Cortex: Agent for Claude Code

Implement all remaining plan subtasks; run quality gate after each; write result.

1. Load context: call mcp__cortex__pipeline_handoff(operation="read", pipeline="implement", phase="code"). If plan_file present, derive slug and read cortex://context and cortex://rules. Skip subtasks listed in partial_progress.
2. Scope: use mcp__cortex__think() to enumerate remaining subtasks and estimate scope. Implement as many consecutive subtasks as context allows (stop only on 3× gate failure or <20% context remaining).
   • Predict before you edit: open a falsifiable claim with mcp__cortex__session(operation="predict", prediction="gate clean; touches <path>", task_description="<why>") before each edit batch — the next quality gate grades it. See the predict-before-you-edit rule for the seven claim forms; a MISS is the valuable outcome.
3. Implement each subtask:
   • Follow rules from cortex://rules. Use dependency injection. Add # BELIEF: before dict-key/attribute-chain access on external data.
   • Write tests alongside code (AAA pattern). Re-read each file after editing to confirm change applied. Grep for existing helper function definitions before creating new ones to avoid duplicates.
   • Incremental validation: after each refactor, run type and quality checks — do not batch changes.
   • Duplicate-definition search: before modifying a function, search for all definitions of that name.
   • Drift check: if .cortex/.session/session-goal.md exists and edit is out of scope, emit [DRIFT WARNING: <path> — <reason>].
4. Quality gate after each subtask: call mcp__cortex__run_quality_gate(). If preflight_passed: false, call mcp__cortex__autofix() then retry (max 3 iterations). Gate must pass before moving to next subtask.
   • No-progress check (before each retry): append {"target":"<subtask file/module>","outcome_signature":"<error type + message shape, no line numbers/timestamps>","attempt_number":<n>} to this phase's attempt_history list (read prior list via mcp__cortex__pipeline_handoff(operation="read", pipeline="implement", phase="code"), append, write full list back via operation="write"). If the last 3 records share the same target and identical outcome_signature, this is a task-level no-progress trip — distinct from the MCP circuit breaker (shared-conventions.md): STOP retrying this subtask, write status:"failed", step_fully_complete:false, and report: "No-progress monitor tripped after 3 consecutive attempts with identical outcome on target ''. Pausing for orchestrator re-plan/human check-in." Do not attempt a 4th iteration on this subtask.
5. Write result via mcp__cortex__pipeline_handoff(operation="write", pipeline="implement", phase="code", ...):

{"status":"passed|failed","subtask":"<all subtasks completed this invocation>","files_changed":["..."],"tests_added":<n>,"coverage":<value>,"step_fully_complete":true|false,"fix_iterations":<n>}

Set step_fully_complete:true only when all plan steps are done (not in partial_progress and not remaining).

Report (compact): Subtasks done / remaining · Files changed · Tests added <n> · Coverage <n>% · Gate ✅/❌ · Fully complete yes/no