cmd-T/Cmd-P and Store ownership rules remain the creation contract in new-harness-entry-rules.md. Rechecked the native first workspace, new tab, new pane, Store Open, Store Try, Models, and edited-default placement journeys: all seven pass, including immediate terminal input after creation. They use simulated sessions and do not launch real harnesses.

Usage's current range was truncated at 1.8× text in both themes. The real-font regression failed before the change. Overview and provider headers now share a range field that grows with text, and controls move below their captions when needed. Keyboard selection, Escape, and return focus are tested in both views. The local-only Usage subtitle is shorter. About explanations and actions stack at enlarged text; its version/status line wraps instead of overflowing.

The update audit found failures that were being presented as "up to date" and a concurrent manual check that returned an early, stale answer. Three new regressions reproduced these before the fix. Results now distinguish current, available, disabled, and failed checks, including absent/malformed platform metadata. Manual and background callers share a request. About and the native menu share one result dialog; a failed result offers Retry in place. About does not claim the build is current before a successful check. Development, disabled, and viewer builds do not contact the desktop update service.

Two further failing regressions showed that Update/Skip could act on a newly arrived offer instead of the version displayed. Both actions now take the reviewed version explicitly. Failed installs retain that version for retry; skipping an older displayed version leaves a newer offer visible. A recheck failure preserves a known offer and its installation error. Valid current metadata clears a withdrawn offer. Escape dismisses results and pending checks, while an installation stays open. Late replies after dismissal or notifier disposal are covered. Already-skipped versions show Close and Update, without the redundant Not now action.

Background polling now matches the six-hour cadence described in About, does not overlap a slow request, and cannot publish after its timer is cancelled. Download hashing, staging, and platform artifact selection still pass their existing tests. All update tests use intercepted manifests, temporary fixture archives, or a fake installer; no real desktop update was performed.

The render fixture captures 52 states at 880×560, in light/dark and 1×/1.8× text. All 36 Usage/About/update images were inspected in /private/tmp/harness-settings-update-review/. Dialog action fonts use the app font explicitly; labels, facts, and controls remain readable at enlarged text. The expanded native Settings journey verifies keyboard navigation into About, the disabled-build result, Escape returning to the button, then Settings closing and terminal input returning without a click. That journey passes, bringing the native reruns in this pass to eight.

Final regression: 2,450 unit/widget tests passed; 6 optional tests skipped. Full analysis has no errors or warnings and the same 18 informational lints. Formatting passes for all 15 touched Dart files. Evidence: /private/tmp/harness-settings-update-full-final.log, /private/tmp/harness-settings-update-analyze-final.log, /private/tmp/harness-settings-update-native.log, and /private/tmp/harness-settings-update-native-launch.log.

Live computer-use remains unavailable with the previously recorded native-pipe startup failure. Native fixtures inject Flutter keys and fake transport; physical keyboard/IME, Linux, real device/account changes, and the broader product/performance audit remain unfinished. No commit or push was made.

Rebuilt the normal lib/main.dart macOS application after the native fixtures, preserving JEV_COMMAND_BAR_URL=http://127.0.0.1:18478. The build and strict deep signature verification pass; git diff --check is clean. Build log: /private/tmp/harness-settings-update-build.log. The rebuilt app was not opened.

Cmd-T/Cmd-P preserve the open picker — 2026-09-20

Rechecked the creation entry contract after the user's request to concentrate on the common keyboard paths. The initial 62 focused tests passed, including Workshop → Blender Open/Try, machine changes, generated versus edited names, source-pane draft ownership, uncertainty, and destination recovery. A further keyboard audit found a separate data-loss interaction: pressing Cmd-T or Cmd-P while search was open closed and recreated it, clearing its query. All four repeat/switch combinations failed before the fix: /private/tmp/harness-picker-retarget-before.log.

The picker now changes destination in place. Repeating a shortcut refocuses its existing input; switching shortcuts preserves the query, text selection, highlighted harness, and project/machine scope. Header and Enter action update together. Capacity and existing-pane membership are refreshed for the chosen destination. Store/orchestration tabs continue to send terminal work to a new tab. Pending creation receipts still prevent a shortcut from replacing an unresolved start.

Seven additional widget journeys cover the four repeat/switch combinations, both directions with a highlighted existing harness, and project filtering with Escape back to the parent. The full-tab model regression now switches the same picker between destinations and verifies capacity and Focus pane versus Open in new tab. The native New Tab/New Pane journeys type a task, switch and repeat shortcuts, start it, then send terminal input without a mouse click.

Validation: 2,457 unit/widget tests passed, 6 optional tests skipped. All seven native creation journeys passed: first workspace, New Tab, New Pane, Store Open, Store Try, Models, and edited-default placement. Full analysis has no errors or warnings and the same 18 existing informational lints. Formatting and git diff --check pass. The normal lib/main.dart macOS build was restored after the fixture, retaining the command-bar URL above, and passes strict deep signature verification. No real harness was started; the native fixture uses fake transport and injected Flutter keys, and does not establish physical AppKit/IME or live-daemon behavior. The rebuilt normal app was not opened.

Evidence:

  • /private/tmp/harness-entry-recheck.log
  • /private/tmp/harness-picker-retarget-focused.log
  • /private/tmp/harness-picker-retarget-full.log
  • /private/tmp/harness-picker-retarget-native.log
  • /private/tmp/harness-picker-retarget-analyze-final.log
  • /private/tmp/harness-picker-retarget-build.log

The broader overnight goal remains active. The next independent audit is Usage's delayed load/scan/toggle/disposal behavior. Code inspection found missing generation guards in UsageLedgerStore and post-disposal checks in UsageLedgerController; controlled timing regressions and any fixes remain to be implemented. Native performance benchmarking still requires isolation from the three development previews last observed (PIDs 21874, 32129, 64557); none were stopped. Live computer-use remains unavailable as previously recorded.

Usage lifecycle, loading states, and repeated aggregation — 2026-09-20

Controlled delayed reads and writes reproduced 12 failures in the old Usage store/controller lifecycle. A late scan could restore figures and a snapshot after Off; delayed preferences or cache reads could overwrite an explicit choice; reopening could race pending persistence; and disposing during load could still notify or start a scan. Failed reads also retained stale figures or joined a failed request when Retry should have started a new one.

