Skip to content

Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.

docs: describe features directly and gate design references (#468)

Ruben Lütke104dfbb4e50a
diff --git a/.githooks/pre-commit b/.githooks/pre-commit
index ef7be3179d..5dc1f015a4 100755
--- a/.githooks/pre-commit
+++ b/.githooks/pre-commit
@@ -91,6 +91,23 @@ if [ -f "$dist_gate" ]; then
fi
fi
+# --- DESIGN-REFERENCE gate (FAIL-CLOSED on a confirmed hit) ------------------
+# Public text describes our features directly and never names another product
+# as their template (AGENTS.md, "Public product descriptions"). Checks only the
+# STAGED additions; rc 1 = confirmed hit, any other rc fails OPEN (CI runs the
+# whole-tree scan as the backstop).
+designref="scripts/ci/check_design_references.py"
+if [ -f "$designref" ]; then
+ "$PY" "$designref" --staged </dev/null
+ if [ "$?" -eq 1 ]; then
+ echo ""
+ echo "pre-commit BLOCKED: DESIGN-REFERENCE gate - staged text names another"
+ echo " product as our design template. Describe the feature directly, or mark"
+ echo " a license header / factual integration line with 'design-ref-allow'."
+ exit 1
+ fi
+fi
+
# --- PRIVATE-KEY gate (FAIL-CLOSED) -------------------------------------------
# Signing private keys and the retired signing passphrase must NEVER be
# committed — they live only in GitHub Actions secrets. Catches a full PEM
diff --git a/.github/workflows/release-gate.yml b/.github/workflows/release-gate.yml
index 7263a183a0..9883366d56 100644
--- a/.github/workflows/release-gate.yml
+++ b/.github/workflows/release-gate.yml
@@ -4,7 +4,7 @@
# A tag publishes only when scripts/ci/release_admit.py confirms: tag ==
# pyproject == jarvis.__version__, a CHANGELOG section exists, the commit is on
# main, and `CI gate` passed on that exact commit (it waits while CI runs).
-# Adapted from Hermes Agent's "admit the claim" stage: a release never builds
+# The "admit the claim" stage: a release never builds
# anything main has not already proven.
name: Release gate
diff --git a/AGENTS.md b/AGENTS.md
index f663decb46..0050c2b855 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -104,7 +104,9 @@ Describe our features and interactions directly. Do not present another product
as a design reference or say our UI or behavior is inspired by, modeled on,
copied from, or made to match it. Preserve required copyright, license and
attribution notices and factual dependency, integration and compatibility
-references; this rule never authorizes concealing code provenance.
+references; this rule never authorizes concealing code provenance. Competitor
+research stays out of the repo. `scripts/ci/check_design_references.py` checks
+staged lines at commit time and the whole tree in CI.
**Proportionality.** You own the validation plan: the smallest set of checks
that can detect a plausible regression from the diff. State the scope, what
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 319384e031..ad109be747 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -44,7 +44,7 @@ versioning per [SemVer](https://semver.org/).
### Added
- An editor for appshots, opened from the corner card
-- Draw traces the way the Codex app does
+- Draw traces as quiet prose lines
- Change floors with the elevator doors and a real button panel
- A clickable, draggable appshot card in the screen corner
- Make the area picker's magnifier a round lens
@@ -57,13 +57,13 @@ versioning per [SemVer](https://semver.org/).
- Bring back the quiet session counts in the project tree
- Drop the workspace pane limit and the 4x4 grid bound
- Drop the running marks from the project tree
-- Restyle the project tree as a quiet Codex-style folder list
+- Restyle the project tree as a quiet folder list
- Arcade floor with ten playable retro cabinets
- The SOUL.md section shows the assistant's whole profile
- Drag a stack of side-by-side seams as one line from a seam's ends
- Record the appshot shortcuts and default the area to both Shift keys
-- Give the area picker ShareX's region-capture look
-- Draw turns as a Claude/Codex-style timeline
+- Give the area picker a cleaner region-capture look
+- Draw turns as a timeline
- Select an area of the screen on its own shortcut
- Tell a finished turn's work as a written report
- Jarvis follows the person as their chosen pet, in 3D
@@ -176,7 +176,7 @@ versioning per [SemVer](https://semver.org/).
### Added
-- A quiet, Claude-style front page: one greeting, one composer, and a voice mode that lives inside the chat with a flowing aurora glow.
+- A quiet front page: one greeting, one composer, and a voice mode that lives inside the chat with a flowing aurora glow.
- Open any agent's chat on the front page from the sidebar; `@` in the Jarvis chat names your agents and connected plugins.
- Reasoning traces drawn as a threaded rail, with a wider, calmer column for agent chats.
- A restyled sidebar with search, your most-used agents, every recent chat, and the update button in its footer.
diff --git a/docs/BUGS.md b/docs/BUGS.md
index 7bde7d3491..ce9e681f23 100644
--- a/docs/BUGS.md
+++ b/docs/BUGS.md
@@ -16073,7 +16073,7 @@ child window, and the real outer window kept affinity 0 (measured on the live
child and additionally re-applies `WDA_EXCLUDEFROMCAPTURE` on every `<Map>` of
the toplevel (bound with `add="+"`, child-widget maps ignored). Checked with a
real Tk root and Toplevel: 0x11 after the first map and after withdraw/show.
-Side effect: the overlays also vanish from other capture tools (ShareX, OBS),
+Side effect: the overlays also vanish from other capture and streaming tools,
which is the documented intent of the module. Guard:
`tests/unit/platform/test_capture_exclusion.py`.
diff --git a/docs/agent-society/MASTERPLAN.md b/docs/agent-society/MASTERPLAN.md
index 479e3371c0..a9bc61f5c6 100644
--- a/docs/agent-society/MASTERPLAN.md
+++ b/docs/agent-society/MASTERPLAN.md
@@ -311,7 +311,7 @@ add the four controls it lacked: authenticated writes (chokepoint), bounded non-
3. **Anti-harm blacklist class** seeded in the risk tiers: mass outbound messaging, credential
probing, repeated actions against non-consenting external endpoints — blocked by pattern, not
by prompt. Voice/chat never accept secrets (AP-2); secrets only via `get_secret`; per-agent
- credentials are keyring-scoped (NEVER Hermes-style per-profile `.env` files — AP-12).
+ credentials are keyring-scoped (NEVER per-profile `.env` files — AP-12).
4. **No spawn tools in any agent tool set**; dispatch is the scheduler's privilege (§2.5).
5. Loop caps, iteration counters, spend meters, stall watchdogs (AP-19), `WorkerKilled`, and the
society master kill switch.
@@ -425,7 +425,7 @@ Decided 2026-09-01:
5. **Section layout:** sidebar | world stage | fixed agents rail; thin top strip (content open)
(§4.1).
6. **Model card:** near-full-screen overlay above the world, three columns Specs | 3D | Chat;
- the chat is the ordinary agent chat (Claude Code / Codex-style timeline); one open chat at a
+ the chat is the ordinary agent chat (the coding-agent timeline); one open chat at a
time (§4.2).
7. ~~No Hermes code or UI is copied.~~ Superseded 2026-10-05: well-separated MIT code may be
adapted with attribution (§7).
diff --git a/docs/agent-society/agent-definition.md b/docs/agent-society/agent-definition.md
index 1aed9c9cd1..993735202d 100644
--- a/docs/agent-society/agent-definition.md
+++ b/docs/agent-society/agent-definition.md
@@ -5,7 +5,7 @@ refines §3.1 (roster), §3.2 (runners), §4.2 (model card), §6 (safety) and an
the build plan left open: how an agent gets its hands (plugins, CLIs, MCPs, skills), how it knows
the Jarvis ecosystem, and how the Obsidian wiki becomes the society's shared memory.
-The reference product is Grok Bot (§1). The whole thing is built in-house (MASTERPLAN §10.7).
+The whole thing is built in-house (MASTERPLAN §10.7).
---
@@ -104,7 +104,7 @@ card; nothing hidden drives behavior.
**Safety & economy** (§3.4)
- `permission_ceiling` (`safe | monitor | ask`; block is block, never a bypass).
