fix(opencode): a context overflow that OpenCode compacts no longer fails the turn#26743

open3

main

fix/issue-26674

View on GitHubTwo-way sync enabledPR last imported 5d ago

Problem

When an OpenCode 1.x provider request exceeds the model's context window, OpenCode publishes session.error with ContextOverflowError, stays busy, auto-compacts and keeps working. T3 ended the turn as failed on that first error, so a delegated task reported a failure to its parent while the OpenCode session went on editing the worktree with nobody watching it.

Reproduced against a real opencode serve 1.18.18 and a fake OpenAI-compatible model that answers one request with the reporter's context_length_exceeded error: OpenCode emits session.error ContextOverflowError → compaction summary → session.compacted → another agent step → idle. T3's adapter finalized the turn at the first event.

Change

In the OpenCode 1.x adapter, a ContextOverflowError for a session with an active turn no longer finalizes that turn. The adapter keeps the failure on the turn, and session.compacted clears it. When the session goes idle, the turn completes normally if compaction succeeded. If compaction is off or itself overflowed, the turn fails with the same provider failure as before. Interrupts and broken event streams keep their existing outcomes. Every other session.error is unchanged. The OpenCode 2 adapter is not touched.

Scope and approval

Fixes pingdotgg/t3code#26674

Triaged via the automated issue flow; GitCafe issue #26674.

Verification

vp test run packages/provider-opencode/src/server/adapter.test.ts: 36 passed on the head.
- keeps a turn running through a context overflow that OpenCode auto-compacts: fails on main (expected 'failed' to equal 'completed'), passes with the fix.
- fails a turn on a context overflow that OpenCode does not compact: guard for auto-compaction off; passes on both.
vp run typecheck in packages/provider-opencode: clean. vp format and vp lint on the changed files: pass.
Not finished locally: the knip, cross-package typecheck and test stages of vp run check:changed. They did not finish within 30 minutes on a heavily loaded host and were stopped. CI runs them.

	Before (main `9826dbb990`)	After (`dd3b029f53`)
Overflow, then compaction, then idle	turn `failed` at the overflow; later output orphaned	turn `completed` after idle, post-compaction output in the turn
Overflow, then idle (no compaction)	turn `failed`	turn `failed`, same failure

This only changes how the server ends a turn, so there are no screenshots. Runtime evidence, synthetic data only:

opencode-1.18.18-overflow-transcript.txt: OpenCode's event timeline for the overflow.
Driver: drive-26674.sh, fake_model-26674.py, timeline-26674.py, opencode-26674.json.

Not checked: a full delegated task end to end in the T3 app with this model. The dev server's model picker did not list the fake provider.

Claude Opus 5.5 (Claude Code, run from T3 Code); reviews by Codex gpt-6.1-sol and Claude Opus 5.5.

Takeover

Taken over from @juliusmarminge (evidence-only round). No code changes; evidence and verification added below. The earlier evidence showed only OpenCode's own event timeline; the captures below record T3's turn outcome before and after, against a real opencode serve.

Evidence

Beyond root recovery, overflow settlement timing, failure precedence and native-child outcomes change as listed in what-else-changed-v2.md. Not covered by the title: with an overflow pending, an explicit /compact whose summary also overflows now emits a completed "Context compacted" item and then fails the turn with the pending overflow (R11). Before, the turn failed at the overflow with no compaction item. OpenCode 1.18.18 returns summarize success without session.compacted in that case; the proof audit reproduced it against real OpenCode with controlled event delivery (audit2-report.md).

Fixture matrix: proof-matrix-26743-pass1.md. Server-only change: captures are ordered NDJSON runtime transcripts and normalized outcome diffs; viewport and theme are N/A.

15/15 matrix rows captured; 10 selected live runs and 26 synthetic fixtures per SHA (31 paired fixtures). No missing matrix rows.

OpenCode 1.18.18, Node v26.12.1, Linux fs-14; pinned before e34067edef, after dd3b029f53. LIVE = production adapter/real SDK/external real opencode serve + fake fake/fake HTTP model. UNIT = synthetic native-event harness.

