agentleFS
friuns2friuns2/openclaw-android-assistant/ui/AGENTS.md

openclaw-android-assistant / ui

AGENTS.md573 starsChanged 34 days agoMITView on GitHub

This directory owns Control UI-specific guidance that should not live in the repo root. - The Gateway owns shared state that other clients or channels can change. Renderer copies are caches; local presentation state belongs to the view. Reuse the owning store or controller and Gateway contract instead of implementing session, configuration, or authorization decisions again in the UI. - Scope cached data and in-flight requests to the actual connection, agent, and session they concern. Before publishing a result, check…

2 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this AGENTS.md into the rules it gives an agent

5 rules. About 0 are common, 5 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    After each automatic refresh, wait three times its duration (at least five seconds, at most 15 seconds) before the next automatic read.

    line 85
  2. Distinctive:

    Do not hand-edit translation memory, locale metadata, or fallback metadata unless a targeted generated-output fix is explicitly requested.

    line 96
  3. Distinctive:

    If locale outputs drift, let the workflow reconcile them or run release prep.

    line 110
  4. Distinctive:

    Keep UI-specific rules here.

    line 145
  5. Distinctive:

    Leave repo-global architecture, verification, and git workflow rules in the root AGENTS.md.

    line 146

≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
AGENTS.md

Control UI Guide

This directory owns Control UI-specific guidance that should not live in the repo root.

State Ownership And Async Results

  • The Gateway owns shared state that other clients or channels can change. Renderer copies are caches; local presentation state belongs to the view. Reuse the owning store or controller and Gateway contract instead of implementing session, configuration, or authorization decisions again in the UI.
  • Scope cached data and in-flight requests to the actual connection, agent, and session they concern. Before publishing a result, check that its owner and request generation are still current. A late result from a prior context must not replace newer intent or populate the newly selected context.
  • Optimistic updates retain enough state for visible recovery, then reconcile with the authoritative result. An old request must not roll back a newer edit. Failed writes follow the Gateway's target and outcome contract, not a fallback account or connection selected by the renderer.
  • Background updates may refresh their own scoped cache; they must not replace the foreground selection or publish another context's state into its view.

Session Roster Refresh

  • Session rosters apply nested Gateway row snapshots through the shared reconciler. lib/sessions/session-list-query.ts owns whether a snapshot preserves a held window: lifecycle, participants, placement, patch/send/steer, run-start/settlement/capacity, and title updates can avoid list reads when membership, lineage, and pin/owner/archive facts stay unchanged and recency does not move backwards. Tree events require the Gateway's complete, access-scoped ancestorSessions snapshots plus any ancestorSessionRefs; references require the held row's admitted content revision. Each row retains its own generation and field receipts. Certified nested rows own their facts; only explicit null clearing receipts may fill omissions from the event envelope. Unknown rows, incomplete ancestor coverage, broad changes, catalog changes, Gateway-owned filters, failed reads, owner-prefix boundary uncertainty, and overlapping reads retain an authoritative refresh. Events never create filtered list membership; Current Work's complete unfiltered active-only window may admit a certified full active row as described below.
  • List callers use compact rows and bounded source attribution. Enrichment flags and inclusion of global/unknown kinds do not change held membership when kind stays unchanged; dashboard filters require matching hasBoard/boardFace receipts. Gallery pagination extends the shared managed window. Full settings come from row descriptors on demand. Applied row traffic retains one fallback through the refresh coordinator after at least 60 seconds. Child-query membership uses childOwnerSessionKeys from the Gateway's retention owner. Parent-only events must certify the complete child window; unheld ancestors need explicit exclusion facts, or an admitted reference resolved through the connection's existing row provenance.
  • Re-adopting cached lineage rows changes presentation without invalidating managed list membership. Fresh descriptor reads and Gateway events retain their authoritative invalidation paths.
  • Descriptor observations apply admitted rows immediately; incomplete ancestor coverage retains one paced authoritative descriptor refresh through the coordinator instead of a read per event.
  • Activity's current-work and unfiltered history views apply admitted row events without refetching the list. History uses the shared row-provenance owner: admitted live rows retain field clocks, including recap updates and clearing receipts, while returned lists remain the sole membership authority. Its excludeSubagents query ignores key-proven child exclusions only with complete access-scoped ancestor coverage; held parent snapshots and certified references still update locally. Missing held parents or supplied unheld ancestors require authority; named unheld parents absent from certified coverage may be invisible. Refreshes merge those live facts instead of replaying/coalescing History packets; missing initial membership retains a catch-up read. Query/connection changes reset provenance. Current Work coalesces only consecutive full active snapshots for one generation, retaining the first receipt and latest tail. Partial events, references, terminal snapshots, and other identities are FIFO barriers. A complete Current Work window below its limit admits a certified active snapshot sampled after the accepted list and that session generation's observed retirement and liveness clocks, with a real session ID and kind, preserving the Gateway's cron-run exclusion. Truncated/full windows, conflicting generations, partial unknown rows, ambiguous settlement, and removals from an incomplete window require an authoritative refresh. History requires forward activity clocks and retains omitted recap enrichment; person/search filters keep authoritative refreshes. Healthy row traffic retains one fallback refresh after at least 60 seconds, including whole-query people and pulse facets. Current Work applies supplied ancestor snapshots and references through the shared row reconciler; pending references retain their prerequisite full row receipts. Its controller retains at most 1,000 fence records, coalesced by session generation; the Current Work reconciler owns their independent retirement, liveness-observation, and generation-authority facts. Only retirements prune stale returned rows. A fresh certified full row can resolve older clocked liveness, while unclocked observations and conflicting generations require a list read. Query/connection changes clear records; an authoritative list clears uncertainty and covered retirements before pending events replay, preserving later retirements. Saturation blocks unseen admission until an authoritative read restores the bounded state.
  • lib/sessions/event-refresh-coordinator.ts owns automatic refresh pacing: collect events in a four-to-five-second window sampled once when armed so browsers spread their reads and subsequent events cannot postpone them. After each automatic refresh, wait three times its duration (at least five seconds, at most 15 seconds) before the next automatic read. Trailing invalidation stays with that owner, including while a request is pending.
  • Explicit refreshes, filter/agent changes, reconnects, and foreground replacements bypass event backoff and absorb pending invalidation. Recheck visibility and current intent after background admission; hidden pages retain one catch-up refresh until visible.

