fix(server): stop OpenCode child sessions

[t3dotgg]

OpenCode runs each subagent in a separate child session. T3 Code only stopped the parent session, so child model requests could continue after Stop.

This change walks the parent session's full child tree and stops each descendant. It keeps Stop pending until child cleanup ends, so a parent abort event cannot release the next prompt early. Provider session teardown uses the same cleanup as a best-effort step.

Tests cover nested and late child sessions, unrelated-session isolation, repeated Stop requests, child failures, prompt gating, configured-server teardown, and the global SDK request limit.

Verified with:

• 83 OpenCode adapter tests
• 4 shared interrupt tests
• Server typecheck
• Targeted lint and format checks

Closes #7769

Made with GPT-5.6 Sol in Codex.

---

Note

Medium Risk
Changes interrupt completion ordering and timing (waits on full child tree, up to 10s timeouts), which affects prompt gating and race behavior around session.error during stop.

Overview
Stop and session teardown now walk the full OpenCode child session tree (via session.children) and abort every descendant, not only the parent—so subagent work cannot keep running after the user stops a turn.

interruptTurn aborts the parent first, races provider abort acknowledgment against turn completion, then runs descendant cleanup before it finishes. Prompts stay blocked until that cleanup completes (including sessions that appear while the parent is aborting). Unrelated sessions are not touched. A failed child list/abort fails the interrupt with ProviderAdapterRequestError. Concurrency for list/abort SDK calls is capped at 8.

Cancellation handling adds an acknowledgment deferred so a parent MessageAbortedError does not complete the interrupt before child sessions are stopped; real provider errors can still settle the turn while cleanup is in flight.

Teardown aborts the parent, then best-effort descendant abort (failures ignored). User docs describe stop semantics and error vs best-effort teardown.

^{Reviewed by Cursor Bugbot for commit c463930. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Stop OpenCode child sessions during interrupt and teardown

• Adds abortOpenCodeDescendants util that walks the session tree via session.children, issuing session.abort for each descendant with a concurrency cap of 8 and surfacing the first failure
• Reworks makeOpenCodeAdapter.interruptTurn into two phases: phase 1 aborts the parent and races provider acknowledgment against completion (10s timeout); phase 2 cleans up all descendants (10s timeout). Next prompts are deferred until cleanup completes
• Extends OpenCodeCancellation with an acknowledgment Deferred and turnSettled flag so the adapter can coordinate on explicit provider acknowledgment before proceeding
• Updates abortOpenCodeSessionForTeardown to best-effort abort discovered children after the parent (1s timeout each)
• Behavioral Change: interruptTurn now fails the interrupt if any descendant abort fails (after parent acknowledgment); previously only the parent session was aborted. Reviewers should check the failure-preference logic in interruptTurn and the abortOpenCodeDescendants error collection in OpenCodeAdapter.ts

^{Macroscope summarized c463930.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.3 KiB	0 B (0.0%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	0 B (0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	0 B (0.0%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.6 KiB	55.6 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.3 KiB	+12 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+3 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+9 B (+0.1%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	56.4 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: ce71c04 · PR result: c463930 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.4 KiB
• Claude decoded thread snapshot: 110.1 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — The adapter now performs recursive, concurrent cleanup of OpenCode child sessions and holds cancellation/next-prompt progress until that cleanup completes, with new timeout and failure behavior. This is a substantial change to an existing production request path rather than a small isolated fix.

^{You can add or adjust custom eligibility rules. Learn more.}