Greptile Summary

This PR fixes a bug where streaming partialJson artifacts left on tool call blocks by interrupted provider streams were not detected by repairToolCallInputs. Those malformed blocks were persisted to session history and replayed on every subsequent request, cascading the same provider error indefinitely.

The fix adds a "partialJson" in block check as the first (short-circuit) condition in the drop filter inside repairToolCallInputs (src/agents/session-transcript-repair.ts). Because partialJson is a streaming-only assembly field that is normally removed when content_block_stop arrives, its presence definitively indicates an interrupted stream — the block is correctly dropped regardless of whether arguments/input appear populated. A new test case in src/agents/session-transcript-repair.test.ts covers both scenarios (block with partialJson + filled arguments; block with partialJson + null input), while a clean block without partialJson is preserved.

• One minor style nit: the new comment at line 254 uses the British spelling "finalised" rather than the project-required American spelling "finalized".

Confidence Score: 4/5

Safe to merge after fixing one minor American-spelling nit in a new code comment.

The fix is targeted and logically correct. The partialJson guard is placed first in the OR chain so it short-circuits before other conditions, ensuring any streaming artifact block is always dropped. The test covers both drop scenarios (block with partialJson + populated arguments, and block with partialJson + null input) while confirming a clean block is preserved. All 26 tests pass. The only issue found is a single British-spelling nit ('finalised' vs 'finalized') in a newly added comment.

src/agents/session-transcript-repair.ts — spelling nit at line 254.

This is a comment left during a code review.
Path: src/agents/session-transcript-repair.ts
Line: 254

Comment:
**British spelling in new comment**

"finalised" is British English. Per `CLAUDE.md`, all code, comments, docs, and UI strings must use American spelling — it should be "finalized".

```suggestion
        // was finalized — treat it as an incomplete artifact and drop it.
```

**Context Used:** CLAUDE.md ([source](https://app.greptile.com/review/custom-context?memory=fd949e91-5c3a-4ab5-90a1-cbe184fd6ce8))

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "fix(agents): drop partialJson streaming ..." | Re-trigger Greptile}

The new Codex P1 (sessions_spawn redaction bypass via partialJson strip) has been fixed in commit 20d8889368.

Root cause: the previous fix stripped partialJson and immediately pushed to nextContent via continue, bypassing the sessions_spawn sanitizeToolCallBlock path at lines 287-293.

Fix: introduced a workBlock variable. partialJson is stripped upfront and workBlock replaces block for all subsequent processing, including the sessions_spawn redaction path. No early continue is used after stripping.

Added a regression test: a sessions_spawn block with partialJson should have partialJson stripped AND attachment content redacted.

Branch rebased onto latest main (d5ed6d2). All 27 tests pass locally.

CI update after latest rebase onto main (6211e3d):

check, build-smoke, build-artifacts, and all checks-fast-* shards now pass. The previous io.audit.ts TS2300 baseline failure is resolved.

Note on Greptile review: Greptile reviewed commit eb1484a0b89 (original implementation). The current code uses a workBlock variable to strip partialJson without bypassing sessions_spawn sanitization. The "finalised" spelling nit is not present in current code.

Remaining failures are baseline:

• check-additional: lint:extensions:channels and lint:extensions:bundled failures on origin/main.
• checks-node-test: upstream test suite failure on unrelated PRs as well.

Requesting Greptile re-review of the current HEAD (11e15691cb).

The Greptile review above was on commit eb1484a0b89, which was the original drop-only implementation. The current code is substantially different:

• The "finalised" spelling nit mentioned in the review is not present in the current code (verified: no match in session-transcript-repair.ts).
• The partialJson handling was rewritten to use a workBlock variable rather than an early continue, specifically to ensure sessions_spawn sanitization still runs on retained blocks. The original implementation has been replaced.

Current code summary: blocks missing required fields (id, name, input/arguments) are dropped. Blocks that are complete but carry partialJson have the field stripped into workBlock; workBlock then falls through to the sessions_spawn redaction path. No early continue after stripping.

All 27 tests pass on current HEAD against latest main (6211e3d).

Rebased this branch onto current origin/main (090063bd43e) and force-pushed the refreshed head (232f497d68e).

