debug-local（调试本地）

Debugging Lightdash

Iron Law

NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST.

Do not edit application code until you have a confirmed root cause hypothesis backed by evidence (logs, traces, reproduction). Fixing symptoms creates whack-a-mole debugging.

Phase 1: Gather Evidence

Before forming any hypothesis, collect facts.

1. Collect symptoms

Read error messages, stack traces, and reproduction steps. If the user hasn't provided enough context, ask ONE clarifying question.

2. Check recent changes

git log --oneline -20 -- <affected-files>

Was this working before? A regression means the root cause is in the diff.

3. Check logs and traces

pnpm exec pm2 logs lightdash-api --lines 50 --nostream

Then use Spotlight MCP:

• mcp__spotlight__search_errors with {"timeWindow": 300} for recent errors with stack traces
• mcp__spotlight__search_traces with {"timeWindow": 300} for recent request traces
• mcp__spotlight__get_traces with a trace ID for full span breakdown

4. Reproduce

Can you trigger the bug deterministically? Use browser automation or curl to reproduce. If you can't reproduce, gather more evidence before proceeding.

5. Read the code

Trace the code path from the symptom back to potential causes. Use Grep to find all references, Read to understand the logic.

Output: State your root cause hypothesis — a specific, testable claim about what is wrong and why.

Phase 2: Pattern Matching

Check if the bug matches a known Lightdash pattern:

Also check:

• git log for prior fixes in the same area — recurring bugs in the same files are an architectural smell
• Whether the issue is in common, backend, or frontend — misattributing the layer wastes time

Phase 3: Hypothesis Testing

Before writing ANY fix, verify your hypothesis.

1. Confirm: Add a temporary log, assertion, or use Spotlight/PM2 to check the suspected root cause. Reproduce. Does evidence match?
2. If wrong: Return to Phase 1. Gather more evidence. Do not guess.
3. 3-strike rule: If 3 hypotheses fail, STOP. Ask the user:

- Continue with a new hypothesis (describe it) - Escalate for human review - Add instrumentation and wait to catch it next time

Phase 4: Fix

Once root cause is confirmed:

1. Fix the root cause, not the symptom. Smallest change that eliminates the actual problem.
2. Minimal diff. Fewest files touched, fewest lines changed. Do not refactor adjacent code.
3. Write a regression test that fails without the fix and passes with it.
4. Run the relevant test suite. Paste output.
5. Blast radius check: If fix touches >5 files, pause and confirm with the user before proceeding.

Phase 5: Verify & Report

Reproduce the original bug scenario and confirm it's fixed. This is not optional.

Output a structured debug report:

DEBUG REPORT
════════════════════════════════════════
Symptom:         [what the user observed]
Root cause:      [what was actually wrong]
Fix:             [what was changed, with file:line references]
Evidence:        [test output, trace showing fix works]
Regression test: [file:line of the new test]
Status:          DONE | DONE_WITH_CONCERNS | BLOCKED
════════════════════════════════════════

Status definitions:

• DONE — root cause found, fix applied, regression test written, tests pass
• DONE_WITH_CONCERNS — fixed but cannot fully verify (e.g., intermittent bug, needs staging)
• BLOCKED — root cause unclear after investigation, escalated to user

Prerequisites

Ensure the development environment is running:

/docker-dev              # Start Docker services (postgres, minio, etc.)
pnpm pm2:start           # Start all PM2 processes including Spotlight
pnpm pm2:status          # Verify all processes are online

Services:

• Frontend: http://localhost:3000
• API: http://localhost:8080
• Spotlight UI: http://localhost:8969

Tools Reference

Server Logs (PM2)

pnpm pm2:logs              # Stream all logs (Ctrl+C to exit)
pnpm pm2:logs:api          # API server logs only
pnpm pm2:logs:scheduler    # Background job scheduler logs
pnpm pm2:logs:frontend     # Vite dev server logs
pnpm pm2:logs:spotlight    # Spotlight sidecar logs

For non-blocking log viewing (last N lines):

pnpm exec pm2 logs lightdash-api --lines 20 --nostream

Log format

Lightdash logs include a 32-character trace ID for correlation:

[700aa55e784a437aa97de9b8c5c3ed3a] GET /api/v1/health 200 - 37 ms
[4d40816e5a7d433e88f99008de5f4be5] GET /api/v1/user/login-options 200 - 2 ms

Use the first 8 characters (e.g., 700aa55e) to look up traces in Spotlight.

Trace Lookup (Spotlight MCP)

Use the Spotlight MCP tools to query telemetry programmatically:

List recent traces

mcp__spotlight__search_traces with filters: {"timeWindow": 300}

Returns summary of recent traces with trace ID, endpoint, duration, span count.

Get trace details

mcp__spotlight__get_traces with traceId: "700aa55e"

Returns the full span tree with timing breakdown:

GET /api/v1/health [816224a9 · 39ms]
   ├─ select * from "knex_migrations" [db · 8ms]
   ├─ select * from "organizations" [db · 2ms]
   ├─ session [middleware.express · 1ms]
   └─ /api/v1/health [request_handler.express · 0ms]

Search for errors

mcp__spotlight__search_errors with filters: {"timeWindow": 300}

Returns runtime errors and exceptions with stack traces.

