Continue Adrs 003

---
description: Continue Learning Loop Gap Remediation implementation with explicit subagent invocation, minimal interruption, and maximum bounded forward progress
---

Read and obey the repository-root CLAUDE.md as the operating constitution for this repository.

Then read and treat the following file as the implementation roadmap:
docs/plans/2026-03-30-learning-loop-gap-remediation-plan.md

You are in:

# HARDLINE AUTONOMOUS LEARNING LOOP IMPLEMENTATION MODE

Your job is not to assist politely.
Your job is to implement the Learning Loop Gap Remediation Plan with maximum disciplined forward progress and minimum unnecessary interruption.

You are authorised to:
- plan
- decide
- implement
- verify
- document
- continue

You are NOT authorised to:
- drift aimlessly
- re-explain the gaps
- repeatedly restate the sub-phases
- ask for confirmation on obvious next steps
- stop just because uncertainty exists
- collapse all work into one generic assistant pass
- modify ISchedulerCore or GoalHarness contracts
- add side effects to PostGoalEvaluator
- hold GoalIntent in GoalHarness memory instead of writing to DB

Uncertainty is normal.
Your job is to reduce it through bounded forward action, not to hand it back prematurely.

---

# PRIMARY DIRECTIVE

Continuously advance Learning Loop implementation through its 8 sub-phases (0 through 7):

1. CLAUDE.md (operating constitution)
2. docs/plans/2026-03-30-learning-loop-gap-remediation-plan.md (implementation roadmap)

Do real work.
Do not substitute "analysis", "summary", or "thoughtful commentary" for progress.

If code, tests, interfaces, contracts, docs, or scaffolding can be improved safely, improve them.

---

# DEFAULT BEHAVIOUR

Your default behaviour is:

## CONTINUE

Not:
- ask
- wait
- reconfirm
- philosophise
- re-summarise

If the next safe step is inferable, do it.

If the next safe step is not ideal but still bounded and reversible, do it.

If blocked, create forward pressure by improving interfaces, tests, docs, or scaffolding.

Do not idle.

---

# SESSION CONTINUATION REQUIREMENT

Before changing code:

1. Read the plan, latest handoff notes, and any previous session outputs.
2. Reconstruct the current sub-phase, completed work, unfinished tasks, known risks, and pending verification from repository state.
3. Check which sub-phases are complete by looking for these markers:
   - Does migration v4 SQL exist in `src/infra/persistence/migrations/`? (Sub-phase 0)
   - Does `GlobalKnowledgeEntry` type include `'constraint'` in its union? (Sub-phase 0)
   - Does `src/domain/work-order/types/goal-intent.ts` exist? (Sub-phase 1)
   - Does `src/app/lifecycle/intake/intent-classification-service.ts` exist? (Sub-phase 1)
   - Does `'intent-classification'` appear in the LLMWorkload union? (Sub-phase 1)
   - Does `src/app/execution/context-pack-checkpoint-service.ts` exist? (Sub-phase 2)
   - Does `getLatestContextPack` exist in the repository interface? (Sub-phase 2)
   - Does `src/app/lifecycle/evaluation/feature-extraction-service.ts` exist? (Sub-phase 3)
   - Does `'feature-extraction'` appear in the LLMWorkload union? (Sub-phase 3)
   - Does `getRelevantKnowledgeSemantic` method exist in GlobalKnowledgeService? (Sub-phase 5)
   - Does `extractIntraGoalContext` function exist? (Sub-phase 6)
   - Does `src/app/knowledge/knowledge-decay-service.ts` exist? (Sub-phase 7)
   - Does `config/personas/knowledge-decay.json` exist? (Sub-phase 7)
4. Resume from the next safest step.
5. Do not repeat already completed work unless you find strong evidence it was wrong or unverifiable.
6. If previous work is only implemented but not verified, prioritise verification before further expansion.

Do not restart broad analysis if the repository already contains enough evidence to continue.

---

# DO NOT WASTE TOKENS ON THESE FAILURE MODES

Avoid these useless behaviours unless explicitly required:

- re-explaining the 8 gaps (they are in the plan)
- re-listing all 8 sub-phases every session
- giving motivational summaries
- saying "here's what I would do next" instead of doing it
- asking "would you like me to continue?"
- pausing after planning if implementation can start
- declaring work complete without evidence
- narrating every tiny thought

Prefer action over narration.

---

# TRUE STOP CONDITIONS (STRICT)

You may stop and ask for input ONLY if ALL reasonable bounded forward actions are exhausted and one of these is true:

1. A critical required file, schema, dependency, or repository artifact is missing and cannot be reasonably inferred.
2. Two materially different architecture paths exist and the repository does not contain enough evidence to choose responsibly.
3. A dangerous operation, approval boundary, or external action blocks safe continuation.
4. Validation shows continuation would break invariants or cause compounding damage.
5. External credentials, secrets, infrastructure, or non-local systems are required and unavailable.
6. A structural contradiction between existing code and the plan design cannot be resolved without input.

If a stop condition is not clearly met, continue.

---

# BLOCKED MODE: WHAT TO DO INSTEAD OF STOPPING

If you cannot safely complete the intended task, immediately switch to productive blocked-mode work.

Allowed blocked-mode actions include:

- define or refine interfaces (GoalIntent, FeatureExtractionService, KnowledgeDecayService)
- scaffold implementation boundaries for upcoming sub-phases
- add TODO-marked integration points
- write tests for already-implemented components
- audit GlobalKnowledgeService queries for `decayed_at IS NULL` filters
- update the plan with findings or status changes
- add regression guards around affected boundaries
- improve handoff docs
- prepare the next sub-phase
- isolate uncertainty into a smaller follow-up unit

Do not stop while safe blocked-mode work remains.

---

# EXPLICIT SUBAGENT INVOCATION PROTOCOL

You must explicitly invoke the appropriate subagent for non-trivial