-- `approval_rules` JSON: Grok-style rules — `require_approval: [patterns]`,
+- `approval_rules` JSON: plain-language rules — `require_approval: [patterns]`,
`always_allow: [patterns]`; require wins; patterns are capability ids or `capability:verb`
(`plugin:gmail:send`). Global blacklist and `always_confirm_tiers` sit above all of it.
- `daily_budget_usd`, `max_concurrent_runs` (default 1 for specialists, 3 for orchestrators).
diff --git a/docs/agent-society/build-plan-m1-m2.md b/docs/agent-society/build-plan-m1-m2.md
index aa71aebfd8..9cf315586a 100644
--- a/docs/agent-society/build-plan-m1-m2.md
+++ b/docs/agent-society/build-plan-m1-m2.md
@@ -166,7 +166,7 @@ Create/modify:
- `src/components/layout/MainView.tsx` — lazy loader (`:87`) + switch case (`:548`) for `"agents"` now import SocietyView; `FULL_BLEED_SECTIONS` (`:397-401`) += `"agents"`. Sticky branch (`:480-524`) deliberately untouched (M3 concern). Section registry untouched — id, labels, nav keys all stay.
**Guards:** `tests/unit/ui/web/test_section_id_parity.py` + `tests/unit/plugins/tool/test_navigate.py` (green by construction — run anyway), vitest for the view, `npm run build` + bundle-consistency + bundle budget (SocietyView is a lazy chunk), light+dark, `check_boot_budget.py` manual run.
-**Exit:** open Agents: board center, Hermes-style roster right, "+" creates an agent, row click expands exactly one chat over the stage with the reasoning trail, second row switches, collapse returns the stage. Old view unmounted but still compiling.
+**Exit:** open Agents: board center, roster right, "+" creates an agent, row click expands exactly one chat over the stage with the reasoning trail, second row switches, collapse returns the stage. Old view unmounted but still compiling.
### Wave 9 — Model card, approvals queue + cards, [agent:*] routines (M/L) → **M2 exit**
diff --git a/docs/agent-society/research/gap-check.md b/docs/agent-society/research/gap-check.md
index 59504e1dee..6f08e7e94d 100644
--- a/docs/agent-society/research/gap-check.md
+++ b/docs/agent-society/research/gap-check.md
@@ -11,7 +11,7 @@ NOT READY FOR PLANNING — one reconciliation pass required first. The reports a
1. Inter-agent protocol: Report A adopts Hermes' single prose send path (message_agent-equivalent, natural-language messages with sender prefix, rendered as chat threads); Report B calls the opposite 'the single most important design decision' — typed enum events (ASSIGN/CLAIM/RESULT/QUERY/ANSWER/HOLD/VETO/DIGEST) on a blackboard, explicitly 'not a chat'. The plan must pick one, or explicitly define typed-envelope-with-prose-payload and say which surface renders what.
2. Group discussions: A adopts hard-bounded natural-language group rooms (2–6 members, <=3 rounds, <=10 messages) as THE consumer-token-budget mechanism and builds the world's meeting-table moment on them; B rejects conversation-centric coordination as the expensive anti-pattern ('not a free-form debate club') and ships no group mechanism at all.
3. What an agent IS at runtime: A = a persistent agent_chat session per agent (a long-lived conversation engine reusing service.py/runners); B = a durable roster row whose work is done by ephemeral SubagentFanoutRunner workers that are TTL-reaped after 60 s. These imply different RAM footprints, token profiles, and persistence stories — both cannot be the primary definition.
-4. Where agent identity/config persists: A extends jarvis/core/config.py with an [agents.*] schema plus Hermes-style agent-account directories (jarvis/agent_accounts.py); B creates a society_agents table in a new data/society.db. Two sources of truth for the same model-card data (model/provider pin, tool grants, permission ceiling, avatar).
+4. Where agent identity/config persists: A extends jarvis/core/config.py with an [agents.*] schema plus per-agent account directories (jarvis/agent_accounts.py); B creates a society_agents table in a new data/society.db. Two sources of truth for the same model-card data (model/provider pin, tool grants, permission ceiling, avatar).
5. Hierarchy and the AP-5/AP-14 no-spawn wall: A says the society is flat specialists under Jarvis and 'the society cannot chain-react' because no society agent can spawn; B's middle tier (orchestrator/kontrollierer agents) explicitly delegates to sub-agents, i.e. some society members DO dispatch workers. The plan must state which tiers may dispatch and where the structural wall sits, because the safety claim differs between the reports.
6. World data feed: A says the world's data spine is the existing three-source merge (/api/sub-agents/tree + /api/missions + /api/outputs); B says the world is a pure projection of the new society_events log. Two competing single-sources-of-truth for the same view.
7. Spatial truth: B persists world_x/world_y in the backend roster and claims the world 'can never desync' because it is a projection; A runs idle wander purely client-side for zero cost. Unresolved: who owns position, whether it survives restart, and whether two open app windows (live + dev instance exist today) show the same world.
diff --git a/docs/ci-pipeline.md b/docs/ci-pipeline.md
index 5c9589c353..30d01471bc 100644
--- a/docs/ci-pipeline.md
+++ b/docs/ci-pipeline.md
@@ -1,11 +1,9 @@
# CI/CD pipeline
How a change travels from a coding agent's branch to a user's machine. The
-design borrows the strongest ideas of the
-[Hermes Agent pipeline](https://github.com/NousResearch/hermes-agent/tree/main/.github/workflows)
-and adapts them to this repository: several coding agents working in
-parallel, a Python + TypeScript desktop app, three operating systems, and
-signed installers.
+design fits this repository: several coding agents working in parallel, a
+Python + TypeScript desktop app, three operating systems, and signed
+installers.
```
agent branch ──► pull request ──► CI (lanes) ──► CI gate ──► merge train ──► main
@@ -133,7 +131,7 @@ minutes:
comment listing the files. `git rerere` replays recorded resolutions.
3. The first green, conflict-free pull request is squash-merged, one per
tick. Main's full post-merge test run is the backstop for changes that
- pass alone and break together — GitHub's non-strict model, as in Hermes.
+ pass alone and break together — GitHub's non-strict model.
Opt out with `no-auto-merge`, `do-not-merge`, `wip` or `needs-human`.
`priority` moves a pull request to the front. Forks and Dependabot never ride
@@ -307,9 +305,9 @@ Native macOS signing imports the publisher certificate into a temporary
keychain and cleans it up even when the build fails. Native OS build and
signing proof still requires hosted runners; local syntax checks cannot prove it.
-## 4. Adapted from Hermes, and what was left out
+## 4. Design choices, and what was left out
-| Hermes idea | Here |
+| Idea | Here |
| --- | --- |
| Orchestrator + change classifier, fail-open lanes | `detect` + `classify_changes.py` |
| One aggregate required check (`all-checks-pass`) | `CI gate` |
diff --git a/docs/handoffs/report-tutorial-review-answers.md b/docs/handoffs/report-tutorial-review-answers.md
index cea9d4ceee..acdb905c82 100644
--- a/docs/handoffs/report-tutorial-review-answers.md
+++ b/docs/handoffs/report-tutorial-review-answers.md
@@ -25,10 +25,10 @@
contributor list — readable at 4K / on pause. It is also **stale** (the public
history has since been anonymized, so the video would show data that no longer
exists in the live repo).
-- **One secondary, non-video finding:** a stray ShareX capture
+- **One secondary, non-video finding:** a stray screenshot-tool capture
`video/pythonw_xpdq2qZtJg.png` shows the **private email
`[email protected]`** twice. It is `.gitignore`d and is **NOT** used in
- the film, but it is sitting in `video/` and should be removed / ShareX repointed.
+ the film, but it is sitting in `video/` and should be removed / the screenshot tool repointed.
---
@@ -71,7 +71,7 @@ That tutorial-that-demos-a-report is the "Report" + "Tutorial" combination.
`scripts/ci/privacy_gate/references/distribution-denylist.txt:134`
(rationale lines 126–133).
- Additionally `.gitignore`d (never tracked at all): the rendered films
- (`git check-ignore video/out/onboarding.mp4` → IGNORED) and the stray ShareX
+ (`git check-ignore video/out/onboarding.mp4` → IGNORED) and the stray screenshot-tool
PNGs in the `video/` root (`video/pythonw_xpdq2qZtJg.png` → IGNORED).
