P5 + P7 — Live Browser Smoke Record (2026-06-12)

Real-browser verification of the launch-critical P5 (skill governance) + P7/D15 (error states + risk taxonomy) work, run against a live local stack: sidecar (Fastify, :3333, anthropic-proxy provider) + vite dev server (:8080) + Playwright. Goal: confirm no real-browser gap that the unit/integration suites couldn't catch.

Verdict: PASS (core goal met)

App boots clean and every screen changed in P5/P7 renders correctly with 0 console errors attributable to the changes. The P5 provenance→badge chain is verified end-to-end in the browser. One residual (live approval-card trigger) is unit-covered.

What was verified

Surface	Phase	Result
Boot → `/home` (Home Cockpit)	regression	Renders full briefing; 0 console errors
`/approvals`	B1	Empty state ("No pending approvals") on the success path — correct; no error-as-empty, no crash
`/skills` (Skills Hub)	P5/D4iv	Renders; see end-to-end badge check below
`/room`	B2	"idle-empty" — SSE subscribed OK, settled (not stuck connecting, not error); the connecting/error/idle distinction works
`/files`	B4	Populated (real load) — no false cold-load error; `loadedPath` gating correct
`/settings/events`	B5	Renders; no error boundary, no false "couldn't load"
`/memory`	B5	Two-mind Memory Center renders; no error boundary
`/settings`	B5	Renders; no unhandled promise rejection (the getTelemetryStatus `.catch` fix held)

Cumulative: 0 console errors across all 8 screens (all: true).

P5/D4iv provenance badge — verified END-TO-END in the browser

Wrote an agent-stamped skill (initiator: agent, source: chat) directly to ~/.waggle/skills/smoke-agent-skill.md.
GET /api/skills returned { initiator: 'agent', source: 'chat', preview: '<body>' } — provenance present and the preview is the clean body, not the YAML frontmatter (confirms the P5-review fix #3 live).
Reloaded /skills → the skill row rendered the "agent · review" badge (rowHasBadge: true, exactly 1 badge), and did NOT also wear the name-heuristic "custom" label (P5-review fix #5). Cleaned up the smoke skill afterward.

Residual (NOT verified live — unit-covered)

A5/A6 in-chat approval card with the risk badge + Always-allow gating was not triggered live: it needs a real workspace chat + a deterministic LLM tool-call to produce an approval_required event, which is flaky. The attempt hit a routing stub (/workspaces/default/chat 404s /context — default is not a real workspace id, pre-existing, unrelated to P5/P7).
- Coverage that stands in: p7-a5-approval-card-risk.test.tsx (badge renders incl. critical), p7-a6-approval-gating.test.tsx (modal represents critical; Always-allow gated by approvalClass), and the A4 classifyGatedToolRisk unit tests. The FE data path is a type-only widening (useChat casts the SSE payload straight to ApprovalRequest), so the fields the A4 server enrichment sends are already present at runtime — the unit tests exercise the render of exactly that shape.
- To close fully (post-launch nicety): drive a chat in a real workspace that elicits a write_file/bash call and assert the card shows the RiskBadge + that "Always allow" is hidden on a critical op.

Environment notes (for the next smoke)

Sidecar: WAGGLE_SKIP_LITELLM=1 node --env-file=.env node_modules/tsx/dist/cli.mjs packages/server/src/local/start.ts → :3333 (anthropic-proxy; ioredis ECONNREFUSED on :6381 is non-fatal noise; Redis not required).
Vite: npm run dev → :8080, proxies /api,/health,/ws → 3333.
build:packages must be green first (hive-mind-core exports only dist/).
Stop servers by port via PowerShell Stop-Process (PID from netstat -ano | grep ":<port> ").
.playwright-mcp/ artifacts are gitignored.

3.9 KiB Raw Blame History