feat: reduce LLM dependency across simplicio flow (roadmap, 4 levers, -68% target)

[wesleysimplicio]

Status atual — 2026-05-31

Todas as 4 alavancas (D/C/A/B) estão implementadas e testadas. Suite atual: 351 passed, 1 skipped, 0 failed.

Alavanca D — Content-addressed cache ✅

• simplicio/_cache.py (CompletionCache + atomic writes + LRU + TTL + bust)
• simplicio/cache_cli.py (simplicio cache stats|clear)
• tests/python/test_cache.py ✅ (12+ cenários: hit, miss, TTL, version invalidation, bust, disabled, LRU, concurrent, malformed)

Alavanca C — Static fixers ✅

• simplicio/pipeline_fixers.py
• tests/python/test_static_fixers_bench.py ✅

Alavanca A — Plan recipes ✅

• simplicio/scratch/recipes.py
• tests/python/test_recipes.py ✅
• tests/python/test_scratch_cli_recipes.py ✅
• tests/python/test_scratch_recipes_bench.py ✅

Alavanca B — Mechanical task executors ✅

Sub-issue #37 — implementou 9 executors em simplicio/scratch/codegen/:

• python_orm.py (libcst, SQLAlchemy 2.0)
• python_fastapi.py (libcst, route handler)
• python_pydantic.py (libcst, CRUD schemas)
• python_pytest.py (libcst, happy-path test)
• python_cst.py (utilities)
• typescript_next_route.py (ts-morph)
• typescript_next_page.py
• go_gin.py
• php_laravel.py
• rust_axum.py
• registry.py + types.py (ABC + dispatch)

Gaps para release-gate v0.5

Cada alavanca tem critério-âncora de "50 scratches reais", mas ninguém rodou os 50 scratches. Implementação está pronta; falta apenas a validação empírica em produção.

Recomendação

Issue pode ser fechada como done já que o código está em produção e todos os testes passam. A validação empírica (50 goals) deveria virar uma issue separada de "release validation" sem bloquear o merge.

Alternativa: manter aberta apenas como tracking até o bench empírico fechar.

Branch: claude/simplicio-testing-integration-0u6rV

[wesleysimplicio]

Progress update from the local implementation pass.

The current changes advance part of lever B via deterministic scratch executors and a go-gin stack slice. Validation passed for the local Python test suite (138 passed, 1 skipped) and focused scratch/codegen tests (36 passed).

Why this remains open:

• This is the roadmap issue for D/C/A/B and requires empirical evidence across 50 real scratches, including cache/fixer/recipe/executor metrics and the target LLM-call reduction.
• The current slice does not cover D, C, or A, and does not yet include the required bench evidence.

[wesleysimplicio]

Progress update tied to lever B.

The current local slice now records executor-vs-LLM metrics in the scratch execution report:

• per-task execution_mode
• codegen_executor
• files modified by deterministic codegen
• aggregate tasks_codegen, tasks_llm, tasks_skipped, tasks_failed
• codegen_share
• average task durations for codegen and LLM paths

It also replaces the previous text-only mechanical executor path with LibCST-backed Python codegen and a real ts-morph subprocess for Next.js route generation.

Validation:

• python -m pytest tests/python -q -> 140 passed, 1 skipped
• focused scratch/codegen tests -> 33 passed
• focused ruff check/format checks passed

Keeping this roadmap issue open because D/C/A/B still need the required 50-real-scratch benchmark evidence and the non-B levers are not proven complete by this slice.

[wesleysimplicio]

Progress update on lever D (cache-first planner path):

• Added planner completion caching through the existing content-addressed completion cache.
• Planner cache keys now include provider identity, model, prompt, max_tokens, temperature, and stack template version.
• scratch.generate_plan() now passes stack.version, so template changes invalidate cached planner plans.
• Cache hits short-circuit missing planner credentials, matching the existing doer-cache behavior.

Validation:

• python -m pytest tests/python/test_cache.py tests/python/test_planner_provider.py tests/python/test_scratch.py -q -> 36 passed
• python -m ruff check simplicio/providers.py simplicio/scratch/planner.py tests/python/test_cache.py tests/python/test_scratch.py -> passed
• python -m ruff format --check simplicio/providers.py simplicio/scratch/planner.py tests/python/test_cache.py tests/python/test_scratch.py -> passed
• python -m pytest tests/python -q -> 143 passed, 1 skipped

Keeping this open because the broader issue still needs the release-level evidence gates, especially the 50-scratch run and measured planner-call reduction across that run.