[chinawch007]

Summary

Enables an active Goal using AI SDK direct tools to wait for managed background Bash without repeated work-model or Goal-evaluator calls. The source Turn yields, and persisted terminal facts resume the same Goal through existing Root admission.

• Verify same-Session, same-State-Root Bash provenance with bounded SQLite/Memory queries. A terminal result may indicate failure, timeout or cancellation; the continuation reads it to determine success. Orphaned tasks report uncertainty without replaying commands.
• Add Desktop/CLI waiting, unknown and awaiting-admission states, with diagnostics preserved across reconnects. Cancellation targets the original Goal, generation and wait; it pauses the Goal while the background task keeps running.

Refs #5926 (PR4 implementation plan).
Refs the original discussion #5920.

Verification

• Root build/typecheck, lint, Desktop/UI Knip, renderer architecture and E2E budget checks passed on macOS. Final build uses Node 24.18.0/npm 11.19.0.
• Current-commit Host/storage regression: 410 passed, 1 subprocess-fixture entry skipped. Desktop preload/controller/IPC: 33 passed. Covers real Bash/PTY, cancellation, source revocation, reconnects, polling and SIGKILL recovery with a command-start counter. Defect regressions were verified failing before their fixes.
• Chrome Storybook: 8 light/dark checks, including narrow layout and cancellation tooltip assertions. Captured UI components are unchanged by the later Host/preload fixes.
• Tracked-file formatting passed. Root format:check includes unrelated untracked local investigation scripts and fails on those files; they are excluded from this PR. Full npm test, Windows/Linux runs and Electron E2E were not run.

Protocol compatibility

The new Goal projection and cancellation operation advance Runtime Host compatibility epoch 226→ 227. Snapshot preflight reserves the wait summary within the existing frame limit, reducing the remaining capacity for queued content. No storage schema migration is required.

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: OpenAI Codex contributed implementation, tests, documentation and review assistance. The commit retains Generated-by: OpenAI Codex.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

The second item remains unchecked because of the root format-command limitation above; tracked-file formatting and the other listed checks passed.

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[Astro-Han]

Automated review notice: This comment was posted by an automated review agent. It is not an independent human review and does not replace one.

Reviewed b4bbc67faa5df6fca89fcf76ed9b06996e098cc3 against main dc4515cd6e7637e8267d59cd4f71ad53f403f9e7. I found no confirmed P0–P2 defect.

The core behavior is demonstrated with real managed Bash processes, SQLite, the existing Goal and Root admission components, and the production AI SDK backend with a scripted model. Pending observations do not invoke the work model or Goal evaluator. My additional expiry and terminal scenarios each admitted exactly one continuation and left the command start count at one; expiry left the background process running. Disabling terminal observation made the terminal scenario fail while expiry still worked, so the completion check depends on the source adapter. Original tests also exercised unsuccessful terminal states, PTY, stale cancellation, source revocation and recovery after killing the owning process. Both unmodified SIGKILL recovery cases passed under the repository's specified Node 24.18.1.

Validation included a fresh install, dependency patches and normal clean root build; Desktop, Host and CLI type checks; renderer and Storybook builds; 589 client tests (586 passed, three skipped); the Host/storage group (360 passed, one skipped, with the two Node 22 harness failures separately passing under Node 24); and 121 neighboring tests (116 passed, four skipped, with one unchanged PTY drain test passing on isolated rerun after an initial failure). All three wait states completed their actual Storybook play assertions at 600, 720 and 1280 pixels in light and dark themes, including the cancellation explanation and reachable button. Architecture, surface inventory, headers, localization, TUI copy, formatting and protocol checks passed. The current test and windows_acp CI checks are green.

P3 integration note: this branch, #6176 and #6195 currently all use protocol epoch 226 while main remains 225. The first merged protocol change takes that epoch; later branches need the next unused value and renewed compatibility checks before merging. A separate P3 comment covers local test portability on Node 22.

