Completed subagents can remain pinned and exempt from automatic settlement because makeSubagentChildThread spreads the parent thread without clearing those settings. This affects both provider-native and MCP children and can keep finished children in background PR discovery indefinitely.
Clear pinnedAt, pinOrderKey, and autoSettleDisabledAt when projecting a child. Keep the parent's settings and the child's repository, worktree, and lineage intact. This corrects newly projected children; it does not migrate existing records.
Validation: the two new provider/MCP regression cases fail before the fix and pass afterward. The full required vp run check:changed passes, including 9 tests and affected-package typechecks.
Model: GPT-6. Harness: Codex.
Added 1 commit
Verdict: Needs Human Approval
Macroscope's review found this PR requires human approval — The shared subagent projection now clears pinning and automatic-settlement opt-out state for every newly created child, with regression coverage for provider and MCP paths. This changes the effective default lifecycle behavior of new subagents and warrants human review.
You can add or adjust custom eligibility rules. Learn more.
makeSubagentChildThread now initializes autoSettleDisabledAt to null. Tests cover provider and MCP child creation, verify that the parent retains its timestamp, and check settlement candidate selection.
Changes: Subagent child fields
| Layer / File(s) | Summary |
|---|---|
Clear child auto-settlement protection packages/provider-core/src/server/subagentProjection.ts, apps/server/src/orchestration-v2/SubagentProjection.test.ts | The child projection initializes autoSettleDisabledAt to null. Tests cover both creation sources, verify the child does not inherit the parent timestamp, and check that the persisted parent retains it. |
Priority: External image Normal
Estimated code review effort: 2 (Simple) | ~10 minutes
Change: Bug fix
Possibly related issues
autoSettleDisabledAt value.Possibly related PRs
#21271 both change makeSubagentChildThread to clear child protection fields. This change also clears autoSettleDisabledAt.Suggested labels: bugfix
Suggested reviewers: only21mil, juliusmarminge, t3dotgg
5f80fNo concrete merge-blocking issue is established by the reviewed change.
| Check name | Status | Explanation |
|---|---|---|
| Title check | Passed | The title clearly and concisely describes the main change: preventing subagents from inheriting parent settlement protections. |
| Description check | Passed | The description explains the problem, the change, preserved behavior, migration scope, and verification results. It does not use the template headings or provide explicit scope or approval details, bu… |
| Linked Issues check | Passed | Check skipped because no linked issues were found for this pull request. |
| Out of Scope Changes check | Passed | Check skipped because no linked issues were found for this pull request. |
Explanation
The description explains the problem, the change, preserved behavior, migration scope, and verification results. It does not use the template headings or provide explicit scope or approval details, but it is otherwise sufficiently complete for this focused bug fix.
Note
This review was completed with usage-based billing: files reviewed beyond your plan's included limits are billed at $0.25/file. View usage-based billing.
Comment @coderabbitai help to get the list of available commands.
Wild, my agent was debugging lagging and slow T3 code on a VM, and apparently Sol wanted to contribute this upstream.
I have a few monster threads open on T3, with autosettle off. I don't want my monster threads to ever vanish. Apparently, after ~1k PRs with tons of subagents, none of those sub-worktrees got settled, and it kept on checking for git status constantly. On a GCP VM machine with a pd-disk, the IO is slowish, and this amplified the delays.
Added 1 commit
Note
Written by gpt-6.1-sol on behalf of Maria
thanks @agentic-jarmo. the original correctly identified that child projections inherit the parent's automatic-settlement opt-out.
i preserved your commit and merged current main into this branch. main already resets both pin fields, so the remaining production change is one line: autoSettleDisabledAt: null. i folded the duplicate regression cases into main's existing sql persistence/settlement test and parameterized it for provider and mcp children. parent protections, child lineage and worktree behavior remain intact. existing children are not migrated.
verification at 5f80f106de76dbebeeda9d188011eec0a3d260b3:
vp run check:changed passed: formatting, lint, knip, graphql checks, 23 package typechecks, and 10 tests.unverified: a fresh live provider/subagent spawn through a client, desktop/mobile, and remote/relay connections. this changes hidden server lifecycle state and no client rendering; no screenshot evidence applies. the existing human-approval requirement remains for maria's review.
Note
Written by gpt-6.1-sol on behalf of Maria
verified as-is at 5f80f106de76dbebeeda9d188011eec0a3d260b3. no code changes or pushes were needed.
ran github/main (6575ce3250) in /home/user/main first, then stopped it before starting the exact pr head. installed each checkout with vp i. used the real dev web client, repository Playwright, /usr/bin/google-chrome, and real Codex gpt-6.1-sol turns through the local proxy. created a local git project and parent thread through the app, pinned the parent, and disabled its auto-settle through its action menu. no mocks, fixtures, or direct database writes.
spawn_agent and T3 MCP delegate_task children completed their real git status --short tasks, but inherited the parent's autoSettleDisabledAt. the child menu showed auto-settle disabled.autoSettleDisabledAt, pinnedAt, and pinOrderKey set to null; the child menus showed auto-settle enabled. parent pin/opt-out values stayed intact. parent lineage, model, branch, and current-checkout context remained correct.FOLLOWUP_OK and cleared the manual active override. the parent remained protected.short transcript from read-only inspection of the app-created persisted records:
| 1 | parent opt-out native child opt-out mcp child opt-out |
| 2 | main timestamp same timestamp same timestamp |
| 3 | pr timestamp null null |
| 4 | child pins, both builds null null |
| 5 | real provider results NATIVE_CHILD_OK MCP_CHILD_OK |
| 6 | pr follow-up FOLLOWUP_OK |
blacksmith checks passed: all 43 existing projection/settlement tests and vp run check:changed (format, lint, knip, graphql, 23 affected-package typechecks, 10 tests). fetched the public pr history into the worker to satisfy the changed-check merge base. inspected the captures and checked uploaded video playback and seeking.
main: completed native child incorrectly inherits disabled auto-settle
pr: equivalent completed native child has auto-settle enabled
pr: completed mcp child also has auto-settle enabled
protected parent, enabled children, and manual settle/un-settle in the real client
wall-clock inactivity auto-settlement requires at least one day, so i did not wait for that sweep or fabricate old timestamps. the live proof covers the changed inheritance and persisted settings; the existing settlement tests cover eligibility and expiry. also unverified: native desktop/mobile clients, macos/windows, other providers, isolated child worktrees, and remote/relay/tunnel connections. existing children are not migrated by this pr.