### 4. Format/type.
@@ -171,7 +171,7 @@ and that screenshot is not used in the film.)
`video/src/intro/onboarding/scenes/CreatorIntro.tsx:29`, displayed near native
width → readable at 4K). **Primary finding.**
- `video/pythonw_xpdq2qZtJg.png` — shows the **private email** twice. **Not in the
- film**; stray ShareX capture, `.gitignore`d.
+ film**; stray screenshot-tool capture, `.gitignore`d.
- `video/public/shot-outputs.png` — shows an internal **German mission prompt**
("…That was just rigid. Could you do a deep dive?…") + first name.
**Not referenced by any scene → not in the film.**
@@ -267,9 +267,9 @@ you…"). The concern is therefore not about the git repo (protected) but about
it shows data that no longer exists live). Either recapture it from the
anonymized repo, blur the commit-author line + contributor sidebar, or
consciously accept the surname being public.
-2. **Remove the stray ShareX capture.** `video/pythonw_xpdq2qZtJg.png` embeds the
- private email; delete it and repoint ShareX out of the repo (a known issue —
- stray `pythonw_*.png` are ShareX captures, not a Jarvis artifact).
+2. **Remove the stray screenshot-tool capture.** `video/pythonw_xpdq2qZtJg.png` embeds the
+ private email; delete it and repoint the screenshot tool out of the repo (a known issue —
+ stray `pythonw_*.png` are screenshot-tool captures, not a Jarvis artifact).
3. **Everything else is fine to ship as-is.** First name "Alex" / "Hey Alex",
the masked-key API screenshot, the clean `app-home.png`, and the thumbnail
carry no secrets and reflect the maintainer's intended public identity.
diff --git a/docs/jarvis-learning.md b/docs/jarvis-learning.md
index bec5c44c38..5595493751 100644
--- a/docs/jarvis-learning.md
+++ b/docs/jarvis-learning.md
@@ -185,7 +185,7 @@ the loop adds no new destination. `enabled = false` switches it off.
## Design reference
-The loop follows the design of Nous Research's Hermes Agent: a bounded
+The loop is a bounded
USER/MEMORY pair, a review that runs after the reply instead of during it,
declarative entries instead of imperatives, and an explicit list of what not
to keep. Jarvis adds the evidence rule of its Society agents (every change
diff --git a/docs/macos-permissions.md b/docs/macos-permissions.md
index 442ee82514..b1cf70cdce 100644
--- a/docs/macos-permissions.md
+++ b/docs/macos-permissions.md
@@ -160,65 +160,9 @@ need to detect silent failures) that was never weighed, and a small set of real
got conflated with it. None of the real constraints requires a banner, a wizard, a readiness gate
or refusing to let the OS show its own dialog.
-## 3. How other apps do it, and what Apple says
-
-### 3.1 Evidence quality (read first)
-
-- Vendor domains (OpenAI help centre and developer pages, Raycast manual, Zoom, Slack,
- 1Password, Loom, CleanShot, Alfred, a dictation app, a dictation app) were **not reachable** from the
- research environment. Most vendor statements below come from **search-result summaries of the
- official page**, university or IT how-to pages that mirror the vendor knowledge base, and
- third-party reviews. They are weaker than a direct read.
-- Pages read directly: the Claude desktop and computer-use documentation, one Claude support
- article, and GitHub issues of public repositories (these quote real dialog text and bundle ids
- but are user reports, not vendor statements).
-- Nothing was observed on a Mac. A "no wall found" statement means "none documented", never
- "none exists".
-- Source types: **direct** (page read), **summary** (search summary of the official page),
- **how-to** (third-party or institutional write-up), **issue** (GitHub issue).
-
-### 3.2 Comparison table
-
-| App | When it asks | Explanation before / after | On denial | Route to System Settings | Confidence, source type |
-|---|---|---|---|---|---|
-| **ChatGPT desktop** | At first use of a feature: Work with Apps (Accessibility), Record and voice (Microphone, Screen & System Audio Recording), Computer Use after an explicit opt-in. No launch wall found. | Help-centre prose plus the OS dialog; an own pre-prompt card was not verified. Computer Use adds app-level approvals independent of TCC ("allow once / Always allow"). | Feature unavailable; Computer Use reports "permissions are still pending" when the grant sits on the wrong client (a helper is a separate TCC client, openai/codex#46776). | Text path to the Privacy & Security pane; a deep link is unverified. | Partly documented. Summary of vendor help pages plus issues (read). |
-| **Claude desktop** | Computer use is off by default; the user flips a Settings toggle and must grant two permissions "before the toggle takes effect". Quick Entry shows an in-app opt-in card at first launch of the updated app (an app-level consent card, not a TCC prompt). | Docs table with one line per permission and a status badge per permission in Settings. | Toggle stays inert until both are granted; if off, Claude says in chat it could do the task if enabled. | "Click the badge to open the relevant System Settings pane" (documented). Restart after a grant: not stated for Desktop. | Documented for computer use (direct). Partly documented for the timing of the OS dialogs and for Quick Entry denial. |
-| **Claude Code** (terminal) | Nothing up front. Computer use is a built-in server, off by default; "the first time Claude tries to use your computer" a prompt appears. Protected folders prompt on first touch (plain OS template). | Inline card in the session with links to the panes and a **Try again** button. | Try again, no nagging. Folder denial surfaces as an unexplained error (issues). | Links in the prompt. "macOS may require you to restart Claude Code after granting Screen Recording." | Documented for the flow (direct). Attribution is mixed: the docs name the terminal app; issues show the Claude binary as its own client (a version path in Automation dialogs, per-update re-prompts). |
-| **Raycast** | Per feature at first use (summary of the manual); feature setup views for Screen Awareness and Dictation. Whether Accessibility is also asked in first-run onboarding is **conflicting** across third-party sources. | Setup view per feature and an inline "Grant Permission" card. | Feature degrades; Screen Recording is optional for Screen Awareness. | Button opens the pane; a changelog entry fixes "would not open System Settings when requesting access for a denied permission" (summary). Auto-restart and resume is asserted by one summary only. | Partly documented. Summary of the manual and changelog; third-party how-tos. |
-| **Zoom** | Camera and microphone at first use; Screen Recording at the first share; Accessibility only when someone requests remote control. No permission onboarding documented. | OS dialog with the usage string; for Accessibility "Open System Preferences" on the dialog, which as far as can be told is the standard macOS dialog, not an own one. | Feature dead; all how-tos tell the user to fix it manually in Privacy & Security. | Button for Accessibility; OS asks to restart Zoom after Screen Recording. | Partly documented. How-tos mirroring the vendor KB. |
-| **Slack** | First huddle (microphone), first video or share (camera, Screen Recording). | OS dialog (inferred). | **Silent degrade**: sharing shows the wallpaper and menu bar. A cautionary example. | Help text, quit and reopen; `tccutil reset ScreenCapture <bundle id>` as a last resort (third party). | Partly documented / inferred. Third-party sources. |
-| **1Password** | Screen Recording at the first QR or setup-code scan; Accessibility only for the opt-in Universal Autofill and shortcuts. The vault works with no permission at all. | OS dialog; the help page is named after the dialog text ("1Password would like to record this screen"). | QR scan unavailable until granted and restarted; core product unaffected. | "Open System Settings" on the dialog, then "Quit & Reopen". | Partly documented. Summary of vendor pages. |
-| **Loom** | The install article says users "will be prompted" when installing the desktop app; microphone and camera when selected (inferred). Exact order not documented. | Unknown. | Manual toggles; `tccutil reset` commands documented as a fix for stuck state (a known pain point). | Help-article steps; no in-app deep link confirmed. | Partly documented. Summary of help articles. |
-| **CleanShot X** | Screen Recording and Accessibility during setup (a screenshot tool whose whole purpose needs Screen Recording, so up front equals first use); microphone when a recording uses it. | Short onboarding; wording unknown. | Capture does not work; the app appears in the Screen Recording list only after a capture attempt. | Quit and reopen needed (third party). | Inferred / partly documented. Third-party reviews. |
-
-### 3.3 Closest comparators (voice, hotkey, insert text)
-
-| App | Finding | Confidence |
-|---|---|---|
-| **a dictation app** | First-launch setup with two cards (Accessibility, Microphone), each with an own priming sentence and an Allow button, then Continue and a microphone test; resumable. Input Monitoring only if the user binds Caps Lock alone; screen capture not needed for dictation; re-checks permissions when brought to the foreground (community guide); warns that a missing Accessibility grant is a silent failure. | Documented (summary of vendor docs); runtime recovery behaviour from a community guide. |
-| **a dictation app** | First-launch wizard for Microphone and Accessibility; a changelog note says a re-prompt after completed onboarding was fixed. | Partly documented. |
-| **Alfred** | Hybrid: a first-launch checklist plus a permanent "Request Permissions..." button in Preferences; Contacts and Automation are asked by macOS only when needed. | Partly documented. |
-| **Codex desktop** | Opt-in toggles (Computer Use, Chronicle) then the OS prompts, then per-app approvals; helper identities are separate TCC clients (issues #46776, #18507); a prompt for a feature the user never enabled is perceived as a bug (#37378, "access data from other apps" about daily with Computer Use off). | Partly documented; issues read. |
-
-### 3.4 What the comparison supports, and what it does not
-
-- Documented only for the Claude products (direct read): the OS dialog hangs off a feature
- toggle or first use, with a feature-local explanation and a link to the right pane. For the
- other apps the evidence is summary, how-to or inference. a dictation app, a dictation app, Alfred,
- CleanShot X and (inferred) Loom run short first-launch permission flows, which is a setup
- wizard of the kind Jarvis removes. No source found documents a persistent app-wide banner or a
- "set up everything" wizard, but absence of documentation is not absence of the feature.
