Plan VoidBot routine for cast swarm

docs/architecture/voidbot-routine-adoption-plan.md

@@ -0,0 +1,151 @@
+# VoidBot Routine Adoption Plan
+
+VoidBot now has the best live reference for the agent routine Ghostlight needs:
+work or rumination on heartbeat, periodic sleep for consolidation, typed
+MessagePack persistence through a CultCache-style contract, and vector-backed
+memory resonance. Ghostlight should adopt the routine without inheriting
+VoidBot's moderation job.
+
+## Imported Organs
+
+- **Heartbeat scheduler:** chooses work, rumination, or sleep/consolidation from
+  readiness, load, status, reaction windows, and pending turns.
+- **Completion-gated cooldown:** a participant cannot act again until its
+  previous heartbeat turn finishes; recovery starts at completion.
+- **Sleep cycle:** scheduled naps are not absence. They compress memory,
+  strengthen useful associations, cool saturated themes, refresh dreams, and
+  keep direct speech rare.
+- **CultCache persistence:** canonical state is MessagePack, routed by typed
+  document definitions, with JSON working projections for inspection/review.
+- **Vector memory:** memories carry semantic vector metadata and Qdrant points
+  so recall, resonance, incubation, and dreams are grounded in retrievable
+  structure.
+- **Thought organs:** analytic lane, associative lane, bridge, memory resonance,
+  incubation, and candidate intervention/dream queues.
+
+## Ghostlight Translation
+
+Ghostlight's swarm is the cast. The heartbeat initiative system therefore
+applies directly to characters taking turns within a scene, not merely to
+background maintenance workers.
+
+There are two scopes using the same timing law:
+
+- **scene heartbeat:** selects which in-scene character acts, reacts, waits,
+  hesitates, refuses, speaks, withholds, moves, or uses an object.
+- **maintenance heartbeat:** wakes offstage agents, cast records, coordinator
+  organs, appraisers, mutators, and fixture builders for rumination, sleep
+  consolidation, or reviewed maintenance.
+
+The point is one initiative physiology with different arenas, not two unrelated
+schedulers wearing each other's hat.
+
+```text
+heartbeat initiative
+  -> choose scene participant or maintenance organ
+  -> scene turn | reaction | wait | rumination | sleep
+  -> read typed state from CultCache-compatible store
+  -> optional Qdrant recall / memory resonance
+  -> produce visible action, event, appraisal, mutation, dream, or maintenance receipt
+  -> completion starts cooldown
+  -> write canonical MessagePack + JSON working projection
+```
+
+## Persistence Shape
+
+Ghostlight can start with the CultCacheTS contract directly for Node-side tools,
+but the Python-heavy validator/tooling surface needs a compatible file contract:
+
+- canonical store: single-file MessagePack envelope array
+- envelope: `key`, `type`, `payload`, `storedAt`
+- schema owner: code definitions, not hand-edited loose JSON
+- working projection: JSON, human-readable, safe for review, optionally stripped
+  of bulky vector values
+- write discipline: use an external lock or one coordinator writer; the
+  single-file store is not a database wearing a fake mustache
+
+Initial document types:
+
+- `ghostlight.agent_state`
+- `ghostlight.scene_state`
+- `ghostlight.initiative_schedule`
+- `ghostlight.scene_heartbeat_state`
+- `ghostlight.training_loop_bundle`
+- `ghostlight.runtime_heartbeat_state`
+- `ghostlight.memory_resonance`
+- `ghostlight.sleep_cycle`
+
+## Memory Vector Shape
+
+Qdrant becomes the retrieval layer for agent memory, not canonical social truth.
+
+Each memory-like record should keep:
+
+- stable memory id
+- kind: episodic, semantic, relationship, goal, value, musing, dream,
+  scene-memory, appraisal, mutation, or lore-grounding
+- summary text used for embedding
+- source hash
+- embedding backend/model/dimensions
+- Qdrant point id and collection
+- salience/confidence/update time
+- source refs into fixtures, lore, or reviewed artifacts
+
+Ghostlight's existing training docs already point toward retrieval-backed memory.
+This plan makes it operational instead of aspirational, the traditional route
+by which a research note stops being furniture.
+
+## Sleep Behavior
+
+During sleep turns:
+
+- do not generate ordinary scene turns unless the character is actually in the
+  scene and selected by the scene heartbeat
+- inspect memory resonance clusters and incubation entries
+- merge duplicate memories
+- prune stale or low-confidence residue
+- create or refresh at least one dream only when it compresses a useful seam
+- update `sleep_cycle.activeDreamThemes`, `lastDreamAt`, and
+  `lastDistillationSummary`
+- write a maintenance receipt for review
+
+Dreams are compressed operational memory, not glitter. If a dream survives, the
+next waking turn should be more coherent because of it.
+
+## Migration Slices
+
+1. **Contract fixture**
+   - Add a small CultCache-compatible MessagePack fixture and validator.
+   - Prove Python can read/write the envelope without losing schema shape.
+
+2. **Runtime heartbeat schema**
+   - Add `schemas/runtime-heartbeat-state.schema.json`.
+   - Include pending turn, sleep cycle, rumination, and completion receipts.
+
+3. **Agent state migration**
+   - Keep current JSON examples as projections.
+   - Add canonical MessagePack test store for one small agent-state fixture.
+
+4. **Memory resonance**
+   - Add a local memory-organ script that embeds memory summaries, writes Qdrant
+     points, builds resonance edges/clusters, and writes incubation state.
+
+5. **Sleep receipt**
+   - Add a sleep/consolidation artifact schema and validator.
+   - Require sleep receipts to name pruned/merged/strengthened memories.
+
+6. **Training-loop integration**
+   - Promote scene initiative schedules into the same heartbeat contract used
+     by the swarm.
+   - Let characters take scene turns through heartbeat initiative, then route
+     appraisal/mutation/coordinator maintenance through the same completion and
+     cooldown law.
+
+## Boundaries
+
+- Do not make Qdrant canonical truth.
+- Do not store raw vector blobs in every human working projection.
+- Do not let sleep turns mutate accepted training fixtures without reviewed
+  maintenance receipts.
+- Do not make every Ghostlight character into a background daemon before the
+  runtime heartbeat store can prove who is awake, asleep, pending, or complete.

