---

name: compact-state description: Extract historical content from STATE.md into cycle files (burst logs, adversary passes, session checkpoints, lessons). Slims STATE.md to <200 lines. disable-model-invocation: false allowed-tools: Read, Write, Edit, Bash, Glob, Grep

Compact State

Extract historical content from a bloated STATE.md into proper cycle-scoped files, leaving a lean status document.

Templates

Read and follow the output format in:

• ${CLAUDE_PLUGIN_ROOT}/templates/burst-log-template.md — burst narrative log
• ${CLAUDE_PLUGIN_ROOT}/templates/convergence-trajectory-template.md — finding progression + per-pass details
• ${CLAUDE_PLUGIN_ROOT}/templates/session-checkpoints-template.md — archived session resume checkpoints
• ${CLAUDE_PLUGIN_ROOT}/templates/lessons-template.md — lessons learned by category
• ${CLAUDE_PLUGIN_ROOT}/templates/blocking-issues-resolved-template.md — resolved blocking issues

Announce at Start

Before any other action, say verbatim:

I'm using the compact-state skill to extract historical content from STATE.md into cycle files.

Prerequisites

• .factory/STATE.md exists
• .factory/ is a git worktree on factory-artifacts
• current_cycle is set in STATE.md frontmatter (or .factory/current-cycle exists)

Procedure

Step 1: Read and Analyze

1. Read STATE.md fully
2. Count total lines
3. Identify the current cycle name from frontmatter current_cycle or .factory/current-cycle
4. Ensure cycles/<cycle>/ directory exists (create if needed)

Step 2: Identify Extractable Content

Scan for these section patterns that should be moved to cycle files:

| Pattern | Target File | |---------|------------| | ## Burst N sections (burst narratives) | cycles/<cycle>/burst-log.md | | ## Pass N sections (adversary pass summaries) | cycles/<cycle>/convergence-trajectory.md | | adversary_pass_* frontmatter fields | cycles/<cycle>/convergence-trajectory.md | | ## Session Resume Checkpoint sections (all except the LAST one) | cycles/<cycle>/session-checkpoints.md | | ## Lessons section content | cycles/<cycle>/lessons.md | | Resolved rows from ## Blocking Issues table | cycles/<cycle>/blocking-issues-resolved.md | | ## Session Chain Summary sections | cycles/<cycle>/session-checkpoints.md |

Step 3: Extract to Cycle Files

For each extractable section:

1. Read the section content from STATE.md
2. Append to the target cycle file (create if it doesn't exist)
   • Each cycle file starts with a header: # <Title> — <Cycle Name>
   • Appended content is timestamped: ## Extracted from STATE.md on <date>
3. Preserve chronological order in the target file

Output Formats

Each cycle file follows its template (see Templates section above). Key points:

• Burst log — one ## Burst N section per burst, with agents, files, versions, summary
• Convergence trajectory — Finding Progression table (one row per pass) + Per-Pass Details sections. Convert adversary_pass_* frontmatter fields to table rows.
• Session checkpoints — one section per archived checkpoint with spec versions, state, and resume prompt
• Lessons — numbered entries organized by category (agent, process, infrastructure) with discovery dates
• Blocking issues resolved — table rows moved from STATE.md with resolution date added

Step 4: Slim STATE.md

After extracting, rewrite STATE.md keeping ONLY:

1. Frontmatter — remove adversary_pass_* fields. Keep: project, mode, phase, status, current_step, awaiting, current_cycle, dtu_required, plugin_version_adopted, historical_cycles
2. Project Metadata table
3. Phase Progress table (one row per phase, with trajectory shorthand in the Finding Progression column)
4. Current Phase Steps — last 5 rows only
5. Decisions Log — all (small, stable)
6. Skip Log — all (small, stable)
7. Blocking Issues — open only
8. Phase Numbering Reconciliation — keep if it exists (explains history)
9. Latest Session Resume Checkpoint — the most recent one only

Replace extracted sections with pointers:

## Historical Content

Burst logs, adversary pass details, session checkpoints, and lessons
have been extracted to cycle files:

- Burst history: `cycles/<cycle>/burst-log.md`
- Convergence trajectory: `cycles/<cycle>/convergence-trajectory.md`
- Session checkpoints: `cycles/<cycle>/session-checkpoints.md`
- Lessons learned: `cycles/<cycle>/lessons.md`
- Resolved blockers: `cycles/<cycle>/blocking-issues-resolved.md`

Step 5: Verify

1. Count lines in the new STATE.md — should be <200
2. Verify all cycle files were created and contain the extracted content
3. Verify no content was lost (extracted sections + remaining STATE.md = original)

Step 6: Commit

cd .factory
git add -A
git commit -m "factory(state): compact STATE.md — extract historical content to cycle files"

Output

Report:

STATE.md compaction complete:
  Before: <N> lines
  After:  <N> lines
  Extracted:
    - <N> burst narratives → cycles/<cycle>/burst-log.md
    - <N> adversary passes → cycles/<cycle>/convergence-trajectory.md
    - <N> session checkpoints → cycles/<cycle>/session-checkpoints.md
    - <N> lessons → cycles/<cycle>/lessons.md
    - <N> resolved blockers → cycles/<cycle>/blocking-issues-resolved.md

Safety

• This skill ONLY moves content between files — it never deletes content
• All extracted content is written to cycle files BEFORE being removed from STATE.md
• If any write fails, abort without modifying STATE.md
• The git commit captures both the slim STATE.md and the new cycle files atomically

Terminology Note

/compact-state (this skill) vs PreCompact hook event — these are distinct mechanisms.

| Concept | What it is | When it fires | |---------|-----------|---------------| | /compact-state (this skill) | A manually invoked skill that extracts historical content from STATE.md into cycle files (burst logs, adversary passes, session checkpoints, lessons) and slims STATE.md to <200 lines. It does NOT invoke the Claude Code /compact command — it reorganizes the STATE.md file, not the conversation context. | Only when an operator or agent explicitly invokes /compact-state. | | PreCompact hook event | A Claude Code harness event fired automatically before the harness performs context compaction. Triggers the precompact-flush PreCompact WASM plugin (precompact-flush.wasm) (S-18.04a deliverable, native WASM per ADR-028 §Decision 2), which persists wave-boundary state to factory-artifacts before context is lost. | Only when the Claude Code harness triggers automatic compaction (e.g., when context usage reaches the configured autocompact threshold). | | PostCompact hook event | A Claude Code harness event fired automatically after compaction completes. Triggers postcompact-reanchor.sh (S-18.05 deliverable), which emits a [PostCompact Re-anchor] block to stdout so the LLM session can re-ground itself after compaction. | Only when the Claude Code harness triggers automatic compaction (same trigger condition as PreCompact). |

Invoking /compact-state does NOT fire the PreCompact hook chain. The precompact-flush WASM plugin fires automatically before Claude Code compacts — not during a manual /compact-state invocation. These are independent paths: /compact-state reorganizes STATE.md content into cycle files; the PreCompact/PostCompact hook chain persists and restores session context across a harness-driven compaction event.

After any session clear or context reset, the mandatory first step before any pipeline work is /rehydrate-wave. See plugins/vsdd-factory/skills/rehydrate-wave/SKILL.md for the full invocation contract (BC-6.24.001 / ADR-026 §Decision 4).