AgentSkillsCN

debug-mode

为 agentdev 会话开启、关闭并管理调试模式。 系统会将所有工具调用、技能激活、钩子触发以及智能体委托操作记录为 JSONL 格式。 当需要调试智能体行为、优化工作流,或分析会话性能时,可调用此技能。

SKILL.md
--- frontmatter
name: debug-mode
description: |
  Enable, disable, and manage debug mode for agentdev sessions.
  Records all tool invocations, skill activations, hook triggers, and agent delegations to JSONL.
  Use when debugging agent behavior, optimizing workflows, or analyzing session performance.

plugin: agentdev updated: 2026-01-20

AgentDev Debug Mode

Debug mode captures detailed session information for analysis, debugging, and optimization. All events are recorded to a JSONL file in claude-code-session-debug/.

Configuration

Debug mode uses per-project configuration stored in .claude/agentdev-debug.json.

Config File Format

Location: .claude/agentdev-debug.json (in project root)

json
{
  "enabled": true,
  "level": "standard",
  "created_at": "2026-01-09T07:00:00Z"
}

Fields:

  • enabled: boolean - Whether debug mode is active
  • level: string - Debug level (minimal, standard, verbose)
  • created_at: string - ISO timestamp when config was created

Enabling Debug Mode

Use the command to create the config file:

code
/agentdev:debug-enable

This creates .claude/agentdev-debug.json with enabled: true.

Or manually create the file:

bash
mkdir -p .claude
cat > .claude/agentdev-debug.json << 'EOF'
{
  "enabled": true,
  "level": "standard",
  "created_at": "2026-01-09T07:00:00Z"
}
EOF

Debug Levels

LevelCaptured Events
minimalPhase transitions, errors, session start/end
standardAll of minimal + tool invocations, agent delegations
verboseAll of standard + skill activations, hook triggers, full parameters

Default level is standard.

Changing Debug Level

Using jq:

bash
jq '.level = "verbose"' .claude/agentdev-debug.json > tmp.json && mv tmp.json .claude/agentdev-debug.json

Output Location

Debug sessions are saved to:

code
claude-code-session-debug/agentdev-{slug}-{timestamp}-{id}.jsonl

Example:

code
claude-code-session-debug/agentdev-graphql-reviewer-20260109-063623-ba71.jsonl

JSONL Format

Each line in the JSONL file is a complete JSON event object. This append-only format is:

  • Crash-resilient (no data loss on unexpected termination)
  • Easy to process with jq
  • Streamable during the session

Event Schema (v1.0.0)

json
{
  "event_id": "550e8400-e29b-41d4-a716-446655440001",
  "correlation_id": null,
  "timestamp": "2026-01-09T06:40:00Z",
  "type": "tool_invocation",
  "data": { ... }
}

Fields:

  • event_id: Unique UUID for this event
  • correlation_id: Links related events (e.g., tool_invocation -> tool_result)
  • timestamp: ISO 8601 timestamp
  • type: Event type (see below)
  • data: Type-specific payload

Event Types

TypeDescription
session_startSession initialization with metadata
session_endSession completion
tool_invocationTool called with parameters
tool_resultTool execution result
skill_activationSkill loaded by agent
hook_triggerPreToolUse/PostToolUse hook fired
agent_delegationTask delegated to sub-agent
agent_responseSub-agent returned result
phase_transitionWorkflow phase changed
user_interactionUser approval/input requested
proxy_mode_requestExternal model request via Claudish
proxy_mode_responseExternal model response
errorError occurred

What Gets Captured

Session Metadata

  • Session ID and path
  • User request
  • Environment (Claudish availability, plugin version)
  • Start/end timestamps

Tool Invocations

  • Tool name
  • Parameters (sanitized - credentials redacted)
  • Execution context (phase, agent)
  • Duration and result size

Agent Delegations

  • Target agent name
  • Prompt preview (first 200 chars)
  • Proxy mode model if used
  • Session path

Proxy Mode

  • Model ID
  • Request/response duration
  • Success/failure status

Phase Transitions

  • From/to phase numbers and names
  • Transition reason (completed, skipped, failed)
  • Quality gate results

Errors

  • Error type (tool_error, hook_error, agent_error, etc.)
  • Message and stack trace
  • Context (phase, agent, tool)
  • Recoverability

Sensitive Data Protection

Debug mode automatically sanitizes sensitive data:

Redacted Patterns:

  • API keys (sk-*, ghp_*, AKIA*, etc.)
  • Tokens (bearer, access, auth)
  • Passwords and secrets
  • AWS credentials
  • Slack tokens (xox*)
  • Google API keys (AIza*)

Analyzing Debug Output

Prerequisites

Install jq for JSON processing:

bash
# macOS
brew install jq

# Linux
apt-get install jq

Quick Statistics

bash
# Count events by type
cat session.jsonl | jq -s 'group_by(.type) | map({type: .[0].type, count: length})'

Tool Usage Analysis

bash
# Tool invocation counts
cat session.jsonl | jq -s '
  [.[] | select(.type == "tool_invocation") | .data.tool_name]
  | group_by(.)
  | map({tool: .[0], count: length})
  | sort_by(-.count)'

Failed Operations

bash
# Find all errors and failed tool results
cat session.jsonl | jq 'select(.type == "error" or (.type == "tool_result" and .data.success == false))'

Timeline View

bash
# Chronological event summary
cat session.jsonl | jq '"\(.timestamp) [\(.type)] \(.data | keys | join(", "))"'

