3.5 KiB
Sprint 168 Plan: Autonomous Production Completion Loop
Context
With stronger generation and hard gates, the final gap is orchestration: tools must iterate automatically until production gates are green, instead of stopping at first draft output.
This sprint delivers a closed-loop production completion workflow and validates it on the PriorityQueue benchmark and one additional non-trivial workload.
Goals
- Add iterative fix loop until compile/test/placeholder gates pass
- Integrate gate-driven retries into taskitem execution flow
- Record deterministic completion traces for training and audit
- Demonstrate full “spec -> production-ready code” completion on benchmark tasks
Steps
Step 1884: Gate-driven remediation planner (10 tests)
Implement remediation planner that maps gate failures to concrete next actions:
- syntax/type failures -> targeted regeneration/patch ops
- missing behavior -> test-guided method/body completion
- placeholders -> forced concrete-type/body refinement pass
Tests (10): failure classification, action selection correctness, deterministic plan order, no-op on green gates, multi-failure prioritization, retry budget handling, loop termination criteria, unsupported failure fallback, structured reason emission, reproducibility.
Step 1885: Closed-loop execution engine for production completion (12 tests)
Add loop:
- generate
- run gates
- remediate
- repeat until green or budget exhausted
Tests (12): converges on solvable sample, exits on budget exhaustion, preserves best-known artifact, deterministic iteration logs, handles intermittent failures, reports per-iteration deltas, no silent success on red gates, timeout safety, parallel-safe temp dirs, cancellation support, final status schema, performance bounds.
Step 1886: Taskitem pipeline integration (overall_ready semantics) (8 tests)
Update planning/execution integration so completion requires green gates:
- tasks remain incomplete if
overall_ready=false - queue readiness reflects gate status
Tests (8): task blocked on gate failure, task completes on all-green, queue-ready counts reflect gating, status propagation into validation report, backward-compatible fields for existing dashboards, deterministic ordering, no false-ready state, explicit blocker reasons.
Step 1887: Benchmark validation (PriorityQueue + secondary workload) (10 tests)
Run benchmark set:
- PriorityQueue (A/B task)
- one additional structured workload (e.g., cache indexer module)
Collect metrics:
- iterations to green
- compile/test pass rates
- placeholder count final
- wall time and token usage
Tests (10): benchmark harness runnable, both workloads reported, green-gate proof for completed run, metric schema completeness, reproducible run IDs, output artifact preservation, comparison vs baseline draft-only mode, failure-mode trace capture, summary generation, non-empty diagnostics on failure.
Step 1888: Sprint 168 Integration Summary (8 tests)
Add editor/src/Sprint168IntegrationSummary.h.
Record: steps_completed=5 (1884-1888), closed_loop_active=true,
production_gate_enforced=true, benchmark_completion_verified=true, success=true.
Tests (8): constructable summary struct, steps_completed==5, loop active, gate enforcement active, benchmark evidence path non-empty, no false-green completion, regression suite pass, success==true.
Architecture Gate
- Completion state must be gate-derived, never text-heuristic-only
- Every failed run must preserve actionable trace/evidence
- Benchmark outputs must be reproducible and auditable from saved artifacts