Migrate mobile specs to the resilient AI interceptor (and modernise the folder)

Status summary#

Done. Scope grew from the original two-spec migration into a full modernisation of specs/mobile/ (one commit per file): fixture signup everywhere, scripted turns instead of silent default-model LLM calls, inlined helpers, terse comments, plus the requests notification-payload product change so rows are findable by their held operations. Every spec in the folder passes locally, including notes/note-composer once main's #2567 fixed the fleet-wide Artifacts 403s. PR #2563 marked ready.

Motivation#

specs/mobile/approvals.spec.ts and specs/mobile/notifications.spec.ts predate createAgentHelper (they landed 2026-07-29/31; the helper landed 2026-08-28 in PR #2543). Each hand-rolls the ritual the helper now owns:

• agent.create() + the agent/configured append pointing the thread's agent at intercepted/driver with a dropped debounce
• a raw itx.ai.intercept(handler) on the spec's main admin session

Two problems with the raw intercept:

1. Not churn-proof. A Durable Object restart mid-spec (routine on cold preview deployments — an environment both specs' comments already fight with) closes the session with 4901 and nothing reinstalls; the next scripted turn dies with No AI interceptor installed.
2. Shared fate. The interception rides the spec's main admin session, so either one dying takes the other down. createAgentHelper installs via installResilientAiInterceptor on a dedicated connection that owns the reconnect-and-reinstall loop.

specs/mobile/live-status.spec.ts already uses the helper (via createMobileFixture) and proves the shape works for mobile specs.

Checklist#

• approvals.spec.ts: replace agent birth + config append + raw itx.ai.intercept with createAgentHelper + createAgent({ path: agentPath }) + responses.set(handler) commit f6c91baf6; the warm-up runScript now runs after createAgent instead of concurrent with the config append
• approvals.spec.ts: update the journal assertion + comments from intercepted/driver to the helper's intercepted/typed same commit
• notifications.spec.ts: same swap for the watched-thread lane (the orphan and elsewhere agents run scripts directly, no turns — untouched) same commit
• both: keep the content-routing handlers (route by last user message) inside responses.set(fn) — the queue's fingerprint replay also gives retried attempts the same script for free handlers moved verbatim into responses.set
• run both specs locally; typecheck/lint/format/knip all green; see log

Decisions (made while fleshing out)#

• Minimal scope: both specs keep their own signup flow and their own connectItxReady admin session. createMobileFixture isn't adopted here — it doesn't expose itx, and both specs lean on it heavily (egress rules, device enrollment, stream waits). Adopting it would mean extending its API and rewriting the popup flow in the suite's two most timing-sensitive specs; not worth coupling to this fix.
• Warm-up run ordering (approvals): the throwaway runScript that pays the cold isolate start currently races the config append; after the swap it starts just after createAgent resolves (create + append now live inside the helper). The ~100ms shift is immaterial to what the warm-up is for.
• Model name: the helper configures intercepted/typed; the specs' journal assertions and prose move off intercepted/driver. The guarantee asserted is unchanged — a model only this process's handler can serve.

Follow-up: approvals simplification pass (requested after the migration)#

• drop test.setTimeout(120_000) — it LOWERED the 240s default (SPEC_TEST_TIMEOUT_MS); the "two REAL agent turns" comment existed only to justify it both deleted; the turns are still real agent turns on a scripted model, nothing changed there
• sign up via helpers.createMobileFixture — deletes the hand-rolled popup flow, connectItxReady, resolveOsBaseUrl, and the try/finally (the echo tunnel is await using now) fixture gained an itx handle; its signup email prefix now derives from slugPrefix (was hardcoded mobile-live-status) and its popup waits carry approvals' explicit timeouts
• remove the journal waits and timing hacks: the 6s rules-cache outwait (post-append chain is always > the 5s staleness bound in project-durable-object.ts #egressRules), the cold-isolate warm-up run, awaitAgentScriptStart, the settle expect.poll (replaced by waiting for each script's narrated outcome in the feed), the outcome-count expect.poll (single journal READ after the UI shows both), and the pre-Notifications root/device waitForEvent sync spinner-waiter + turnPending cover the whole turn; verified twice locally at ~19s

Observed, out of scope#

• createMobileFixture.signUpToProject hardcodes uniqueSignupEmail("mobile-live-status") for every caller — should be derived from slugPrefix. Trivial, but touching the fixture used by other green specs doesn't belong in this diff.

Implementation log#

• typecheck / lint / knip / format all green in the worktree (lint re-run after oxfmt reflow).
• Local spec runs needed three attempts, none related to this diff:
  1. pnpm spec --project mobile <files> — playwright's --project is variadic and swallowed the file args; use --project=mobile.
  2. "Process from config.webServer exited early", twice: the dev.ts start --detach webServer entry exits immediately when the dev server is already booting/restarting but not yet health-checkable (second occurrence was the dev server's own post-spec-run auto-restart, pid changed under the run). Retry once the server answers /api/health.
• Green runs: approvals passed in the paired run (1 passed, notifications in that run failed at the sign-in popup before any migrated code — cold auth-worker latency); notifications alone passed 35.6s.