The tests used a scripted model rather than a paid provider. Native Windows/macOS GUI behavior, an authenticated remote Host and long-running production workloads were not exercised locally. The 360-pixel browser experiment is below the Desktop's enforced 600-pixel minimum and is not used as a product finding.

[Astro-Han]

[P3] Keep the crash harness runnable on the supported Node 22 line

The package still declares ^22.19.0 || >=24.0.0, but the new child process argument --test-isolation=none is rejected by Node 22.23.2 before either crash fixture runs. Both new crash-recovery tests fail locally for that reason; the same unmodified tests pass under Node 24.18.1. CI uses Node 24, so this does not invalidate its green result or establish a production runtime failure. Make the harness compatible with the declared contributor runtime, or state/enforce that these tests require Node 24.

[Astro-Han]

Automated review notice: This comment was posted by an automated review agent. It is not an independent human review and does not replace one.

Reviewed range: dc4515cd..b4bbc67f, i.e. merge-base with main (the current main tip) up to head b4bbc67faa5df6fca89fcf76ed9b06996e098cc3. That is three commits: c5a56cf feat, 3808356 CLI localization, b4bbc67 docs inventory. This is a first review, so there are no prior findings to track.

Verdict: I found no P0–P2 defects. There are four P3 items.

What I checked

• Source authorization (ShellRunEventWaitAccess). A wait is accepted only when all of these hold:

  • the ShellRun belongs to the same Session in the same Host stores, and its visibility is not user;
  • the ShellRun has a source Run whose invocation is a runtime/ai-sdk route on the same Turn;
  • exactly one immutable, non-partial Bash function_call, T1 dispatch and T2 outcome exist, in that order, with matching ids, args hash and operationId, run_in_background: true, and a shell_run result whose ref matches the wait.

  The live record is re-read after the proof reads. Registration additionally rejects code_mode and non-ai-sdk invocations. The wait never runs, restarts or kills a command. Cwd, env and sandbox stay owned by the original Bash dispatch, and expiry or cancellation leaves the task running, as documented.

• Storage. readToolProvenanceEvents has matching predicates in SQLite and Memory: in practice only function_call/function_response carry content.id, and both also match refs.toolCallId and toolDispatch.providerToolCallId. Both stores measure the budget before decoding, and both are covered by the dual-provider test.

• Restart behaviour. A task that was running when the Host died recovers as orphaned, which maps to source_unavailable without a re-dispatch. A task that was already terminal recovers as satisfied. The receipt key is a digest of the terminal fact, so delivery is idempotent through the existing #6147 delivery key.

• Interaction with #6147.

  • pause now also requires the wait's waitId to still be the active gate, so a stale observer cannot pause a newer wait.
  • Admission-time revocation now persists a specific diagnostic.
  • goal.event-wait.cancel is serialized on the Session admission lane plus the Goal mutation lane. It is fenced on goalId, generation, waitId and goal revision, and it reports already_admitted / already_cancelled correctly.
  • The projection hides the wait once it is delivered or cancelled, or the Goal is no longer active.
  • A plugin cannot shadow WaitForEvent, and the tool is hidden unless the Goal is active or waiting and the mode is not code mode.

• Output and prompt bounds. The continuation prompt carries only refs, never task output, and stays within 4000 chars / 12000 bytes after escaping. The projection's resourceId/reasonCode are truncated with truncateUtf8, whose output (marker included) fits the decoder limits. The worst-case snapshot reservation includes the wait summary.

