The room is the world model. Actions are the transition function. Views are the observation function. The audit log is the trajectory. What's missing? The reward signal. That's adaptive salience.

— kernel.ts, closing comment

Phase 1: Unification — ✅ Complete, verified live

• log_index table with backfill from existing audit entries
• CEL dependency extraction (hybrid AST walking + regex for pipe expressions)
• Dual-write to _actions, _views, _agents scopes alongside legacy tables
• Temporal query endpoints: /history/:scope/:key, prefix queries
• Structural event indexing (register/delete actions, views, agents)
• Observable vocabulary: state._actions.add_concept.description evaluable as CEL

Phase 2: Log-Based Temporal — ✅ Complete, verified live

• View value sampling piggybacked on dashboardPoll (60s interval)
• kind: "sample" audit entries indexed under _samples scope
• /samples/:viewId endpoint for temporal sparkline data
• marked.js (v15) + mermaid.js (v11.4) loaded in dashboard shell
• MarkdownSurface renders markdown + detects mermaid code blocks

Phase 3: Adaptive Salience — ✅ Complete, verified live

• Six structural signals: recency, dependency, authorship, directed, contest, delta
• Per-room configurable weights via _config.salience_weights
• _salience system view injected in agent context with transparent scoring
• /salience endpoint for full map with per-key signal breakdown
• Salience is observable — agents can inspect their own relevance map

Phase 4: Clean Break — ✅ Complete, verified live

• Dashboard poll reads state table only (single SQL query via poll-v8.ts)
• Surfaces.tsx decoupled from PollData — uses minimal inline interfaces
• Dashboard.tsx and Replay.tsx build SurfaceContext directly from v8 data
• v8ToLegacyPoll() adapter created, used, and deleted
• dashboardPoll() deleted (old 6-query legacy poll)
• PollData type deleted — frontend consumes PollDataV8 exclusively

Phase 5: Legacy Table Removal — ✅ Complete, verified live

• views table: DROPPED — all CRUD via _views scope only
• actions table: DROPPED — timer cooldown state in _timer field of _actions scope value
• agents table: DROPPED — tokens migrated to unified tokens table (82 agents), last_seen_seq in _agents scope, heartbeat via tokens.last_used_at + scope
• timers.ts: removed references to dropped tables
• replay.ts: already v8-clean (reconstructs from _audit scope, no legacy reads)
• Schema: DROP TABLE IF EXISTS views/actions/agents in migration chain

A sync room is exactly two things:

State — a finite partial function from located keys to versioned values. The current projection of all truth.

Log — an append-only sequence of events. The complete causal history.

Everything else is convention and interpretation.

state(room_id, scope, key) → { value, revision, hash, updated_at }
log(room_id, seq)          → { kind, ts, agent, payload }

The audit trail (now: log) is optimised for sequential replay. The log_index table makes it queryable by affected key:

CREATE TABLE log_index (
  room_id TEXT NOT NULL,
  seq     INTEGER NOT NULL,
  scope   TEXT NOT NULL,
  key     TEXT NOT NULL
);
CREATE INDEX idx_li_key ON log_index(room_id, scope, key);
CREATE INDEX idx_li_seq ON log_index(room_id, seq);

This replaces proposed state_history and view_journal tables. The log IS the history.

All legacy entity tables (actions, views, agents) have been dropped. The database now contains:

The state table is the single source of truth for all domain data. The tokens table handles all authentication. Everything else is MCP auth plumbing.

State and Log. One is the current projection, the other is causal history. Everything else — actions, views, agents, messages — is convention interpreted by scope prefix.