Skip to content

Copilot Session Review

Context

Long Copilot sessions can accumulate repeated context, broad file reads, follow-up prompts, and validation loops. The project needs a lightweight retrospective that improves the workflow without adding permanent instrumentation to every session.

Details

Use the repository agent speckit.session-review.agent.md by running /speckit.session-review once when a development session is complete.

The agent follows a cost-conscious two-pass process:

  1. Query Chronicle for compact aggregates covering turns, duration, compaction, prompt size, file references, and GitHub references.
  2. Read only the target session's relevant turns and no more than five recent, relevant activity reports.
  3. Identify repeated follow-ups, corrections, failed validations, oversized pastes, and repeated context reads.
  4. Add three to five prioritized recommendations to the matching French and English activity reports.

Local session storage does not provide authoritative per-model token totals. In that case, report turn count, duration, prompt size, compaction, and repetitions as proxies. Never invent token figures.

Operational Rules

  • Trigger manually at session closure, never after every message.
  • Keep the matching activity report as the source of the retrospective; do not create a separate business report.
  • Preserve existing tags, business sections, test results, and report metadata.
  • Use the French report as the first drafting language and keep the English mirror synchronized.
  • Evaluate effectiveness over several sessions by comparing turns, follow-ups, duration, and available token measurements.