Event Correlation

bash
# Find tool invocation and its result
INVOCATION_ID="550e8400-e29b-41d4-a716-446655440001"
cat session.jsonl | jq "select(.event_id == \"$INVOCATION_ID\" or .correlation_id == \"$INVOCATION_ID\")"

Phase Duration Analysis

bash
# Calculate time between phase transitions
cat session.jsonl | jq -s '
  [.[] | select(.type == "phase_transition")]
  | sort_by(.timestamp)
  | .[]
  | {phase: .data.to_name, timestamp: .timestamp}'

Agent Delegation Timing

bash
# Find slowest agent delegations
cat session.jsonl | jq -s '
  [.[] | select(.type == "agent_response")]
  | sort_by(-.data.duration_ms)
  | .[:5]
  | .[]
  | {agent: .data.agent, duration_sec: (.data.duration_ms / 1000)}'

Proxy Mode Performance

bash
# External model response times
cat session.jsonl | jq -s '
  [.[] | select(.type == "proxy_mode_response")]
  | .[]
  | {model: .data.model_id, success: .data.success, duration_sec: (.data.duration_ms / 1000)}'

Disabling Debug Mode

Use the command:

code
/agentdev:debug-disable

Or manually update:

bash
jq '.enabled = false' .claude/agentdev-debug.json > tmp.json && mv tmp.json .claude/agentdev-debug.json

Or delete the config file:

bash
rm -f .claude/agentdev-debug.json

Cleaning Up Debug Files

Remove All Debug Files

bash
rm -rf claude-code-session-debug/

Remove Files Older Than 7 Days

bash
find claude-code-session-debug/ -name "*.jsonl" -mtime +7 -delete

Remove Files Larger Than 10MB

bash
find claude-code-session-debug/ -name "*.jsonl" -size +10M -delete

File Permissions

Debug files are created with restrictive permissions:

  • Directory: 0o700 (owner only)
  • Files: 0o600 (owner read/write only)

This prevents other users from reading potentially sensitive session data.

Example Session Output

jsonl
{"event_id":"init-1736408183","timestamp":"2026-01-09T06:36:23Z","type":"session_start","data":{"schema_version":"1.0.0","session_id":"agentdev-graphql-reviewer-20260109-063623-ba71","user_request":"Create an agent that reviews GraphQL schemas","session_path":"ai-docs/sessions/agentdev-graphql-reviewer-20260109-063623-ba71","environment":{"claudish_available":true,"plugin_version":"1.4.0","jq_available":true}}}
{"event_id":"550e8400-e29b-41d4-a716-446655440001","timestamp":"2026-01-09T06:36:25Z","type":"tool_invocation","data":{"tool_name":"TodoWrite","parameters":{"todos":"[REDACTED]"},"context":{"phase":0,"agent":null}}}
{"event_id":"550e8400-e29b-41d4-a716-446655440002","correlation_id":"550e8400-e29b-41d4-a716-446655440001","timestamp":"2026-01-09T06:36:25Z","type":"tool_result","data":{"tool_name":"TodoWrite","success":true,"result_size_bytes":156,"duration_ms":12}}
{"event_id":"550e8400-e29b-41d4-a716-446655440003","timestamp":"2026-01-09T06:36:26Z","type":"phase_transition","data":{"from_phase":null,"to_phase":0,"from_name":null,"to_name":"Init","transition_reason":"completed","quality_gate_result":true}}
{"event_id":"550e8400-e29b-41d4-a716-446655440004","timestamp":"2026-01-09T06:36:30Z","type":"agent_delegation","data":{"target_agent":"agentdev:architect","prompt_preview":"SESSION_PATH: ai-docs/sessions/agentdev-graphql-reviewer...","prompt_length":1456,"proxy_mode":null,"session_path":"ai-docs/sessions/agentdev-graphql-reviewer-20260109-063623-ba71"}}
{"event_id":"end-1736408565","timestamp":"2026-01-09T06:42:45Z","type":"session_end","data":{"success":true}}

Troubleshooting

Debug File Not Created

  1. Check if debug mode is enabled:

    bash
    /agentdev:debug-status
    
  2. Verify config file:

    bash
    cat .claude/agentdev-debug.json
    
  3. Verify the directory is writable:

    bash
    ls -la claude-code-session-debug/
    

jq Commands Not Working

  1. Install jq: brew install jq or apt-get install jq
  2. Verify JSONL format (each line should be valid JSON):
    bash
    head -1 session.jsonl | jq .
    

Large Debug Files

Debug files can grow large in verbose mode. Use minimal level for lighter capture:

Update config:

bash
jq '.level = "minimal"' .claude/agentdev-debug.json > tmp.json && mv tmp.json .claude/agentdev-debug.json

Or clean up old files regularly:

bash
find claude-code-session-debug/ -name "*.jsonl" -mtime +3 -delete

Integration with Other Tools

Viewing in VS Code

The JSONL format works with JSON syntax highlighting. For better viewing:

  1. Install "JSON Lines" VS Code extension
  2. Use "Format Document" on each line individually

Importing to Analytics

bash
# Convert to CSV for spreadsheet import
cat session.jsonl | jq -rs '
  (.[0] | keys_unsorted) as $keys
  | ($keys | @csv),
  (.[] | [.[$keys[]]] | @csv)' > session.csv

Streaming to External Service

bash
# Tail and send to logging service
tail -f session.jsonl | while read line; do
  curl -X POST -d "$line" https://logging.example.com/ingest
done

Commands Reference

CommandDescription
/agentdev:debug-enableEnable debug mode (creates config file)
/agentdev:debug-disableDisable debug mode (updates config file)
/agentdev:debug-statusCheck current debug mode status