set:forensics

Triage what failed in a completed orchestration run using filtered views over Claude Code session transcripts and orchestration logs.

WARNING — never read raw session jsonl files unbounded. A single session jsonl can exceed 300 KB and a run typically has 50+ sessions. Always go through the CLI's filtered views first; only fall back to raw Read on a specific line range when the CLI has narrowed the suspect.

Triage order

Always follow this sequence:

1. set-run-logs <run-id> discover — confirm the scope of resolved sources. Verify the main session dir and every -wt-<change> worktree dir was found, plus the orchestration dir.

2. set-run-logs <run-id> digest — get the error map. Aggregates tool errors, non-zero bash exits, stop-reason anomalies, user interrupts, permission denials, and crash suspects across every session. Output is grouped by change → session → tool with snippets.

3. set-run-logs <run-id> session <uuid> --errors-only — for each suspect session in the digest, view its tool-call timeline filtered to errors. Use the short UUID prefix (≥6 chars) from the digest. Only AFTER you've narrowed via this view should you read raw jsonl.

4. set-run-logs <run-id> grep <pattern> — targeted probe for a specific error signature across all sessions. Emits ONLY the matching message.content text (not raw jsonl lines), so it's safe to use in agent context. Default cap is 50 matches — pass --limit to raise.

5. set-run-logs <run-id> orchestration — cross-reference timing. Shows per-change dispatch counts, last gate status, terminal status, gate-outcome counts. Use this to map a suspect session back to its orchestration timeline (when it dispatched, which gate triggered, what the terminal status was).

6. Raw jsonl as last resort. If the CLI's filtered views still leave a question unanswered, locate the specific jsonl from discover output and Read only the line range you need. Never Read a session jsonl without an offset/limit.

Common subcommand options

• --json — emit JSON instead of markdown (for piping to jq or programmatic consumers).
• session --tool <name> — filter a session timeline to a single tool (e.g. --tool Bash).
• grep -i — case-insensitive search.
• grep --tool <name> — restrict grep to one tool's tool_use / tool_result blocks.
• grep --limit <N> — raise the default 50-match cap (use sparingly to protect context).

When to use

• After any orchestration run that didn't merge cleanly.
• When the user asks "what happened in run X?" / "why did change Y fail?" / "did agent Z get stuck?".
• Before reading any session jsonl by hand.
• When investigating a finding from the verify gate — find the agent session that produced the failing commit.

When NOT to use

• For an in-flight run — the sentinel owns live monitoring; this targets completed runs.
• For cross-run comparison — use set-compare instead.
• For project-type-specific analysis — this CLI is project-agnostic.

Quick reference

set-run-logs <run-id> discover                          # what's there
set-run-logs <run-id> digest                            # error map
set-run-logs <run-id> session <uuid> --errors-only      # one session's errors
set-run-logs <run-id> grep <pattern> [--tool <name>]    # targeted probe
set-run-logs <run-id> orchestration                     # engine-level summary

Pair every drill-down with the orchestration view to align session events with dispatch / gate timing. Don't read raw jsonl until the CLI has narrowed the suspect.