i18n Rules

  • Foreign-language files in ui/src/i18n/locales/*.ts are stable, source-owned lazy-module adapters; their translations are generated from canonical grouped memory in ui/src/i18n/.i18n/*.tm.jsonl.
  • Do not hand-edit translation memory, locale metadata, or fallback metadata unless a targeted generated-output fix is explicitly requested.
  • English source lives in ui/src/i18n/locales/en.ts, its static en-agents.ts dependency, and lazy en-*.ts registrar catalogs. scripts/lib/control-ui-i18n-catalog.ts owns complete ordered composition and raw source-hash dependencies for generation, verification, and Vite; it reads .catalog data without runtime registration. Related wiring:
    • scripts/control-ui-i18n.ts
    • scripts/lib/control-ui-i18n-catalog.ts
    • scripts/lib/control-ui-i18n-sync-plan.ts
    • ui/config/control-ui-locales.ts
    • ui/src/i18n/lib/types.ts
    • ui/src/i18n/lib/registry.ts
  • Register lazy English synchronously at each lazy consumer, including Settings search before a destination page loads. Keep startup/shared copy eager. Preserve the shared en object and sibling namespaces; leave empty whole-subtree anchors in en.ts when extraction would change flattened source order and grouped translation-memory aliases. Never import the host-only catalog owner into the runtime.
  • Contributor flow: update English strings and locale adapters/wiring, run keyless pnpm ui:i18n:baseline, and commit source files plus any changed raw-copy baseline. Do not include catalog fallback metadata, locale metadata, or translation memory in a source PR; CI rejects mixed source/generated diffs outside canonical release/YYYY.M.PATCH branches or an explicitly detected complete canonical-memory ownership migration.
  • pnpm ui:i18n:verify is deterministic and keyless. pnpm lint and the changed-check UI lane run it. It validates English catalog shape, runtime locale wiring, and raw-copy baseline drift; foreign catalog parity belongs to the post-merge bot and strict generated-output gate.
  • Translation flow: the serialized control-ui-locale-refresh workflow translates after merge, opens an isolated generated PR, and enables auto-merge for its exact head. pnpm ui:i18n:sync remains the authenticated maintainer/release repair path; do not run it without provider auth when new keys exist.
  • pnpm release:prep runs the locale sync before release freeze, then pnpm ui:i18n:check remains the strict generated-output/release gate with zero fallbacks.
  • Prioritization report: pnpm ui:i18n:report [--surface <name>] [--locale <locale>] [--top <n>] shows current hardcoded-copy focus areas and locale fallback metadata. It is not a drift gate; use pnpm ui:i18n:check for that.
  • If locale outputs drift, let the workflow reconcile them or run release prep. Do not manually translate, merge, or hand-maintain generated translation memory or locale metadata.

CSS / Template Linting

  • pnpm lint:ui:styles runs stylelint over ui/src stylesheets and Lit css templates (postcss-lit). pnpm lint includes it; error-class rules only, oxfmt owns formatting. Config: config/stylelint.config.mjs.
  • Icons: shared 24x24 Lucide icons go through strokeIcon() in ui/src/components/icons-tools.ts so stroke presentation attributes stay inline and render inside shadow roots. Icon bodies are svg\`fragments, neverhtml``` (wrong namespace renders nothing).
  • pnpm lint:ui:lit is an opt-in lit-analyzer diagnostic for template bindings (slow, ~9 min; known baseline of pre-existing findings). It is not a CI gate.

