The dominant production failure was a dead dev-server process behind a
'running' DB flag (idle-stop / OOM / crash / host restart), which the UI
trusted and embedded -> permanent 502 until a manual restart.
- dev-container.ts: add isDevServerListening() fast liveness probe; stop the
container entrypoint from auto-running 'npx next dev --webpack' (it competed
with the managed server, forced the wrong bundler/cwd, and doubled memory);
drop the fake state='running' seed row; bump dev container memory 1g -> 2g.
- ensure route: verify a 'running' row is ACTUALLY listening and resurrect it
if dead, instead of trusting the flag; never bounce a healthy server.
- preview page: call ensure on every mount and on refresh (verify + heal),
force an immediate anatomy refetch on (re)start so a dead frame swaps to
'warming up' without the 5s lag.
Backstopped by the partial unique index + startDevServer idempotency, so heals
can never duplicate or thrash a server.
- Add partial unique index on (project_id, port) for active dev servers so the
SELECT-then-INSERT race can no longer create duplicate 'Port 3000' rows.
- Make startDevServer race-safe: on unique violation, adopt the winning row
instead of duplicating.
- ensure route no longer marks a server 'running' before it binds the port;
the readiness probe flips starting->running only after the port answers.
Kills the '502 -> broken CSS -> works' refresh loop.
- Deduplicate previews per-port in sortDevPreviewsFrontendFirst as a defensive
backstop for the dropdown.
- Revert iframe _refresh query-param hack (was forcing cold recompiles).
- New POST /api/projects/[id]/dev-server/ensure: checks if dev server is running,
queries last known config from fs_dev_servers, fires startDevServer +
probeDevServerReadiness in background, returns immediately
- Preview pane calls ensure on mount when anatomy is loaded but no server is running
- Distinguishes state='running' (show iframe) from state='starting' (show warming-up UI)
- WarmingUpState: indigo spinner, elapsed timer, 'View last deployed version' link if available
- ensureCalledRef prevents duplicate calls per mount
- The 5s anatomy poll handles the starting→running transition automatically
Stop button fix:
- Plumb AbortSignal end-to-end: callVibnChat → Gemini SDK (config.abortSignal) / OpenAI fetch → executeMcpTool (/api/mcp fetch)
- Treat abort as clean user stop (not fatal error); partial reply persisted with '(stopped by user)'
Classifier fix:
- Add timeout/gateway/5xx/connection-error vocabulary to diagnose intent
- Prevents 'I get a gateway timeout' from falling through to feature_build (40 rounds) and looping
Prompt / agent behaviour:
- Render verification is now scope-aware: small edits stop at green healthCheck; no browser_console/curl audit on healthy server
- Sanitize stale '### Phase Checkpoint' walls from loaded history so old threads stop biasing new turns
- Next.js dev command updated to --no-turbopack for container stability (per-route lazy compile caused cold-start 503s)
- New public page prompt: agent checks middleware allowlist in the same turn
- Scope discipline and QA-tool gating carried forward from prior session
Code cleanup:
- Remove duplicate AgentPhase declaration (TS2440)
- Remove dead checkpoint emit branch and orphan 'checkpoint' phase value
- Remove unused MAX_TOOL_ROUNDS constant
Preview pane (build status):
- 4-state machine: initial-load / building (with elapsed timer) / build-failed / not-running
- pollMs 0 → 5 000ms so dev-server recovery and build completion auto-update without refresh
- anatomy route + use-anatomy type: inFlightBuild gains createdAt for elapsed timer
