[chat-widgets] In-chat widget system — kickoff (LifeOps MVP)

[lalalune]

This is the coordination thread for the in-chat widget system workstream of
the LifeOps Personal Assistant MVP (GitHub project 15). Chat is the primary
surface; we are moving functionality from views/plugins into chat, rendered as
widgets that show information and accept input.

Research doc (committed in-repo, read it before touching anything):
packages/docs/ongoing-development/research/02-chat-widget-system.md

Decisions adopted

1. No new registry, no new widget framework. The system exists:
   • Inline widget registry: packages/ui/src/components/chat/widgets/inline-registry.tsx
     (registerInlineWidget — the plugin extension point; plugin-task-coordinator
     already uses it).
   • Typed connector-agnostic protocol: packages/core/src/types/interactions.ts
     • packages/core/src/messaging/interactions/ (parse/serialize/layout/callback).
   • Out-of-band sensitive-request pipeline with a hosted cloud page:
     packages/core/src/sensitive-requests/, packages/cloud/api/v1/sensitive-requests/,
     packages/ui/src/cloud/public-pages/pages/sensitive-requests/.
     MVP work is fixing, testing, and verifying this — not building parallel
     infrastructure. If you find yourself writing a new registry, stop and re-read
     the research doc.
2. Closed widget vocabulary. form, choice, followups, task,
   secret/oauth (out-of-band), [CONFIG:pluginId]. Confirmations = choice.
   Pickers = form field types (date/time being added). Lists = markdown or
   choice. Status cards = the task-card pattern + slot widgets. Plugins
   parameterize these; they do not mint new marker kinds for MVP.
   registerInlineWidget remains the display-renderer escape hatch — a
   registration adds zero prompt text.
3. Context efficiency stays structural. Markers are emitted by action code
   (zero prompt cost); the model-facing guide is a dynamic: true,
   relevance-gated provider excluded from default composition
   (packages/core/src/runtime.ts:4088). Nothing you add may inject
   per-widget prompt text into default composition.
4. Secrets never transit chat as text. In-app → inline secure form
   (SensitiveRequestBlock); connectors → single link-out button to the hosted
   authenticated page. This is doctrine, already implemented — our job is
   proving it on real connectors.
5. Fail fast, no fabricated affordances. Malformed blocks render as plain
   text; a widget must never render a healthy-looking control for a broken
   pipeline (the dead /forms/:id connector link is the standing violation —
   first issue below).
6. UI doctrine applies: minimal UI, no card chrome, orange accent only, no
   blue, uniform top bar, no suggestion chips outside the [FOLLOWUPS] widget.

Issues (most important first)

#	Priority	Title
1	P0	[chat-widgets] Connector [FORM] link-out points at a nonexistent /forms/:id page — suppress the dead URL, use free-text fallback
2	P0	[chat-widgets] Live-LLM scenario coverage for in-chat widget round-trips (FORM submit, CHOICE pick, [CONFIG] emission)
3	P1	[chat-widgets] Add date/time/datetime field types to the interaction FORM vocabulary
4	P1	[chat-widgets] Slim the uiCatalog provider: markers first, GenUI catalog split into its own dynamic provider
5	P1	[chat-widgets] Verify settings-in-chat end to end: [CONFIG:pluginId] card + SETTINGS action with rendered evidence
6	P1	[chat-widgets] Prove the sensitive-request hosted-page flow on a real connector (DM link-out → hosted page → agent sees the result)
7	P2	[chat-widgets] Refresh WIDGET_MATRIX.md to match the code (stale D1 divergence, coverage table)

Sequencing notes: #1 is independent and small — take it first if you're new to
the area. #2 is the heart of the workstream; #3 and #5 build on its harness, so
coordinate if you take them in parallel. #7 should land after #1 so the docs
describe the fixed behavior once, not twice.

How to claim work

1. Comment on the issue: "claiming — <your session/agent id>".
2. Move the card on project 15 to In Progress.
3. Branch fix/<issue#>-<slug> or feat/<issue#>-<slug> off develop; PR back
   to develop. Sync (git fetch origin && git rebase origin/develop) before
   opening/updating the PR; run bun run verify + the relevant test lanes.
4. One issue per PR. If your work reveals an out-of-scope defect, file a new
   issue with evidence instead of widening the PR.
5. If you stall or hand off, comment with state + branch name so the next agent
   can resume. Push early, push often — unpushed work does not exist.

Evidence rules (binding — read PR_EVIDENCE.md)

Evidence is posted inline in the GitHub issue/PR (not committed to
.github/issue-evidence/):

• MP4 for videos (renders inline on GitHub) — a full walkthrough of the flow,
  start to finish, on the real surface (dashboard chat, mobile overlay, real
  Telegram/Discord where the issue touches connectors).
• JPG for screenshots (size) — before AND after, desktop AND mobile; for
  bugs, the BEFORE must show the wrong behavior.
• Real-LLM trajectories for anything touching agent/provider/prompt/marker
  behavior — captured with scenario-runner against a live model, hand-reviewed,
  and the relevant excerpts quoted in the PR.
• Backend structured logs ([ClassName] …) and frontend console+network
  logs proving the real code path fired.
• Domain artifacts — the scheduled-task row, the settings write, the stored
  secret (masked) — inspected and shown, not asserted.
• Each evidence type is attached or explicitly marked N/A — <reason>. Never
  blank. Stale evidence (captured before a develop rebase that changed
  behavior) must be re-captured.

No page/flow a PR touches may ship needs-work/broken in the app visual
audit (bun run --cwd packages/app audit:app) where applicable.

Filed issues

• [chat-widgets] Connector [FORM] link-out points at a nonexistent /forms/:id page — suppress the dead URL, use free-text fallback #14321 — P0 [chat-widgets] Connector [FORM] link-out points at a nonexistent /forms/:id page — suppress the dead URL, use free-text fallback
• [chat-widgets] Live-LLM scenario coverage for in-chat widget round-trips (FORM submit, CHOICE pick, [CONFIG] emission) #14322 — P0 [chat-widgets] Live-LLM scenario coverage for in-chat widget round-trips (FORM submit, CHOICE pick, [CONFIG] emission)
• [chat-widgets] Add date/time/datetime field types to the interaction FORM vocabulary #14323 — P1 [chat-widgets] Add date/time/datetime field types to the interaction FORM vocabulary
• [chat-widgets] Slim the uiCatalog provider: markers first, GenUI catalog split into its own dynamic provider #14324 — P1 [chat-widgets] Slim the uiCatalog provider: markers first, GenUI catalog split into its own dynamic provider
• [chat-widgets] Verify settings-in-chat end to end: [CONFIG:pluginId] card + SETTINGS action with rendered evidence #14325 — P1 [chat-widgets] Verify settings-in-chat end to end: [CONFIG:pluginId] card + SETTINGS action with rendered evidence
• [chat-widgets] Prove the sensitive-request hosted-page flow on a real connector (DM link-out → hosted page → agent sees the result) #14326 — P1 [chat-widgets] Prove the sensitive-request hosted-page flow on a real connector (DM link-out → hosted page → agent sees the result)
• [chat-widgets] Refresh WIDGET_MATRIX.md to match the code (stale D1 divergence, coverage table) #14327 — P2 [chat-widgets] Refresh WIDGET_MATRIX.md to match the code (stale D1 divergence, coverage table)

---

Design doc · MVP board