compact-ops

compact-ops

The PreCompact hook already saves a state file automatically whenever /compact runs. This skill is the manual fallback: use it when the agent should write a richer, self-authored state file right before compaction (the agent knows nuances an LLM summarizer can miss).

Strict procedure profile

• Strictness: strict-procedure. The state file content and completion receipt are the deliverable.
• Hard gate: if the session id cannot be detected, do not create a guessed state file name. Stop and report that session id detection failed.
• Forcing function: fix the destination path, then read the saved file back and verify that the required headings exist.
• Completion receipt: report the state file path, main saved items, unverified items, and the instruction to run /compact.

Procedure

1. Get the session id and destination path.
   • Run bash "${CLAUDE_PLUGIN_ROOT}/scripts/get-session-id.sh" (bundled with this plugin).
   • If it fails, stop and report that preparation is incomplete because the session id is unavailable.
2. Compute the destination:
   • CWD_HASH=$(printf '%s' "$PWD" | md5sum | cut -c1-12)
   • Destination: ~/.claude/compact-ops/state/${CWD_HASH}/${SESSION_ID}.md
   • Create the directory if needed.
3. Review the task list, any active plan, and files currently being edited.
4. Write the state file with these headings in this exact order:

# Compact Prep State
## Active Plan
## Current Phase
## TaskList Summary
## Session Decisions
## Constraints and Blockers
## Worker Topology
## Skills Invoked
## Editing Files
## Failed Attempts
## Recovery Notes

5. Read the state file back and verify every heading exists.
6. Tell the user: Preparation complete. Please run /compact.

Content to save

• Active plan file path and current phase or step.
• In-progress task list and relevant notes.
• Decisions made during the session, user choices, and rejected alternatives with rationale.
• Constraints, blockers, and incomplete verification.
• Worker topology (subagents / panes) or Not used.
• Skills and slash commands invoked earlier in the session. This is an invocation record, not proof the skill is currently active.
• Files being edited and notes about unsaved or unverified work.
• Failed attempts, tool errors, and rejected approaches that should not be repeated.
• Recovery notes for the post-compaction agent: session_id, branch, key commands, validation results.
• If the user passed arguments to this skill, treat them as priority recovery notes.