Capture tool calls in live OpenAI Responses projector

[bgmcmullen]

We previously were not logging Codex tool calls, they would appear if backfilled but would not be captured live.

This PR adds the functionality to exchange-projector.js to capture the tool calls.

[philcunliffe]

Dual-agent review — request_changes

• Verdict: request_changes
• Risk class: medium
• Auto-merge advisory: 👎 thumbs down — verdict is request_changes; needs human-gated follow-up

Advisory only: no merge was attempted.

Reviewer reconciliation: The request_changes verdict is driven by one Codex major (Behavioral Correctness) plus medium blast radius. Independent verification found Codex's finding to be a false positive under the current recorder invariant (is_sse ⟹ response_body=null, recorder.js:190,260), which makes the body+stream mixed state it requires unreachable. It remains valid as a latent-fragility flag. The remaining Claude findings are edge-case test-coverage gaps, not defects. Practical recommendation: safe to land; optionally add the four edge-case tests and a guard/comment documenting the recorder invariant.

Risk capstone

Cross-reference: reviewer findings vs high-risk surfaces

Source	Finding (severity, evidence)	Intersects
Codex	Behavioral Correctness — stream-only tool calls dropped when body has text, major, exchange-projector.js:263-267,386	Targets (openAiResponsesMessages) + Risks bullet 1. Verified false positive: recorder invariant (isSse ⟹ response_body=null, recorder.js:190,260) makes the mixed body+stream state unreachable. Valid as a latent-fragility flag, not a present bug.
Claude	Tool-call drop paths untested, minor, exchange-projector.js:771,781	Risks bullet 2
Claude	toolOutputText object branches untested, minor, exchange-projector.js:808	Risks bullet 2
Claude	Malformed-JSON args untested, minor, exchange-projector.js:791	Risks bullet 2
Claude	Stream call-id dedup untested, minor, exchange-projector.js:427	Risks bullet 2

Codex review

Fix Validations

Live OpenAI Responses tool calls were missing

• Status: incomplete
• Evidence: hypaware-core/plugins-workspace/codex/src/exchange-projector.js:266, hypaware-core/plugins-workspace/codex/src/exchange-projector.js:267, hypaware-core/plugins-workspace/codex/src/exchange-projector.js:386
• Assessment: The new stream parser can capture response.output_item.done tool calls, but the top-level dispatcher only uses it when the parsed response_body yields zero assistant messages. Any capture that has both a final body and SSE events can still drop stream-only tool calls.

Findings

1) Behavioral Correctness

• Severity: major
• Confidence: high
• Evidence: hypaware-core/plugins-workspace/codex/src/exchange-projector.js:263, hypaware-core/plugins-workspace/codex/src/exchange-projector.js:266, hypaware-core/plugins-workspace/codex/src/exchange-projector.js:267, hypaware-core/plugins-workspace/codex/src/exchange-projector.js:386
• Why it matters: A live SSE exchange with response_body containing text plus stream_events containing response.output_item.done will project the text from the body and skip the stream parser, so the tool call is still not logged.
• Suggested fix: When streamEvents.length > 0, run the stream projection and merge/dedupe it with body-projected messages, or pass the parsed top-level body into the stream merge path so stream-only tool calls are appended even when body text exists.

No Finding

2. Contract & Interface Fidelity
3. Change Impact / Blast Radius
4. Concurrency, Ordering & State Safety
5. Error Handling & Resilience
6. Security Surface
7. Resource Lifecycle & Cleanup
8. Release Safety
9. Test Evidence Quality
10. Architectural Consistency
11. Debuggability & Operability

Evidence Bundle

• Changed hot paths: openAiResponsesMessages; responsesInputMessages; responsesAssistantMessagesFromBody; responsesAssistantMessagesFromStream; toolUseBlockFromPayload; toolResultBlockFromPayload
• Impacted callers: hypaware-core/plugins-workspace/codex/src/exchange-projector.js:184
• Impacted tests: test/plugins/codex-exchange-projector.test.js:131, test/plugins/codex-exchange-projector.test.js:197, test/plugins/codex-exchange-projector.test.js:275, test/plugins/codex-exchange-projector.test.js:311, test/plugins/codex-exchange-projector.test.js:351
• Unresolved uncertainty: I did not verify the gateway capture invariant for SSE response_body; the new tests only cover SSE with response_body: '', so the mixed body-plus-stream case remains unproven.