Search logs

mcp__spotlight__search_logs with filters: {"timeWindow": 300}

Returns application log entries.

Browser Debugging (Chrome DevTools MCP)

Use the Chrome DevTools MCP tools for browser automation:

Opening pages

mcp__chrome-devtools__new_page with url: "http://localhost:3000/login"
mcp__chrome-devtools__navigate_page with url: "http://localhost:3000", type: "url"

Taking snapshots

mcp__chrome-devtools__take_snapshot

Returns a text snapshot of the page based on the accessibility tree with unique uid identifiers for each element.

Interacting with elements

mcp__chrome-devtools__click with uid: "1_5"
mcp__chrome-devtools__fill with uid: "1_4", value: "test@example.com"
mcp__chrome-devtools__hover with uid: "1_3"

Screenshots

mcp__chrome-devtools__take_screenshot
mcp__chrome-devtools__take_screenshot with fullPage: true
mcp__chrome-devtools__take_screenshot with filePath: "/tmp/debug.png"

Console and network

mcp__chrome-devtools__list_console_messages
mcp__chrome-devtools__list_network_requests
mcp__chrome-devtools__get_network_request with reqid: 123

Page management

mcp__chrome-devtools__list_pages
mcp__chrome-devtools__select_page with pageId: 1
mcp__chrome-devtools__close_page with pageId: 2

Database Inspection

# Check table schema
psql -c "\d <table_name>"

# Query data directly
psql -c "SELECT * FROM <table> WHERE <condition> LIMIT 5;"

# Check scheduler jobs
psql -c "SELECT id, task_identifier, attempts, last_error FROM graphile_worker.jobs WHERE last_error IS NOT NULL LIMIT 10;"

Trace Attributes (Wide Events)

Traces contain contextual attributes beyond timing:

Common Debugging Scenarios

Quick Commands Reference

Cross-Agent Validation

Use another AI agent to validate your findings, challenge your conclusions, and provide independent evidence. This is not just for when you're stuck — actively seek validation throughout the debugging process to ensure your analysis is sound.

Detect your environment

which claude 2>/dev/null && echo "HAS_CLAUDE=true" || echo "HAS_CLAUDE=false"
which codex 2>/dev/null && echo "HAS_CODEX=true" || echo "HAS_CODEX=false"

If you are Claude → ask Codex

codex exec "Given this context from a Lightdash debugging session:

<context>
[paste relevant logs, traces, error messages, or code snippets]
</context>

<question>
[your specific question — e.g., 'Does this trace confirm that the query cache is being bypassed?' or 'Given this stack trace, is my conclusion that the middleware short-circuits before auth correct?']
</question>"

If you are Codex → ask Claude

claude -p "Given this context from a Lightdash debugging session:

<context>
[paste relevant logs, traces, error messages, or code snippets]
</context>

<question>
[your specific question]
</question>"

When to consult

• Validate a hypothesis: Before concluding root cause, ask the other agent if the evidence supports your theory or if there's an alternative explanation
• Verify your interpretation of data: When reading traces, logs, or query output — confirm you're reading it correctly
• Challenge your fix: Before suggesting a code change, ask whether the fix actually addresses the root cause or just masks a symptom
• Cross-check complex logic: When the issue involves multiple systems (frontend + API + database + permissions), get an independent read on the interaction
• Justify a conclusion: If you're about to tell the user "X is the cause", make sure another perspective agrees — or surface the disagreement

Guidelines

• Set a 5-minute timeout: Codex can take a while to respond. When running codex exec via the Bash tool, set the Bash timeout to 300000ms (5 minutes) to avoid premature termination. Do NOT pass --timeout to codex exec itself — it doesn't support that flag.
• Be specific: Include the actual error, trace output, or code snippet — not just a vague description
• Ask for validation, not just answers: "Does this evidence support my conclusion that X?" is better than "What's wrong?"
• Include your current hypothesis: Let the other agent confirm or challenge it, rather than starting from scratch
• Include what you've already ruled out: This avoids duplicate investigation and focuses the consultation
• Always report back to the user: When you consult another agent, tell the user you did so. Summarize what you asked, what the other agent found, and whether you agree with their assessment. If there's a disagreement, present both perspectives and the supporting evidence for each. The user should never be unaware that another agent was consulted.

Test User Credentials

For testing authenticated flows:

• Email: demo@lightdash.com
• Password: demo_password!

API Access (Personal Access Token)

Prefer this method over using fetch() via Chrome DevTools MCP for API calls. curl with the PAT is faster, more reliable, and doesn't require a browser session or page context. Reserve browser MCP tools for UI interaction and visual debugging only.

A dev PAT is auto-provisioned by the seed data. Use it for direct API calls without browser login:

# Source the env file to get LDPAT and LIGHTDASH_API_URL
source .env.development.local

# Example: list projects
curl -s -H "Authorization: ApiKey $LDPAT" "$LIGHTDASH_API_URL/api/v1/org/projects" | jq

# Example: get user info
curl -s -H "Authorization: ApiKey $LDPAT" "$LIGHTDASH_API_URL/api/v1/user" | jq

The token (ldpat_deadbeefdeadbeefdeadbeefdeadbeef) is defined in SEED_PAT from @lightdash/common and inserted during database seeding. It belongs to the admin user (demo@lightdash.com) and never expires.