Epiphany/notes/voidbot-routine-adoption-plan.md

VoidBot Routine Adoption Plan

VoidBot has grown the reference agent routine Epiphany should steal without cosplaying as a Discord moderator. The useful organs are:

• heartbeat turns that choose work, rumination, or sleep/consolidation
• cooldown that starts after a heartbeat turn completes, not when it launches
• nap windows for memory consolidation, pruning, dream residue, and incubation
• canonical MessagePack state through a CultCache-style typed document store
• CultCache-native inspection/debug tooling rather than long-lived JSON shadow state deliberately
• code-owned schemas instead of schema drift hiding in loose files
• semantic memory vectors in Qdrant with source hashes and compact dimensions
• memory resonance, incubation queues, analytic/associative thought lanes, and a bridge that decides speech, draft, hold, or silence
• repo identities with Persona state, Discord roles, pending mentions, jurisdiction, proposal authority, and a CTB/Final-Fantasy-style initiative scheduler

Project-Persona Direction

VoidBot is no longer only a moderator with retrieval. It now has the small working version of the Epiphany pitch: repos grow Personas.

The live pattern is:

• a repo identity has a Discord role and registered allowed channels
• that identity has repo-local Persona state, not just base Void personality
• first contact can birth the Persona through Epiphany repo terrain/personality scouting and reviewable startup artifacts
• role/display-name mentions become pending obligations, not direct hidden jobs
• an initiative scheduler chooses which Persona or system agent gets the next turn
• heat can prioritize identities, repos, participant kinds, turns, channels, or all participants
• active turns freeze initiative until the owner-Codex job or moderation lock completes
• Personas may advocate for repo needs and propose changes, while implementation authority still requires the configured human/review path

Epiphany should absorb that as native product direction. The user should talk to projects: "Aqua, what does AquaSynth need?", "Nibu, what is wrong with this lore arc?", "Mimir, what seam is drifting?" The project should schedule modeling, research, verification, memory maintenance, or a Persona response from its own typed state. The human should not have to deliver a complete implementation brief just to make the system start thinking.

Target Shape

Epiphany should treat this as a shared agent routine substrate, not as another specialist lane. The substrate owns scheduling, persistence, and memory maintenance; the coordinator owns authority.

coordinator pressure / user objective / swarm messages
  -> heartbeat scheduler
  -> work turn | rumination turn | sleep turn
  -> typed state read through CultCache-compatible store
  -> optional Qdrant memory recall / resonance
  -> bounded result, selfPatch, statePatch, dream, or callback
  -> completion receipt starts cooldown
  -> canonical MessagePack write + typed inspection/debug views

Scheduler Contract

Heartbeat state keeps:

• participant identity, role, speed, load, status, readiness, reaction bias, and pending turn
• pacing policy with work recovery, idle recovery, sleep multiplier, and completion-gated cooldown
• sleep cycle state: enabled, cadence, nap duration, current nap, next nap, dream count, active dream themes, and last distillation summary
• history of work, rumination, sleep, completion, and refused-repeat events

Turn kinds:

• work: coordinator-approved lane work, specialist result, Persona surface, or implementation continuation.
• rumination: role-local memory hygiene, selfPatch, stale habit detection, or low-pressure thought incubation.
• sleep: inward maintenance. The agent should distill/prune memory, strengthen resonance clusters, update incubation, write or refresh dreams, and speak only through Persona if something genuinely needs to surface.

No lane may receive another heartbeat while pending_turn.status == running. The completion receipt is the only normal path that starts cooldown. This is the part that keeps the heart from turning into a tiny manager with a stopwatch and no shame.

Repo Personas use the same law. A Persona can be heated because a human mentioned it, a repo is hot, a stream is live, or a group needs attention; that heat changes opportunity, not authority. The Persona still freezes while thinking, still returns through typed receipts, and still needs an explicit grant before crossing from proposal into canonical repo mutation.

Persistence Contract

Epiphany needs a Rust version of the CultCache contract because the app-server and epiphany-core are Rust and should not route core state through a Node helper just to persist its own mind.

Minimum Rust surface:

• DocumentType<T>: type id plus serde/schema metadata.
• CultCacheEnvelope: { key, type, payload, stored_at }.
• BackingStore: pull_all, push, delete, optional push_all.
• SingleFileMessagePackBackingStore: atomic whole-file MessagePack store with external lock expectation.
• CultCache: register types, load stores, validate payloads through typed serde structs, get/get_required/get_all/put/update/delete/snapshot.
• Typed inspection/debug views should come from CultLib/CultCache tooling so all wire-compatible runtimes share the same schema and display assumptions. Repo-local JSON projections are temporary compatibility shims only.

The schema should live in annotated Rust structs using serde plus generated JSON Schema where useful. MessagePack is canonical. JSON is the window, not the house.

Shared Persona-State Contract

Epiphany should not invent its own rival Persona ontology.