- This comparison is therefore not evidence for the "no wizard" decision; that decision rests on
- the Apple HIG quotes (3.5).
-- Not supported: "no app asks at launch". Claude desktop shows a first-launch opt-in card;
- a dictation app, a dictation app, CleanShot X and Alfred have short first-launch flows scoped to what
- the product cannot work without; Raycast onboarding is uncertain. State it as "none documented
- as a persistent wall".
-- Two-layer consent for agent features (an app-level toggle, then per-app approvals, then the
- OS dialog) appears in the Claude and Codex products. Jarvis already has its own risk tiers
- for tool use; this change adds no new app-level approval layer.
-- Helper-identity pitfall (issues only): when a helper does the capture, the grant must be on the
- helper. Ask from, and check, the process that will call the API (4.7, attribution).
+## 3. What Apple says
+
+The decision in section 4 rests on Apple's own guidance and APIs (3.5 to 3.8). Sections 3.1 to 3.4 held a survey of other products and were removed; the numbering of the remaining sections is kept so existing references stay valid.
### 3.5 Apple reference points
diff --git a/docs/superpowers/specs/2026-05-30-chats-conversation-manager-design.md b/docs/superpowers/specs/2026-05-30-chats-conversation-manager-design.md
index 918d0f7845..3eeabb09cb 100644
--- a/docs/superpowers/specs/2026-05-30-chats-conversation-manager-design.md
+++ b/docs/superpowers/specs/2026-05-30-chats-conversation-manager-design.md
@@ -1,7 +1,7 @@
# Chats Conversation Manager — Design
**Date:** 2026-05-30
-**Status:** Approved (scope confirmed by maintainer: unified text+voice history, two-pane ChatGPT-style layout)
+**Status:** Approved (scope confirmed by maintainer: unified text+voice history, two-pane layout)
**Author:** Claude Code session
## Goal
diff --git a/docs/superpowers/specs/2026-05-31-chat-mic-dictation-design.md b/docs/superpowers/specs/2026-05-31-chat-mic-dictation-design.md
index bf3caec132..90904b4ed9 100644
--- a/docs/superpowers/specs/2026-05-31-chat-mic-dictation-design.md
+++ b/docs/superpowers/specs/2026-05-31-chat-mic-dictation-design.md
@@ -8,7 +8,7 @@
A microphone button inside the chat composer that runs the project's **own STT
model** and streams the transcript **live into the chat text input** (text
-appears progressively as the user speaks, Claude-style). The transcript is
+appears progressively as the user speaks). The transcript is
**not** auto-dispatched to the brain — it lands in the input box, where the user
edits and sends it manually (Enter / send button).
@@ -90,7 +90,7 @@ Server handler: new `elif cmd.action == "stt_dictate":` branch in
### Empty state (already done in main, mirror here)
-The empty chat is now Claude-style: centered mascot + greeting, **no** prompt
+The empty chat is now calm: centered mascot + greeting, **no** prompt
cards. This worktree's baseline already has that. No further change unless the
user wants the composer centered in the empty state (separate follow-up).
diff --git a/docs/superpowers/specs/2026-06-16-cu-modal-awareness-design.md b/docs/superpowers/specs/2026-06-16-cu-modal-awareness-design.md
index 5bc00027a9..90a7fd6768 100644
--- a/docs/superpowers/specs/2026-06-16-cu-modal-awareness-design.md
+++ b/docs/superpowers/specs/2026-06-16-cu-modal-awareness-design.md
@@ -16,7 +16,7 @@ mission dies in the circling-guard cap.
User voice command (German, confidence 0.965):
-> "Ich bin verwirrt in ShareX … navigiere durch die Einstellungen, damit die
+> "Ich bin verwirrt in meinem Screenshot-Tool … navigiere durch die Einstellungen, damit die
> Screenshots direkt angezeigt werden."
The CU loop ran:
@@ -175,7 +175,7 @@ cap, and it carries no per-unit-counter reset hazard (BUG-032 class avoided).
In-window modals with no separate top-level HWND (Electron / web overlays,
custom-painted in-process dialogs) are **not** detected by HWND-level probing.
-The ShareX case is a native Win32 dialog and is covered. For web/Electron
+The screenshot-tool case is a native Win32 dialog and is covered. For web/Electron
overlays the behavior is unchanged from today (no regression) — the model still
sees them in the screenshot and can interact with them. Documented, not fixed.
diff --git a/jarvis/codex_app_server.py b/jarvis/codex_app_server.py
index 244f4ba769..51d95b4aae 100644
--- a/jarvis/codex_app_server.py
+++ b/jarvis/codex_app_server.py
@@ -3,7 +3,7 @@
This module owns the local JSONL protocol, process lifecycle, and the dedicated
Codex identity used by subscription voice. The user signs in directly inside a
Jarvis-owned ``CODEX_HOME``; OAuth tokens are never copied, hard-linked, or
-borrowed from an ordinary Codex/IDE profile. Credentials are forced into that
+taken over from an ordinary Codex/IDE profile. Credentials are forced into that
profile's file store, and live ``account/read`` is the authoritative account and
plan check.
API-key environment variables are removed so a subscription request cannot
diff --git a/jarvis/society/focus.py b/jarvis/society/focus.py
index 599e87f879..2011dcb19a 100644
--- a/jarvis/society/focus.py
+++ b/jarvis/society/focus.py
@@ -7,7 +7,7 @@ labels, one-liners and MCP server names are matched too, so a freshly
connected MCP server is reachable by its own name without a table edit.
``derive_approval_rules`` reads the same text for an approval boundary
-("only after approval", "nur nach Freigabe") and turns it into Grok-style
+("only after approval", "nur nach Freigabe") and turns it into plain-language
require-approval patterns on the focus tools' sending verbs.
"""
diff --git a/jarvis/tasks/schema.py b/jarvis/tasks/schema.py
index d060f2b7fb..f1f621c589 100644
--- a/jarvis/tasks/schema.py
+++ b/jarvis/tasks/schema.py
@@ -280,7 +280,7 @@ class PluginGrant(BaseModel):
class AgentAction(BaseModel):
"""An agentic brain turn — the task runs ``prompt`` and the brain decides
- how to combine the enabled plugins to reach the goal (Claude-style
+ how to combine the enabled plugins to reach the goal (
scheduled task). The toggled plugins become the turn's tool allowlist;
each grant's ``scope`` gates what the unattended run may do.