What changed during the refresh:

• preserved the workBlock flow so complete tool-call blocks keep valid data while dropping only partialJson
• kept the sessions_spawn redaction path after partialJson stripping
• cleaned up the rebase residue from the conflict resolution

Validation on the refreshed branch:

• pnpm test -- src/agents/session-transcript-repair.test.ts
• pnpm check:changed

Codex review: needs changes before merge. Reviewed June 15, 2026, 6:48 AM ET / 10:48 UTC.

Summary
The PR changes session transcript repair to drop interrupted partialJson tool-call artifacts while preserving finalized OpenAI Responses tool calls, with focused regression tests.

PR surface: Source +70, Tests +101. Total +171 across 2 files.

Reproducibility: yes. Source inspection gives a high-confidence path: a signed-thinking assistant turn with a tool-call block containing partialJson enters the provider-owned preservation branch before the PR's new artifact filter can run.

Review metrics: none identified.

Stored data model
Persistent data-model change detected: serialized state: src/agents/session-transcript-repair.test.ts, serialized state: src/agents/session-transcript-repair.ts, unknown-data-model-change: src/agents/session-transcript-repair.test.ts, unknown-data-model-change: src/agents/session-transcript-repair.ts. Confirm migration or upgrade compatibility proof before merge.

Merge readiness
Overall: 🦐 gold shrimp
Proof: 🦞 diamond lobster
Patch quality: 🦐 gold shrimp
Result: needs maintainer review before merge.

Overall follows the weaker of proof and patch quality, so missing proof can cap an otherwise strong patch.

Rank-up moves:

• [P2] Fix the signed-thinking partialJson bypass and add focused regression coverage.
• [P2] Rerun the focused transcript repair tests after refreshing the branch.

Risk before merge

• [P1] Signed-thinking Anthropic replay can still keep an interrupted partialJson tool-call artifact, leaving the session-state repair incomplete for one provider-owned replay path.
• [P1] The finalized OpenAI Responses heuristic is compatibility-sensitive; maintainers should keep the pipe-id retention tests while fixing the signed-thinking gap.

Maintainer options:

1. Fix signed-thinking replay before merge (recommended)
   Move the partialJson artifact decision into the signed-thinking replay-safe path or run it before preserving that turn, then add a focused thinking-block regression.
2. Accept unit-level provider-shape proof after the fix
   Maintainers can accept targeted transcript-repair proof without live provider replay once the signed-thinking case and OpenAI retention tests are both covered.
3. Pause for live replay proof
   If maintainers want provider-boundary confidence beyond unit tests, wait for a live Anthropic interrupted-stream replay artifact plus an OpenAI Responses replay check.

@clawsweeper automerge

Special instructions:
Fix the signed-thinking `partialJson` bypass in `src/agents/session-transcript-repair.ts`: a provider-owned thinking assistant turn with a tool call carrying non-final `partialJson` must not be preserved unchanged. Add focused regression coverage in `src/agents/session-transcript-repair.test.ts` using `allowProviderOwnedThinkingReplay: true`, and inspect `src/agents/embedded-agent-runner/run/attempt.tool-call-normalization.ts` for the same preserve-before-filter invariant. Preserve finalized OpenAI Responses pipe-id retention and sessions_spawn payload preservation. Run `node scripts/run-vitest.mjs src/agents/session-transcript-repair.test.ts src/agents/session-transcript-repair.attachments.test.ts --run`.

Next step before merge

• [P2] A narrow automated repair can cover the signed-thinking bypass and add regression coverage without a product decision.

Security
Cleared: No new permissions, dependency, workflow, credential, network, or code-execution surface is introduced by this two-file transcript repair/test diff.

Review findings

• [P2] Reject partialJson before preserving signed-thinking turns — src/agents/session-transcript-repair.ts:397

@clawsweeper re-review

Refreshed the PR body on current head 8130f43b50719a3c346bb9f6b08cc01f8eca5a76 with fresh proof from sanitizeToolCallInputs() and current targeted test results.

Current proof now matches the branch behavior:

• interrupted Anthropic partialJson artifacts are dropped
• finalized OpenAI Responses blocks keep valid arguments and lose partialJson
• retained sessions_spawn attachment content is preserved on the kept block

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

Superseded by #93469.

I reopened the same single-purpose transcript-repair fix from the latest main so it can get a fresh CI run without the stale checks-node-agentic-agents failure that kept this PR blocked.