Session Wrap-Up

Run a candid, compact retrospective of the current session. Base it on the conversation, commands, tool output, edits, verification, and unresolved work; do not invent friction to make the report look substantial.

Interrogate The Session

Ask yourself:

• Where did I lose time, take a false start, or repeat work?
• What did I misunderstand or assume too early?
• What did the user's questions or corrections reveal about unclear scope, location, status, or explanation?
• Which instructions were missing, hard to discover, ambiguous, or in tension?
• Which code, documentation, architecture, or terminology did I have to rediscover?
• Which tool, dependency, environment, build, or test failure slowed the work?
• What nearly went wrong, even if the final result was correct?
• What remains incomplete, uncertain, or insufficiently verified?
• What worked especially well and is worth preserving or standardising?
• What single change would most improve a similar session next time?

Verify the factual premise behind user questions and corrections against the available evidence. Treat a correction as strong evidence of an expectation or communication gap even when its diagnosis is not confirmed.

Before reporting the outcome, perform only the cheap read-only checks needed to confirm the current state. For repository work, normally inspect working-tree status and the changed-file summary, then reconcile them with verification already recorded in the session. Do not rerun expensive tests solely for the wrap-up. State when relevant evidence is unavailable.

Distinguish observed facts from diagnosis and speculation. Say "No material friction observed" when that is the honest conclusion.

Classify Findings

Assign each meaningful issue to the narrowest useful category:

• Agent practice: planning, assumptions, sequencing, communication, tool choice, or verification discipline.
• Repository knowledge: AGENTS.md, documentation, architecture inventory, glossary, examples, or code discoverability.
• Reusable workflow: a missing or weak skill, script, task, template, check, or automation.
• Environment or tooling: dependencies, permissions, local setup, external services, or product limitations.
• Task definition: unclear scope, acceptance criteria, ownership, or product boundary.

Keep Chatto, Authling, shared-framework, and repository-wide recommendations separate. Place a proposed improvement with the product or infrastructure that owns it.

Recommend Improvements

For each material finding, propose the smallest change likely to prevent the same class of friction. Include:

• the evidence and impact;
• whether the root cause is known or inferred;
• the concrete improvement and likely owner or location;
• the expected payoff and any trade-off;
• a priority of now, soon, or only if repeated.

Prefer durable fixes over reminders. Examples include tightening a scoped instruction, adding a focused test, exposing an existing task, improving an error message, documenting a hidden invariant, or extending a relevant skill. Do not recommend new process or documentation when a simpler code or tooling fix would remove the problem.

Keep the wrap-up propose-only. Do not edit files, create issues, send feedback, or start follow-up work unless the user explicitly asks.

Report Format

Return one concise report with these sections:

Outcome

State the session goal, result, and verification level. Call out incomplete or unverified work without repeating the entire implementation handoff.

Friction And Near Misses

List only evidence-backed findings. For each, give the category, what happened, and its practical impact. Separate agent mistakes from repository or tooling problems. Treat direct user corrections as high-signal evidence of an expectation or communication gap, not automatic proof of their diagnosis. Do not infer dissatisfaction from tone alone.

What Worked

Capture practices, instructions, tools, or repository affordances that made the session more reliable or efficient.

Improvements

Prioritise at most five concrete proposals. Name the likely owner or file when known, and mark inferred diagnoses clearly.

Recommended Next Step

Choose the single highest-leverage action. If no follow-up is warranted, say so. End with one short question only when the user's perspective would materially change the recommendation.

Guardrails

• Be specific and candid without assigning blame or performing false humility.
• Do not exaggerate routine exploration into a systemic problem.
• Do not hide the agent's own errors behind tooling or instruction complaints.
• Do not expose secrets, personal data, or sensitive command output.
• Do not claim that a proposed improvement is validated when it has not been tried.