motel-debug（汽车旅馆调试）

Motel Debug

You are in debug mode. Debug with runtime evidence, not guesswork.

Agents guess based on code alone. You need actual runtime data. Motel is the local OpenTelemetry server that collects traces and logs — use it as your evidence loop.

Default local server details:

• Base URL: http://127.0.0.1:27686
• OTLP traces: POST /v1/traces
• OTLP logs: POST /v1/logs
• Query API: GET /api/*
• OpenAPI: GET /openapi.json
• Header: Content-Type: application/json
• Auth: none by default

If the user provides a different motel URL, use that instead of the default.

Workflow

1. Verify motel is running — and start it if not

Check GET /api/health. If it returns 200, continue.

If it fails (connection refused, timeout, non-200), motel isn't running. Start it as a background daemon — do not launch the TUI, which is interactive and will block your shell:

motel start

motel start ensures a managed daemon process is running, writes a lockfile under .motel-data/, and returns a JSON status blob. It's idempotent — safe to call repeatedly. If motel isn't on PATH, fall back to bunx @kitlangton/motel start.

After starting, re-check GET /api/health (may take 1–2s to become ready). If it still fails, read .motel-data/daemon.log for the error and surface it to the user.

Other lifecycle commands, for reference:

motel status   # JSON status (running? pid? workdir?)
motel stop     # stop the managed daemon for this workdir

Discover reporting services with GET /api/services when needed.

2. Generate hypotheses

Before touching any code, generate 3-5 specific hypotheses about why the bug occurs. Be precise — "the cache key doesn't include the user ID" is better than "something is wrong with caching."

3. Instrument with tagged debug blocks

Add the minimum instrumentation needed to confirm or reject all hypotheses in parallel. Every debug block must:

• Be wrapped in #region motel debug / #endregion motel debug markers
• Include a debug.hypothesis attribute linking it to a specific hypothesis
• Use whatever tracing/logging mechanism the codebase already has (spans, structured logs, annotations — not raw fetch calls)

Tag every piece of debug instrumentation with structured attributes so you can query it later. Reuse these keys:

Choose log placements based on your hypotheses:

• Function entry with parameters
• Function exit with return values
• Values before and after critical operations
• Branch execution paths (which if/else ran)
• State mutations and intermediate values
• Suspected error or edge-case values

Guidelines:

• At least 1 instrumentation point is required; never skip instrumentation
• Do not exceed 10 — if you think you need more, narrow your hypotheses
• Typical range is 2-6

4. Reproduce the issue

• If a failing test exists, run it directly
• If reproduction is straightforward (CLI command, curl, simple script), write and run it yourself
• Otherwise, ask the user to reproduce — provide clear numbered steps and remind them to restart if needed
• Once a reproduction pathway is established, reuse it for all subsequent iterations

5. Analyze evidence

Query motel for the debug instrumentation:

curl "http://127.0.0.1:27686/api/spans/search?service=<service>&attr.debug.hypothesis=<id>"
curl "http://127.0.0.1:27686/api/logs/search?service=<service>&attr.debug.session=<session>"
curl "http://127.0.0.1:27686/api/traces/search?service=<service>&attr.debug.hypothesis=<id>"

For each hypothesis, evaluate: CONFIRMED, REJECTED, or INCONCLUSIVE — cite specific spans, logs, or attribute values as evidence.

6. Fix only with evidence

Do not fix without runtime evidence. When you fix:

• Keep all debug instrumentation in place — do not remove it yet
• Make the fix as small and targeted as possible
• Reuse existing architecture and patterns; do not overengineer

7. Verify the fix

Reproduce the issue again with instrumentation still active. Compare before/after evidence:

• Cite specific log lines or span attributes that prove the fix works
• If the fix failed: revert code changes from rejected hypotheses (do not let speculative fixes accumulate), generate new hypotheses from different subsystems, add more instrumentation, and iterate
• Iteration is expected. Taking longer with more data yields better fixes.

8. Clean up

Only after the fix is verified and the user confirms there are no remaining issues:

• Run the cleanup script or remove blocks manually (see Cleanup section below)
• Run git diff to confirm only the intentional fix remains

Instrumentation Rules

Wrap every temporary debug block in these exact markers:

// #region motel debug
// temporary debug instrumentation
// #endregion motel debug

Use whatever the codebase already provides for tracing and logging. The markers are language-comment wrappers — adapt the comment syntax for non-JS/TS files (e.g. # #region motel debug for Python).

Do not:

• Log secrets, tokens, passwords, or raw PII
• Remove instrumentation before post-fix verification succeeds
• Use setTimeout, sleep, or artificial delays as a "fix"
• Let code changes from rejected hypotheses accumulate — revert them

Query Patterns

Two filter prefixes for attribute search:

curl http://127.0.0.1:27686/api/health
curl http://127.0.0.1:27686/api/services

# Trace search
curl "http://127.0.0.1:27686/api/traces/search?service=<service>&operation=<text>&attr.debug.session=<session>"

# Span search (supports traceId to scope to one trace)
curl "http://127.0.0.1:27686/api/spans/search?service=<service>&traceId=<trace-id>&attr.debug.hypothesis=<id>"
curl "http://127.0.0.1:27686/api/spans/search?service=<service>&attrContains.ai.prompt.messages=<phrase>"

# Log search (supports severity filter, case-insensitive body search)
curl "http://127.0.0.1:27686/api/logs/search?service=<service>&severity=ERROR&body=<text>"
curl "http://127.0.0.1:27686/api/logs/search?service=<service>&attrContains.debug.label=<substring>"

# AI call search (compact summaries with previews)
curl "http://127.0.0.1:27686/api/ai/calls?model=gpt-5.4&sessionId=<session>"
curl "http://127.0.0.1:27686/api/ai/calls?text=<phrase>&status=error"

# AI call detail (full prompt/response payloads)
curl "http://127.0.0.1:27686/api/ai/calls/<span-id>"

# AI stats
curl "http://127.0.0.1:27686/api/ai/stats?groupBy=model&agg=total_input_tokens"

curl http://127.0.0.1:27686/openapi.json

List and search responses include meta.nextCursor when more data is available.

Motel gives you trace-correlated data — you can see which span a debug log belongs to, the parent operation, timing, and the full trace tree. Use GET /api/traces/<trace-id>/spans and GET /api/spans/<span-id>/logs to navigate the correlation.

For AI/LLM calls, use /api/ai/calls for compact searchable summaries (with prompt/response previews and token usage), and /api/ai/calls/<span-id> for full payloads.

Effect

If the target repo uses Effect, read references/effect.md before changing runtime wiring or adding instrumentation.

Cleanup

Use the bundled script at scripts/clear-motel-debug.ts when you want deterministic cleanup. It removes every block between #region motel debug and #endregion motel debug in JS/TS files and fails on unmatched markers.

If you cannot run the script, delete every marked block manually and then grep for #region motel debug to confirm none remain.