Apply later-turn provider environment to Pi and ACP - #3299
Merged
ymichael merged 1 commit intoSep 9, 2026
Merged
Conversation
ymichael
added a commit
that referenced
this pull request
Sep 9, 2026
## Human comments ## What was wrong `plugins/provider-pi/src/bridge/bridge.skill-command.test.ts` coupled a deterministic input-to-Pi-prompt transformation matrix to full bridge process lifecycle. Per-test setup initialized a new isolated bridge harness, each case's first prompt started a fresh thread/session and spawned a fake Pi RPC Node child, and teardown stopped that child; the final case started two sessions. The 11-case file therefore paid for 12 serial child startups even though the behavior under test is a typed input conversion. The packages shard runs 89 package tests with Turbo `--concurrency=4`, while each package starts its own Vitest worker pool. That nested CPU/process oversubscription is the systemic trigger: child construction crossed Vitest's 5 s case ceiling in unrelated [PR #3295 attempt 1](https://github.com/get-bb/bb/actions/runs/34285876225/job/102261224349), [PR #3295 attempt 2](https://github.com/get-bb/bb/actions/runs/34285876225/job/102263472294), and [PR #3299 attempt 2](https://github.com/get-bb/bb/actions/runs/34287966835/job/102269943782). Attempt 1 timed out the no-arguments case; attempts 2 and 3 timed out the two-turn unchanged-input case. PR #3295 subsequently merged a file-wide 30 s timeout override for this test. That masks the repeated-startup cost instead of removing it. Scheduler contention exposes the flake, but the root cause is the avoidable process lifecycle around pure transformation assertions. ## What changed The existing typed input conversion now lives in `turn-input.ts`, and `bridge.ts` calls it for start and steer input. The exact transformation matrix and raw schema-boundary rejection cases run directly in `turn-input.test.ts`. The old process-backed feature test and its newly merged 30 s override are removed. Feature behavior remains covered at the narrowest deterministic boundary: native skill invocation, no-argument invocation, command-position movement, whitespace, text/file/image composition, multiple skills, malformed ranges, text mismatch, provider commands, raw slash text, and invalid protocol inputs. Generic real-process bridge lifecycle and input delivery remain covered by the existing conformance, lifecycle, local-file, framing, and session suites. Keeping another child solely around the skill conversion is redundant: the bridge now directly calls the tested typed function, and the existing integration suites cover delivery of its output to Pi. No timeout, polling deadline, retry budget, wrapper limit, protocol assertion, or production behavior is increased or weakened. This does not change host-daemon wire behavior, so `HOST_DAEMON_PROTOCOL_VERSION` is unchanged. There is no CLI, SDK, configuration, or documentation surface change. ## How you verified The mandatory fresh-main gate ran on enrolled Intel host `host_nwqfteeqz4` at frozen SHA and merge-base `124cf2e8888a396726f3f00271074b6a622bdb42`. After #3295 merged, the final commit was rebased onto `bb26b4fd9a054d50e564642416319711bc4db16a` and deliberately removed its timeout-only workaround. - Unchanged-main focused baseline: 11 passed in 14.05 s of test time; ordinary process cases took about 1.1 s and the two-session case took 2.218 s. - Unchanged-main 32-burner stress: 7 of 11 cases hit the unchanged 5 s ceiling; the first two were reported at 6.175 s and 6.177 s. - An intermediate one-shared-child version was still not robust: one identical 32-burner round passed, but another exceeded the existing suite-hook ceiling before either bridge assertion ran. That proved the feature-specific process was unnecessary residual exposure. - Final identical 32-burner stress with no feature child and no timeout override: all 11 deterministic/protocol cases passed in 89 ms of test time. - Final-tree `pnpm exec turbo run typecheck --filter=bb-plugin-provider-pi --force` — 4 Turbo tasks passed. - Final-tree full provider-pi Turbo test reached all 11 affected cases in 81 ms; a separate existing `bridge.lifecycle.test.ts` local-image process case timed out at 5.262 s while many unrelated process cases ran 2–3× slower. An earlier unloaded full run with the same production extraction passed 28 files plus 1 skipped and 156 tests plus 1 skipped. The separate recurrence is outside this focused fix and is reported as a limitation. - `pnpm exec turbo run build --filter=@bb/host-daemon --force` — 5 Turbo tasks passed. Provider-pi has no package-local build task. - `pnpm exec turbo run test --filter=@bb/plugin-build --force` — 10 files and 137 tests passed (1 skipped), including bundling and importing the built-in provider-pi server and host artifacts. - `pnpm exec oxfmt --check plugins/provider-pi/src/bridge/bridge.ts plugins/provider-pi/src/bridge/turn-input.ts plugins/provider-pi/src/bridge/turn-input.test.ts` - `git diff --check` - Post-run process and temp-directory audits found no surviving test, build, load, fake-Pi, watcher, or worktree process and no related test temp directory. Every stress/load/test/build/watch command used a tracked process group, external deadline, and cleanup trap. The 32-burner stress is intentionally synthetic and models nested worker oversubscription rather than treating contention as the root cause. Hosted CI on final head `ecb35451f9407c9de5c6e9cd29cb41854e42d027` passed every required check, including the packages shard and both package-smoke jobs. > AGENT GENERATED
ymichael
force-pushed
the
bb/fixer-apply-later-turn-provider-environment-to-p-thr_g4d6wqfyk4
branch
from
September 9, 2026 04:06
c084b3a to
79db9d1
Compare
ymichael
deleted the
bb/fixer-apply-later-turn-provider-environment-to-p-thr_g4d6wqfyk4
branch
September 9, 2026 04:10
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Human comments
What was wrong
PR #3035 made plugin-contributed provider environments available on every command, but Pi reconciled only model and reasoning changes while ACP ignored later
turn/startoptions. Their long-lived child processes therefore kept the environment from session creation even after BB accepted and recorded new credentials or endpoints, so a later turn could use the wrong account or fail with stale credentials.What changed
Pi now compares the effective turn environment with its construction environment and uses its existing rollback-safe rebuild and resume path when values differ. ACP now retains its construction parameters, merges current launch and turn environments, and restarts through its existing resume path; it reports context loss only when fallback changes the provider thread id. Identical environments do not trigger replacement, and legacy empty replay payloads keep their existing no-change behavior.
The rebase preserves current main's deterministic Pi skill-input coverage and drops the now-obsolete timeout-only commit for the deleted process-backed test. No public contract, CLI, documentation, or host-daemon wire behavior changed, so no protocol bump is required.
How you verified
pnpm exec turbo run test typecheck --filter=bb-plugin-provider-pi --filter=@bb/provider-bridge-acp --forcepassed all eight Turbo tasks.bb plugin build plugins/provider-piandbb plugin build plugins/provider-acpemitted their packages successfully.git diff --check origin/main...HEADpassed. The three-line fake-Pi fixture addition follows formatter output; the remainder of that legacy fixture is unchanged.No linked issue; this is a follow-up to #3035.