Findings

1. P3 – protocol epoch collision. The epoch goes from 225 to 226. Open PRs #6176 and #6195 also use 226, so whichever merges later must re-bump. The change also adds no // 226: history line above the 225 entry (packages/runtime-host/src/protocol/index.ts:111).
2. P3 – crash-recovery test fails on Node 22.
   • goal-root-authority.test.ts:3326 spawns its fixture with --test-isolation=none. Node 22.23.2 rejects that flag (bad option), and engines.node allows ^22.19.0.
   • Locally, both "ShellRun process crash recovery from {running,terminal}" tests fail on Node 22. They pass when the flag is switched to --experimental-test-isolation=none.
   • CI uses Node 24 and is green. Pick the flag based on process.version, or skip on Node 22.
3. P3 – repeated provenance reads on every poll (performance; inferred from the code, not measured). Each observation calls authorize and then adapter.read, and each of those runs the full ShellRunEventWaitAccess.read. That is per poll: readRunInvocation, two json_extract scans over every row of the source Run, and two readShellRun. The proof is immutable once complete, so it could be verified once per wait and only the live ShellRun binding rechecked afterwards. At a 15 s poll interval this is acceptable, but cost grows with the size of the source Run's history.
4. P3 – misplaced JSDoc. In packages/cli/src/session-driver.ts:249, cancelGoalEventWait was inserted between the controlGoal JSDoc and controlGoal itself, so that doc now attaches to the wrong member.

Tests run (local, Node 22.23.2, built core/storage/mcp/runtime/runtime-host only)

• runtime-host suites (shell-run-event-wait, goal-root-authority, event-wait-manager, goal-protocol, goal-coordinator, interactive-run-composer, canonical-session-projection): 411 pass, 1 skip, 2 fail. The 2 failures are finding 2 and pass with the flag adjusted.
• storage tool-provenance-query: 2/2 pass.
• runtime shell-run-manager: one failure on the first run, then clean on rerun. I attribute this to load from a concurrent build.

CI and mergeability: test passed (22m37s) and windows_acp passed. The PR is mergeable; merge state is BLOCKED only on required review.

[Astro-Han]

P3: Epoch 226 collides with open PRs #6176 and #6195, so whichever merges later must re-bump. Also add a // 226: history line describing this change (the Goal projection eventWait field and the goal.event-wait.cancel operation).

[Astro-Han]

P3: Node 22 (allowed by engines as ^22.19.0) rejects --test-isolation=none with bad option, so both crash-recovery tests fail locally on Node 22.23.2. They pass with --experimental-test-isolation=none. Pick the flag based on process.version.

[Astro-Han]

P3 (performance, inferred from the code, not measured): every poll runs this full proof twice, once from authorize and once from adapter.read. Each run does two json_extract scans over the whole source Run. The proof is immutable once complete, so it could be cached per wait, with only the live ShellRun binding rechecked on later polls.

[Astro-Han]

P3: This member was inserted between the controlGoal JSDoc and controlGoal, so that doc now attaches to cancelGoalEventWait.

[Astro-Han]

Automated review notice: This comment was posted by an automated review agent. It is not an independent human review and does not replace one.

Heads-up on the Runtime Host compatibility epoch: #6176 has merged, so main is now at 226 (RUNTIME_HOST_COMPATIBILITY_EPOCH = 226 on 6104132). This PR changes the Host protocol and currently declares 226, so it needs the next unused value (227 at the time of writing; #6195, #6206, #6091 and #6163 are all in the same position, and whichever merges first takes 227). Please rebase, bump the epoch and add the history comment in packages/runtime-host/src/protocol/index.ts.

[chinawch007]

Thanks for the detailed review. I rebased onto main at 0d19f2423 and pushed the updates in e57275c2d.

1. Protocol epoch: bumped to 227, preserving main's 226 history entry and adding an entry for the Goal eventWait projection and goal.event-wait.cancel. The epoch check against the rebased main passes.
2. Node 22: the crash harness now selects --experimental-test-isolation=none on Node 22 and --test-isolation=none on newer supported versions. Both real SIGKILL recovery cases pass on Node 22.19.0 and 24.18.0, including the assertions that the command is not redispatched.
3. JSDoc: restored the comment to controlGoal and added a separate description for cancelGoalEventWait.
4. Repeated provenance reads: agreed that there is avoidable work. Tracing the normal observation path shows four full checks: initial authorization, dispatch-time authorization, adapter read, and post-read authorization. I have left caching out of this revision because its performance impact has not been measured. Any optimization should cache only immutable proof facts and retain current Goal control, resource visibility, and identity revalidation.

