Observability for AI Debugging

Use this reference to inspect whether an AI agent can diagnose a runtime failure with the project's existing logging architecture. Logs are the primary surface; traces, metrics, health endpoints, profiles, and test artifacts extend it when the project actually uses them.

This file owns the bounded inspection procedure. The Agent Work Loop core and high-impact observability judgment owns applicability, Ready / Partial / Blocked interpretation, pass/block decisions, and scoring.

Inspection Question

Can the agent find the real framework, run one small reproduction, retrieve readable output, correlate the failing path, and verify a hypothesis? For core files, name the trigger -> boundary/decision -> failure/recovery -> result chain. Do not infer coverage from imports, call counts, api/, or tests/.

1. Discover the Existing Logging Architecture

• Route: read scoped instructions, logging/observability docs, runbooks, and test guidance. Check content: reactive Observable guidance is not runtime observability.
• Framework: identify the runtime framework, project logger facade, adapters, and underlying library.
• Wiring: find initialization, profiles, levels, middleware/interceptors, buffering, rotation, and flush. An unused dependency proves nothing.
• Output: locate stdout/stderr, files, IDE channels, test capture, CI artifacts, or query systems; record format and encryption. For multi-process apps, map process/component-to-file instead of naming one directory.
• Correlation: identify request, trace, session, task, test, job, or run id and verify it crosses the failing boundary.
• Safety: record redaction, sensitive data, production access, retention, and whether the current agent can read the output.

Prefer the established framework and test capture. Recommend a new logging library only when no coherent route exists or the current route cannot satisfy the diagnostic need.

2. Prove an Executable Debug Route

Workload	Smallest route	Required evidence
REST or RPC service	Focused handler test, or local service plus one safe request with a stable request id.	Response plus captured/runtime logs for the same id.
Unit or integration test	One named test using the real runner and profile.	Exit/assertion plus captured logs or stderr.
CLI, worker, or job	One bounded invocation with fixed local input.	Result, diagnostic output, run/job id, and final state.
UI, IDE, desktop, or E2E	Focused launch/test profile with a test-owned log path.	Relevant process logs, console/trace/screenshot, correlation id, and visible result.

The route should form:

start or focused test -> stable id -> readable filtered output
-> failing boundary/decision -> assertion, response, or state check

Do not invent a curl command, port, test script, log path, or environment flag. Missing credentials, private dependencies, platform support, startup commands, or readable sinks are debugging constraints, not minor documentation gaps.

3. Return Direct AI-Debug Evidence

• Discoverable: framework, command, sink, component map, and safety boundary are findable; commands/paths support the declared platforms.
• Runnable: the current agent can execute the focused route.
• Readable: output is machine-readable; encrypted or remote-only output needs a safe query/decryption route.
• Correlatable: one stable id joins trigger, boundary, failure/decision, recovery, and result.
• Verifiable: evidence can confirm or refute a hypothesis and support recheck.
• Safe and reversible: access and temporary probes respect least privilege, cleanup, secrets, and production boundaries.

Return each gate separately for one bounded scenario and profile. Do not widen the scope: a focused test and a live or production route are different scenarios. The model classifies the combined result.

Map the result by failure state. If no failure was observed but an affected core or high-impact path has no focused diagnostic route, return the gap to relevant-check. If a failure was observed but missing, unreadable, or uncorrelated output blocks diagnosis, return it to failure-repair. Logger imports, log-call counts, or generally “few logs” remain search leads. A finding must name the missing chain segment and the reproduction, diagnosis, or verification outcome it blocks.

4. Recommend the Earliest Missing Move

Prefer, in order: document the real command; expose existing local/test capture; enable existing request/trace middleware; preserve correlation across boundaries; add one focused failure-path test; link concise architecture/access guidance; then consider a new logger, tracer, collector, or debug endpoint. Do not scatter entry/exit logs across every function.

What Belongs in AGENTS.md

Load agents-md-review.md when instructions are in scope. Keep only non-inferable facts:

• actual logging facade and architecture link;
• exact start, reproduction, and focused-test commands;
• output location/format, profile switch, test capture, and component map;
• correlation id and how to supply/filter it;
• encrypted, remote-only, access, redaction, and platform constraints.

Logging style rules do not make a project AI-debuggable without a runnable consumption route. Keep credentials and long framework tutorials out of root instructions.

Review Result

scenario/profile -> framework -> command -> output/correlation
-> six gate results -> missing segment/impact -> move/re-check

Use this result as evidence for Controlled Execution or Change Validation; it does not by itself prove delivery or longitudinal improvement. Follow Agent Work Loop for those claim boundaries.