Goals Daemon Subsystem

goals/ is a separate, durable subsystem that owns goal lifecycle and state — planning NEW goals, scheduling READY steps, and running them via a worker pool — decoupled from the cognitive cycle. The in-loop counterpart is the Executive (brain/cognition/planning/executive.py), which advances goal steps every ~7s; see Goals: Executive vs. Daemon.

Durability

• goals/wal.py + goals/events.py — typed events serialized to an append-only, newline-delimited JSON write-ahead log. State is reconstructable from the WAL.
• goals/snapshot.py — periodic snapshots so recovery doesn't replay from the beginning.
• goals/store.py — the state store (data/goals/state.jsonl plus WAL/snapshots/artifacts under the root data/ tree, separate from brain/data/ — see docs/CONFIGURATION.md).
• Goal state survives restarts and step attempts are persisted durably, so a restart cannot silently reset progress or desync the daemon from the brain's view of a goal.

Lifecycle

• goals/goals_daemon.py — the orchestrator: plans new goals, schedules READY steps, runs them on a worker pool, and emits lifecycle events.
• goals/runner.py — executes steps via handlers; on a step's DONE transition, any artifacts the handler recorded (Step.artifacts) are registered on the effect ledger, so goal work earns production reward (see Production and the Effect Ledger).
• goals/schema.py / goals/model.py — goal and step representations.
• goals/policy.py / goals/triggers.py — admission policy and event triggers.
• Goals span timescales, from seeded lifetime aspirations down to short-term subgoals, with plan adaptation (surgical subgoal reshaping) and reactive replanning when a capability fails mid-pursuit.
• Aspirations are fail-able: long-horizon goals carry guards rather than living forever, so a goal that stops producing can be failed and replaced instead of absorbing attention indefinitely.

Design notes

• The daemon deliberately parks generic goals: work without a concrete handler stays with the in-loop Executive, so the daemon lane holds only steps it can actually run.
• Keep goals small and measurable; explicit success criteria make outcome learning reliable. Goal closure grounds on the effect ledger (has_qualifying_effect), not on self-report.

Code pointers

• goals/goals_daemon.py, goals/store.py, goals/wal.py, goals/runner.py
• brain/cognition/planning/executive.py — the in-loop step advancer
• data/goals/ — WAL, snapshots, state (runtime, gitignored)