"""
diff --git a/jarvis/ui/web/frontend/src/App.tsx b/jarvis/ui/web/frontend/src/App.tsx
index a708a2f9e7..c37b1f7b0a 100644
--- a/jarvis/ui/web/frontend/src/App.tsx
+++ b/jarvis/ui/web/frontend/src/App.tsx
@@ -127,7 +127,7 @@ export default function App() {
/*
The quick switcher's chord (Ctrl+Space; ⌥+Space on a Mac — both can be
changed or switched off under Settings → Keyboard shortcuts) opens the
- Spotlight-style "type a section, press Enter" launcher. Listened for in the
+ "type a section, press Enter" launcher. Listened for in the
CAPTURE phase so it also works while a terminal pane has focus: xterm
handles keys on its own textarea and would otherwise send the chord to the
agent. Pressing it again closes the switcher, the way Spotlight's own
diff --git a/jarvis/ui/web/frontend/src/components/QuickSwitcher.tsx b/jarvis/ui/web/frontend/src/components/QuickSwitcher.tsx
index c0fd9e9eba..4433e48299 100644
--- a/jarvis/ui/web/frontend/src/components/QuickSwitcher.tsx
+++ b/jarvis/ui/web/frontend/src/components/QuickSwitcher.tsx
@@ -2,7 +2,7 @@
* The quick switcher's Spotlight window — Ctrl+Space anywhere, type where you
* want to go, Enter.
*
- * Modelled on the Mac's Spotlight bar: one large field floating high on the
+ * One large field floating high on the
* screen over a frosted panel, results directly underneath with the first one
* already selected, the app still visible behind it (no dimming scrim). The
* results themselves are `QuickSwitchList`, the same list the sidebar search
diff --git a/jarvis/ui/web/frontend/src/components/agentchat/AgentTimeline.tsx b/jarvis/ui/web/frontend/src/components/agentchat/AgentTimeline.tsx
index b9568f507d..1052c67f43 100644
--- a/jarvis/ui/web/frontend/src/components/agentchat/AgentTimeline.tsx
+++ b/jarvis/ui/web/frontend/src/components/agentchat/AgentTimeline.tsx
@@ -35,7 +35,7 @@ export function AgentTimeline({
/** Sender faces by agent id, where the caller has a roster. */
agentsById?: Record<string, InternalParticipant>;
/**
- * The front page's conversation look (2026-10-01, after the Claude app):
+ * The front page's conversation look (2026-10-01):
* your words in a soft rounded bubble on the right, the assistant's answer
* as plain text with no name-and-model header above it, and a quiet
* centred time stamp wherever the conversation paused. The Agentic IDE
@@ -292,7 +292,7 @@ function Prose({ block }: { block: TextBlock }) {
<div
data-testid="agent-text"
className={cn(
- // Compact reading size, like the Claude app: headings stay close to body
+ // Compact reading size: headings stay close to body
// size and set apart by weight and spacing, not by scale.
"prose prose-neutral max-w-none text-[15px] leading-[25px] text-foreground dark:prose-invert dark:text-foreground [overflow-wrap:anywhere]",
"[&>div>:first-child]:mt-0 [&>div>:last-child]:mb-0",
diff --git a/jarvis/ui/web/frontend/src/components/agentchat/TraceTimeline.test.tsx b/jarvis/ui/web/frontend/src/components/agentchat/TraceTimeline.test.tsx
index 30f46fcf4f..51a5c0a5c9 100644
--- a/jarvis/ui/web/frontend/src/components/agentchat/TraceTimeline.test.tsx
+++ b/jarvis/ui/web/frontend/src/components/agentchat/TraceTimeline.test.tsx
@@ -9,7 +9,7 @@ const step = (over: Partial<ThinkingStep>): ThinkingStep => ({
id: "s", kind: "tool", labelKey: "thinking.step_tool", status: "done", startedTs: 1000, ...over,
});
-it("reopens a stored voice turn as a Claude/Codex-style timeline", () => {
+it("reopens a stored voice turn as a timeline", () => {
const steps: ThinkingStep[] = [
step({ id: "b", kind: "brain", labelKey: "thinking.step_brain", detail: "grok · grok-4.3", durationMs: 900 }),
step({ id: "p", kind: "thought", labelKey: "thinking.step_thought", detail: "I'll check the **wiki** first." }),
diff --git a/jarvis/ui/web/frontend/src/components/agentchat/TraceTimeline.tsx b/jarvis/ui/web/frontend/src/components/agentchat/TraceTimeline.tsx
index 5a69e6efae..76dd1646a0 100644
--- a/jarvis/ui/web/frontend/src/components/agentchat/TraceTimeline.tsx
+++ b/jarvis/ui/web/frontend/src/components/agentchat/TraceTimeline.tsx
@@ -160,7 +160,7 @@ function Chevron({ open }: { open: boolean }) {
function ActivityView({ entry, renderers }: { entry: ActivityEntry; renderers: TimelineRenderers }) {
const [open, setOpen] = useState(false);
- // A stretch is always exactly ONE line, as in the Codex app. While a call
+ // A stretch is always exactly ONE line. While a call
// runs, that line IS the running call; when it settles, the line becomes
// the summary again. Drawing the running call as an extra line made the
// whole chat grow and shrink by a line on every call (the bounce bug).
diff --git a/jarvis/ui/web/frontend/src/components/agentchat/WorkTrace.css b/jarvis/ui/web/frontend/src/components/agentchat/WorkTrace.css
index cfc375add8..a3e9409e0d 100644
--- a/jarvis/ui/web/frontend/src/components/agentchat/WorkTrace.css
+++ b/jarvis/ui/web/frontend/src/components/agentchat/WorkTrace.css
@@ -31,7 +31,7 @@
position: relative;
min-width: 0;
}
-/* Since 2026-10-02 the rail draws no thread: traces read like the Codex app,
+/* Since 2026-10-02 the rail draws no thread: traces read as quiet prose,
prose and quiet lines with nothing joining them. */
.trace-rail > .trace-step::before,
.trace-rail > .trace-step::after {
@@ -98,7 +98,7 @@
}
/* The words of the step that is working: a slow light sweep over muted ink,
- the way the Claude app marks "Thinking". */
+ the way a live "Thinking" mark reads. */
.trace-shimmer {
background-image: linear-gradient(
90deg,
@@ -170,7 +170,7 @@
border-radius: 0.5rem;
}
/* Work lines and reasoning are set in the size of the surface's own text,
- as in the Codex app: 15/25 on the front page, 13/21 in agent chats. */
+ 15/25 on the front page, 13/21 in agent chats. */
/* !important: the typography plugin's .prose sets its own size. */
.trace-text {
font-size: var(--trace-text-size, 15px) !important;
diff --git a/jarvis/ui/web/frontend/src/components/agentchat/WorkTrace.test.tsx b/jarvis/ui/web/frontend/src/components/agentchat/WorkTrace.test.tsx
index 3537734b32..62a5f61bb7 100644
--- a/jarvis/ui/web/frontend/src/components/agentchat/WorkTrace.test.tsx
+++ b/jarvis/ui/web/frontend/src/components/agentchat/WorkTrace.test.tsx
@@ -136,7 +136,7 @@ describe("work trace", () => {
it("renders edit input, a diff and output on demand", () => {
render(<WorkTrace {...props} blocks={[tool("edit",{name:"Edit",input:{file_path:"app.ts",old_string:"oldValue",new_string:"newValue"},output:"File updated"})]} />);
- // Codex-style: the edit is one line with its size; the diff is one tap away.
