/agent-session-forensics

name: agent-session-forensics
description: Use when diagnosing agent session history, interrupted tool loops, missing tool results, timing bottlenecks, or subagent trace correlation.

Agent Session Forensics

When To Use

• Agent session ended mid-tool-call or cannot resume.
• Tool call appears in assistant turn but no corresponding result turn follows.
• Need to correlate tool calls, tool results, and timing metadata.
• Diagnosing slow LLM calls, duplicate user turns, or malformed results.
• Debugging experimental MCP data-layer sessions.

Requirements / Checks

• Locate session directory and history files before editing anything.
• Prefer read-only inspection first.
• Have `jq` or equivalent JSON tooling available.
• Ask before modifying, truncating, or deleting any session or history file.

Workflow

1. **Inventory session files** — find metadata, current history, rotated previous history, and related subagent histories.

2. **Count and list last turns**:

   jq 'length' history.json                                      # total messages
   jq '.[-10:] | .[] | {role, stop_reason}' history.json        # last 10 turns
   jq '.[] | select(.role=="assistant") | .tool_calls[].id' history.json  # tool call IDs
   jq '.[] | select(.role=="user") | .tool_results[]?.tool_call_id' history.json  # results

3. **Correlate tool call IDs** — every `tool_call` in an assistant turn must have a matching `tool_result` in the immediately following user turn. Find the first gap.

4. **Check timing for slow calls**:

   jq '.[] | select(.timing) | {role, duration_ms: .timing.duration_ms}' history.json

5. **Identify failure pattern** — see table below.

6. **Repair (if approved)** — write a backup first (`cp history.json history.json.bak`), then make the smallest possible fix at the last valid correlation boundary.

Common Failure Patterns

| Symptom | Likely cause | Repair | |---|---|---| | Tool call with no result turn | Session interrupted mid-tool | Truncate after last matched pair | | Two consecutive user turns | Duplicate message insertion | Remove the duplicate | | `tool_result` with no prior `tool_call` | Corrupted or manually edited history | Remove orphan result | | Empty `content` on assistant turn | Model returned no text + no tools | Usually safe to truncate | | Session loops without progress | Missing result causes re-prompt | Inject minimal synthetic result |

Safety Constraints

• Do not edit session JSON without backing up the original first.
• Treat history files as sensitive: prompts, tool arguments, credentials, and file contents may appear.
• Do not infer user intent from stale history when current user instructions conflict.
• Do not repair by deleting broad ranges — find the last valid tool-call/result correlation boundary.

Validation / Done Criteria

• Report the names of every session file inspected.
• Every tool-call ID is accounted for (matched or flagged as unmatched).
• Any proposed repair names the backup path and truncation boundary.
• No session file is mutated without explicit user approval.

References

• `references/history-diagnostics.md`