The LQ.AI Atlas LQ.AI's documentation, bound to the code it describes
295 documents
docs/LQVern/HANDOFF-2026-06-19-pr6b-build-next.md

Handoff — 2026-06-19 · WS5 6b (chat tool-loop UI) spec+plan DONE, ready to BUILD · 6a merged

Repo: ~/Code/lq-ai (canonical; NEVER ~/Desktop; Bash cwd resets — prefix every command cd ~/Code/lq-ai && …). main HEAD = 205efd9 (PR6a merged; origin == tucuxi). Migration head 0054, EXPECTED_PATHS 133 (unchanged by 6b — frontend-only). 6b branch = feat/pr6b-chat-tool-loop-ui (off 205efd9, pushed origin + tucuxi). Currently holds two doc commits: the spec (7962491) and the plan (ac1aa82). Next session: BUILD the plan. origin/main is PROTECTED — PR + GitHub merge only; sync tucuxi after. NEVER git push origin main.

Read these memories FIRST: [[project-legal-research-mcp-milestone]], [[project-pr6-transparency-posture-narrative]] (has the WS5 6a→6b→6c→6d decomposition + the D6 forward-consistency rule), [[project-pr6-release-completion-gate]] (the end-of-PR6 release gate), [[feedback-commit-trailer-model]]. This file is the session-specific pointer.

NEXT SESSION STARTS HERE — build PR6b

The plan is complete and committed: docs/superpowers/plans/2026-06-19-pr6b-chat-tool-loop-ui.md (7 tasks; deterministic TS is verbatim; the Svelte component/route chrome is concrete skeletons that mirror RefusalMessageBubble.svelte). Spec: docs/superpowers/specs/2026-06-19-pr6b-chat-tool-loop-ui-design.md.

Recommended execution = INLINE (superpowers:executing-plans) — like 6a, the load-bearing verification is visual (do the gate cards render, do the flows work) + the Vitest units; a subagent can't see the rendered UI. (Kevin chose inline-vs-subagent at the plan handoff; default to inline unless he says otherwise.) Tasks 1–3 are clean Vitest TDD; Tasks 4–5 are the stateful ChatPanel/route wiring (verify via build + visual).

What 6b builds (frontend-only, NOT security-gated → self-merge after CI green): consume PR5b's two terminal SSE events in the chat UI —

  • tool_confirmation_required → an in-message approve/deny card (ToolGatePrompt.svelte, variant confirm) → POST /api/v1/chats/{id}/tool-calls/{pending_call_id} {decision} → resume the same assistant bubble (shared consumeIntoMessage helper handles chained gates).
  • mcp_authorization_required → a connect card → same-tab redirect to authorize_url?return_url=<chat URL> → PR4d returns with ?mcp_connected → a "Continue" button re-sends the last user message.
  • D6 (Task 6, MANDATORY): flip the 6a narrative's availability claims "coming next" → "shipped" (explorer AVAILABILITY block + station 5/7 markers in governed-tool-flow.html, Learn section-17 sentence, README sentence). Do NOT let 6b merge without this — it's the honesty rule Kevin set.

Session-specific gotchas (don't relearn)

  1. Web tests: cd web && npx vitest run <file> and npm run check:lq-ai (the scoped svelte-check; CI gate = svelte-check + Vitest). Web is a pre-built static bundle — docker compose up -d --build web to view a UI change (never docker compose down -v).
  2. Visual verification without a full dev flow: a headless Chromium is cached at /Users/kevinkeller/Library/Caches/ms-playwright/chromium_headless_shell-1223/chrome-headless-shell-mac-arm64/chrome-headless-shell — used in 6a to --screenshot/--dump-dom a static HTML file (--headless --disable-gpu --window-size=1280,860 --screenshot=out.png --virtual-time-budget=2500 file://…). Reuse it to render ToolGatePrompt (both variants) in a tiny static harness for the PR screenshots (Playwright the npm package is NOT installed; only the browser binary).
  3. Anchors already read (in the plan verbatim): parser.ts (terminal return pattern), types.ts frame interfaces (after line 391), ChatPanel.svelte send/consume (~534-621; messagesApi.sendMessageStream + consumeMessageStream; messagesStore, streamingMessageId, sendError, draftAssistantId), api/messages.ts (apiStreamRequest pattern; resumeToolCall goes here), api/client.ts apiStreamRequest (throws errorFor(status) on non-2xx → resume 409/410 = expired), MessageBubble.svelte assistant branch (~130-160; render ToolGatePrompt after content), chats route web/src/routes/lq-ai/chats/+page.svelte (owns the URL for the ?mcp_connected return handler; reuse the project toast convention — likely svelte-sonner).
  4. Commit trailer: git commit -s + Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>. Stage explicitly, never git add -A. Push origin + tucuxi.

After 6b merges (sync tucuxi, then continue WS5)

  • 6c — external-source citations / source-kind (C4) in api/app/citation/ + MessageCitation (migration) + provenance pills + the case-law results panel (moved here from 6b — it needs 6c's structured citation data). Backend + frontend. Per D6, 6c updates the narrative for now-real provenance.
  • 6d — the skills/case-law-research/ SKILL + the C5 frontmatter parser (or docs-only — pin in 6d planning) + retire the OpenWebUI MCP stub (DE-341) (migrate web/backend/open_webui/routers/configs.py + utils/middleware.py off utils/mcp/client.py → gateway path, delete the stub). Per D6, 6d does the final milestone-completion honesty pass over the whole narrative.
  • THEN the release gate ([[project-pr6-release-completion-gate]]): fresh-clone Docker works per README → build+push GHCR images → rebuild the signed/notarized macOS launcher (docs/lq-ai-macos-launcher-playbook.md) → verify for an external user → tag v0.5.0 (current tags reach v0.4.2).

The loop (used all milestone — keep it)

brainstorm (forks via AskUserQuestion w/ recommendation) → spec → writing-plans → build (inline for UI/visual work; subagent-driven for backend TDD; independent review before ship) → run gates yourself (evidence before claims) → ship (commit -s + trailer; push origin + tucuxi; PR vs main; CI; self-merge if frontend/api-only, Kevin merges security-gated; sync tucuxi). DEs → PRD §9.