The stores now guard results with a choice revision, share concurrent loading and scanning, and serialize provider persistence across reopened stores. Switching Off clears memory immediately and queues snapshot deletion after earlier writes. Failed/unavailable scans clear figures, timestamps, and cached sources. An Off/On waits for the old physical read and discards its result before rescanning; the scanner API cannot cancel a read already in progress. Disposal prevents later state changes, notifications, or new scanning.

Usage distinguishes reading preferences, an initial scan, a rescan with previous figures, failure, and Off. Initial pending/failed/disabled sources no longer show zero tokens, cost, or sessions as if those were measured results. Changing the date range uses in-memory data and does not rescan the disk.

Repeated overview aggregation was measured with 60,000 synthetic turns in warm debug CPU work: median 30,716 μs and p95 88,298 μs before the change. Caching each range and its cutoff date gives a median below the timer's 1 μs resolution and p95 1 μs afterward. Provider changes and midnight invalidate the relevant results. This only measures repeated overview lookups, not cold parsing, first aggregation, app launch, or native input/frame latency.

Enlarged-text renders exposed a daily-chart overflow and shorter bars floating above the baseline. Chart height/width now use measured label sizes; small bars align at the bottom. A dedicated geometry regression checks 1:2 bar heights, the shared baseline, and readable labels at 2.5× text.

Validation:

  • 2,480 unit/widget tests passed; 7 optional tests skipped. The added benchmark is opt-in. After the final zero-session-count refinement, all 80 focused Usage tests passed again.
  • Added 18 lifecycle tests and five UI/chart tests. The four UI workflows cover both themes at 1.0×/1.8× text and 880×560, with real font metrics and icon fonts. All 24 captured loading/failure/rescan/Off states were inspected.
  • Eight native journeys passed together: Settings/About, first workspace, Cmd-T, Cmd-P, Store Open, Store Try, Models, and edited-default placement. The new native Usage journey initially sent keys before its menu's first frame. After making that fixture wait for the rendered menu, it passed separately, including keyboard range selection, Retry, and Off during scan.
  • Static analysis has no errors or warnings and the same 18 informational lints. Formatting and git diff --check pass.
  • Rebuilt the normal lib/main.dart macOS app after native fixtures with JEV_COMMAND_BAR_URL=http://127.0.0.1:18478; strict deep signature verification passes. The rebuilt normal app was not opened.

Evidence:

  • /private/tmp/harness-usage-lifecycle-before.log
  • /private/tmp/harness-usage-full.log
  • /private/tmp/harness-usage-final-focused.log
  • /private/tmp/harness-usage-native.log
  • /private/tmp/harness-usage-native-retry.log
  • /private/tmp/harness-usage-performance-before.log
  • /private/tmp/harness-usage-performance-after.log
  • /private/tmp/harness-usage-lifecycle-review/
  • /private/tmp/harness-usage-analyze-final.log
  • /private/tmp/harness-usage-build.log

All Usage data in tests is synthetic; no real transcript was read or harness started. Native fixtures use fake transport and injected Flutter keys. Live computer-use and isolated native performance measurement remain limited as previously recorded. No commit or push was made. The broader goal remains active: next inspect scanner cache validity while OpenCode writes to SQLite WAL, unreadable-database states, and cold parsing/aggregation performance. These are inspection targets, not verified bugs or completed work.

OpenCode live reads, incomplete data, and scan responsiveness — 2026-09-20

The next audit used temporary SQLite databases exclusively. Six initial regressions failed: a committed update in WAL was missed when the main file's mtime/size stayed unchanged; locked and corrupt databases were called missing; an unsupported schema returned a successful empty scan; a broken sibling was silently omitted from totals; and sessions with only cache-write tokens were excluded by the query.

OpenCode now queries SQLite's committed snapshot for each requested scan, without using the main file's metadata as proof that its contents are unchanged. The aggregate query runs in a worker isolate. The store's five-minute cache still avoids rescanning on every visit. Tests hold a writer open, update its WAL, check an uncommitted transaction is excluded, then checkpoint/truncate and update again. They also cover locks and recovery, corruption, absent versus misconfigured directories, unsupported schemas, duplicate session ownership, cache-write-only usage, and a genuinely empty supported database.

An unreadable source is a read failure, distinct from a missing installation. When other databases are readable, their figures remain visible with an Incomplete badge and explanation. Overview totals carry the same warning. The warning stays during Retry, and disappears only after a complete read. Partial results clear the old disk snapshot rather than being restored later as a complete fresh result. Empty incomplete scans do not claim zero sessions or zero tokens; successful empty scans still can.

A report-only benchmark uses 30,000 synthetic sessions and a 1 ms periodic timer on the caller isolate. Across five measured scans after one warmup, the median maximum event-loop gap fell 44,440 → 5,828 μs. Median total scan time was 44,320 → 60,631 μs. The tradeoff is responsiveness while waiting for the worker, not higher query throughput. This is debug scheduling evidence on a shared development machine, not a native frame/input measurement or the cost of subsequent aggregation/rendering.

Validation:

  • 2,493 desktop unit/widget tests passed; 8 optional tests skipped. Nine scanner, two lifecycle, and two UI regressions were added; the four existing theme/text-scale journeys now exercise partial data and retry too. The added scheduling benchmark is opt-in.
  • Three native macOS journeys pass: delayed Usage retry/range/Off, real SQLite worker reads and recovery through the Refresh button, and Settings/About keyboard ownership with return to the terminal. The real-database journey only touches its temporary fixture. It never reads production transcripts.
  • Thirty-two Usage states were captured in both themes at 1.0×/1.8× text; the 12 new or changed partial-data/rescan images were inspected. The unchanged loading/Off states were inspected in the preceding pass.
  • Static analysis has no errors or warnings and the same 18 existing info findings. Formatting and git diff --check pass.
  • The normal macOS lib/main.dart app was rebuilt after native testing with JEV_COMMAND_BAR_URL=http://127.0.0.1:18478 and passes strict deep signature verification. It was not opened. No commit or push was made.

