git:20261019.b4caf4b to git:20261027.dca7fe1

- description: "Verifies RudderStack AI events locally: captures what an instrumented app or MCP server sends, validates it against the AI events v0.3 schema, and checks joins, counts, identity and content. Use when testing AI events, or when AI or MCP events look wrong or missing."

- allowed-tools: "Bash(python3 *), Bash(python *), Bash(npm *), Bash(npx *), Bash(node *), Bash(uv *), Bash(env *), Bash(curl *), Bash(tar *), Bash(lsof *), Bash(kill *), Bash(go build *), Bash(go run *), Bash(date *), Read, Write, Edit, Glob, Grep"

- Proves the instrumentation sends what `ai-events.yaml` promises, before any real traffic. Runs offline: a stub or mock RudderStack client, a mock model, no provider keys. Scripts and the checklist: `references/verification.md`.

- 1. **Read the plan.** `ai-events.yaml` gives the surfaces, identity sources, content capture setting, legacy events and event rules to check. No plan: check against the spec only, and say so in the report.

- 2. **Scratch directory.** Create one outside the repo (for example `$TMPDIR/ai-events-verify/`). Captures, schemas, scripts and the virtualenv go there, never into the user's source tree.

- 3. **Get the schemas.** `ai-events-v0.3.schema.json` from the `@rudderstack/ai` npm release that ships spec v0.3, or the Go module's `spec/` directory (it is not in the Python wheel), and the client-event schema from `rudder-ai-events-instrument`.

- 4. **Capture** what each planned surface sends (skip surfaces the app does not have):

- - **A. Stub client**, when the app has a test runner that can call its routes or job functions: `configure()` with an in-memory client (Go: an `analytics.Client` stub passed to `Configure` / the MCP wrapper), then drive the surfaces.

- - **B. Mock data plane and mock model**, when it has no such tests, or is a script or cron job: run the app unchanged with its RudderStack data plane URL and provider base URL pointed at a local mock server, then send sample requests or run the job.

- - **C. Live Events**, after A or B passes: a development source in RudderStack, inspected in Live Events or through `rudder-mcp-workflow`.

- - **D. MCP servers**: the stub or the mock data plane, driven by an MCP client script (list tools, valid and invalid calls, `_meta` with an `interaction_id`).

- - Browser events: a recorder in place of `rudderanalytics`, appended to the same capture.

- 5. **Run the app's own tests** too: instrumentation must not break them.

- 6. **Validate** with `check_events.py` and the flags that match the plan: schema, `specVersion: 0.3.0` and a known SDK (`context.wrapperLibrary.name` `@rudderstack/ai`, `rudderstack-ai` or `ai-sdk-go`) per event, one server `ai_response_received` per `interaction_id` (and at most one client copy), client copies of the prompt, answer and feedback unless the plan opts out, joins on `interaction_id` and `conversation_id`, feature, surface, user ID and session, no content unless the plan turns capture on, system runs without a user ID, MCP events without argument values and with planned server names and intents, legacy events still sent. Compare the printed background runs and MCP calls with the plan.

- 7. **Fix and repeat.** A failure in app wiring goes back to `rudder-ai-events-instrument`; a wrong plan assumption goes back to `rudder-ai-events-setup` (plan back to `draft`). A gap that matches a known limit (setup skill, "Known limits") is reported, not fixed.

- 8. **Record the time.** On the first PASS, set `metadata.time_to_integrate.verified_at` in `ai-events.yaml` (`date -u +%Y-%m-%dT%H:%MZ`), and ask the user for `hours_spent` and `notes`; leave them `null` if they do not say.

- Capture: A stub client | B mock data plane | C Live Events

- Checks: schema PASS, one answer per turn PASS, joins PASS, identity PASS, content PASS, background n/a, mcp PASS, legacy PASS # n/a for surfaces the app lacks

- Known limits hit: <e.g. Flask route: no client_disconnected>

- Time to integrate: started <started_at>, verified <verified_at>, hours <hours_spent or "not recorded">

- Next: <fixes, or: point the app at a dev source and watch Live Events>

- | Finding | Usual cause |

- |---|---|

- | Several `ai_response_received` for one turn | model calls outside `aiInteraction` / `ai_interaction` / `RunInteraction` (Go: a call made with a context that is not the interaction's), a stream not returned or bound, a Python thread pool without `propagate`, a LangChain handler on the model constructor |

- | `ai_user_action` without a matching answer | the client sent a different ID than the server used; check the `x-interaction-id` hand-off |

- | `prompt_text` / `response_text` present | `captureContent` on, globally or per interaction |

- | `userId` on a background run | `aiTask` / `ai_task` not used, or a plain interaction opened inside the job |

- | No events from a script or job | missing `flush()` before exit (Go: `client.Close()`); a `false` from Node.js/Python `flush()` means a client did not finish within 10 s |

- | No `total_cost_usd` | expected unless the plan registers prices or the provider reports cost |

- | Legacy event missing | flag defaulting to off, or the old call was removed |

- | Server answer with no client copy | the answer copy is not sent on every final state (error, stop), or a sampling wrapper or ad blocker dropped it; in production it can also mean the user left: the server answer then has `client_disconnected: true` (not detectable on Django or Flask) |

- | `specVersion` is `0.2.0` | SDK 0.4.0 installed instead of the release that ships spec v0.3 |

- | `trace_id` missing | expected without a tracing source; with OpenTelemetry planned: the SDK's span processor is not registered (Node.js, Python), `WithTraceIDFrom(rudderotel.TraceID)` is missing (Go), or no span started inside the interaction |

- | `response_text` or breakdown entries missing on a large answer | the event exceeded 30,000 bytes of properties: the SDK drops captured text first, then `tool_breakdown` / `usage_breakdown` entries |

- | No `mcp_*` events | `wrapMcpServer` / `wrap_mcp_server` / `WrapMCPServer` / `ruddermcpgo.WrapServer` called after the server connected, a 2.x server factory returning unwrapped servers, a different server instance wrapped, or the process exited before its client flushed |

- | `request_context` in `argument_keys`, or a tool fails on it | the tool already declares a `request_context` argument (the wrapper then leaves it alone), or the server is on an unsupported MCP SDK |

- | `mcp_tool_called` not joined to an answer | the calling agent sends no `_meta` `com.rudderstack/interaction_id` (expected for hosted MCP tools) |

- | Several `mcp_client_seen` for one client | several server processes or a restart (state is per server name, per process) |

- | MCP counters or `mcp_client_seen` shared across two servers | both servers use the same name: give each distinct server its own |

- | Feedback on one side only | the browser handler or the endpoint's `trackAction` is missing, and the plan does not opt that side out |

- | Mock data plane captures nothing, app logs connection resets | the mock cannot read the gzip batch (use the reference mock as written), or the app exited before its batch was sent |

- | No `anonymousId` / `context.sessionId` from test cookies | cookie values not in the JavaScript SDK's encoding (`RUDDERSTACK_AI_DEBUG=1` logs "unreadable cookie"); Go: the request not passed to `WithValues` / `InteractionOptions.Request` |

- - Captured events can hold personal data when capture is on: keep captures in the scratch directory and delete them when done.

- - `references/verification.md`: schema sources, stub clients for Node.js and Python, the mock data plane and model server, an MCP client driver, `check_events.py`, and the checklist.