notes/ghostlight-current-system-map.md

@@ -169,6 +169,13 @@ reaction windows; it does not decide what the actor wants or how other
 participants appraise the event. Every affected participant still appraises and
 mutates before the scheduler selects the next projected actor.
 
+VoidBot's newer heartbeat routine is the reference shape for turning that seam
+into a living swarm routine. For Ghostlight, the swarm is the cast: the same
+heartbeat initiative law selects characters taking turns within a scene, while
+offstage or maintenance heartbeats handle rumination, sleep consolidation,
+memory resonance, and reviewed upkeep. See
+`docs/architecture/voidbot-routine-adoption-plan.md`.
+
 ## Pruned Receipts
 
 Old prototype fixture receipts and model-runner files have been removed

notes/ghostlight-implementation-plan.md

@@ -43,6 +43,10 @@ not carry live next-action lists.
    - Training readiness is per organ. A receipt-complete scene loop does not
      automatically count as branch compiler, IF reviewer, visual artifact, or
      accepted full-fixture coverage.
+   - Adopt VoidBot's heartbeat routine as the cast initiative substrate:
+     characters take scene turns through heartbeat initiative, cooldown starts
+     after turn completion, and offstage agents use slower rumination/sleep
+     turns for memory consolidation instead of generating fake scene motion.
 
 4. Keep the Pallas fixture as the current scaffold reference.
    - Current fixture set:

state/evidence.jsonl

@@ -62,3 +62,4 @@
 {"ts": "2026-05-05T00:31:04+00:00", "type": "implementation", "status": "ok", "note": "Enabled Corvid visual replay plan and visual-enabled review receipts. Explicit hatch_3c_state plus a draft Blackbox Aviary 3C scene_set fixed branch spatial leakage; narrative/lore/spatial/visual reviewers now accept the fixture as draft, with visual artifacts recorded as collateral rather than core model training data."}
 {"ts": "2026-05-05T00:49:06+00:00", "type": "implementation", "status": "ok", "note": "Rebuilt a reusable BFL image batch script after the old runner had been cleaned out. Corvid BFL test generated 16 ignored local images from succinct abstract prompts; one proof prompt moderated until simplified, confirming Flux prompt batches should avoid emotionally loaded/ambiguous safety language and keep one strong visual symbol per frame."}
 {"ts": "2026-05-05T01:04:51+00:00", "type": "implementation", "status": "ok", "note": "Published Corvid Collective First Exodus to AetheriaLore as a cinematic illustrated Ink page. Site player now supports data-ink-mode=cinematic with section click-through, background image crossfades, animated text transitions, hidden state chips by default, and cohesive branch-card choices."}
+{"ts": "2026-05-06T00:25:00+00:00", "type": "architecture", "status": "ok", "note": "VoidBot's new routine should become Ghostlight's cast swarm timing and memory substrate: characters take scene turns through heartbeat initiative, offstage agents use slower rumination/sleep turns for consolidation, canonical state moves toward CultCache-style MessagePack with JSON projections, and memory resonance/incubation/dream state should be backed by Qdrant vectors. Captured the plan in docs/architecture/voidbot-routine-adoption-plan.md."}

state/map.yaml

@@ -80,6 +80,7 @@ live_architecture:
     - Actions carry scale, recovery, initiative cost, interruptibility, commitment, and local affordance basis.
     - Reaction windows are event-triggered opportunities with eligible actors, urgency, expiry, and allowed action scales; they are not automatic responses or hidden-state access.
     - Coordinator overrides must record a reason before the artifact can claim training value.
+    - VoidBot's newer heartbeat routine is the reference for turning this seam into the cast swarm's live timing law: characters take scene turns through heartbeat initiative, cooldown begins after turn completion, and offstage agents use slower rumination/sleep turns for memory consolidation, resonance, incubation, and dreams. See `docs/architecture/voidbot-routine-adoption-plan.md`.
   coordinator:
     - The coordinator is a trainable soft organ: it stages scenes, chooses next beats and acting agents, invokes specialist machinery, carries unresolved hooks, emits glue prose, and proposes game-engine-shaped state changes.
     - Future coordinator turns should be sandboxed as their own promptable worker outputs, with exact visible input, raw output, parsed coordinator artifact, review labels, and interventions preserved. Codex acts as meta-coordinator: it scopes, launches, reviews, repairs, and wires Ghostlight organs, but should not silently collapse the trainable coordinator organ back into omniscient chat.