Stylesheet Policy

  • No universal targets or pseudo-elements after a :has() compound, no :has() with ::placeholder (measured ~9/~8 ms subtree restyles per insertion with 534 messages), and no descendant after a sibling-relative :has(+ …) on repeated items (it restyled every position-rail tick per transcript row); use owner-set classes, named children, or a custom property on the :has() subject. Stylelint enforces these. No :has() on .shell, .content, .chat-thread, .chat-split-view, :root, html, or body compounds, including modifiers, :is()/:where() list subjects, and nested & forms: every insertion below a :has() subject schedules a global :has restyle of that subtree (stylelint cannot resolve nesting). Third-party global CSS goes through the Vite PostCSS pipeline; ui/config/control-ui-web-awesome-page-rule.ts drops Web Awesome's never-matching :is(html, body):has(wa-page) rule.
  • Cursors: links and controls that open a new tab use the pointer; state-changing controls keep the default arrow.
  • Colors: stylesheet colors flow through custom-property tokens defined in ui/src/styles/base.css; color-no-hex enforces this. Exempt surfaces (token definitions, lobster-pet.css sprite artwork, --theme-chip-* preview swatches) each carry a stated contract. Lit css\`` templates are not yet gated — prefer tokens there too.
  • Breakpoints: max-width media conditions use the canonical ladder 400/560/640/768/900/1100/1320px (plus the 932×500 landscape-phone compound); stylelint's allowed-list enforces it. New thresholds round up to the next rung. Don't add rungs without updating the config comment and this note.
  • Duplicate selectors are lint errors; deliberate topic-section reopens use stylelint-disable-next-line no-duplicate-selectors -- <reason>.
  • Dead CSS: node --import tsx scripts/audit-control-ui-dead-css.mts reports class selectors with no production reference (AST-based, understands X--${...} stems and classMap). Advisory, not a gate — verify by hand before deleting; extend the script's stem detection rather than its allowlist when it misses a dynamic family.
  • Native CSS nesting: opportunistic only — nest when already rewriting a section; no conversion sweeps.
  • @layer is deliberately not used: the shared light-DOM stylesheet's precedence relies on import order plus specificity, page CSS imports lazily per component, and measured no-descending-specificity hits are within-file — layering the import manifest would flip unlayered-vs-layered precedence across ~48 lazily imported page files for no measured win. Revisit only with computed-style parity proof across all routes on the mocked dev server (PR #123156/#123160 show the evidence pattern).

Gateway Coupling

  • The Control UI ships from and with its Gateway: one install, one version (product decision, 2026-08-16). UI code never carries gateway-version compatibility — no fallbacks to older methods when a current core method is missing, no version-conditional behavior for older gateways.
  • Method-advertisement checks (isGatewayMethodAdvertised) remain only as feature gates for config/plugin-dependent surfaces, never as version compat.
  • The handshake rejects gateway-served same-origin skew. The admission-exempt paths (pnpm ui:dev, custom gateway.controlUi.root, cross-origin/connection-settings dialing) are unsupported for version mismatch without enforcement: they carry no compat code and fail visibly at the first missing method, by design. Tightening admission to reject them at connect is a server-side product change owned separately.

Build Chunking

  • ui/config/control-ui-boot-modules.json is generated from ready /new and /chat captures. Each route records only fetched modules reachable from the HTML entry and the dynamic entries it requested (static imports, plus dynamic imports that resolve into an already-fetched chunk without a request), so chat-only code co-located in a fetched common chunk stays out of New Session. Shared modules and each route's exclusive modules get separate control-ui-boot-* groups in ui/config/control-ui-chunking.ts, reducing requests without pulling chat-only code into New Session. Its entries record contains the dynamic entry points actually requested by each route; control-ui-boot-preloads.ts follows their static dependencies to emit inert route preload templates, which the Gateway activates for the requested route. CSS hints preload bytes without changing stylesheet insertion order. Regenerate with pnpm ui:boot-manifest:gen when boot-path surfaces change materially; it builds into a temporary directory with all measured boot groups disabled and inactive preload templates so stale entries cannot feed back into the capture. Rebuild with pnpm ui:build afterward to verify grouped output. Do not hand-edit the manifest.

Live Verification

  • The Gateway serves the prebuilt bundle from dist/control-ui; editing ui/src changes nothing live until pnpm ui:build. Confirm the served /assets/index-*.js hash changed before trusting a live result.

Scope

  • Keep UI-specific rules here.
  • Leave repo-global architecture, verification, and git workflow rules in the root AGENTS.md.

Visual Proof

  • For substantial UI design changes, follow the Control UI E2E skill's UI stress test to review states in an HTML gallery and collect feedback per example.
  • Visual proofs never include the Discord invitation card: the mock and E2E harness seed its canonical browser dismissal before rendering. Dedicated invitation behavior tests may opt into a fresh visitor with communityInviteDismissed: false, but do not capture invitation screenshots or videos.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.