lessons-triage

lessons-triage

Read .agents/skills/lessons-triage/SKILL.md and follow the workflow steps.

Claude Code-specific additions

How to read the transcript

• Claude Code's transcript is at ~/.claude/projects/<project-slug>/<session_id>.jsonl (1 line = 1 event, JSONL)
• Traces of skill invocations can be identified via the Skill tool's tool_use events, or the <command-name> tag
• For a large transcript, use Grep to narrow down to skill-related event lines, then Read the surrounding context

Determining past triage sessions (outcome: self)

Since the currently running session hasn't fired SessionEnd yet, it doesn't exist in the queue. The prefilter's self determination target is a session in which lessons-triage was run in the past, identified via an execution marker in the transcript:

• The <command-name>lessons-triage</command-name> tag (a trace of /lessons-triage being invoked)
• A Skill tool invocation event for lessons-triage

Any session whose transcript contains either of these is treated as outcome: self and made a discard candidate (identifiable via Grep).

HITL approval (policy's externally-visible gate = batch-confirm)

The batch confirmation in step 4 is done via AskUserQuestion (do not set the answers parameter). With gate=batch-confirm (default), present all lessons as the options of a single question, and use multiSelect: true to have the user select as a batch which lessons to file. If the number of lessons exceeds the option limit for a single question, it's fine to split across multiple rounds, but treat all of them as "one batch-confirmation round" — do not revert to per-lesson sequential approval. Only run gh issue create for the selected lessons; discard the unselected ones.

Only fall back to per-lesson confirmation (file / file with edits / discard) when gate=ask (tightened by policy).