fix(ai): preserve executable Anthropic tool calls on terminal stop

[probepark]

What

Restores execution of Anthropic tool calls that stream partially and then finish normally. dev is red because of this.

Why

#4612 (d97b79ef) removed terminal-stop-aware finalization:

-if (output.stopReason === "length" || !sawTerminalStopReason) {
-	for (const block of output.content) {
-		if (block.type === "toolCall" && truncatedToolCalls.has(block)) {
-			block.incompleteArguments = true;
+for (const block of output.content) {
+	if (block.type === "toolCall" && truncatedToolCalls.has(block)) {
+		block.incompleteArguments = true;

Membership in truncatedToolCalls records that a block passed through transient partial JSON while streaming. It is not evidence that the finished call is incomplete. Flagging unconditionally marks completed, valid tool calls incompleteArguments: true, and that blocks them from executing — so the agent silently loses the ability to run those tools on an explicit tool_use stop.

#4612 also changed duplicate active-index handling to fail immediately, which lost orphan finalization evidence.

Measured against the commit immediately before d97b79ef: anthropic-truncated-toolcall.test.ts went from 10 pass / 0 fail to 5 / 5.

How

Terminal stop reasons now suppress truncation flags for normally finalized partial calls, while two cases are still flagged because they genuinely never completed:

• calls with complete JSON that were never finalized
• duplicate-index orphaned incomplete calls, marked before the duplicate-envelope error is emitted

Duplicate replacement events are finalized before that error, so #4612's envelope failure behavior is preserved.

#4612's hardening is untouched: malformed and non-string toolcall_delta are still rejected, and terminal Codex/function inputs are still validated. This only changes what counts as evidence of incompleteness.

Source only — no test file was modified.

Testing

suite	result
anthropic-truncated-toolcall	10 pass / 0 fail (was 5/5)
anthropic-stream-envelope	31 / 0
increment guards + codex stream	74 / 0
managed-attempt-transaction	64 / 0

Whole packages/ai suite is back to its exact pre-#4612 failure set — same 12 names, no new entries. (Those 12 predate #4612 and are tracked separately in #4722.)

Negative control: reverting only the source diff reproduces exactly the five named failures (5 pass / 5 fail); reapplying returns 10/0.

Note

I reviewed, approved and merged #4612, and I did not run this suite — it is named after the behavior the change affected. Same miss as #4757. I am now checking for suites named after the behavior a PR removes or changes before approving.

GJC verdict

gajae.pr-review-verdict.v1 merge-approved sha256:843ef7f2076048c196efdce6b02b216d53ca7a07ecbc9f4749104b9f5996c7f8 reviewer:human reviewer-id:Yeachan-Heo evidence:exact-head-6c993cfe-current-dev-861b000cda-focused-tests

---

• Target branch is dev
• Focused tests pass with a negative control
• CHANGELOG updated under ## [Unreleased]
• Verdict matches the exact PR head

[snowykr]

Verdict

CHANGES_REQUESTED

Summary

This narrowly targets the correct regression—normally completed Anthropic tool_use calls should remain executable after transient partial JSON. However, the new terminal-stop relaxation also allows genuinely incomplete or integrity-compromised tool calls to cross the existing execution guard. The stream must remain fail-closed unless the tool JSON is strictly complete and the content-block lifecycle is valid.

Findings / Required Changes

1. [P1] Keep incomplete tool JSON non-executable even when a terminal stop reason is present.
   packages/ai/src/providers/anthropic.ts:2405-2423, 2906-2924 records incomplete input_json_delta buffers, but the final marking now runs only for length or a missing terminal stop reason. An explicit tool_use (and also other explicit stops) can therefore leave best-effort-repaired incomplete JSON without incompleteArguments, bypassing the agent loop's execution guard. Mark every syntactically incomplete or unclosed tool JSON buffer as incompleteArguments: true regardless of terminal reason; relax only for strictly complete JSON. Add closed- and unclosed-buffer cases with explicit terminal stops that assert non-execution.

2. [P1] Do not publish executable replacement calls after a duplicate active content-block index.
   packages/ai/src/providers/anthropic.ts:2055-2068, 2421, 2928 now defers the duplicate-index error until finalization. That lets the replacement block emit toolcall_end before the stream fails, whereas the previous implementation rejected the malformed envelope immediately. Fail at duplicate-index detection, or suppress all subsequent executable tool lifecycle events and mark exposed calls non-executable before reporting the integrity error.

3. [P2] Ensure affected-path PR CI runs the Anthropic streaming regression suites.
   scripts/ci-dev-affected.ts:1038-1053 selects same-basename tests for provider source changes, but anthropic.ts has no such test; the relevant coverage is in packages/ai/test/anthropic-truncated-toolcall.test.ts and packages/ai/test/anthropic-stream-envelope.test.ts. Add an explicit behavioral-owner/risk-canary mapping so this path runs those suites before merge.

CI / Verification

• No project code, tests, linters, or formatters were executed as part of this review.
• The verification review inspected the affected-path CI selection and relevant test sources; it found that the provider change does not schedule the focused Anthropic stream suites.
• No needs-human-ci failure was considered.