The rebase also preserves main's shared observer/Goal-wrapper split, with strict Goal decoding for projection callbacks.

Validation on macOS after a clean build: 437 Host/storage tests passed, 1 subprocess-fixture entry skipped; 197 Desktop/CLI tests passed. The two Node 22 recovery tests also passed separately. Build, typecheck, lint, Desktop/UI Knip, renderer architecture, and protocol checks passed. format:check passed in a clean snapshot of the tracked files; existing untracked investigation scripts still cause it to fail in the original checkout. The full test suite and cross-platform runs were not rerun.

Prepared with assistance from OpenAI Codex.

[Astro-Han]

Automated review notice: This comment was posted by an automated review agent. It is not an independent human review and does not replace one.

Incremental review at e57275c2dcafa6214de5aa8ab3d963075eea6734, following b4bbc67f: a rebase and one six-file follow-up (+21/-12). I found no new confirmed P0–P2 issue. The feature still waits for the original managed background Bash task and resumes its Goal through existing Root admission, without restarting the command.

I separated the rebase from the implementation changes. Of the 61 PR paths, 51 files are byte-identical to the previously reviewed head, including all Desktop/UI changes. Applying only main's patches to the old composition/prompt files reproduces the three changed files exactly. Before the final follow-up, the added/removed author lines match across 60 of 61 paths; the exception is the protocol epoch/history update.

The follow-up adapts the source adapters to main's shared observer record type and retains strict Goal decoding for onChanged at packages/runtime-host/src/server/event-wait-manager.ts:1478–1504. Authentication and strict decoding remain at the Goal manager's storage, authorization and callback boundaries. The earlier Node 22 test-flag P3 and misplaced CLI JSDoc P3 are fixed. Both original SIGKILL recovery tests now pass on Node 22.23.2 without a loader changing their flags.

Validation on this head:

• Fresh dependencies/patches, the normal clean root test build, Host/CLI type checks, Biome on the six follow-up files, the surface inventory, git diff --check and protocol guard passed.
• 418 Host/storage tests: 417 passed, one fixture skip; 589 Desktop/CLI tests: 586 passed, three skipped. These include source authorization, wait state/projection, cancellation, actual Bash/PTY and crash recovery, plus the client/preload/IPC paths.
• I repeated the actual production-backend/SQLite/Bash/Goal/Root checks for expiry and terminal completion. Each made three work-model calls in total, one Goal evaluation and one continuation admission; the command-start counter stayed at one. Pending waiting made no extra work-model/evaluator call, and expiry left the task running until explicitly released.
• Current main advanced to 27016471f6f74132c7592f2cde744a2be8ca2123 with the finish-hook/step-cap fix. I also loaded the exact runtime turn module from the clean merge result into the same Host fixture and reduced its step cap to two. Both scenarios still passed. Disabling only terminal observation preserved expiry but made terminal completion fail before the wait deadline, confirming the continuation depends on that observation.

P3 — coordinate epoch 227. The guard correctly reports 226 → 227, and the history now retains main's entries. Current open #6091 (efbe29e6), #6192 (07c7a8bb), #6195 (4f9a8762) and #6197 (0bcb095a) also use 227. The first merged protocol change consumes it; later ones need the next unused value and another check. The PR description still says 225 → 226 and should reflect the current 226 → 227 change. The previously noted repeated provenance-read cost is unchanged; I did not benchmark large histories.

Exact-head CI test and Windows ACP pass. The static merge with current main is clean. The additional merged-runtime check above is not a separate full merged-tree build. I did not repeat browser/Storybook or native GUI checks, run the full repository suite, or use authenticated paid providers. The model was scripted; Bash, persistence, admission and continuation were real.