Skip to content

fix(server): deliver child results to running Claude turns - #39528

Merged
juliusmarminge merged 1 commit into
pingdotgg:mainfrom
shethshlok:fix/claude-active-notice-delivery
Nov 20, 2026
Merged

juliusmarminge merged 1 commit into
pingdotgg:mainfrom
shethshlok:fix/claude-active-notice-delivery

Conversation

@shethshlok

Copy link
Copy Markdown
Contributor

Fixes #39441.

Claude completion notices were queued until the parent turn ended. The adapter still advertised that steering interrupts tools, even after #39018 switched steering to Claude's next priority. Update the adapter capability so eligible notices can reach the active turn at a tool boundary.

Reproduction and verification

In two fresh, isolated Linux web environments at 1280 × 800, I started a real Claude Haiku 5.5 parent and delegated to a synthetic Replay child. While the child worked, the parent ran sleep 18; printf 'FOREGROUND_OK\n' in foreground Bash. I compared the visible timeline and the server's run and notice records.

Build What happened
Before The child finished while Bash was running. Run 1 ended, then T3 delivered the result in run 2.
After T3 recorded one notice in run 1. Bash returned FOREGROUND_OK, and Claude acknowledged the child result in the same turn.

The live captures compare db51b6719f11 with that revision plus the fix. The PR branch is based on 194726492b45, where the updated integration test failed on the base adapter and passed with the patch. vp run check:changed passed with 483 tests. The live check used the web client and covered delegated completion. Native clients were outside this pass.

Evidence

The 38-second walkthrough starts with the baseline screenshot, pauses on the patched run while the parent is still working, then plays the key part of the actual recording at 0.55×. It ends on the completed patched timeline.

issue-39441-claude-notice-before-after.mp4
Before and after screenshots

Before · result delivered in a second run

Baseline timeline with two parent runs

After · result delivered in the original run

Patched timeline with one parent run and foreground Bash output

Model used for the patch and PR: GPT-6.2 Sol. Harness: Codex in T3 Code.

@github-actions github-actions Bot added vouch:unvouched Author is not yet trusted in the VOUCHED list. size:XS 0-9 changed lines (additions + deletions). labels Nov 20, 2026
@coderabbitai

coderabbitai Bot commented Nov 20, 2026

Copy link
Copy Markdown
Contributor

Explain Change →

No actionable comments were generated in the recent review.

Walkthrough

Claude’s adapter now sets activeSteeringInterruptsTools to false. The automatic-delivery integration test distinguishes scheduled messages from child-completion and other steering deliveries, and checks their offers and resulting run behavior.

Changes: Claude active-turn notice steering

Layer / File(s) Summary
Adapter capability and delivery regression coverage
apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts, apps/server/src/orchestration-v2/ClaudeAutomaticDelivery.integration.test.ts
The adapter sets activeSteeringInterruptsTools to false. The test checks that scheduled messages queue behind the active turn, while child completion and other steering deliveries receive a second offer at next priority. It also verifies the distinct run and delivery outcomes for scheduled messages and child completion.

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Bug fix · Severity of issue fixed: Medium

Possibly related PRs

  • pingdotgg/t3code#15892: Changes the same Claude capability to route background notices into the queue rather than the active turn.
  • pingdotgg/t3code#39444: Handles held results and unread steers in Claude turns, including steers routed into the active turn.
  • pingdotgg/t3code#35088: Changes Claude steer priority and unread-steer handling, related to the capability updated here.

Suggested labels: bugfix

Suggested reviewers: yash-singh1, juliusmarminge, macroscopeapp

Merge Risk: Minimal · up to 4de75

This change lets Claude parent notices from delegated tasks reach the running turn at a tool boundary instead of waiting until the turn ends. The adapter capability change is small and covered by an updated integration test. No actionable merge-blocking risk was found.

Pre-merge checks: Passed 3 · Inconclusive 1

Failed Failed checks: 1 inconclusive

Check name Status Explanation Resolution
Linked Issues check Inconclusive Issue #39441 has four coding requirements. (1) Steer Claude parent notices into a running turn: the summary says ClaudeAdapterV2.ts changes turns.activeSteeringInterruptsTools from true to false, and … Provide readable diff hunks for apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts and ClaudeAutomaticDelivery.integration.test.ts, or make the blobs available in the checkout. Confirm whether any Codex steering regression test wa…
Passed Passed checks: 3 passed

Check name Status Explanation
Title check Passed The title "fix(server): deliver child results to running Claude turns" describes the main change: the adapter capability for steering interrupts tools is changed so child completion notices reach the …
Description check Passed The description covers the problem (completion notices queued until the parent turn ends, and the stale activeSteeringInterruptsTools capability), the change (flipping the capability in ClaudeAdapterV…
Out of Scope Changes check Passed The change summary lists two files: the ClaudeAdapterV2.ts capability flag (true to false, one line) and the ClaudeAutomaticDelivery.integration.test.ts assertions (+26/-26). Both directly implement i…
Full details: Title check

Explanation

The title "fix(server): deliver child results to running Claude turns" describes the main change: the adapter capability for steering interrupts tools is changed so child completion notices reach the active Claude turn at a tool boundary. It follows the conventional commit format and is specific to the behavior change.