Evidence:

  • /private/tmp/harness-scanners-before.log
  • /private/tmp/harness-scanners-first-fix.log
  • /private/tmp/harness-scanners-partial-ui-final.log
  • /private/tmp/harness-scanners-performance-before.log
  • /private/tmp/harness-scanners-performance-after.log
  • /private/tmp/harness-scanners-full.log
  • /private/tmp/harness-scanners-native.log
  • /private/tmp/harness-scanners-analyze-final.log
  • /private/tmp/harness-scanners-build.log
  • /private/tmp/harness-usage-partial-review/

The broader goal remains active. Live computer-use and isolated native timing have the limitations already recorded. The next useful work is controlled JSONL file-read failures and cold parsing/first-aggregation profiling; current Claude/Codex parsers still run on the caller isolate and swallow filesystem read failures as empty entries. Those observations need concrete reproductions before choosing changes. Device/account changes and the remaining product audit are also still outstanding.

JSONL failure recovery, live writes, and first aggregation — 2026-09-20

Temporary Claude/Codex transcripts reproduced ten failures: denied file reads were silently cached as empty successful files (and stayed empty after access returned), unreadable roots were called missing, a broken sibling could hide valid data, and a file caught halfway through a Unicode character failed the whole refresh. These are now handled by a shared file lifecycle while each provider keeps its existing parsing and deduplication rules.

The shared reader distinguishes missing roots from listing/read failures, retains readable sources with an incomplete-data warning, and never caches a failed source as empty. Snapshot format 3 invalidates version 2 caches that may contain that mistake. A migration regression uses a fresh old snapshot and verifies the next refresh really scans, rather than accepting its cached zero.

Reads stop at the captured file size. A streaming transformer holds only the last four bytes, omitting a valid unfinished UTF-8 suffix while still sending all earlier bytes through the strict UTF-8 decoder and normal line splitter. Earlier complete usage records survive concurrent writes; a subsequent refresh picks up the finished record. Corrupt completed text remains a failed source. Tests cover every byte cut through two-, three-, and four-byte characters, CRLF, a valid last line without a newline, malformed and out-of-range suffixes, both transcript roots, appended/deleted files, and permission restoration. Permission-denial cases skip only on Windows or a host that can bypass their temporary file permissions; they all ran on this macOS host.

The timing fixture uses 20,000 synthetic usage turns and a 2 MiB tool result. It found the larger aggregation pause in repeated Claude model-name matching. A bounded cache (128 names per provider, maximum 512-character keys) now retains normalization results, including unknown names. Token-dependent tier arithmetic still runs separately for each entry; neither rates nor model-name matching rules changed. Cold file reads retain the standard chunked decoder.

Final debug medians, five measured runs after one warmup:

Work Before After
Claude cold scan 137.717 ms 129.314 ms
Codex cold scan 86.577 ms 90.045 ms
Claude cached scan 0.567 ms 0.665 ms
Codex cached scan 0.600 ms 0.794 ms
Claude aggregation 21.288 ms 4.671 ms
Codex aggregation 2.128 ms 2.076 ms

The corresponding median maximum caller event-loop gap during Claude aggregation was 21.383 → 4.732 ms. Cold-scan gaps stayed around 5–6 ms. These are shared-machine debug observations, not a native-frame or full-app input latency claim. The final comparison is performance-before versus performance-final; an intermediate per-line decoder was discarded after timing showed unnecessary overhead.

Validation:

  • 2,513 desktop unit/widget tests passed; 10 optional tests skipped. Added 18 transcript/Unicode regressions, one old-cache migration regression, one per-turn pricing regression, and two opt-in benchmark cases.
  • Five native macOS journeys pass: delayed Usage retry/range/Off, live SQLite reads, live Claude transcripts, live Codex transcripts, and Settings/About keyboard ownership with terminal focus restored. Both JSONL journeys use real temporary files and the actual Refresh control to recover from an interrupted Unicode write and a corrupt sibling.
  • Static analysis has no errors or warnings and the same 18 existing info findings. Formatting and git diff --check pass.
  • Rebuilt normal lib/main.dart after the native fixture, retaining JEV_COMMAND_BAR_URL=http://127.0.0.1:18478. Strict deep signature verification passes. The rebuilt app was not opened. No commit or push was made.

Evidence:

  • /private/tmp/harness-jsonl-before.log
  • /private/tmp/harness-jsonl-focused.log
  • /private/tmp/harness-jsonl-stream-fix.log
  • /private/tmp/harness-jsonl-performance-before.log
  • /private/tmp/harness-jsonl-performance-final.log
  • /private/tmp/harness-jsonl-full.log
  • /private/tmp/harness-jsonl-native.log
  • /private/tmp/harness-jsonl-analyze-final.log
  • /private/tmp/harness-jsonl-build.log

All transcript data is synthetic and confined to temporary directories. No user transcript was read, no account changed, and no real harness started. Native keys remain Flutter-injected; live computer-use and isolated native performance measurement retain the previously recorded limits. The broader goal is active. Next audit account changes and reconnect/inventory responses, with attention to old replies restoring stale workspace or onboarding state. The rest of the device, platform, and product audit remains outstanding.

Account lifetime, sign-out recovery, and readable authentication — 2026-09-20

The initial viewer-auth probes reproduced ten failures in fourteen cases. An old refresh could restore credentials after sign-out, overwrite a new account, or clear it when the old refresh failed. Cancelling browser sign-in during server binding, authorization, exchange, or persistence could still publish a URL or save credentials. An older attempt's cleanup also owned the replacement attempt's callback reference.

Viewer authentication now gives each login and account session a revision. Refresh completion checks that revision before saving or clearing anything; its cleanup only releases its own in-flight future. Credential writes and deletions are serialized, and token readers wait for pending writes. A cancelled save clears its incomplete credentials before a newer queued login is allowed to write. Callback cancellation closes its listener, treats an early outcome as data until it is awaited, and cannot cancel a replacement.

Desktop sign-out previously ignored CLI failure and enabled Sign in while the CLI could still be clearing the saved session. It now owns the child process, checks its exit, drains output, and terminates a timed-out child before returning failure. App state joins repeated sign-out requests, clears the visible account, and waits for credential and connection cleanup before allowing sign-in. Failure offers a focused Retry sign out action with a plain explanation. Failed terminal/connection closure cannot strand the remaining cleanup. Disconnecting a development fixture never runs real CLI logout. A delayed machine response can no longer change signed-out cache status.

