Warm-Hive PR6b — Build + Review + Live Smoke (2026-06-18)

Verdict: PASS. PR6b reskins + feature-fills 5 shipped apps (Launcher · Storage/Files · Power surfaces · Evolution · App-sextet) in place. Built by 5 parallel agents, adversarially reviewed, and live-smoked. Review: 0 real HIGH (the 7 "HIGH" were PASS confirmations of the no-fabrication/regression contract), 1 actionable design-fidelity fix, both that + a live-smoke finding fixed + re-verified. Gates green. 0 PR6b-attributable console errors. Ready for commit.

Scope (PR6 sub-PR b — the bulk; reskin-in-place, no rebuild)

Screen	Change	Files
B1 Launcher (06)	Variation B (memory-sharing explainer) behind an A/B toggle; Variation A unchanged	LauncherApp.tsx (+test)
B2 Storage (07)	Variation A ("Where it lives": on-disk tree + storage-type cards) via StorageAndFilesApp A/B wrapper; FilesApp Source column (D11)	StorageApp, StorageAndFilesApp (new, +test), FilesApp, FilesRoute
B3 Power (08)	Shared warm primitives + per-model spend bars + daily-budget (Usage), Vault Renew, color-stratified risk badges (Approvals). D8: no route consolidation	power/power-primitives (new, +test), TelemetryApp, VaultApp, ApprovalsApp, AutomationCenterApp
B4 Evolution (12)	"By skill" version-ladder over real run data; D13-gated provenance badge	memory/EvolutionTab.tsx (+test)
B5 Sextet (16)	Room two-col, Artifacts 3-col grid (per-kind tints), Profile "what Waggle knows" (D20), Agents swarm CTA; Mission Control = CockpitApp warm sweep	RoomApp (+test), ArtifactCenterApp, UserProfileApp (+test), AgentsApp, CockpitApp

Build method

5 parallel build agents on disjoint file-sets (surgical/preserve-behavior + no-fabrication mandates); main-agent integration of the shared sweep. The concurrent agents shared one working tree → transient cross-agent tsc "errors" during the run resolved clean at the end.

Gates (after fixes)

Gate	Result
`tsc -p apps/web/tsconfig.app.json`	0 errors
FE vitest (full)	1130 passing (2 PR5 settings tests are load-flaky — pass 3/3 in isolation in ~1s; they render the heavy SettingsApp and time out only under full-suite concurrency. Not PR6b-related.)
eslint (changed files)	0 errors

No-fabrication — all six traps defused at build time (review-confirmed)

B1 provenance chip prefixed "example ·" (not a real recall).
B2 (D11) Source column = "—" — grounded: FileEntry (core schema) carries no creator field anywhere. memoryCount gated to "—" when absent (no fake "142 memories").
B3 (D9/D13-class) per-model bars from real /api/cost/summary.byModel (absent → gated); budget from /api/settings; Vault "Renew" reuses the real upsert (no rotate endpoint exists — not invented); Approvals risk derived from tool name (no risk field in payload, not fabricated).
B4 (D13) absolute quality % gated off (only delta_accuracy exists → honest +N.Npp); "GEPA · judged by 3 models" never rendered (no judge roster stored).
B5 (D20) Profile facts only from real stored fields (empty → section hidden); Room participants from live SSE only (design's fake "Jonas P · invited" omitted); Artifact source = coarse real class (not invented "made by Claude Code").

Two findings fixed (review + live smoke each caught one the other missed)

Mission Control was edited on dead code (live smoke). /settings/mission-control renders CockpitApp — MissionControlApp.tsx is explicitly KILLED (rendered nowhere). B5 had relabeled the dead component (invisible + a §3.3 dead-code edit). Fix: reverted the dead-file edit; applied the D21 warm-token sweep to the live CockpitApp (8 raw emerald/amber/green → --healthy/--attention). The "Services-lead tab relabel" doesn't apply — CockpitApp is a single-scroll dashboard, no fleet/team/activity tabs (recon mis-mapping, noted).
Artifact icons lost their per-kind palette (review). B5 flattened all card icons to uniform honey-wash; design §16 uses per-kind tints. Fix: added a KIND_TINT map (9 kinds → warm work/healthy/intel/honey washes — design-faithful and D21-compliant) + applied. tsc + full vitest green after.

Live smoke (real sidecar :3333, Vite :8080) — 0 console errors on every surface

B1 /launcher — A/B toggle; Variation A (live detect/launch preserved) + Variation B (3-node flow + 3 cards + "example ·" provenance). (01)
B2 /files — A/B toggle; "Where it lives" (real tree: palindrome.py 222 B, dirs "—") + Files browser (storage-type tabs preserved + Source column "—"). (02,03)
B3 /settings/usage — Daily Budget editor + real Tool Usage breakdown; per-model bars gated off (no spend). (04)
B4 /memory Evolution — "By skill / By run" toggle (existing run-review preserved). (05)
B5 /room (empty state preserved, no fake participants); /settings/profile "What Waggle knows" facts from real data (06); /artifacts 3-col grid (data fetch timed out under sidecar load — env, not code; per-kind tints code-verified) (07); Mission Control = CockpitApp (warm).

Open / not blocking

/artifacts data fetch timed out (10s) under sidecar load during the smoke — environment, not PR6b; the error state + Retry render correctly. Per-kind tints verified in code + tsc/vitest.
The 2 load-flaky PR5 settings tests (pass in isolation) — pre-existing harness flakiness, documented.
Next: PR6c (04 All-workspaces + 15 Habit).

5.5 KiB Raw Blame History