[mindfn]

What

A cherry-pick (git cherry-pick -x) of upstream zts212653#1531, merged as 1789afb026 and approved in review 5324110998 at f6413758c1. It fixes zts212653#1532: on a Host without ChatGPT Pro (F247 personal Chrome), the reply poller checks for the Host Adapter every second, forever, and with debug logging on it writes a [F247] personal Chrome assistant return poll deferred line each time. Landy noticed this in our running instance.

While the adapter answers HOST_UNAVAILABLE, the poller now backs off exponentially: 2s, 4s, 8s, 16s, 32s, then one check every 60s. The backoff is timed on a monotonic clock. It logs one line when it starts backing off and one when it resumes. The first answer restores the one-second cadence.

Files: the poller, a new personal-chrome-assistant-return-poller-backoff.test.ts, and one line of packages/api/package.json that adds the new test to the test:f247-chrome-host-spike lane. The cherry-pick applied cleanly: the poller and the original poller test are byte-identical to upstream's pre-fix versions, and the lane line is unchanged on develop_base.

Verification on this branch (8c7b7014c7, on develop_base 7d57371444)

• New backoff tests: 4/4. Adapter tests: pass.
• Original personal-chrome-assistant-return-poller.test.ts: 5/6, with one pre-existing failure on develop_base that this PR doesn't cause:
  • The failing test is retains a no-Redis restart return without head-blocking newer authorized returns.
  • Its fixture calls MessageStore.append with no sender, and develop_base's MessageStore rejects that: append requires one valid MessageFrom sender identity.
  • The same test fails identically against develop_base's own poller without this commit.
  • This file runs only in the test:f247-chrome-host-spike lane, not public CI, so the failure has gone unnoticed. The fixture needs a from identity; the owner of the MessageFrom change should take it.
• tsc --noEmit (API) is clean. Biome is clean on the changed files.
• pnpm gate --no-rebase (the fork-internal merge criterion) is running; its result will be posted here.

After merge

The running instance picks this up only after the operator pulls develop_base and restarts. Cats don't restart the instance. That restart can also pick up anything else merged into develop_base by then, such as #191.

[宪宪/claude-opus-5-5🐾]

🤖 Generated with Claude Code

[mindfn]

pnpm gate --no-rebase on 8c7b7014c7: fails at step 4/6 (full type check including tests). The failure already exists on develop_base; this PR doesn't cause it.

I ran the gate with REDIS_URL, NODE_ENV and npm_config_production stripped, so it could not reach the running instance's Redis.

• Steps 1–3 pass, including the build (public direct lane, F299 provider census OK).
• Step 4 fails in packages/web:

  src/stores/__tests__/chatStore-delivered-connector-source.test.ts(83,9): error TS2353: Object literal may only specify known properties, and 'mentions' does not exist in type '{ id: string; from?: MessageFrom | undefined; ... }'
  

  • This PR changes no web files (git diff --stat develop_base..HEAD -- packages/web is empty), so the result is the same on develop_base itself. Every fork PR will fail the gate here until the fixture is fixed.
  • The test stages after step 4 did not run.

Both pre-existing breakages trace back to fork-only F117 changes to the message shape:

• The web fixture above still passes mentions. git log -S mentions on the web store types points to 63921725d8 / f8071e44ae.
• The API fixture in personal-chrome-assistant-return-poller.test.ts (5/6, noted in the PR description) appends without a sender. develop_base has required one since f426fb7902 (feat(F117): register the unified message delivery lifecycle).

I've reported both to the F117 owner thread.

This PR's own checks pass: backoff tests 4/4, adapter tests, API tsc --noEmit, Biome on the changed files, and the build. Whether to merge before the F117 fixtures are fixed is the maintainer's call. After a merge, the fix takes effect when the operator restarts the running instance.

[宪宪/claude-opus-5-5🐾]

[mindfn]

Rebased onto develop_base 52b80e1f6c, which now includes #194, the F117 fixture fix. The new head is 8e659d3049. The diff is unchanged: 3 files, +207/−6.

Gate evidence. I ran pnpm gate --no-rebase, with REDIS_URL and NODE_ENV stripped, on a local branch equal to 7d57371444 + this fix + #194's a3dd94f789. That tree is byte-identical to the rebased head (git diff 9ef5ce79fa 8e659d3049 is empty).

• Steps 1–3: pass (dependencies, Biome toolchain, full build).
• Step 4, full type check including tests: pass. This is the step that failed before test(F117): two fixtures follow the F117 message shape (unblocks fork gate Step 4) #194.
• Step 5, public repo test suite: cut off by the managed job's 3-hour wall-clock limit (timed_out, SIGTERM). There were 0 failures in the part that ran.
  • The time limit was hit because this machine keeps suspending (a sleep 20 took 18 minutes of wall time), not because of a hang in the tests.
  • Upstream public CI ran the full suite green on the identical change (fix(f247): back off the personal Chrome reply poll while the Host Adapter is unavailable zts212653/clowder-ai#1531 at f6413758c1).
• Focused lane on the rebased head: 21/21 (adapter, original poller suite now 6/6, backoff 4/4).

Merging is the maintainer's call. After a merge, the fix takes effect once the operator restarts the running instance.

[宪宪/claude-opus-5-5🐾]