Matrix row / harness	Before e34067edef	After dd3b029f53	Normalized diff
R1 LIVE root recovery	R1-before.ndjson	R1-after.ndjson	R1-outcome.diff
R2 LIVE compaction disabled	R2-before.ndjson	R2-after.ndjson	R2-outcome.diff
R3 LIVE summary overflow	R3-before.ndjson	R3-after.ndjson	R3-outcome.diff
R4 LIVE stop during recovery	R4-before.ndjson	R4-after.ndjson	R4-outcome.diff
R5 UNIT EOF after overflow	R5-before.ndjson	R5-after.ndjson	R5-outcome.diff
R6 UNIT known inactive session	R6-before.ndjson	R6-after.ndjson	R6-outcome.diff
R7 UNIT foreign/related isolation	R7-before.ndjson	R7-after.ndjson	R7-outcome.diff
R8 UNIT missing session ID	R8-before.ndjson	R8-after.ndjson	R8-outcome.diff
R9 LIVE APIError + UNIT error guards	R9-before.ndjson	R9-after.ndjson	R9-outcome.diff
R10 UNIT native child recovery/failure	R10-before.ndjson	R10-after.ndjson	R10-outcome.diff
R11 UNIT explicit compact	R11-before.ndjson	R11-after.ndjson	R11-outcome.diff
R12 UNIT healthy/next-turn controls	R12-before.ndjson	R12-after.ndjson	R12-outcome.diff
R13 UNIT replace/clear orderings	R13-before.ndjson	R13-after.ndjson	R13-outcome.diff
R14 UNIT admission/legacy idle	R14-before.ndjson	R14-after.ndjson	R14-outcome.diff
R15 UNIT finalizer precedence/race	R15-before.ndjson	R15-after.ndjson	R15-outcome.diff

Original adapter tests: before 34 passed, after 36 passed. Head's two new tests against before source: 1 failed (recovery: failed ≠ completed), 1 passed (uncompacted overflow). Guard capture cases complete 26/26 per SHA; separate transcript audit 202/202.

LIVE logs include native SSE through idle, model request logs, full emitted T3 outcomes, and adapter snapshots at native error/compaction/idle boundaries. Queued event-observer timing alone is not used to infer emission timing.

Limits: no orchestration persistence, parent delegate notification, client rendering, external-server ownership cleanup, or OpenCode 2 claim. Viewport/theme N/A; synthetic data only.

Drivers, every generated per-run config, raw SSE/model logs and preliminary runs: raw-evidence-and-drivers.tar.gz. environment-and-limits.md.

Unverified required rows: none. Guard rows use labelled UNIT recordings; live native-child and EOF routes are outside this evidence.

Verification log

Mode: evidence-only, round 1. No code changes, no review rounds this round.
Proof audit pass 1: 15-row matrix (R1–R4, R9 live and paired; the rest labelled unit guards).
Capture: 15/15 rows at both SHAs; 10 live runs, 26 synthetic guard fixtures per SHA.
Proof audit pass 2: all 30 transcripts and 15 diffs read; SHA headers match; diffs recompute exactly; independent R1 rebuild at dd3b029f53 gives zero normalized delta (21/21 outcome records). R1–R13 and R15 sufficient; the R14 inventory omitted the retained provider-assigned assistant item and failureItemOrdinal 101→102, corrected in what-else-changed-v2.md. R11 reproduced live: supplemental proof.
Tests: adapter.test.ts 34 passed at e34067edef, 36 passed at dd3b029f53; the recovery regression test fails on e34067edef.
CI on dd3b029f53: 40/40 checks green, mergeable.
Unverified surfaces: none (server only).
Proof audit and capture: Codex gpt-6.1-sol (high) via T3 Code; lead: Claude Opus 5.5 (Claude Code, run from T3 Code).

🤖 Generated with Claude Code

Checks

23 successful · 17 skipped

All checks passed

Added 3 commits6d ago

test(opencode): turns survive context overflows that OpenCode auto-compacts (red)6370fdc
Merge remote-tracking branch 'origin/main' into fix/issue-26674dd3b029
fix(opencode): a context overflow that OpenCode compacts no longer fails the turnf7c3dc1

MmacroscopeappBotapproved these changes6d ago

MmacroscopeappBot6d ago

Approvability

Verdict: Approved at dd3b029

Macroscope's review found this PR approvable — OpenCode turns now remain active through recoverable context overflows, while overflows that are not compacted still fail. The change is limited to turn lifecycle reporting and does not enable new provider actions or change product defaults.

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

Ggithub-actionsBot6d ago

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	5.0 KiB	5.0 KiB	−6 B (−0.1%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.3 KiB	1.2 KiB	−6 B (−0.5%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	21.6 KiB	21.6 KiB	0 B (0.0%)	24.4 KiB	✅
Codex	Live turn messages	2	2	0 (0.0%)	8	✅
Claude	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.3 KiB	1.3 KiB	0 B (0.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	21.9 KiB	21.9 KiB	0 B (0.0%)	24.4 KiB	✅
Claude	Live turn messages	2	2	0 (0.0%)	8	✅

Baseline: 29f33af · PR result: dd3b029 · 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: 108.5 KiB
Claude decoded thread snapshot: 108.8 KiB

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

CcoderabbitaiBot6d ago

Review in Change Stack →

No actionable comments were generated in the recent review.

