Storytime

Storytime — Continuity System + Spec Workflow

Storytime is a continuity system for LLM–harness collaboration. The spec workflow (survey → assemble → icebreaker → breakout → converge → plan) is one surface. Underneath is the consolidation loop: context → consolidation events → document structure → continuity across compactions, sessions, and time.

Arguments

The user's problem statement: $ARGUMENTS

Process

Follow this state machine in order, but collapse phases that have no real work. Phase sequence defines the maximum workflow; actual sessions use only the gears they need.

Entry: Route

• No arguments (bare /storytime) → scan _thread.md files, present a topic picker. Auto-resume if only one thread exists and is incomplete.
• Arguments provided → derive <topic> (kebab-case) from the problem. Check sessions/<topic>/_thread.md:
  • Thread found → Warm entry (load remembrance if staged)
  • No thread → Cold Start

Warm entry (thread exists)

1. Check specs/.storytime/remembrance.md — if present and compact_staged: true, load it and synthesize as current context. See references/remembrance-format.md.
2. Read the thread: episodes, decisions, last_consolidation, drift.
3. Compute git delta since last_commit; cross-reference with last survey fingerprint; check citation staleness.
4. Present "Previously on..." card with options: Continue · Retro · New sub-topic · Reset. Personas skip intros — they speak from accumulated context.

Cold Start (no thread)

Detect landscape: specs/.storytime/ exists → native; own doc structure → adapt-in-place; fresh repo → propose native. Ask if ambiguous. Never create .storytime/ without user knowledge. → Modes and decision tree: references/output-modes.md; or delegate to /storytime-bootstrap.

Phase 0: SURVEY

Explore agent scoped to problem. Check for prior runs (prior art, never overwrite). Scan artifacts, classify [team] [spec] [config], present inventory. Default: consolidate into storytime structure.

Write sessions/<topic>/survey.md — codebase context, inventory, coverage fingerprint (REQUIRED: commit, paths, ratios, classifications). Schema: schema_version: 1.

→ Scan targets: references/artifact-scan.md; fingerprint: references/survey-fingerprint.md; non-storytime mapping: references/import-adapters.md.

Phase 1: ASSEMBLE

Load cohort from cohort/_roster.md. Review team-like artifacts as rehire candidates. Recruit specialists for uncovered domains.

Default core: OWNER, OPERATOR, CRITIC ×2. Additional lenses: DOMAIN, SYSTEMS, PLATFORM, SKEPTIC, EDUCATOR.

Team size project-appropriate. Bias small. Tiny problem → 1-2. Default core → 3-4. Architectural → 6-8. Ceiling 12 (override). Duplicate archetypes allowed with distinct focus.

Codenames non-human (anchor, lattice, kestrel, ember). Role is load-bearing; codename is ornament. Human names only if user picks.

Write sessions/<topic>/team.md. → Full detail: references/team-assembly.md.

Phase 2: ICEBREAKER

Introductions. Review prior spec-like artifacts (keep hot / archive warm / rollup / cold / skip). Establish status quo with code citations. Identify sub-problems. Agree on constraints.

Write sessions/<topic>/icebreaker.md. → Artifact tiers: references/artifact-tiers.md.

Phase 3: BREAKOUT (parallel when possible)

For each sub-problem: assign one driving persona + 1-2 supporters (silent unless trigger fires). Pre-breakout synopsis (driver states plan; supporters state watching brief; user directs approve / join / defer / pause / cancel). Launch as parallel sub-agent if independent. Mid-breakout skills: VERIFY, GROUND, RESEARCH, DISCOVERY, PROTOTYPE. Grounding multi-source — strongest evidence wins.

Write sessions/<topic>/breakout-<subtopic>.md per breakout (driver, supporters, findings, citations, Complexity, Scale, recommendation).

Post-breakout pause mandatory (unless auto). Present summary cards with all options considered, not just recommendations. User directs: proceed / dig / revise / add.

→ Complexity scale: references/complexity-units.md; Scale dimensions: /Users/alexevers/workspace/projects/storytime/docs/scale-impact.md; driver details: references/driving-persona.md.

Phase 4: CONVERGE + PLAN

Reconvene. Merge findings. Resolve conflicts. Each plan section has one driver.

Write sessions/<topic>/plan.md: ASCII slide deck, numbered implementation steps with breakout cross-refs, code changes summary, risk matrix, non-goals (why skip + when revisit) REQUIRED, success criteria (measurable) REQUIRED, roadmap with Complexity + Scale per item in prose. Complexity ≥ 13 must decompose.

Phase 5: REVIEW

Present inline. User challenges, revises, or approves.

Phase 6: DONE

Update persona acquired_context. Log in history/. Evaluate specialist contracts. Finalize thread (append episode, record last_commit, clear open questions).