+ // The edit is one line with its size; the diff is one tap away.
const edit = screen.getByRole("button", { name: /^Edited app\.ts \+1 −1/ });
fireEvent.click(edit);
expect(screen.getByLabelText("Changes").textContent).toContain("− oldValue");
@@ -237,7 +237,7 @@ describe("work trace", () => {
});
describe("rail look", () => {
- it("reads like the Codex app: prose, quiet lines, no thread, no bullets", () => {
+ it("reads as prose: quiet lines, no thread, no bullets", () => {
const { container } = render(<WorkTrace {...props} status="running" blocks={[thought, tool("a", { name: "exec_command", input: { command: "npm test" } })]} />);
expect(screen.getByTestId("work-trace").getAttribute("data-look")).toBe("rail");
expect(container.querySelector(".trace-dot, .trace-node-live")).toBeNull();
diff --git a/jarvis/ui/web/frontend/src/components/agentchat/traceEntries.test.ts b/jarvis/ui/web/frontend/src/components/agentchat/traceEntries.test.ts
index 094b916d26..e9048326d8 100644
--- a/jarvis/ui/web/frontend/src/components/agentchat/traceEntries.test.ts
+++ b/jarvis/ui/web/frontend/src/components/agentchat/traceEntries.test.ts
@@ -43,7 +43,7 @@ function replay(events: Array<[string, Record<string, unknown>]>): ThinkingStep[
return steps.map((s) => (s.status === "active" ? { ...s, status: "done" } : s));
}
-describe("trace timeline (Codex-style)", () => {
+describe("trace timeline", () => {
it("alternates the model's words with one line per stretch of calls", () => {
const line = buildTimeline([
text("n1", "I'll look at the blog first."),
diff --git a/jarvis/ui/web/frontend/src/components/home/ChatKindMark.tsx b/jarvis/ui/web/frontend/src/components/home/ChatKindMark.tsx
index 09424c655a..976b90ed89 100644
--- a/jarvis/ui/web/frontend/src/components/home/ChatKindMark.tsx
+++ b/jarvis/ui/web/frontend/src/components/home/ChatKindMark.tsx
@@ -7,7 +7,7 @@ import type { ChatRow } from "@/components/home/chatRows";
* The small mark in front of a history row that says what kind of chat it is:
* an open ring for a typed chat, sound bars for a voice chat — the same bars
* the composer's voice button wears. Both sit in one 14 px box in the same
- * muted tone, the way the Claude app marks its chat and code sessions, so the
+ * muted tone, marking chat and code sessions apart, so the
* column reads as one list with two clearly different kinds in it.
*/
export function ChatKindMark({
diff --git a/jarvis/ui/web/frontend/src/components/home/ChatStage.tsx b/jarvis/ui/web/frontend/src/components/home/ChatStage.tsx
index 90b32e4f3a..84feb084a4 100644
--- a/jarvis/ui/web/frontend/src/components/home/ChatStage.tsx
+++ b/jarvis/ui/web/frontend/src/components/home/ChatStage.tsx
@@ -204,7 +204,7 @@ function ChatStageContent() {
// chat is coding sessions and never shows a spoken thread.
if (surface === "jarvis" && voiceThreadId && !activeSessionId) return <VoiceThreadStage />;
- // The front page opens the way the Codex app does: the greeting alone in
+ // The front page opens with the greeting alone in
// the middle of the page, the composer already where it will stay — at the
// bottom — so the first send moves nothing.
if (!hasContent && isJarvis) {
diff --git a/jarvis/ui/web/frontend/src/components/home/Greeting.tsx b/jarvis/ui/web/frontend/src/components/home/Greeting.tsx
index 2650d78c37..f5c96bde9e 100644
--- a/jarvis/ui/web/frontend/src/components/home/Greeting.tsx
+++ b/jarvis/ui/web/frontend/src/components/home/Greeting.tsx
@@ -33,7 +33,7 @@ export function Greeting({
data-testid="home-greeting"
>
{/* The user's pet above, the line under it in regular weight — the
- Codex app's opening, quieter than a bold headline beside a logo.
+ a quiet opening, softer than a bold headline beside a logo.
The pet follows the voice: it listens, thinks and talks along. */}
<PetMark size={48} reactive className="mb-3" />
<h1 className="text-2xl font-normal tracking-tight text-foreground [text-wrap:balance]">
diff --git a/jarvis/ui/web/frontend/src/components/home/RecentChats.tsx b/jarvis/ui/web/frontend/src/components/home/RecentChats.tsx
index 8cae8c9998..898aeb526d 100644
--- a/jarvis/ui/web/frontend/src/components/home/RecentChats.tsx
+++ b/jarvis/ui/web/frontend/src/components/home/RecentChats.tsx
@@ -40,7 +40,7 @@ export function RecentChats() {
useHistoryPolling(loadSessions);
- // Every chat is listed, the way the Claude app's column lists them: the
+ // Every chat is listed in one column: the
// sidebar scrolls instead of hiding the history behind "Show all"
// (maintainer, 2026-10-01). The archive dialog stays for searching it.
const shown = recentRows;
diff --git a/jarvis/ui/web/frontend/src/components/home/VoiceComposer.tsx b/jarvis/ui/web/frontend/src/components/home/VoiceComposer.tsx
index d24b1a555b..8e1b06d666 100644
--- a/jarvis/ui/web/frontend/src/components/home/VoiceComposer.tsx
+++ b/jarvis/ui/web/frontend/src/components/home/VoiceComposer.tsx
@@ -11,7 +11,7 @@ import { cn } from "@/lib/utils";
/**
* Voice mode's composer — the chat composer's card, spoken instead of typed
- * (2026-10-01, after the Claude app's voice mode).
+ * (2026-10-01).
*
* Same shape as the front page's typed composer (components/agentchat/
* AgentComposer, minimal): one rounded row. Where the text would be, the
diff --git a/jarvis/ui/web/frontend/src/components/home/VoiceGlow.tsx b/jarvis/ui/web/frontend/src/components/home/VoiceGlow.tsx
index bb4c3f7e10..3873db16b4 100644
--- a/jarvis/ui/web/frontend/src/components/home/VoiceGlow.tsx
+++ b/jarvis/ui/web/frontend/src/components/home/VoiceGlow.tsx
@@ -8,7 +8,7 @@ import { createGlowRenderer } from "@/components/home/voiceGlowRenderer";
import { createVoiceFollower } from "@/components/home/voiceFollower";
/**
- * The light behind voice mode's composer — the Claude app's voice glow, in
+ * The light behind voice mode's composer — a soft voice glow, in
* this theme's one signal hue.
*
* An aurora rises from the bottom edge: a soft haze with a flowing crest,
diff --git a/jarvis/ui/web/frontend/src/components/layout/Sidebar.tsx b/jarvis/ui/web/frontend/src/components/layout/Sidebar.tsx
index 21134c4aa4..35a8da1a1e 100644
--- a/jarvis/ui/web/frontend/src/components/layout/Sidebar.tsx
+++ b/jarvis/ui/web/frontend/src/components/layout/Sidebar.tsx
@@ -328,7 +328,7 @@ export function Sidebar({
// Lit while any hub section is on screen — the profile button IS the hub's
// entry point now, so it carries the "you are here" state for all of them.
const hubActive = (SETTINGS_HUB_IDS as readonly string[]).includes(active);
- // Rows read like the Claude app's column: regular weight, light ink at
+ // Rows read quietly: regular weight, light ink at
// rest (muted grey made every entry look disabled), the icon in the same
// ink, a lift on hover. Section labels and tail rows stay muted.
const rowClass = "flex min-h-8 w-full items-center gap-3 rounded-lg px-3 text-base text-foreground transition-colors hover:bg-secondary focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring";
diff --git a/jarvis/ui/web/frontend/src/components/society/data.ts b/jarvis/ui/web/frontend/src/components/society/data.ts
index 9cab866862..e5f7fbd120 100644
--- a/jarvis/ui/web/frontend/src/components/society/data.ts
+++ b/jarvis/ui/web/frontend/src/components/society/data.ts
@@ -91,7 +91,7 @@ export interface AgentStats {
lastActiveMs: number | null;
}
-/** Grok-style rules under the ceiling (agent-definition §3.4); require wins. */
+/** Plain-language rules under the ceiling (agent-definition §3.4); require wins. */
export interface ApprovalRules {
requireApproval: string[];
alwaysAllow: string[];
diff --git a/jarvis/ui/web/frontend/src/lib/quickSwitch.ts b/jarvis/ui/web/frontend/src/lib/quickSwitch.ts
index ebec0cfe1e..836823eeb2 100644
--- a/jarvis/ui/web/frontend/src/lib/quickSwitch.ts
+++ b/jarvis/ui/web/frontend/src/lib/quickSwitch.ts
@@ -1,5 +1,5 @@
/**
- * The quick switcher's destinations and ranking — the Spotlight-style "type a
+ * The quick switcher's destinations and ranking — the "type a
* section's name, press Enter, be there" launcher.