Visual review caught a separate accessibility issue: at 880×560 with 1.8× text, the decorative fleet diagram pushed the action and recovery message below the viewport. Short windows now omit that diagram when text is enlarged. The action and explanation remain visible, and retry/success restore keyboard focus. The capture fixture was also corrected to synchronize its global palette with the requested theme; its initial light captures mixed light text styles with a dark palette. All twelve final captures were inspected.

Validation:

  • Full suite: 2,541 passed, 10 optional tests skipped. Added 17 viewer lifecycle cases, three CLI logout cases, seven sign-out state/keyboard/ visual cases, and one delayed-inventory case.
  • After the viewport adjustment: all 25 sign-in/sign-out widget regressions pass, followed by all seven sign-out cases with synchronized light/dark captures. Capture states are pending, failed, and ready at 1.0× and 1.8×.
  • Nine native macOS journeys pass: sign-out retry and keyboard focus, Settings, first workspace, New Tab and New Pane creation, Store Open and Try, Models, and edited creation destination. Creation tests deliver terminal input without a mouse click and cover product-prefix isolation.
  • Static analysis: no errors/warnings, the same 18 informational findings. Formatting and git diff --check pass.
  • Normal lib/main.dart rebuilt with the existing command-bar URL http://127.0.0.1:18478; strict deep signature verification passes. The rebuilt app was not opened. No commit or push was made.

Evidence:

  • /private/tmp/harness-auth-lifecycle-before.log
  • /private/tmp/harness-auth-recovery-focused.log
  • /private/tmp/harness-auth-full.log
  • /private/tmp/harness-auth-visual.log
  • /private/tmp/harness-auth-visual-final.log
  • /private/tmp/harness-auth-review/
  • /private/tmp/harness-auth-native.log
  • /private/tmp/harness-auth-analyze-final.log
  • /private/tmp/harness-auth-build.log

Accounts, credentials, CLI processes and harnesses in these tests are fakes; viewer callback tests use ephemeral loopback listeners and in-memory storage. No real account was changed. Native keys are Flutter-injected, so physical AppKit/IME and live account interaction retain their earlier verification limits. The broader goal remains active. Next inspect runtime session expiry, saved workspace restoration across account changes, and concurrent machine inventory responses; the remaining platform/device/performance audit continues.

Workspace restoration and concurrent inventory — 2026-09-20

Thirteen initial account/workspace regressions reproduced lost restoration, stale layout reads after sign-out, and callbacks from replaced connections. Four additional inventory cases reproduced older replies replacing newer rows, loading state, or retry errors. These cases now pass.

Sign-out and runtime expiry share live-workspace cleanup. They close all terminal sessions on their original transports, clear account-owned inventory and history, and leave one fresh empty tab without saving it over the previous desk. Sign-in waits for pending cleanup; cancellation while waiting prevents later login. Layout reads check account and layout revisions. Saved tabs, active selection, pinning, layout presets, and zoom restore after signing in. The persisted desk remains computer-local intent, not an account-namespaced store. Replaced WebSocket connections lose callback ownership immediately.

Machine inventory now has request ownership as well as account ownership. Only the latest request can publish rows, finish loading, or clear its error. API responses retain their own freshness, and shared-machine fallback cannot cross account changes. Retry retains the public refresh path, whose boolean result tells callers whether its response was applied. An early version bypassed this path; the full suite caught three regressions, which were fixed.

Runtime expiry now presents one Sign in action with its explanation, without mislabeling it as a failed sign-in. Four captures were inspected at 880×560: dark/light themes and 1.0×/1.8× text. A visual fixture initially left a timer alive after successful sign-in; explicit fixture disposal fixed all four failures. The first native restoration fixture waited for an indeterminate reconnect spinner to settle; it now waits for sign-in completion and checks keyboard tab selection directly, with a bounded timeout.

Validation:

  • Focused account/connection cases: 80 passed. Focused inventory cases: 80 passed. Retry/public-path regression pass: 71 passed. Final inventory-result and expiry-screen checks: 29 passed.
  • Full suite: 2,565 passed, 10 optional tests skipped.
  • Native macOS: 10 passed, covering expiry/restoration, sign-out recovery, Settings, first workspace, Cmd-T/Cmd-P immediate input, Store Open/Try, Models, and edited draft placement.
  • Static analysis: no errors or warnings, the same 18 informational findings.
  • Normal lib/main.dart rebuilt with command-bar URL http://127.0.0.1:18478; strict deep signature verification passes. The app was not opened. No commit or push was made.

Evidence:

  • /private/tmp/harness-workspace-account-before.log
  • /private/tmp/harness-workspace-account-focused.log
  • /private/tmp/harness-inventory-before.log
  • /private/tmp/harness-inventory-focused.log
  • /private/tmp/harness-workspace-account-retry.log
  • /private/tmp/harness-workspace-account-final-focused.log
  • /private/tmp/harness-workspace-account-full-final.log
  • /private/tmp/harness-workspace-account-review/
  • /private/tmp/harness-workspace-account-native-final.log
  • /private/tmp/harness-workspace-account-analyze-final.log
  • /private/tmp/harness-workspace-account-build.log

Account and terminal fixtures are synthetic; API tests use an ephemeral loopback server and memory storage. The temporary retry-path regression bypassed discovery overrides and attempted local read-only discovery; no real account changed and no real harness started. Native keys are Flutter-injected; live AppKit/IME and platform/device checks retain their recorded limits. The broader goal remains active. Next inspect restored panes whose machines no longer appear in a successful inventory and their reconnect guidance.

Clear recovery for unavailable restored panes — 2026-09-20

A successful inventory with no row for a saved pane still displayed an indefinite "Waiting for this machine" spinner. Five initial UI regressions failed, covering a fresh empty reply, a recovered-but-absent machine, a cached empty reply, failed reads, and keyboard Retry.

App state now records whether an inventory has completed for the current account. A missing fresh row says the machine is unavailable; cached data says its status could not be confirmed; failed reads say machines could not be loaded. Initial loading retains its waiting state. All completed states offer Retry, which stays mounted while a request is pending and preserves its position and keyboard focus. Repeated activation joins the current retry. The progress/icon space has a fixed size. Account changes reset completion. Unknown agent headers use the established Untitled Pane placeholder instead of exposing a raw identifier, and retained-output notices no longer promise an answer from an absent machine.

Validation:

  • 64 focused checks pass, including keyboard activation during/after refresh, unchanged Retry geometry, stale-versus-fresh guidance, and account reset. The pane fixture now listens to notifier updates as the real shell does; its earlier static host was corrected.
  • 12 small-pane captures at 440×240 were inspected: missing/cached/failed in both themes at 1.0× and 1.8× text. All content and Retry remain visible. Initial captures lacked a Material ancestor; the final fixture includes the app's normal surface and removes the debug banner.
  • Full suite: 2,581 passed, 10 optional tests skipped.
  • Native: nine launch/account scenarios passed in the combined run. The new recovery assertion first assumed Retry was the first Tab stop; the workspace also has pane-header controls. The corrected test traverses the real order, activates Retry with Enter, and verifies another inventory request. It passes separately, retaining saved-tab selection checks. Ten native scenarios therefore pass across these final runs.
  • Analyzer: no errors or warnings; the existing 18 informational findings. Formatting and git diff --check pass. A missing import for the shared placeholder constant was caught and corrected before the final runs.
  • Normal lib/main.dart rebuilt with command-bar URL http://127.0.0.1:18478; strict deep signature verification passes. The rebuilt app was not opened, and no commit or push was made.

Evidence:

  • /private/tmp/harness-missing-machine-before.log
  • /private/tmp/harness-missing-machine-focused-final.log
  • /private/tmp/harness-missing-machine-review/
  • /private/tmp/harness-missing-machine-full-final.log
  • /private/tmp/harness-missing-machine-native-final.log
  • /private/tmp/harness-missing-machine-native-recovery.log
  • /private/tmp/harness-missing-machine-analyze-final.log
  • /private/tmp/harness-missing-machine-build.log

This pass uses fake inventory and account fixtures and in-memory storage. No real account changed or real harness started. Native keys are Flutter-injected; physical input, live account/daemon behavior, and platform coverage retain the previously recorded limits. The broader product goal remains active. Continue with reconnect transitions, retained terminal state, and the remaining device/platform/performance checks.

Reconnect discovery and retained terminal input — 2026-09-20

Five new discovery probes failed before the fix: a reconnect reused pending capability discovery from the old connection; its late success/failure could replace new capability state; an old inventory timeout marked the reconnected machine offline; an offline transition still accepted pending discovery; and a background inventory reply overwrote the fresh reconnect snapshot. Two further probes reproduced a second reconnect waiting behind obsolete recovery and an old stream ID remaining writable after connection replacement.

Machine discovery now has a connection revision in addition to its account and machine-object guards. A disconnect or new connection releases old inventory/capability futures so fresh discovery can begin immediately. Each completion checks its originating revision. Reconnect clears negotiated capabilities until this connection answers, and old completion handlers cannot release a newer request's ownership. Background inventory uses the same guard. Pending-agent recovery has explicit ownership and checks its revision both before and after discovery; a second reconnect can begin without waiting for the first recovery loop to finish.

A replacement connection also releases prior stream IDs immediately. The session and renderer remain in place with retained output, scroll, selection, and draft state, while input waits for the replacement stream's first keyframe. Taken-over sessions remain under the existing explicit-retry policy.

Validation:

  • The initial probes failed 5/5, then 2/2. The final focused reconnect, discovery, offline, and retained-view pass has 37 passing checks. The final shared discovery fixture separately has 19 passing checks.
  • Full suite: 2,588 passed, 10 optional tests skipped.
  • Native workspace: 11 passed together. The new journey verifies retained renderer, output, scroll, selection, suppressed input before negotiation, fresh screen arrival, unchanged focused pane, and input resumption without a mouse click. Main Cmd-T/Cmd-P, Store Open/Try, first-run, Settings, account recovery, and edited draft placement remain covered in that run.
  • Native terminal: 2 passed, exercising keyframe replacement and switching agents without stale/duplicate native terminal views.
  • Analyzer: no errors or warnings; 18 existing informational findings after correcting six new brace-style findings. Formatting and diff checks pass.
  • Normal lib/main.dart rebuilt with command-bar URL http://127.0.0.1:18478; strict deep signature verification passes. The app was not opened. No commit or push was made.

Evidence:

  • /private/tmp/harness-reconnect-discovery-before.log
  • /private/tmp/harness-reconnect-handoff-before.log
  • /private/tmp/harness-reconnect-discovery-focused-final.log
  • /private/tmp/harness-reconnect-discovery-final.log
  • /private/tmp/harness-reconnect-discovery-full.log
  • /private/tmp/harness-reconnect-discovery-native.log
  • /private/tmp/harness-reconnect-terminal-native.log
  • /private/tmp/harness-reconnect-discovery-analyze-final.log
  • /private/tmp/harness-reconnect-discovery-build.log

The direct computer-use retry returned empty app/browser inventories and Sky Computer Use native pipe startup failed. Native checks therefore remain Flutter-injected input with synthetic transport; they do not establish physical AppKit/IME, live credentials/daemon behavior, or other-platform parity. No real agent or account was changed. Immediate discovery is verified by request ordering under held fake replies, not a network latency claim. The broader goal remains active. Continue the performance and interaction audit, especially sustained command-dock use and large inventories, with the existing physical/platform limits kept explicit.

Command-dock query rendering profile — 2026-09-20

Exercised the real SwarmScreen with 2,000 and 10,000 synthetic harnesses: repeated Cmd-T, broad/narrow typing, another character with the same matches, arrows, and Cmd-P → Cmd-T retargeting. The fixture asserts that no machine transport is used. Optional CPU sampling connects only to its own loopback VM service and saves samples outside the workspace. Run the benchmark explicitly; it is excluded from the normal *_test.dart suite.

The first CPU sample involved Flutter rebuild work in roughly 44% of samples; search filtering appeared in roughly 7%. Debug/test overhead and truncated stacks limit interpretation. A focused regression then reproduced five visible ListTile controls rebuilding when only query match text changed. Row text now listens for match-query/help-mode changes separately from the bounded control cache. Unchanged controls keep their widgets; row metadata, availability, selection, actions, geometry, palette, and font still invalidate as needed. The regression now sees zero unchanged control rebuilds, updated matches, and retained input focus. Existing arrow movement still rebuilds only two rows.

A paired experiment temporarily restored query-driven control invalidation, then removed it before final verification. Consecutive-character timings:

Synthetic harnesses Controls invalidated: median / p95 Controls retained: median / p95
2,000 11.059 / 12.485 ms 9.513 / 11.888 ms
10,000 16.568 / 19.649 ms 17.648 / 35.046 ms

These are 50-sample headless debug widget elapsed measurements after eight warmups. Broad open/filter timings also have noisy tails. The consistent result is reduced rebuild work; the mixed timings do not establish a general speedup, native display latency, or release-build performance. No ranking or visible interaction rules were changed for this optimization.

Verification:

  • 88 focused dock, Store, project, draft, and placement checks passed.
  • The initial full run found a race in the shared-viewer fixture: a recorded server request was treated as proof its reply had reached the UI. The fixture now awaits the actual displayed response. Its four tests pass, followed by the full suite: 2,589 passed, 10 optional tests skipped.
  • All 25 native macOS workspace journeys passed together, including first workspace, shared views, Cmd-T/Cmd-P creation, edited placement, Store Open/Try/Models, immediate terminal input, Settings, machine management, account recovery, rename, stop, fork, and restart.
  • Analyzer: no errors/warnings; the same 18 existing informational findings. Two new informational findings were corrected before the final analysis.
  • The optional profiler and cleanup path passed in a final 2,000-row run.
  • Normal lib/main.dart rebuilt with the existing command-bar URL at port
    1. Deep strict code-signature verification passed. App not opened; nothing committed or pushed.

Evidence:

  • /private/tmp/harness-dock-row-before.log
  • /private/tmp/harness-dock-query-focused.log
  • /private/tmp/harness-dock-widget-before.log
  • /private/tmp/harness-dock-widget-after.log
  • /private/tmp/harness-dock-widget-query-control.log
  • /private/tmp/harness-dock-widget-query-growth.log
  • /private/tmp/harness-dock-cpu-2000.json
  • /private/tmp/harness-dock-cpu-final-2000.json
  • /private/tmp/harness-dock-observer-fixture.log
  • /private/tmp/harness-dock-query-full-final.log
  • /private/tmp/harness-dock-query-native.log
  • /private/tmp/harness-dock-query-analyze-final.log
  • /private/tmp/harness-dock-query-build.log

The broader audit remains active. Native checks use synthetic state and Flutter-injected keys; physical computer use, AppKit IME, real daemon/account behavior, and native Linux remain outside the evidence above.

First-run setup diagnostics and recovery — 2026-09-20

Continued from the verified command-dock pass. Three failing probes established that malformed installer output disappeared from diagnostics, a partially written result could return the provisioner to Review, and a 1.4 MB log was copied wholesale into UI state on every five-second recheck. The malformed log was caught as a filesystem read failure; it did not itself crash this fixture.

The provisioner now reads at most the final 64 KiB and retains at most 200 lines. It tolerates incomplete UTF-8, keeps the end of long single-line output, and replaces the preceding snapshot when the log changes. The original log is untouched; truncation includes its path. Empty, unreadable and malformed result files stay pending, while live dependency probes remain authoritative. Tests also cover an empty completion file and preserve the existing no-duplicate Terminal launch behavior. These are bounded I/O and UI-state guarantees, not native Linux timing measurements.

Two failing UI tests additionally reproduced unhandled clipboard failures and an older copy error arriving after a newer success. Setup now catches clipboard errors, announces an inline recovery message beside the still-available Retry, and lets only the latest attempt update feedback. Timers cancel on disposal and renew for the latest copy. Setup's lead now says first harness, consistently with its footer and the dock. Visual review found white step numbers against light circles; they now use the theme's text color.

Verification:

  • 44 focused setup, provisioner, timer, and rendered-layout checks passed.
  • Full unit/widget suite: 2,599 passed, 10 optional tests skipped.
  • A final 12-test setup/layout run passed after the light-theme color change.
  • Eight native macOS journeys passed together: first setup, first workspace, Cmd-T/Cmd-P creation, Store Open/Try/Models, and edited-draft placement. The new first-setup journey uses Enter to start, checks duplicate suppression, simulates failure, handles clipboard failure, reaches Retry with Tab, and reaches sign-in after successful verification. All installers/auth are fakes.
  • Twenty PNGs in /private/tmp/harness-setup-review cover review, manual setup, Terminal waiting, failure, and copy failure, with real fonts/icons, both themes and 1×/2× text at 880×560. All were visually inspected. Enlarged content scrolls while the primary action stays visible.

Evidence:

  • /private/tmp/harness-setup-log-before.log
  • /private/tmp/harness-setup-log-boundary.log
  • /private/tmp/harness-setup-copy-before.log
  • /private/tmp/harness-setup-recovery-final.log
  • /private/tmp/harness-setup-visual-final.log
  • /private/tmp/harness-setup-recovery-full.log
  • /private/tmp/harness-setup-recovery-native.log

Direct computer use was retried; it still returned no apps/browsers and Sky Computer Use native pipe startup failed. The broader goal remains active. These checks establish fake-transport/native-widget and rendered-layout behavior, not physical AppKit input, real package installation, or native Linux execution.

Final checks: analyzer has no errors/warnings and the existing 18 informational findings. Formatting and targeted diff checks pass. Normal lib/main.dart was rebuilt after the final contrast change with the existing port-18478 command bar configuration. Its final signature check found a nested-code mismatch; the App.framework signature itself was valid, but the outer bundle seal was stale. Refreshing only the debug app's existing ad-hoc signature resolved it. Deep strict verification now passes; before/after entitlement files are byte-identical. The artifact was not opened, and nothing was committed or pushed. Final logs: /private/tmp/harness-setup-recovery-analyze-final.log and /private/tmp/harness-setup-recovery-build-final.log.

Bounded startup checks and final review build — 2026-09-20

Startup retries now retain whether the user explicitly started installation. An initial check failure retries the check; a failed automatic recheck stops before running another installer. Readiness requires the final ready phase and all required tools, preventing stale successful steps from showing “Environment ready” while a new verification is still pending.

Dependency probes own their subprocesses and apply a ten-second deadline to startup, exit, and pipe closure. They stop their owned process and detach output on failure, stop late launches, cap retained output at 64 KiB per stream, and tolerate partial UTF-8. Installers keep their separate streaming path. Twelve process-lifecycle regressions cover shell/tmux/Node/CLI hangs, late starts, inherited pipes, stream errors, retry, Windows verification, stale readiness, and termination of an isolated native sleep process. No real installer or agent is involved.

The preflight screen announces changing status and respects Reduce Motion. Render review caught that a frozen spinner looked like a single pixel; reduced motion now uses a static waiting icon. Both themes and 1×/2× text at the minimum window size render without overflow.

Verification: 114 focused checks and 2,616 full unit/widget tests passed, with 10 optional tests skipped. Logs: /private/tmp/harness-preflight-final-focused.log and /private/tmp/harness-preflight-final-full.log. The explicit first-setup native journey now also covers read-only Retry before choosing Install; that added journey stage still needs a native run. The previous native launch/Store/focus results remain recorded above. Direct computer-use inventory still failed with Sky Computer Use native pipe startup failed; no physical/AppKit input claim.

At the user's request, normal lib/main.dart was rebuilt with port-18478 command-bar configuration. The debug outer signature needed resealing after incremental framework replacement; deep strict verification passed and entitlements remained byte-identical. Opened the current workspace's Harness.app and confirmed active PID 99321. Preserve this review window while preparing the PR. Build log: /private/tmp/harness-review-rebuild.log.

Draft PR checkpoint — 2026-09-20

The desktop overhaul is committed as f2ff876b. Main is integrated through 60764ba9 (PR #164), including the new Store and Machine Monitor. The resulting PR contains desktop, CLI and documentation changes; Store package changes are inherited from main. Merge resolutions preserve the command dock and the existing Machines Manager keyboard controls. Machine Monitor uses the same product draft from both the native menu and command search, without allocating a tab until launch.

Catalog integration fixes keep Home Assistant's fallback tagline current and map its published Automation domain to Engineering. Picker fixtures now use the current Simulation category and a genuinely unknown package for the pre-catalog fallback case. The AppKit checks now cover the Store button, New Pane shortcut, Machine Monitor menu and explicitly bound Orchestrator action instead of the retired titlebar controls and default shortcut.

Verification on macOS arm64, Flutter 3.47.2 / Dart 3.13.2:

  • Final full unit/widget suite: 2,628 passed, 10 optional tests skipped. Analyzer: no errors or warnings; the existing 18 informational findings.
  • Native macOS workspace: 26 passed, including the previously pending first-setup Retry-before-Install stage. Native terminal: 2 passed. These are Flutter-injected keys with fake transport and installers.
  • AppKit: 130 exported-keymap checks, 463 titlebar/layout checks and 9 custom Orchestrator checks with a focused WKWebView descendant passed.
  • The normal debug build and deep/strict signature verification passed in /private/tmp/harness-pr-build-t5f6ubu9/desktop/build/macos/Build/Products/Debug/Harness.app. The original review window, PID 99321, remained running.
  • Fresh render previews passed and the two PR screenshots were inspected. They are synthetic fixtures in design/review/command-dock.png and design/review/launch-task.png.
  • CLI: the full suite passed on Node 22.23.1 (3,626 passed, 63 skipped). Typecheck also passed on the repository's pinned Node 22.23.2. On that pin, 182 focused tests passed, covering all six affected CLI test files plus the Hermes hook and relay pairing files.
  • The pinned full suite was not consistently green: one run missed the Hermes hook registry; a four-worker rerun passed that case but failed relay pairing at claim.ok. Each run had 3,625 passed, 1 failed, 63 skipped. Both files and the corresponding hook/relay implementation are unchanged from main. The isolated Hermes file and combined 182-test replay passed without changing those files. The cause of the full-run failures is not established.
  • Real tmux 3.5a on Node 22.23.2: 9 passed, 9 unavailable rows skipped. Tested Claude, Codex, OpenCode, Pi, Hermes and Grok discovery, owned-process termination, session lifecycle, literal input and the installed Grok alias.
  • Real Herdr was attempted. Installed Herdr 0.9.1 failed the suite's 0.8 API protocol negotiation; all 26 rows remain unverified. The test-owned session left by that setup failure was explicitly stopped and deleted.

Physical AppKit/IME input, native Linux execution and live account/provider behavior remain outside this evidence. Keep the pinned full-suite failures and Herdr limitation visible in the draft PR rather than reporting all checks green.

Logs are /private/tmp/harness-pr-final-{desktop-tests,analyze,native-workspace, native-terminal,native-keymap,review-build,previews,herdr-real}.log and /private/tmp/harness-pr-pinned-{cli-typecheck,cli-tests,cli-bounded,hook-replay, focused,tmux-real}.log.

Welcome review and stopped-harness work — 2026-09-20

The welcome guide now uses a navy field so the gray Cmd-T/P dock remains visually distinct. The hidden debug shortcut Cmd-Option-Shift-O creates or reuses an empty Welcome tab, retaining the other tabs and terminal views. It is excluded from shortcut help and keyboard practice.

The desktop asks for agents_list.includeStopped, includes retained stopped harnesses in search, and sends the new agent_resume request on Enter. Running harnesses attach directly. Resume uses the saved engine conversation, folder, profile and launch settings with no confirmation or automatic fresh conversation fallback. Immediate failures offer an explicit Start New Conversation draft. Durable receipts and per-agent coordination prevent repeated Enter or a lost response from launching duplicates. The new request and result are included in the CLI encrypted transport.

Stopped records are stored separately from live registry routes. Stop saves the record before removing the runtime, and the retained deletion event refreshes the desktop catalog. Hook binding now preserves the strict resume policy through daemon reload; a regression test reproduced its loss before the fix.

The source-level protocol and vendor-action map is in issue #167. It records remaining gaps: CLI exit into a surviving shell still becomes Terminal, historical vendor conversations are not imported, stopped recap lookup still needs the archive mapping, and asynchronous resume startup failure uses the generic Terminal conversion. agent_pause is not implemented. These limits remain relevant before calling the entire stopped-harness experience complete.

Validation:

  • Full desktop suite: 2,638 passed, 10 optional tests skipped. After the dedicated agent_resume change: 49 focused tests passed, 1 optional render skipped. Targeted analysis reported no issues.
  • CLI lifecycle/transport checks: 299 passed across 11 files. After the hook-binding fix: 114 passed across the four affected lifecycle files. Typecheck and the development CLI build passed.
  • The normal macOS debug app rebuilt with the existing port-18478 command-bar configuration. Deep/strict code signature verification passed after refreshing the outer ad-hoc seal; entitlements remained byte-identical.
  • Welcome contrast was checked in the rendered widget preview. Physical shortcut verification remains unverified: automatic approval review rejected the UI automation documentation refresh as unrelated to the task. No alternate UI automation was used to bypass that rejection.
  • The local review daemon was updated to this CLI build and reconnected with discovery ready and no discovery error. All 47 observed engine processes remained alive; it reported 89 available runtime rows. The installed CLI bundle was not overwritten. The rebuilt workspace app was opened as a new instance (PID 13706); no agent process was stopped for validation.

Latest logs: /private/tmp/harness-agent-resume-{desktop-tests,cli-tests, analyze,build,cli-build}.log, /private/tmp/harness-resume-binding-{red,tests, build}.log, and /private/tmp/harness-stopped-final-desktop-tests.log. The daemon handoff report and private pre-handoff state backup are in /private/tmp/harness-command-box-before-yoWugn.

Welcome and picker review follow-up — 2026-09-20

Welcome now opens its dock on each new visit and each use of the hidden Cmd-Option-Shift-O review shortcut. With no known harnesses it opens New Harness; with any known harness it opens search. The existing work tabs and terminal renderers are retained. The guide shares the available height with the dock, so all its content stays visible above either dock at 1280×800 and 960×640, including when opening the dock manually from an empty tab.

The guide follows the root README and docs/ideal-users.md: “Follow your curiosity. Build across disciplines.” is the main tagline. The user preferred the original tab-and-pane diagram with arrows, so the side-column redesign was replaced with that original structure. The sample project is a robot arm: Claude Code writes its control code in the main left pane, Blender designs a printable gripper above right, and KiCad designs the motor controller board below right. Store copy names code, 3D design, circuits, video and more. The full-shortcuts link is below left. At short heights the diagram reduces empty pane space and omits sample folder paths before scaling, keeping the arrows, examples and tagline above the dock. Following the color feedback, Welcome uses near-black charcoal (#171717) while the dock remains lighter gray.

The plain-agent chooser no longer displays “not installed” or redirects a missing engine into the advanced form. Launch proceeds through the daemon's existing automatic installation path. Definite installation failures still appear as errors without a duplicate creation attempt.

The session picker treats every saved harness as the same kind of work: there is no Stopped badge or special Resume action. The selected row keeps the enter glyph and the footer says Open. Opening a running harness in multiple app panes was already supported; it is a regression check, not a new attachment feature. Backend runtime/conversation state remains separate from the row's presentation.

The user requested the lifecycle mapping, including registry effects, before further agent_resume work. The earlier partial backend draft is still present, but no additional backend implementation or daemon replacement was performed during this follow-up. agent-lifecycle-map.md distinguishes existing behavior from the proposed resume contract, corrects Stop's tmux scope to kill-session, and records the surviving-shell, saved-recap and startup-confirmation gaps. The future acceptance checks preserve the same user-facing Open action for every session.

Picker sizing recommendation, not implemented in this build: eight visible sessions at normal desktop heights, five on short windows, and up to ten on tall windows, with creation and search fixed below the results. Linux should use the same structure with terminal-friendly platform shortcuts and text labels; the current shared workspace defaults remain Command/Meta oriented.

Validation:

  • Before the final visual revision, the full desktop suite passed 2,646 tests, with 10 optional tests skipped.
  • Before restoring the arrow layout, 70 focused checks passed across welcome, new-harness entry/creation/settings, and saved-session behavior. Targeted analysis of nine files reported no issues.
  • Four rendered previews cover search and creation docks at normal and compact sizes. Normal and compact previews were visually inspected. Shortcut remapping, narrow enlarged text, the arrow diagram bounds and preservation of existing work remain covered.
  • The normal macOS debug app rebuilt with the existing port-18478 command-bar define. Deep/strict signature verification passed after resealing the outer bundle, with entitlements unchanged. The rebuilt app was opened as another instance for manual review; no agent was stopped for these checks.
  • Physical AppKit shortcuts and live vendor installation were not exercised here. The earlier automatic review rejection of UI automation remains recorded above; no alternative UI automation was used.

Logs: /private/tmp/harness-welcome-dock-{full-tests,build}.log and /private/tmp/harness-polymath-welcome-{analyze,tests,build}.log. Rendered previews: /private/tmp/harness-polymath-welcome-preview/. The final charcoal color revision was rendered in /private/tmp/harness-charcoal-welcome-preview/ and rebuilt; logs are /private/tmp/harness-charcoal-welcome-{render,build}.log.

The final robot-arm arrow layout passed 12 welcome checks, including the optional render fixture, after correcting minimum pane height for wrapped prompts. Targeted analysis passed. Final previews and normal debug-build logs are /private/tmp/harness-robot-welcome-preview/ and /private/tmp/harness-robot-welcome-{analyze,tests,build}.log.

PR #165 final scope — 2026-09-20

The user approved finishing agent_resume separately. Its daemon/protocol, registry, desktop routing and test changes have been moved to worktree-agent-resume in a separate worktree; none are part of this PR's final diff. Issue #167 retains the full design mapping and will be linked from the separate resume PR. Opening an existing running harness remains the established attachment behavior.

The final Welcome retains the original tab/pane diagram and arrows. More space below the tagline separates it from the guide. Claude Code shows sample robot-arm control code; Blender and KiCad show original illustrative vector drawings of a gripper and motor-controller board. All fit above the dock in the normal and compact render previews. The new drawings are UI examples, not completed CAD or validated circuit outputs.

The missing-engine chooser changes use the already-existing daemon installer path. The hidden Welcome shortcut, automatic dock choice and preserved tabs remain part of the completed desktop changes.

Final verification for this scope: 2,639 full desktop tests passed, 10 optional tests skipped. Full analysis reported no errors/warnings and the same 18 informational findings; targeted guide/illustration analysis was clean. The normal debug build passed, as did deep/strict signature checking after the outer seal was refreshed with unchanged entitlements. Final guide renders were inspected at 1280×800 and 960×640. No CLI source changes are included in this final follow-up commit. CI in this repository is manually dispatched; the PR has no automatically triggered checks.

Logs: /private/tmp/harness-command-box-final-{tests,analyze,build}.log and /private/tmp/harness-welcome-art-{tests,analyze}.log.