Axis Coverage

Axis	Verdict	Coverage
A1. Intent / Policy / Contract	APPROVED	The change is narrow and aligns with the stated regression fix; no public-contract or scope finding.
A2. Architecture / Correctness / Failure	CHANGES_REQUESTED	Incomplete JSON can bypass the execution guard after an explicit terminal stop.
A3. Security / Privacy / Trust	CHANGES_REQUESTED	Best-effort parsing of malformed tool arguments can reach tool execution.
A4. Verification / Tests / CI	CHANGES_REQUESTED	Affected-path PR CI does not select the focused Anthropic stream regression suites.
A5. Context / Compatibility / Platform	CHANGES_REQUESTED	Deferred duplicate-index failure can expose replacement tool lifecycle events before the integrity error.

[probepark]

@snowykr all three closed at dbfb812b, each with a negative control.

[P1] Incomplete JSON stays non-executable regardless of stop reason. You were right that I relaxed on the wrong axis — I keyed on the stop reason when the question is JSON completeness. Incomplete buffers are now flagged before toolcall_end whatever the terminal reason, and unclosed active tool blocks are flagged even when their JSON happens to be complete. Added explicit tool_use terminal coverage for both directions: complete JSON stays executable, incomplete JSON is blocked.

Negative control: re-gating incomplete-buffer marking on length fails 3 tests including the explicit terminal-stop incomplete case.

[P1] No executable replacement after a duplicate active index. Duplicate indices throw immediately again. The replacement block is marked ambiguous and non-executable before the integrity error, so nothing executable escapes ahead of the failure. Envelope coverage now asserts the exact sequence: 1 toolcall_start, 1 delta, 0 toolcall_end, 1 error, 0 done. I updated those envelope assertions deliberately — they previously pinned the deferred-failure behavior my earlier patch introduced.

[P2] Provider changes now schedule the Anthropic stream suites. This is the one that matters most, and thank you for finding it — it is the systemic cause of the whole incident.

mappedTestsFor() selects same-basename tests only, and there is no anthropic.test.ts, so changing packages/ai/src/providers/anthropic.ts scheduled nothing. That is why #4612 broke dev undetected, and why my own fix would not have been caught either. Added a narrow BEHAVIORAL_OWNER_TESTS mapping to anthropic-truncated-toolcall and anthropic-stream-envelope, with a planner regression asserting both are selected alongside the existing root/native tasks.

Negative control: removing the mapping fails that planner test (90/1) with only root-check and native selected.

suite	result
anthropic-truncated-toolcall	10 / 0
anthropic-stream-envelope	31 / 0
increment guards + Codex stream	74 / 0
managed-attempt-transaction	64 / 0
ci-dev-affected	91 / 0

For context on why this PR exists: I reviewed, approved and merged #4612, and I did not run the suite named after the behavior it changed. #4757 is the same mistake on a different PR. Your P2 fixes the class rather than the instance.

[Yeachan-Heo]

Fix-forward pushed as 2fc51b2. The three prior blockers are addressed: incomplete/unclosed JSON remains non-executable under explicit stops; duplicate active indexes fail before replacement lifecycle publication and quarantine both complete orphan and replacement calls; anthropic.ts behavioral-owner mapping schedules anthropic-truncated-toolcall.test.ts and anthropic-stream-envelope.test.ts. Local evidence: bun --cwd=packages/ai run check; 50 Anthropic regression tests pass; 66 Codex guard tests pass; 72 downstream managed/execution tests pass; 91 affected-selector tests pass. Please re-review this exact head.

[Yeachan-Heo]

APPROVE exact head 2fc51b2. Resolved snowykr review 4980233998: explicit terminal stops execute only strictly complete, closed JSON; incomplete/unclosed buffers remain incompleteArguments; duplicate active indices abort before replacement lifecycle publication and quarantine complete orphan/replacement calls; packages/ai affected-path mapping selects both Anthropic stream suites. Verified current head with package AI check, 50 Anthropic regression tests, 66 Codex guard tests, 72 downstream execution/managed tests, 91 affected-selector tests, and base/original-patch negative controls.

[Yeachan-Heo]

Rebased the reviewed fix-forward onto current origin/dev 861b000. New exact head: 6c993cf; merge-base is current dev; exact diff digest remains 843ef7f2076048c196efdce6b02b216d53ca7a07ecbc9f4749104b9f5996c7f8. The prior approval is stale by policy. Please re-review this exact head; the needs-human verdict is intentionally reset until that review lands.

[Yeachan-Heo]

APPROVE exact rebased head 6c993cf against current dev 861b000. The three requested blockers are closed: explicit-stop execution requires strictly complete closed JSON; duplicate active indices abort before replacement publication and quarantine complete orphan/replacement calls; affected-path mapping selects both Anthropic stream suites. Local evidence: bun --cwd=packages/ai run check; 50 Anthropic regression tests; 66 Codex guard tests; 72 downstream managed/execution tests; 91 affected-selector tests; 141 combined exact-head focused/selector assertions; base and original-patch negative controls.