| name | analyze-coding-session |
| description | Diagnose a coding-agent session with InferShape and produce evidence-backed findings plus a privacy-safe repair packet. |
Analyze Coding Session
Use this skill when a coding agent stalled, repeated exploration, churned through patches, escaped scope, or claimed completion without sufficient post-change proof.
This is an executable adapter for InferShape. It does not summarize inference engineering theory and does not invent findings with another model.
Prerequisites
- Node.js 22 or newer.
- The
infershape CLI available on PATH, or INFERSHAPE_BIN set to its absolute executable path.
- A JSON or JSONL session trace accepted by InferShape.
Workflow
- Preserve the original trace as read-only input.
- Run the adapter with the same arguments you would pass to InferShape.
- Request JSON, Markdown, HTML, and repair-packet outputs when handing work to another agent.
- Treat a false-completion or unresolved-failure exit as a blocked delivery, not as a warning to ignore.
node skills/analyze-coding-session/scripts/run.mjs \
analyze session.jsonl \
--repo . \
--json-out reports/session.json \
--markdown-out reports/session.md \
--html-out reports/session.html \
--repair-out reports/repair.json \
--fail-on-false-completion
Evidence
Keep these artifacts together:
- normalized session report;
- human-readable diagnosis;
- standalone local HTML timeline;
- repair packet with verified facts and exact next actions;
- adapter and InferShape exit status.
Do not claim the underlying product is correct merely because the session was analyzed. Use the running-product proof skill for executable delivery evidence.
Failure handling
- Exit
127: InferShape was not found. Install it or set INFERSHAPE_BIN.
- Exit
2: an explicit InferShape quality gate failed; preserve the report and keep the task blocked.
- Exit
1: invalid input or operational failure; fix the trace/configuration before retrying.
- Never replace a failed run with an uncited narrative summary.
