Bridge Agent Log Analyzer

Overview

Inspect one persisted Bridge session and turn its on-disk logs into a behavior report.

This skill is for analyzing the agent itself:

• what it tried to do
• which tools it called
• which calls ran in the background
• where it failed
• how it recovered or retried
• which artifacts it read or produced

This skill is not for UI reconstruction. Ignore frontend display behavior unless the user explicitly asks for it.

Default Workflow

1. Identify the target session or log identifier.
2. Run scripts/inspect_bridge_session.py first.
3. Read only the raw files needed to explain the behavior:
   • session-meta.json
   • agent/history.json
   • agent/context.json
   • agent/state.json
   • agent/toolcalls/*/meta.json
   • agent/toolcalls/*/output.log
   • subagent equivalents under agents/<subagent-id>/agent/
4. Reconstruct the execution chronologically.
5. Separate direct evidence from inference.

Quick Start

Inspect the most recent session:

python3 .agents/skills/bridge-agent-log-analyzer/scripts/inspect_bridge_session.py

Inspect a specific session directory ID:

python3 .agents/skills/bridge-agent-log-analyzer/scripts/inspect_bridge_session.py --session-id <session-id>

Inspect using any UUID or identifier that appears inside that session's logs:

python3 .agents/skills/bridge-agent-log-analyzer/scripts/inspect_bridge_session.py --session-id <message-or-toolcall-id>

The helper script is the index. Use it to find the real session directory, runtime toolcalls, failures, and likely recovery paths before opening raw files.

Primary Evidence Sources

• history.json
  • user-visible messages and task state transitions
• context.json
  • assistant decision points, tool calls, and tool results
• toolcalls/*/meta.json
  • runtime command, async promotion, exit code, timing, environment
• toolcalls/*/output.log
  • actual command output, error text, tracebacks, downloads, progress
• state.json
  • model, reasoning effort, last-round token and step summary
• session-meta.json
  • session title, environments, created/updated timestamps

Use workspace files only when the agent explicitly read or referenced them in logs.

What To Extract

Session overview

• real session directory ID
• whether the requested identifier was a direct session ID or a reverse lookup hit
• environment labels
• effective start/end time based on history and toolcall timestamps

Behavior timeline

Use context.json as the main execution timeline.

• one assistant message is one decision point
• assistant content is a visible reply or intermediate narration
• assistant tool_calls are intended actions
• later tool messages are immediate tool results
• toolcalls/* supplies runtime metadata for those tool calls

Tool execution

Summarize:

• tool name
• order of invocation
• arguments or command preview
• runtime status
• sync vs async/background execution
• exit code
• duration
• output preview

Failures and recovery

Focus on:

• failed toolcalls
• error-like tool results
• retries of the same tool
• diagnostics between a failure and a retry
• mismatches between summary files and raw logs

Artifacts and final state

Extract:

• files the agent explicitly read
• result files written and later inspected
• final kept state when the logs make it clear

Skill evidence

Only mention skill usage when there is concrete evidence such as:

• a tool opening a SKILL.md
• a command explicitly targeting a skill path

Report this as inferred skill usage.

Privacy Boundary

Do not output agent thinking.

• ignore reasoning
• ignore encrypted_reasoning
• do not decode or summarize hidden reasoning

This skill explains behavior from logs, not hidden chain-of-thought.

Reporting Format

Default to a concise Markdown report with:

1. Session overview
2. Behavior timeline
3. Tool execution summary
4. Failures and recovery
5. Artifacts and final state
6. Inferred skill usage
7. Gaps and uncertainty

For each important step, include:

• trigger or input
• assistant action
• tools called
• outcome
• evidence source

Failure Modes

• Missing history.json: rely on context.json, toolcalls/*, and state.json
• Missing context.json: restrict to user-visible history plus runtime toolcalls if present
• Missing toolcalls/*: fall back to context.json tool results
• Missing tool result for a tool_call_id: report it as unmatched, not failed
• Sessions with subagents: analyze each agent separately, then combine

Resources

• scripts/inspect_bridge_session.py
  • behavior-focused summary for one session
• references/log-sources.md
  • artifact map and evidence grading