clear

Persist session state then seamlessly resume work. The user should not notice this happening.

Emergency mode (compacting/prompt-too-long): Skip reads, distill from in-context memory only, go to Step 3.

---

Steps

Step 0: Reset gate + get date

date +"%m%d%y %H:%M" && rm -f ".claude/logs/.quality-gate-active" ".claude/logs/.session-blocks-$(date +"%m%d-%H")" ".claude/logs/.tool-call-count" ".claude/logs/.compaction-occurred"

Step 1: Read state (parallel, skip in emergency)

Read simultaneously: .claude/memory.md + Daily Notes/MMDDYY.md

Step 2: Distill session (from in-context memory)

Extract and compress using restorable compression — preserve retrieval paths so the resumed session can restore full context from compressed form:

1. Task — one sentence
2. Done/remaining — 2-4 bullets, conclusions not process
3. Decisions — one line each, WHAT+WHY not HOW
4. Learnings — rules/facts for knowledge-nominations
5. Files touched — full paths of every file read or modified (not just modified — include key files read that informed decisions). These are retrieval anchors for the resumed session.
6. Active references — URLs, API endpoints, external resources consulted. Drop content, keep the pointer.
7. Next action — precise, actionable instruction including which file(s) to read first

Step 3: Write handoff to daily note

Append (or create new daily note if none exists):

## Session Handoff — HH:MM

**Task:** [one sentence]
**Done:** [bullets]
**Remaining:** [bullets]
**Decisions:** [bullets]
**Files:** [full paths — both modified and key reads]
**Refs:** [URLs, external resources — pointers only, no content]
**Next:** [precise action + which file(s) to read first]

Step 4: Update memory.md (only if changed)

New priorities, threads, decisions → edit. Nothing changed → skip.

Step 5: Promote and nominate learnings (only if discovered)

Two-tier promotion:

Tier 1: Immediate promotion to knowledge-base.md (high-confidence rules):

• User overrides (explicitly corrected something)
• Empirical facts (verified through testing or data)

Write directly with [Source: User directive MMDDYY] or [Source: Empirical MMDDYY].

Tier 2: Nominate to knowledge-nominations.md (lower-confidence):

• Agent inferences (patterns observed but not confirmed)
• Hypotheses (things that seem true but need more evidence)

Append: - [MMDDYY] /clear: [learning] | Evidence: [source]

Rule: When in doubt, promote. A rule in knowledge-base.md that gets corrected later is better than a rule in nominations that never gets seen.

Step 6: Auto-resume (restorable decompression)

Do NOT output a resumption prompt. Do NOT ask the user anything. Instead:

1. Re-read .claude/memory.md and .claude/knowledge-base.md (compressed context reload)
2. Re-read the daily note handoff you just wrote (for the Next action)
3. Restore from retrieval anchors — re-read the file(s) specified in the Next field and any critical files from the Files list that the next action depends on. This is the decompression step: the handoff told you what happened; re-reading the files restores how to continue.
4. Immediately execute the Next action — pick up exactly where you left off

The user should experience a brief pause, then work continuing seamlessly. No visible "clearing" or "resuming" messages. Just keep working.

Target: 5-7 tool calls, <30 seconds. Emergency: 2-3 calls, <15 seconds.