fix(llma): extract pydantic-ai tool calls from output parts

[shauryapednekar]

Problem

Pydantic AI's current OpenTelemetry instrumentation stores generation output in gen_ai.output.messages, which PostHog maps to $ai_output_choices. Tool calls in that output are represented as message parts with type: "tool_call" and a top-level name.

PostHog already extracts tool calls from several provider shapes, but not this Pydantic AI message-parts shape, so affected $ai_generation events do not get $ai_tools_called or $ai_tool_call_count populated.

Pydantic AI instrumentation docs: https://pydantic.dev/docs/ai/api/models/instrumented/

Changes

• Teach AI tool-call extraction to read parts[] arrays on output choices and message wrappers.
• Extract part.name when part.type === "tool_call".
• Reuse existing sanitization, ordering, duplicate preservation, and per-event cap behavior.
• Add tests for direct extraction, malformed parts, stringified JSON, and the full OTel ingestion path.

How did you test this code?

I tested this locally in Docker.

pnpm --filter=@posthog/nodejs format:check
pnpm --filter=@posthog/nodejs lint
pnpm --filter=@posthog/nodejs build

All passed.

pnpm exec jest --runInBand --forceExit \
  src/ingestion/ai/tools/extract-tool-calls.test.ts \
  src/ingestion/ai/otel/attribute-mapping.test.ts \
  src/ingestion/ai/process-ai-event.test.ts

Result: 3 suites passed, 288 tests passed.

git diff --check

Passed.

Publish to changelog?

Do not publish to changelog.

Docs update

No docs update. This is a backend ingestion compatibility fix for an already documented tools normalization field.

🤖 Agent context

I used Codex to help me with this PR.

Codex helped inspect the existing tool-call extractor patterns, implement the narrow Pydantic AI parts[].type === "tool_call" extraction path, and add focused tests.

The implementation stays local to the existing extractor and reuses the current sanitizer, ordering, duplicate preservation, and count behavior. The added tests cover wrapped and unwrapped output shapes, malformed parts, mixed text/tool parts, stringified JSON, and the full OTel ingestion path.

[greptile-apps]

Prompt To Fix All With AI

Fix the following 1 code review issue. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 1
nodejs/src/ingestion/ai/tools/extract-tool-calls.test.ts:681-709
This is the only new test in `processAiToolCallExtraction` that isn't parameterised. Given that the team always prefers parameterised tests, this test (along with the existing non-parameterised neighbours) could be rolled into an `it.each` table in the same style used for `extractToolCallNames` above, making it easy to add further provider shapes later.

```suggestion
    it.each([
        [
            'Pydantic AI OTel tool_call parts from stringified JSON',
            JSON.stringify([
                {
                    role: 'assistant',
                    parts: [
                        { type: 'text', content: 'Let me check.' },
                        { type: 'tool_call', id: 'call_abc', name: 'get_weather', arguments: '{"city":"NYC"}' },
                        { type: 'tool_call', id: 'call_def', name: 'search_docs', arguments: '{"q":"weather"}' },
                    ],
                },
            ]),
            'get_weather,search_docs',
            2,
        ],
    ])('%s', (_description, outputChoices, expectedToolsCalled, expectedCount) => {
        const event = createEvent('$ai_generation', { $ai_output_choices: outputChoices })
        const result = processAiToolCallExtraction(event)
        expect(result.properties!['$ai_tools_called']).toBe(expectedToolsCalled)
        expect(result.properties!['$ai_tool_call_count']).toBe(expectedCount)
    })
```

_{Reviews (1): Last reviewed commit: "fix(llma): extract pydantic-ai tool call..." | Re-trigger Greptile}

[carlos-marchal-ph]

Looks good to me, thanks for the contribution! I'm gonna go ahead with the test change I suggested and then merge it.

[carlos-marchal-ph]

No need to make this a tabular test since it has a single input

[deployment-status-posthog]

Deploy status

Environment	Status	Deployed At	Workflow
dev	✅ Deployed	2026-05-27 10:09 UTC	Run
prod-us	✅ Deployed	2026-05-27 10:20 UTC	Run
prod-eu	✅ Deployed	2026-05-27 10:24 UTC	Run