Storytime Remember

Storytime Remember — Stage Remembrance

Writes specs/.storytime/remembrance.md atomically as a wakeup document

• consolidation prompt. Loaded post-/compact as the first action of any skill to carry session state across the boundary.

Arguments

Optional tier: $ARGUMENTS (nap | shift | compact, defaults to compact)

When the model invokes this

Per V1-002 / V1-017, the model self-detects pause signals and proposes tiers. This skill is how the model stages remembrance. Triggers:

• Nap — 1 pause signal fires (repetition, confusion, context-delta). Quick refresh; compact_staged: false. May not even surface to user unless tutorial tier.
• Shift — 2+ signals or framing-loss alone. Refresh remembrance AND propose a frame change (driver swap, mode shift). User confirms.
• Compact — token budget approaching, unresolved shift, or explicit user request. Finalize remembrance; propose /compact to user.

When the user invokes this

• "Remember this" or "stage remembrance" — write current session state to remembrance.md without triggering a pause tier decision. Explicit checkpoint.
• "Prepare for /compact" — same as compact tier, user-initiated.
• "Nap" or "take a breath" — explicit nap; quick refresh.

Process

1. Determine tier from argument (default compact if explicit invocation; determined by signals if self-detected).

2. Gather session state:

   • Current HEAD via git rev-parse HEAD.
   • All _thread.md files touched this session (scan specs/.storytime/ sessions/*/ for recently modified threads).
   • Active phase per thread (from last_completed_phase fields).
   • Active personas (codenames engaged in recent discussions).
   • In-flight work (partial artifacts, unsealed decisions).
   • Open questions (from thread open_questions fields and session artifacts).

3. Compose the remembrance body:

   • Wakeup — 3-5 sentences summarizing the workday. Workday-shaped, covering all active topics proportionally.
   • Consolidation prompt — imperative instructions to the re- engaging model. Name specific files to load. Include "do NOT" section for things to avoid re-running.
   • State pinned — bullet-point current state: active decisions, changed files, in-flight work, open questions, personas, last pause.

4. Write atomically per V1-018:

   write specs/.storytime/remembrance.md.tmp
   fsync (shell: sync)
   mv remembrance.md.tmp remembrance.md
   

5. Update _thread.md for each active thread:

   • Set remembrance_staged: true
   • Set remembrance_path: ../../remembrance.md
   • Append consolidation event to thread log (per consolidation-format)

6. Surface to user based on tier:

   • Nap (non-tutorial): silent write + [remembrance refreshed] appended to response.
   • Nap (tutorial): surface — teach the user what just happened.
   • Shift: propose the frame change. "I'm in a rut on X. Want to swap driver to @critic [forge] and re-read the thread with their lens? Remembrance is staged either way."
   • Compact: propose /compact. "Staging remembrance at [file:line]. Running /compact next. Remembrance will load automatically post-compact. Confirm?"

Output

• specs/.storytime/remembrance.md (new or overwritten, atomic)
• Consolidation event appended to active _thread.md files
• Inline confirmation message to user

Tier-specific behavior

nap

Signals: 1, localized. Fastest tier. Body may reuse existing wakeup if nothing substantially new; updates updated: timestamp and State pinned bullets only. Always compact_staged: false.

shift

Signals: 2+ or framing-loss. Proposes a frame change alongside the remembrance refresh. The frame-change proposal is part of the surface, not part of the remembrance body. compact_staged: false.

compact

Finalize remembrance — full wakeup rewrite, complete consolidation prompt, full state pinned. compact_staged: true. Proposes /compact to user after staging. If user skips /compact, remembrance stays staged until next pause tier decision.

Format reference

See ${CLAUDE_PLUGIN_ROOT}/skills/storytime/references/remembrance-format.md for the full schema, body structure, and validation rules.

Rules

1. Atomic writes only. Every remembrance write uses tmp + fsync + mv. Never overwrite in place.
2. Workday-shaped. Cover the whole session, not just one topic.
3. Specific consolidation prompts. Name exact file paths. No vague "read the relevant files" language.
4. Thread ledger updated. Every write updates remembrance_staged and remembrance_path in active threads' frontmatter.
5. Tutorial mode surfaces naps. Otherwise naps can be silent.
6. Never proposes /compact without user confirmation. The pause tier is the LLM's judgment; the action is the user's.

Related

• references/remembrance-format.md — artifact schema
• references/consolidation-format.md — consolidation event shape
• references/automation.md — tier gating
• Post-commit hook — records commit consolidation events that feed remembrance state