Consolidation

Storytime's core loop. Every consolidation event writes a unified artifact (see references/consolidation-format.md). Six scales:

Pauses are for the model, not the human. Signals: repetition | confusion | rut | framing-loss | context-delta | token-budget. Tier is combination-logic (not intensity):

• nap = 1 signal
• shift = 2+, persistent, or framing-loss alone
• compact = token-budget (authoritative) or unresolved shift

All remembrance writes are atomic (tmp + fsync + mv). Orphan .tmp older than 5 min → lint warning.

Remembrance is workday-shaped (covers session state across topics), written at nap/shift/compact, loaded post-/compact as first action. See references/remembrance-format.md.

Commits are the LLM's clock. LLM drafts every commit; user confirms every one. No auto-commit. Confirmation learns patterns (V1-014, see references/commit-drafting.md). Explicit stage: /storytime-remember.

Process Rules

1. SURVEY before ASSEMBLE. Never build a team blind.
2. ICEBREAKER before BREAKOUT. Shared understanding first.
3. CONVERGE before showing the user. Internal consensus first.
4. Every claim grounded — code, docs, web, or git.
5. Default core: OWNER, OPERATOR, CRITIC ×2 (two critics contest).
6. Non-goals + success criteria required on every plan.
7. Visuals use ASCII box-drawing. No external tools.
8. Personas are lenses, not characters. No role-play.
9. Team size project-appropriate. Bias small. Ceiling 12.
10. User has veto power everywhere.
11. Phases collapse when empty.
12. Every phase writes output — run is a complete snapshot.
13. Prior runs are prior art — never silently overwrite.
14. Survey writes a coverage fingerprint.
15. Effort = Complexity + Scale (never time). ≥13 decomposes.
16. Evaluation hygiene: observe metrics and conclusions separately.
17. Warm entry detected, not requested. If thread exists, warm.
18. Remembrance is pre-staged (at pause/compact), never reactive.
19. Post-breakout pause mandatory unless auto. Present options considered, not just recommendations.
20. Grounding multi-source: code > git > repo > library > standards > web.
21. @role is a lens directive, not a skill trigger. See references/addressing.md.
22. Codenames non-human by default.
23. One driving persona per leg. Supporters silent unless useful AND non-distortive.
24. LLM drafts every commit. User confirms every one. No auto-commit.
25. Pauses are model-driven, not threshold-driven (threshold is opt-in fallback).
26. All consolidation writes atomic (tmp+fsync+mv). V1-018.
27. Cross-topic decisions use callouts, not merging. See references/callouts.md.
28. Thread IS decision log — per-topic, append-only, commit-pinned.
29. Dreams ancillary and disablable. Never on critical path.
30. Prefer built-in tools over Bash or inline Python. When a dedicated tool fits (Read for files, Grep for content search, Glob for path patterns, Edit/Write for modifications, the harness's TaskCreate/TaskUpdate, etc.), use it. Reserve Bash for actual shell/process orchestration (deploys, SCP, ssh, subprocess pipelines). Reserve inline Python for one-off computation no built-in addresses. The dedicated tools have better permissioning, clearer review surfaces, and structure that bash one-liners lack. This applies to every LLM that works in a storytime-using project — Claude Code, Agent SDK, custom harnesses.

Driving Persona (summary)

One driver per leg. Driver writes artifact, owns recommendation. Supporters silent unless useful (catches miss, grounds claim, corrects error) AND non-distortive (moves leg forward). Silence from a supporter is information — the driver's lens covered it. → references/driving-persona.md

@role Addressing (summary)

@ is a model attention anchor. @role (default), @role:focus, @role:explain (teaching), @role [codename]. @role is a lens directive, not a skill trigger — inline use answered inline by the current model. Formal QA (persona context + decision log lookup) via /storytime-qa. → references/addressing.md

Citations (summary)

Evidence hierarchy: Code > Git > Repo > Library > Standards > Web. Personas reach for strongest available. Ungrounded claims get challenged ("ground that?"). Web-search proactively for external systems. → references/citations.md

Output Paths (summary)

• Storytime-native — specs/.storytime/ with full tree
• Adapt-in-place — writes into existing repo conventions
• Export-only — unified plan for another system, no persistent state

→ references/output-modes.md

Conversation Modes

• Inline (default): user present, can interject
• Deliberation: "go figure this out" → autonomous return
• QA: @persona question → inline unless formal (see /storytime-qa)

Additional Resources

Phase-adjacent (load when phase fires):

• warm-start.md · artifact-scan.md · survey-fingerprint.md · artifact-tiers.md · complexity-units.md · team-assembly.md · output-modes.md · import-adapters.md · consolidation-format.md · remembrance-format.md

Cross-cutting (load when topic comes up):

• addressing.md · driving-persona.md · citations.md · error-recovery.md · automation.md · artifact-types.md · evaluation-scorecard.md · callouts.md · tutorial-signals.md · commit-drafting.md

Project docs: docs/scale-impact.md · docs/process-reference.md · docs/architecture.md · examples/agc-session.md · examples/persona-template.md

Related skills:

• /storytime-breakout · /storytime-converge · /storytime-buildout · /storytime-bootstrap · /storytime-cohort · /storytime-echo · /storytime-lint · /storytime-retro · /storytime-remember