The shared cross-runtime public/person-shaped payload contract is now the GameCult Persona state shape, with Ghostlight and VoidBot as source lineage:

• payload schema id: https://gamecult.dev/cultnet/gamecult.persona_state.v0.schema.json
• payload schema version: gamecult.persona_state.v0
• intended CultNet document type: gamecult.persona_state.v0

That means Epiphany Persona, VoidBot repo Personas, and Ghostlight characters can exchange Persona state without turning every Epiphany work organ into a public personality simulation.

The shared contract uses candidateActions for generic forward pressure. VoidBot can expose the same list as voidbotProjection.candidateInterventions when repo-Persona routine language wants that noun, but the projection does not own the portable Persona contract. Persona documents also carry provenance and a public presentation surface so imported state can say where it came from, when it was exported, whether it is canonical/projection/import data, and how the public person should be rendered.

Source-specific fields belong in anchoredThought.extensions. That bag is preservation glue, not authority. Consumers may keep it round-tripped, but they must not let it steer shared Persona behavior unless they explicitly understand the originating contract.

The affect layer keeps semantic bones where they matter: social bonds name subject/object/kind/trust/tension, status reads name target/kind/confidence, and doctrine stances name principle/stance/action implication. Needs remain anchored thoughts because they behave like pressure records.

The baseline also requires a public presentation block, annotates timestamps as JSON Schema date-time, and gives every custom enum path a companion custom-label field. candidateActions.actions are typed action records now: action type, target, optional delivery target, readiness, risk, urgency, confidence, evidence, timestamps, and expiry. privateNotes stay simple for v0 interchange; if they begin carrying provenance or authority, promote them into typed private-note records instead of letting raw strings grow a steering wheel.

CultNet is not just ergonomic framing. It also carries CultLib-style auth and session semantics:

• shared connection key for both peers
• AES-GCM encrypted auth/session payloads
• server-side session-signing secret
• signed verify/reconnect tokens

So the target shape is:

• canonical work-organ state in MessagePack through CultCache-compatible storage
• typed CultNet document replication for live exchange
• Persona payloads for public/person-shaped agents

Narrow work-organ views stay derived and local. The shared person-state truth is Persona-shaped.

Initial document types:

• epiphany.thread_state
• epiphany.agent_memory
• epiphany.heartbeat_state
• gamecult.persona_state.v0
• epiphany.swarm_message
• later: planning, graph checkpoint, memory resonance, incubation, dreams

Vector Memory Contract

Epiphany already has Qdrant/Ollama retrieval for workspace source. Extend that pattern to agent memory:

• collection namespace: epiphany_memory_<instance_id>
• point id: stable memory id or hash of document type/key/path
• vector model: default qwen3-embedding:0.6b
• payload: agent id, role id, memory kind, source hash, schema version, summary, salience, confidence, updated time, source refs
• canonical MessagePack state stores vector metadata and optionally compact inline vectors; Qdrant stores the retrieval vector
• CultCache inspection views may strip vector values while preserving metadata

Memory resonance is not proof. It is "these things rhyme" evidence for rumination, Persona surfacing, and sleep consolidation.

Migration Slices

1. Contract extraction

   • Add Rust CultCache-compatible module in epiphany-core.
   • Add fixtures proving MessagePack round-trip, typed serde/schema validation, and CultCache inspection compatibility.

2. Heartbeat state migration

   • Move state/agent-heartbeats.json into a typed MessagePack document. Landed: the tracked JSON state file is retired and heartbeat init/status/tick/complete now route through the native CultCache store.
   • Add sleep_cycle, memory_resonance, and incubation fields.
   • Keep the current CLI contract stable while the backing store changes.

3. Organ and Persona memory migration

   • Store lean Epiphany work-organ memory as epiphany.agent_memory.
   • Store public/person-shaped Persona state as gamecult.persona_state.v0.
   • Use CultLib/CultCache inspection tools for human/debug review instead of preserving per-role JSON as a parallel source of truth.
   • Attach semantic vector metadata to episodic, semantic, relationship, goal, value, musing, and dream entries.

4. Memory organ

   • Add a bounded memory maintenance command that embeds missing memory vectors, refreshes Qdrant points, builds resonance edges/clusters, updates incubation, and writes dream residue during sleep turns.

5. Coordinator integration

   • Coordinator sees heartbeat sleep state, pending turns, resonance summaries, and memory maintenance results.
   • Coordinator may route work, accept/refuse selfPatch, or let the swarm sleep.

6. Aquarium visibility

   • Expose sleep state, dreams, resonance clusters, incubating thoughts, and pending turns as inspectable Aquarium surfaces.

Non-Goals

• Do not replace CRRC compaction yet.
• Do not make Qdrant canonical truth.
• Do not let sleep turns edit implementation code.
• Do not auto-accept specialist findings because a dream sounded good.
• Do not build a general database shrine before the typed MessagePack seam proves it deserves to live.