Full details: Description check

Explanation

The description covers the problem (completion notices queued until the parent turn ends, and the stale activeSteeringInterruptsTools capability), the change (flipping the capability in ClaudeAdapterV2), and verification (live before/after reproduction, updated integration test, check:changed run, screenshots and walkthrough). It links issue #39441 under Scope. Minor gaps: the Scope section does not explain why the change is a bug fix rather than a feature beyond the linked issue, and the template asks for the Change section to name the affected components. Verification is detailed and includes what was not checked (native clients).

Full details: Out of Scope Changes check

Explanation

The change summary lists two files: the ClaudeAdapterV2.ts capability flag (true to false, one line) and the ClaudeAutomaticDelivery.integration.test.ts assertions (+26/-26). Both directly implement issue #39441, which concerns Claude steering behavior and the activeSteeringInterruptsTools capability. The test edits support the same objective. No unrelated files or behavior appear in the summary. The summary is the only available evidence for file scope because the diff could not be read.

Full details: Linked Issues check

Explanation

Issue #39441 has four coding requirements. (1) Steer Claude parent notices into a running turn: the summary says ClaudeAdapterV2.ts changes turns.activeSteeringInterruptsTools from true to false, and the PR description reports a live before/after run where the child result arrived in run 1. This supports the requirement, but the reported live check covers only the web client and delegated completion, not command notices through every path. (2) Review whether clearing activeSteeringInterruptsTools is safe given #39018: the PR description states the change follows the #39018 switch to the next priority. The author's review is claimed, not shown in code, and the tool could not read the diff to confirm the delivery-path behavior. (3) Regression coverage for Claude and Codex: ClaudeAutomaticDelivery.integration.test.ts was updated to check next-priority offers for child completion, user steering, and promoted queued messages, and it reportedly fails on base and passes with the patch. No Codex steering coverage change appears in the summary. The repository diff tool returned no content (changed blobs missing from the checkout), so the adapter hunk and test hunks could not be read directly.

Resolution

Provide readable diff hunks for apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts and ClaudeAutomaticDelivery.integration.test.ts, or make the blobs available in the checkout. Confirm whether any Codex steering regression test was added or updated in this PR, or whether existing Codex coverage already asserts next-priority steering at the reviewed head. Confirm that command notices, not only delegated completions, follow the changed path.

Finishing Touches

Generate unit tests (beta)

  • Create a new PR
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts
Recent review info

Run configuration

Commits

Reviewing files that changed from the base of the PR and between 1947264 and 4de75f1.

Files selected for processing: 2

  • apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts
  • apps/server/src/orchestration-v2/ClaudeAutomaticDelivery.integration.test.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 9 remain after this review.

@juliusmarminge juliusmarminge added the takeover:in-progress A takeover lead is working on this PR label Nov 20, 2026
@juliusmarminge

Copy link
Copy Markdown
Member

Evidence

Run a real Claude Haiku parent, delegate an asynchronous child that sleeps 5 seconds and returns CHILD_RESULT_OK, then run foreground Bash sleep 18; printf 'FOREGROUND_OK\n'. Do not poll the child. Both children finished while Bash was running, and Bash completed normally on both sides.

Environment: isolated web dev servers, fresh synthetic data, Chromium 1280×800, light theme, real claude-haiku-5-5. Recordings crop out the private workspace panel.

Before After
194726492b 4de75f17f4
Before: delayed notice After: same-run acknowledgement
Before recording After recording
Original run says NO_CHILD_NOTICE_YET; a second run says CHILD_NOTICE_ACK. Original run prints FOREGROUND_OK and CHILD_NOTICE_ACK; no second delivery run.

Reproduction driver, identical prompt, base runtime transcript, head runtime transcript.

The transcripts confirm child completion inside the foreground command interval on both commits. Excluding the READY setup turn, base uses two parent runs and head uses one.

Not shown:

  • Other eligible automatic notice types were not exercised; the captured proof covers successful delegated-child completion.
  • Desktop and mobile clients were not exercised; this comparison used the real web client and Claude adapter.

@juliusmarminge juliusmarminge added takeover:ready Takeover done; auto-lands when gates pass and removed takeover:in-progress A takeover lead is working on this PR labels Nov 20, 2026
@juliusmarminge
juliusmarminge merged commit be91767 into pingdotgg:main Nov 20, 2026
72 checks passed
@juliusmarminge juliusmarminge removed the takeover:ready Takeover done; auto-lands when gates pass label Nov 20, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

size:XS 0-9 changed lines (additions + deletions). vouch:unvouched Author is not yet trusted in the VOUCHED list.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[Bug]: Claude parents get delegated-task and command notices only after their running turn ends

2 participants