Skill: session-wrap

Skill: session-wrap

Purpose: Prompt for session summary when user indicates they're stopping work, approaching context limits, reaching milestones, or when open plan items, draft decisions, or uncaptured lessons are detected.

Trigger Indicators:

User Intent Signals:

• User says "stopping", "done for today", "that's it for now", "wrapping up", "I'm done"
• Context approaching 180K tokens (~90% of limit)
• PR just pushed
• Milestone completed

Active Work Signals:

• docs/plans/*.md contains unchecked - [ ] items (mid-plan indicator)
• knowledge/decisions/*.md has ADRs with Status: Proposed or Status: Draft (open decisions)
• Recent commits have lesson-worthy keywords (fix, solved, implement, pattern, debug, refactor) but no corresponding lesson file in knowledge/lessons-learned/
• Rules were captured to knowledge/rules.md or knowledge/me.md this session: surface "N rule(s) captured this session — run /kmgraph:recall to review rules.md for drift or conflicts."

Block Conditions:

• Stop hook coordination flag exists: /tmp/.kg-session-summarized-{PPID}-{date} — if present, do NOT prompt (Stop hook already fired; avoid double-prompting)

ECC Compatibility Note: This skill is coordinated with the Stop hook to prevent duplicate prompting. On Claude Code, the /tmp flag mechanism is used for inter-process coordination. On other ECC platforms, equivalent synchronization mechanisms (environment variables, shared state, or hook phase detection) should be used to prevent re-prompting if the Stop hook has already triggered this workflow.

Snapshot Awareness:

• Check for snapshot flag: /tmp/.kg-snapshot-{YYYY-MM-DD} (today's date)
• If snapshot flag exists: session summary already has partial content. Adjust prompt:

  "You took a session snapshot earlier — want to complete the wrap-up and save the final summary?"

• If no snapshot flag: use standard wrap-up prompt.

Behavior: When triggered, directly dispatch to session-summary-agent with conversational language that addresses the user, never exposing internal mechanics.

Example Triggers:

User: "Alright, I've pushed this to the branch. Wrapping up for today."
→ "Before you go — want a quick note on what we worked on today?"

[Skill detects unchecked plan steps]
→ "You're mid-plan — want to mark off what we finished before wrapping up?"

[Skill detects lesson-worthy commits without corresponding lesson]
→ "Looks like some meaningful work wasn't captured — want to save a note before finishing?"

User-Facing Language Rules:

• Address the user directly (never expose internal tool names or agent names)
• Conversational, inviting language
• No technical jargon about "dispatching", "agents", or internal file paths
• Single, clear question per prompt
• Treat the session summary as optional but valuable