External image Recent review info

External image Run configuration

Source mode: Auto
Repository: pingdotgg/t3code
Branch: fix/issue-26674
File: .coderabbit.config.ts
Commit: dd3b029
Layers: Repository TS config
Review profile: CHILL
Plan: Team
Run ID: 3c33cc5d-8230-450e-a328-239a2224cfbd

External image Commits

Reviewing files that changed from the base of the PR and between 29f33af7d7e81ced8d0c3098ea0fe0ff1895d6a7 and dd3b029f5325ae8e67ca89f37c426bdbe92449e6.

External image Files selected for processing: 2

packages/provider-opencode/src/server/adapter.test.ts
packages/provider-opencode/src/server/adapter.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 3 remain after this review.

Walkthrough

The OpenCode adapter now defers a context-overflow failure while the session may compact. Compaction clears the pending failure; if the turn reaches idle without compaction, finalization marks it failed. Tests cover both outcomes.

Changes: Context-overflow turn handling

Layer / File(s)	Summary
Record and clear overflow state `packages/provider-opencode/src/server/adapter.ts`	Active turns track a context-overflow failure. The adapter records the error without immediately finalizing the turn, then clears it on `session.compacted`.
Finalize turns and verify outcomes `packages/provider-opencode/src/server/adapter.ts`, `packages/provider-opencode/src/server/adapter.test.ts`	If overflow remains recorded when a turn completes, finalization marks it failed and updates the session’s `lastError`. Tests cover completion after compaction and failure without compaction.

Priority: External image Normal

Estimated code review effort: 2 (Simple) | ~12 minutes

Change: Bug fix · Severity of issue fixed: Medium

Sequence Diagram(s)

mermaid

1	sequenceDiagram
2	participant OpenCodeSession
3	participant Adapter
4	participant ActiveOpenCodeTurn
5	participant finalizeTurn
6	OpenCodeSession->>Adapter: Send ContextOverflowError
7	Adapter->>ActiveOpenCodeTurn: Record overflow failure
8	alt Session compacts
9	OpenCodeSession->>Adapter: Send session.compacted
10	Adapter->>ActiveOpenCodeTurn: Clear overflow failure
11	OpenCodeSession->>Adapter: Send idle status
12	Adapter->>finalizeTurn: Finalize turn as completed
13	else Session does not compact
14	OpenCodeSession->>Adapter: Send idle status
15	Adapter->>finalizeTurn: Finalize turn as failed with overflow failure
16	end

Possibly related issues

pingdotgg/t3code#26674 — The change defers a context-overflow failure during compaction and allows the session to continue. It does not implement the issue’s alternative of aborting the session.

Possibly related PRs

pingdotgg/t3code#24734 — This PR uses the OpenCode error-message formatting for a recorded context-overflow failure.

Suggested reviewers: t3dotgg

Merge Risk: External image Minimal · up to dd3b0

The adapter waits through overflow recovery and reports failure if the session idles without compaction. No merge-blocking risk was established.

Pre-merge checks: ^Passed 4

Check name	Status	Explanation
Title check	Passed	The title clearly and concisely describes the primary fix: preventing a context overflow from failing an OpenCode turn when OpenCode compacts and continues.
Linked Issues check	Passed	The changes address the coding objective in `#26674`. The OpenCode 1.x adapter records an active turn's `ContextOverflowError` without finalizing the turn, clears that failure on `session.compacted`, an…
Out of Scope Changes check	Passed	The whole-PR diff changes only `packages/provider-opencode/src/server/adapter.ts` and its test file. The adapter logic and tests directly support `#26674`. No unrelated adapter, OpenCode 2, or process-m…
Description check	Passed	The description covers the problem, implementation, scope, approval reference, focused verification, observed results, limitations, and agent attribution. It also documents additional evidence and unv…

Full details: Linked Issues check

Explanation

The changes address the coding objective in #26674. The OpenCode 1.x adapter records an active turn's ContextOverflowError without finalizing the turn, clears that failure on session.compacted, and fails the turn with the recorded provider failure only when the turn reaches completion without compaction. The new tests cover both auto-compaction followed by completion and overflow followed by idle without compaction. The issue's abort-and-stop alternative is not required because the PR implements the first expected outcome.

Full details: Out of Scope Changes check

Explanation

The whole-PR diff changes only packages/provider-opencode/src/server/adapter.ts and its test file. The adapter logic and tests directly support #26674. No unrelated adapter, OpenCode 2, or process-management changes are present.

Full details: Description check

Explanation

The description covers the problem, implementation, scope, approval reference, focused verification, observed results, limitations, and agent attribution. It also documents additional evidence and unverified checks. An explicit maintainer approval comment is not included, but the description is otherwise complete and sufficiently detailed.

Finishing Touches

📝 Generate docstrings

Commit to this branch
Create a new PR

🧪 Generate unit tests (beta)

Commit to this branch
Create a new PR

Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

JJulius Marminge6d ago

Taking over this one