Claude review

Claude review

Five parallel review subagents (guidance compliance, shallow bug scan, historical
context, contract & callers, comments & tests) examined the PR at head b482bfc.
Four of the five found no defects: the change is style-compliant (no semicolons,
JSDoc-only types, @import at top, no @typedef), introduces no logic bugs,
re-introduces no historically-fixed bug, keeps the only export
(createCodexExchangeProjector) and its private-helper renames internally
consistent, and the new tool_use/tool_result block shapes are already
supported by the downstream gateway part-expander (message_projector.js:534-542,
columns tool_args/tool_result_for exist) and the kernel types
(AiGatewayProjectedMessage.content: string | JsonObject[]). The live helpers
mirror codex/src/backfill.js byte-for-byte, so turn-1 response rows hash-equal
turn-2 input-replay rows. Full suite: 759 tests pass.

The surviving findings are all test-coverage gaps on branches this PR introduced.
The implementation in each case was independently verified correct; these raise
regression-detection risk rather than describing a present defect, hence minor.

Tool-call drop paths (undefined returns) are untested

• Severity: minor
• Confidence: 90
• Evidence: hypaware-core/plugins-workspace/codex/src/exchange-projector.js:771, :781
• Why it matters: A function_call missing name/call_id (or function_call_output missing call_id) is silently skipped via if (block) …, and the load-bearing call_id ?? id fallback is never exercised — every test supplies call_id, so a regression that started dropping valid calls would pass.
• Suggested fix: Add cases for (a) a function_call with only id (no call_id) → still captured keyed on id; (b) a function_call missing name → dropped, not emitted broken.

toolOutputText object/wrapper/stringify branches are untested

• Severity: minor
• Confidence: 88
• Evidence: hypaware-core/plugins-workspace/codex/src/exchange-projector.js:808
• Why it matters: Every function_call_output test uses a plain string output; the structured-output path ({ output | content | text } unwrap and JSON.stringify fallback) — the exact case the JSDoc says it handles, and a common real Codex shape — has zero coverage.
• Suggested fix: Add a function_call_output with output: { output: 'x' } (expect content: 'x') and one with output: { foo: 1 } (expect content: '{"foo":1}').

Malformed-JSON tool arguments are untested

• Severity: minor
• Confidence: 85
• Evidence: hypaware-core/plugins-workspace/codex/src/exchange-projector.js:791
• Why it matters: The raw-string fallback in normalizeToolInput (parsed === value ? value) is what protects against truncated/garbled streamed args, but no test feeds a JSON-looking-but-invalid arguments string (e.g. '{"cmd":').
• Suggested fix: One case with arguments: '{"cmd":' → expect input === '{"cmd":' (preserved verbatim, not dropped, no throw).

Stream call-id dedup branch is untested

• Severity: minor
• Confidence: 80
• Evidence: hypaware-core/plugins-workspace/codex/src/exchange-projector.js:427
• Why it matters: The new logic avoids double-emitting a tool call present in both response.output_item.done and the response.completed body, but the "prefers completed body" test has no output_item.done, so toolUsesByCallId is empty and the dedup loop never runs — a regression that emitted the tool twice would pass.
• Suggested fix: One SSE case where output_item.done carries call_id: 'c1' AND the completed body's output[] also contains function_call call_id: 'c1' → assert the tool appears exactly once.

---

Reports: /Users/phil/workspace/hypaware/.git/worktrees/dual-review-pr83/dual-review/pr-83

[philcunliffe]

Not concerned with codex's finding