*
* Pure on purpose: what the switcher lists and in which order is the whole
diff --git a/jarvis/ui/web/frontend/src/lib/societyApi.ts b/jarvis/ui/web/frontend/src/lib/societyApi.ts
index dc11d07385..b82fe07960 100644
--- a/jarvis/ui/web/frontend/src/lib/societyApi.ts
+++ b/jarvis/ui/web/frontend/src/lib/societyApi.ts
@@ -93,7 +93,7 @@ export interface SocietyEnvelope {
payload: Record<string, unknown>;
}
-/** Approval rules — Grok-style, require wins (agent-definition §3.4). */
+/** Approval rules — plain-language, require wins (agent-definition §3.4). */
export interface ApprovalRules {
require_approval: string[];
always_allow: string[];
diff --git a/jarvis/ui/web/frontend/src/store/agentChat.ts b/jarvis/ui/web/frontend/src/store/agentChat.ts
index a1d3791675..a609a28273 100644
--- a/jarvis/ui/web/frontend/src/store/agentChat.ts
+++ b/jarvis/ui/web/frontend/src/store/agentChat.ts
@@ -54,7 +54,7 @@ import { reuseHistoryRows } from "@/lib/historyRequests";
* DRAFT: for a fresh chat it is what the next session is created with; with
* a session open, changing a pick patches that session on the backend and
* the same draft mirrors it. The draft survives a reload (localStorage) so
- * the composer opens on what you used last, like the Claude app does.
+ * the composer opens on what you used last.
*
* The provider list is the backend catalog (`/api/agent-chat/catalog`)
* joined with the Agents tab's credential truth (`/api/jarvis-agent/status`):
diff --git a/jarvis/ui/web/frontend/src/views/ChatsView.tsx b/jarvis/ui/web/frontend/src/views/ChatsView.tsx
index 4ca41029c5..d71aaf7c1a 100644
--- a/jarvis/ui/web/frontend/src/views/ChatsView.tsx
+++ b/jarvis/ui/web/frontend/src/views/ChatsView.tsx
@@ -483,7 +483,7 @@ function MessageBubble({ message }: { message: ChatMessage }) {
}
/**
- * Empty state — Claude-style: calm, centered, no suggestion cards. A quiet
+ * Empty state — calm, centered, no suggestion cards. A quiet
* mascot, a centered greeting and a one-line subtitle. The composer below is
* the focus (it carries the new mic / dictation button). Deliberately minimal —
* the user explicitly asked to drop the canned prompt cards.
diff --git a/scripts/ci/check_design_references.py b/scripts/ci/check_design_references.py
new file mode 100644
index 0000000000..d57e372918
--- /dev/null
+++ b/scripts/ci/check_design_references.py
@@ -0,0 +1,198 @@
+#!/usr/bin/env python3
+"""CI gate: public text never presents another product as our design template.
+
+AGENTS.md ("Public product descriptions"): tracked docs, comments, docstrings,
+test descriptions, video copy and changelog text describe our features
+directly. They never say a feature is inspired by, modeled on, adapted from or
+drawn "the way" another product does it. Required license and attribution
+notices and factual integration references stay; this gate never asks anyone
+to hide where code came from.
+
+The patterns are deliberately narrow — a product name next to a design verb,
+a "<Product>-style" adjective, or "like the <Product> app" — so an integration
+reference ("import the Claude Desktop config", "the Codex CLI worker") never
+trips it. A line that must keep such wording (a license header, a quoted
+user utterance) carries the inline marker ``design-ref-allow``; whole files
+that are third-party text are listed in ``_ALLOWED_PATHS``.
+
+Modes::
+
+ python scripts/ci/check_design_references.py # whole tree (CI)
+ python scripts/ci/check_design_references.py --staged # added lines (pre-commit)
+
+Exit 1 = a confirmed hit; anything else is tooling trouble and the pre-commit
+hook fails open on it.
+"""
+
+from __future__ import annotations
+
+import fnmatch
+import os
+import re
+import subprocess
+import sys
+from pathlib import Path
+
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+from check_no_new_german import SCAN_EXT, parse_added_lines # noqa: E402
+
+_INLINE_ESCAPE = "design-ref-allow"
+
+# Third-party text, the files that define or test this rule, and code that
+# carries an upstream license header for what it adapted.
+_ALLOWED_PATHS: tuple[str, ...] = (
+ "**/THIRD_PARTY_NOTICES*",
+ "CODE_OF_CONDUCT.md",
+ "jarvis/skills/builtin/skill-creator/references/*",
+ "jarvis/ui/web/frontend/src/components/society/AgentSymbol.tsx",
+ "scripts/readme-video/src/identity/AgentSymbol.tsx",
+ "scripts/ci/check_design_references.py",
+ "tests/unit/ci/test_check_design_references.py",
+ "**/dist/*",
+ "*package-lock.json",
+)
+
+# Product names are matched case-sensitively, so "a linear fade", "the cursor"
+# or "an arc" never count; the verbs around them ignore case.
+_PRODUCTS = (
+ r"ShareX|CleanShot(?: X)?|Raycast|Alfred|Spotlight|Grok(?: Bot)?"
+ r"|Hermes(?: Agent| Bot| Desktop)?|Nous Research|Codex(?: app| desktop)?"
+ r"|Claude(?: app| desktop)?|ChatGPT(?: app)?|Cursor"
+ r"|Linear|Notion|Cowork|Superhuman|a dictation app(?: Flow)?|a dictation app|Granola|Arc|Warp"
+)
+_APPS = r"Codex|Claude|ChatGPT|Grok|Cursor|Raycast|CleanShot|ShareX"
+# Every rule needs one of these names, so a line without any is skipped cheaply.
+_ANY_PRODUCT = re.compile(_PRODUCTS)
+
+_RULES: tuple[tuple[str, re.Pattern[str]], ...] = (
+ (
+ "design verb next to a product",
+ re.compile(
+ r"\b(?i:inspired by|modell?ed (?:on|after)|adapted from|borrowed from|copied from"
+ r"|design reference|reference product|follows the design of)\b[^.\n]{0,60}?\b(?:"
+ + _PRODUCTS
+ + r")\b"
+ ),
+ ),
+ (
+ "'<Product>-style' adjective",
+ re.compile(
+ r"\b(?:Codex|Claude|ChatGPT|Cursor|Raycast|Hermes|Grok|Spotlight|Notion|Linear"
+ r"|ShareX|CleanShot)[- ](?i:style|like|inspired)\b"
+ r"(?! `?(?:event|stream|payload|format|apply_patch))"
+ ),
+ ),
+ (
+ "behaves like another app",
+ re.compile(
+ r"\b(?i:like|after|as in|the way|judged against)(?: the)? (?:" + _APPS + r")"
+ r"(?: X)? (?:app|does|draws|shows|marks|opens|lists|keeps)\b"
+ ),
+ ),
+ (
+ "another app's look",
+ re.compile(
+ r"\b(?:" + _APPS + r")(?: X)?(?: app)?['’]s (?i:look|design|layout|style|composer"
+ r"|column|opening|voice mode|voice glow|region-capture look|annotate tool"
+ r"|Quick Access Overlay)\b"
+ ),
+ ),
+ (
+ "names a screenshot tool",
+ re.compile(r"\b(?:ShareX|CleanShot)\b", re.IGNORECASE),
+ ),
+)
+
+
+def is_allowed(path: str) -> bool:
+ norm = path.replace("\\", "/")
+ return any(fnmatch.fnmatch(norm, pat) for pat in _ALLOWED_PATHS)
+
+
+def is_scanned(path: str) -> bool:
+ return Path(path).suffix.lower() in SCAN_EXT
+
+
+def line_hits(text: str) -> list[str]:
+ """Names of the rules ``text`` breaks (empty when it is clean)."""
