Skip to content

Phase B -- Goal lifecycle correctness

Status: Done

Goal

Make daemon goal pursuit honest: honor profile limits, handle MAX_STEPS explicitly, eliminate zombie active goals, unify custom GoalStrategy persistence with ExistenceLoop, and wire easy profile knobs (temperature, per-agent cost cap).

Why (problems addressed -- bullet list with severity)

  • P0: Profile max_steps ignored -- AgentCycleRunner builds Agent(...) without max_steps=profile.max_steps; DaemonAgentAdapter.pursue_goal() creates Task(instruction=...) with default 25 (src/hive/daemon/agent_cycle.py, src/hive/runtime/bridge.py).
  • P0: MAX_STEPS zombie goals -- DaemonAgentAdapter maps only FAILED to steps_failed=1; MAX_STEPS yields success=False with steps_failed=0, so neither complete nor abandon branch runs (agent_cycle.py lines 313--411).
  • P0: Active goal + AgentStatus.IDLE after partial pursuit -- agent forgets progress next heartbeat (feeds Phase C).
  • P1: Custom GoalStrategy skips validation -- agent_cycle.py saves goal directly without ExistenceLoop._validate_goal() / duplicate-active check.
  • P1: ExistenceLoop vs GoalStrategy save divergence -- existence saves inside generate_goal(); strategy path saves in agent_cycle.py with different hooks/events.
  • P2: temperature, max_cost_usd, max_tokens from AgentProfile not passed to runtime Agent or provider (src/hive/agents/profile.py, agent_context.py provider cache).
  • P2: assess_conditions bias -- only runs after pursuit with outcome.steps_done; MAX_STEPS / idle paths skew suffering stats.
ID Finding
LOOP-GOAL-01 Profile max_steps not wired
LOOP-GOAL-02 MAX_STEPS / zombie active goals
LOOP-GOAL-03 GoalStrategy bypasses validation
LOOP-GOAL-04 Temperature / per-agent budget not wired
LOOP-GOAL-05 GeneratedGoal metering only on generation path (partial)

Current state (files)

Area Location Behavior today
Pursuit entry src/hive/daemon/agent_cycle.py New Agent per cycle; no profile step limit
Bridge src/hive/runtime/bridge.py Task(instruction=...) only
Runtime loop src/hive/runtime/agent.py max_steps = task.max_steps or self._max_steps; returns TaskStatus.MAX_STEPS
Outcome mapping bridge.py success = status == COMPLETED only
Goal generation src/hive/agents/existence.py Validates + saves goal; returns GeneratedGoal with spend
Custom strategy src/hive/agents/goal_strategy.py Protocol only; caller saves raw string
Profile src/hive/agents/profile.py max_steps=20, temperature, max_cost_usd defined
Telemetry src/hive/logging/trace.py, GoalLog Events for completed/abandoned; gap for max_steps / parked

Proposed changes (numbered)

  1. Wire profile limits into pursuit:
  2. Pass max_steps=profile.max_steps, temperature=profile.temperature, and max_cost_usd=profile.max_cost_usd (if Agent supports) when constructing runtime Agent in agent_cycle.py.

  3. Pass max_steps into Task(...) in DaemonAgentAdapter.pursue_goal() (or set on Agent ctor consistently).

  4. Define explicit MAX_STEPS policy (document in docs/guide/daemon-mode.md):

  5. Recommended default: treat MAX_STEPS as continuable -- keep goal active, set status IDLE, emit goal_progress / max_steps_reached telemetry; Phase C adds transcript resume.
  6. Alternative (config flag): daemon.max_steps_policy: abandon | continue for operators who prefer auto-abandon.

  7. Map TaskStatus.MAX_STEPS in bridge.py to a distinct GoalOutcome flag (e.g. hit_step_limit: bool).

  8. Fix zombie / fall-through handling in agent_cycle.py:

  9. After pursuit, if not success / waiting_approval / abandoned, branch on hit_step_limit or explicit outcome.status.

  10. Never leave active goal with unlogged indeterminate outcome.

  11. Unify goal persistence:

  12. Extract shared save_generated_goal(agent_id, objective, *, validate=True) used by ExistenceLoop and GoalStrategy paths.
  13. Apply _validate_goal() for custom strategies unless GoalContext.skip_validation opt-in for trusted plugins.

  14. Emit consistent GoalLog + goal_generated hook in one place.

  15. Provider temperature: Ensure create_runtime_provider / cached provider calls pass profile temperature in generate_with_metadata (audit src/hive/daemon/loop.py, model providers).

  16. Telemetry:

  17. Add GoalLog event max_steps or extend event enum in logs.
  18. Update tests/test_narrative_in_prompt.py / add tests/test_goal_lifecycle.py for wiring assertions.

Non-goals

  • Cross-heartbeat transcript (Phase C).
  • Changing default profile max_steps value (20 vs 25 mismatch with runtime default -- document only).
  • Rewriting ExistenceLoop prompt strategy.

Risks / rollback

Risk Mitigation
Continue policy increases LLM spend on stuck goals Pair with profile max_steps + Phase D budget; optional cap on continuations per goal
Stricter validation breaks custom strategies skip_validation flag on GoalStrategy registration
Abandon policy loses work Default to continue; abandon opt-in

Rollback: revert bridge outcome mapping; feature-flag max_steps_policy.

Acceptance criteria (testable)

uv run pytest tests/test_goal_lifecycle.py tests/test_narrative_in_prompt.py tests/agents/test_existence.py -v
uv run pytest tests/test_daemon_integration.py -v -k goal
  • [x] Agent with profile max_steps: 3 stops pursuit at 3 tool/model steps (mock provider counts steps).
  • [x] TaskStatus.MAX_STEPS no longer leaves goal active with no recorded outcome (per chosen policy).
  • [x] Custom GoalStrategy that returns duplicate/rejected goal does not create store row when validation enabled.
  • [x] Agent constructed in daemon uses profile temperature in provider call (mock asserts kwargs).
  • [x] docs/guide/prompt-assembly.md and docs/guide/daemon-mode.md state max_steps + MAX_STEPS policy.

Suggested implementation order

  1. Failing tests for max_steps wiring + MAX_STEPS outcome.
  2. Bridge + agent_cycle outcome branches.
  3. Shared goal save/validate helper.
  4. Temperature / max_cost_usd wiring.
  5. Docs + telemetry.

Estimate

M (2--3 days).

Dependencies (prior phases)

None (start in parallel with Phase A). Blocks Phase C (policy must exist before resume) and Phase D (spend on all outcomes).