logfire

name	logfire
description	This skill should be used when the user asks to query Logfire logs, inspect Logfire traces, fetch SQL trace data, get spans for a trace, inspect LLM spans, inspect tool calls, or check token usage from Logfire. Also use it to triage an ARENA spawn run — diagnose why a run failed, timed out, or took too long, get per-stage turns/cost/tokens, or compare a prod trace to a local one.
version	1.0.0

Logfire

Shared wrapper for querying Logfire trace data via SQL.

Setup

1. Copy .claude/skills/logfire/scripts/.env.example to .claude/skills/logfire/scripts/.env
2. Set LOGFIRE_TOKEN with your Logfire read token

Usage

./.claude/skills/logfire/scripts/logfire.sh "SQL_QUERY"

ARENA Trace Workflow

1. Identify candidate traces by spec id, workdir suffix, or ARENA span names. Query both message and attributes because useful identifiers may live in either place.
2. Group candidates by trace_id and compare MIN(start_timestamp), MAX(end_timestamp), and COUNT(*) to distinguish planner, skeleton, feature, and sanity runs.
3. Fetch the selected trace ordered by start_timestamp ASC. If the payload is large, redirect to a temp file and summarize with Python instead of reading the full JSON into chat.
4. Logfire ingestion can lag. If a local terminal shows the process is still alive but the trace appears stale, re-query after a short wait and inspect child processes to tell an ingestion delay from a blocked subprocess.

ARENA emits manual spans in addition to Claude SDK spans:

• arena.pipeline: parent span for a complete spawn. Useful attributes include arena.spec.id, arena.spec.name, arena.spec.custom_prompt, arena.spec.framework, arena.spec.database, arena.spec.deployment_model, arena.spec.vulnerability_tokens, arena.spec.defense_tokens, arena.workdir.name, arena.pipeline.outcome, and aggregate cost/turns.
• arena.agent_stage: child span for each agent invocation. Useful attributes include arena.agent.stage, arena.agent.model, arena.agent.max_turns, arena.agent.allowed_tools, arena.agent.disallowed_tools, arena.agent.wallclock_seconds, arena.agent.max_fixup_attempts, arena.agent.outcome, arena.agent.turns, and arena.agent.cost_usd.

Triaging a spawn run

To diagnose why a run failed, timed out, or was slow — per-stage turns/cost/tokens, tool histograms, failure root-cause, and prod-vs-local comparison — follow references/trace-triage.md. It has copy-paste queries keyed off the span schema above.

Response Format

Logfire returns columnar format:

{
  "columns": [
    { "name": "trace_id", "datatype": "Utf8", "values": ["abc123", "def456"] },
    { "name": "span_id", "datatype": "Utf8", "values": ["span1", "span2"] }
  ]
}

Common Queries

# Test connection
./.claude/skills/logfire/scripts/logfire.sh "SELECT 1 LIMIT 1"

# Find recent ARENA traces for a spec/workdir
./.claude/skills/logfire/scripts/logfire.sh "SELECT trace_id, MIN(start_timestamp) AS first_seen, MAX(end_timestamp) AS last_seen, COUNT(*) AS record_count FROM records WHERE start_timestamp > now() - interval '2 hours' AND (message ILIKE '%leaky-query%' OR CAST(attributes AS TEXT) ILIKE '%leaky-query%' OR CAST(attributes AS TEXT) ILIKE '%e58aba16%' OR message IN ('arena.pipeline', 'arena.agent_stage')) GROUP BY trace_id ORDER BY first_seen DESC LIMIT 20"

# Find pipeline spans directly
./.claude/skills/logfire/scripts/logfire.sh "SELECT trace_id, span_id, start_timestamp, end_timestamp, attributes->'arena.spec.id' AS spec_id, attributes->'arena.pipeline.outcome' AS outcome FROM records WHERE message = 'arena.pipeline' ORDER BY start_timestamp DESC LIMIT 20"

# Find stage spans for a trace
./.claude/skills/logfire/scripts/logfire.sh "SELECT start_timestamp, end_timestamp, attributes->'arena.agent.stage' AS stage, attributes->'arena.agent.outcome' AS outcome, attributes->'arena.agent.model' AS model, attributes->'arena.agent.turns' AS turns, attributes->'arena.agent.cost_usd' AS cost FROM records WHERE trace_id = 'TRACE_ID' AND message = 'arena.agent_stage' ORDER BY start_timestamp ASC"

# Get spans for a trace
./.claude/skills/logfire/scripts/logfire.sh "SELECT span_id, parent_span_id, start_timestamp, end_timestamp, message, attributes FROM records WHERE trace_id = 'TRACE_ID' ORDER BY start_timestamp ASC LIMIT 1000"

# Get LLM spans
./.claude/skills/logfire/scripts/logfire.sh "SELECT DISTINCT span_id, start_timestamp, attributes FROM records WHERE trace_id = 'TRACE_ID' AND (attributes->>'gen_ai.request.model' IS NOT NULL OR attributes->>'gen_ai.response.model' IS NOT NULL) ORDER BY start_timestamp DESC"

# Get tool calls
./.claude/skills/logfire/scripts/logfire.sh "SELECT message, attributes->'tool.name' as tool_name, attributes->'tool.arguments' as tool_args FROM records WHERE trace_id = 'TRACE_ID' AND attributes ? 'tool.name' ORDER BY start_timestamp ASC"

# Get LLM token usage
./.claude/skills/logfire/scripts/logfire.sh "SELECT attributes->'gen_ai.usage.input_tokens' as input_tokens, attributes->'gen_ai.usage.output_tokens' as output_tokens, attributes->'gen_ai.request.model' as model FROM records WHERE trace_id = 'TRACE_ID' AND attributes ? 'gen_ai.usage.input_tokens'"