+ if not _ANY_PRODUCT.search(text) or _INLINE_ESCAPE in text:
+ return []
+ return [name for name, rx in _RULES if rx.search(text)]
+
+
+def find_violations(
+ lines: list[tuple[str, int, str]],
+) -> list[tuple[str, int, str, str]]:
+ out: list[tuple[str, int, str, str]] = []
+ scanned: dict[str, bool] = {}
+ for path, lineno, text in lines:
+ if path not in scanned:
+ scanned[path] = is_scanned(path) and not is_allowed(path)
+ if not scanned[path]:
+ continue
+ hits = line_hits(text)
+ if hits:
+ out.append((path, lineno, hits[0], text.strip()[:140]))
+ return out
+
+
+def _git(args: list[str]) -> subprocess.CompletedProcess[str]:
+ return subprocess.run( # noqa: S603
+ ["git", *args],
+ capture_output=True,
+ text=True,
+ encoding="utf-8",
+ errors="replace",
+ check=False,
+ creationflags=getattr(subprocess, "CREATE_NO_WINDOW", 0) if os.name == "nt" else 0,
+ )
+
+
+def tracked_lines() -> list[tuple[str, int, str]]:
+ out: list[tuple[str, int, str]] = []
+ for path in _git(["ls-files"]).stdout.splitlines():
+ path = path.strip()
+ if not path or not is_scanned(path) or is_allowed(path):
+ continue
+ try:
+ content = (Path.cwd() / path).read_text(encoding="utf-8", errors="replace")
+ except OSError:
+ continue
+ out.extend((path, n, line) for n, line in enumerate(content.splitlines(), start=1))
+ return out
+
+
+def staged_lines() -> list[tuple[str, int, str]]:
+ diff = _git(["diff", "--cached", "--unified=0", "--no-color", "--diff-filter=AM"]).stdout
+ return parse_added_lines(diff)
+
+
+def main(argv: list[str] | None = None) -> int:
+ argv = list(sys.argv[1:] if argv is None else argv)
+ try:
+ sys.stdout.reconfigure(encoding="utf-8", errors="replace") # type: ignore[union-attr]
+ except (AttributeError, ValueError):
+ pass
+
+ staged = "--staged" in argv
+ violations = find_violations(staged_lines() if staged else tracked_lines())
+ scope = "staged changes" if staged else "whole tree"
+ if not violations:
+ print(f"design-reference gate OK ({scope}).")
+ return 0
+
+ print("DESIGN-REFERENCE GATE FAILED - public text names another product as our template.\n")
+ for path, lineno, rule, text in violations:
+ print(f" [x] {path}:{lineno} ({rule})")
+ print(f" {text}")
+ print(
+ f"\n{len(violations)} line(s). Describe our feature directly instead "
+ "(AGENTS.md, 'Public product descriptions').\n"
+ "A license header or a factual integration line may carry an inline "
+ f"'{_INLINE_ESCAPE}' marker."
+ )
+ return 1
+
+
+if __name__ == "__main__":
+ raise SystemExit(main())
diff --git a/scripts/ci/run_gates.py b/scripts/ci/run_gates.py
index c567bc2757..38c37129d1 100644
--- a/scripts/ci/run_gates.py
+++ b/scripts/ci/run_gates.py
@@ -46,6 +46,7 @@ def _gates(base: str) -> list[Gate]:
Gate("private-keys", ((PY, ci + "check_no_private_keys.py"),)),
Gate("dist-consistency", ((PY, ci + "check_dist_consistency.py"),)),
Gate("brand-logos", ((PY, ci + "check_brand_logos.py"),)),
+ Gate("design-references", ((PY, ci + "check_design_references.py"),)),
Gate(
"mirrors",
(
diff --git a/scripts/ci/run_tests_parallel.py b/scripts/ci/run_tests_parallel.py
index b20c58057e..830cc60173 100644
--- a/scripts/ci/run_tests_parallel.py
+++ b/scripts/ci/run_tests_parallel.py
@@ -5,7 +5,7 @@ Why this exists: one monolithic ``pytest tests/`` process took 75-95 minutes
on a 4-core runner and routinely hit the job timeout. A single wedged test ate
the whole budget, and module state leaked across the ~2 000 test files.
-How it works (the idea is borrowed from Hermes Agent's per-file runner):
+How it works (a per-file runner):
1. **Discover** every ``test_*.py`` / ``*_test.py`` under the roots.
2. **Shard** (``--shard 2/6``): longest-processing-time bin packing over the
diff --git a/tests/unit/ci/test_check_design_references.py b/tests/unit/ci/test_check_design_references.py
new file mode 100644
index 0000000000..3704f7a9d4
--- /dev/null
+++ b/tests/unit/ci/test_check_design_references.py
@@ -0,0 +1,78 @@
+"""Tests for the CI gate that keeps other products out of our design story.
+
+The gate lives in scripts/ci/ (not an importable package), so that directory
+goes on sys.path first, the way CI runs the script directly.
+"""
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_SCRIPTS_CI = Path(__file__).resolve().parents[3] / "scripts" / "ci"
+sys.path.insert(0, str(_SCRIPTS_CI))
+
+import check_design_references as gate # noqa: E402
+
+
+ "line",
+ [
+ "Give the area picker ShareX's region-capture look",
+ "The editor follows CleanShot X's annotate tool",
+ "Draw traces the way the Codex app does",
+ "A quiet, Claude-style front page: one greeting, one composer.",
+ "// Rows read like the Claude app's column: regular weight",
+ "* Left, top to bottom (maintainer, modelled on Grok Bot's New Bot dialog)",
+ "How it works (the idea is borrowed from Hermes Agent's per-file runner):",
+ "Rebuilt after a pipeline inspired by Hermes Agent",
+ "The quick switcher is a Spotlight-style launcher.",
+ "# --- the Codex-like look: filled discs, a glossy orb",
+ "Approval rules — Grok-style, require wins.",
+ ],
+)
+def test_design_references_are_flagged(line: str) -> None:
+ assert gate.line_hits(line), line
+
+
+ "line",
+ [
+ "Import MCP servers from the Claude Desktop config.",
+ "clients (Claude Desktop, Cursor, VS Code, Codex) launch a command",
+ "Parse a Codex-style `apply_patch` payload.",
+ "Synthesize a tool_use Claude-style event so the timeline renders it.",
+ "Modelled on `agentAccountsApi.ts` (same send<T> shape).",
+ "a linear fade modeled on the cursor position along an arc",
+ "Ask for something big, and Jarvis sends background agents like Claude Code.",
+ "The Codex CLI worker reads the same profile as the Claude worker.",
+ "Spotlight search reads the per-volume metadata store kept by mds.",
+ "adapted from Nous Research's avatar.tsx design-ref-allow",
+ ],
+)
+def test_integration_and_neutral_lines_pass(line: str) -> None:
+ assert gate.line_hits(line) == [], line
+
+
+def test_third_party_files_are_skipped() -> None:
+ lines = [
+ ("CODE_OF_CONDUCT.md", 1, "inspired by Mozilla's code of conduct, see ShareX"),
+ ("jarvis/ui/web/frontend/public/THIRD_PARTY_NOTICES.txt", 2, "ShareX"),
+ ("docs/appshots.md", 3, "The picker copies ShareX."),
+ ("assets/brand/logo.png", 4, "ShareX"),
+ ]
+ assert [v[0] for v in gate.find_violations(lines)] == ["docs/appshots.md"]
+
+
+def test_added_lines_from_a_staged_diff_are_scanned() -> None:
+ diff = (
+ "diff --git a/docs/x.md b/docs/x.md\n"
+ "--- a/docs/x.md\n"
+ "+++ b/docs/x.md\n"
+ "@@ -1,0 +4,2 @@\n"
+ "+The card stacks as in CleanShot X.\n"
+ "+The card stacks upward.\n"
+ )
+ violations = gate.find_violations(gate.parse_added_lines(diff))
+ assert [(v[0], v[1]) for v in violations] == [("docs/x.md", 4)]
diff --git a/tests/unit/ui/test_orb_controls_pet.py b/tests/unit/ui/test_orb_controls_pet.py
index 74f8371a1d..616957a11f 100644
--- a/tests/unit/ui/test_orb_controls_pet.py
+++ b/tests/unit/ui/test_orb_controls_pet.py
@@ -298,7 +298,7 @@ def test_concurrent_toggles_never_lose_the_remembered_volume(pipeline: _Pipeline
assert pipeline._volume == pytest.approx(0.7)
-# --- the Codex-like look: filled discs, a glossy orb, proportions --------------------
+# --- the look: filled discs, a glossy orb, proportions ---------------------------
def test_the_pen_and_the_pill_are_filled_without_a_ring() -> None: