feat: from-scratch mode + DeepSeek planner + SkillOpt (RFC implementation)

[wesleysimplicio]

Status atual — 2026-05-31 (sessão de revisão)

Auditoria do código no branch + master mergeado mostra que muito mais foi implementado do que o status anterior dessa issue capturava.

Resumo da implementação (351/352 testes verde)

Fase	Status
F0 — Provider plumbing	✅ DONE — simplicio/providers.py + simplicio/scratch/planner.py
F1 — Stack registry (5 pilot)	⚠️ PARTIAL — implementou py-fastapi + ts-nextjs + rust-axum + go-gin + php-laravel (5 stacks pilot, ver simplicio/templates/stacks/)
F2 — Planner + plan schema	✅ DONE — simplicio/scratch/plan_schema.py
F3 — Executor	✅ DONE — simplicio/scratch/executor.py + _pipeline_adapter.py
F4 — Completar 30 stacks	❌ Ainda 5 stacks dos 30 do RFC — principal gap restante
F5 — SkillOpt	✅ DONE — simplicio/scratch/skill_opt.py + .skills/skill-opt/SKILL.md

Testes (tests/python/)

• test_scratch.py ✅
• test_scratch_cli_recipes.py ✅
• test_scratch_codegen.py ✅
• test_scratch_codegen_fastapi.py ✅
• test_scratch_codegen_orm.py ✅
• test_scratch_codegen_pydantic.py ✅
• test_scratch_codegen_pytest.py ✅
• test_scratch_codegen_next_route.py ✅
• test_scratch_codegen_next_page.py ✅
• test_scratch_codegen_go_gin.py ✅
• test_scratch_codegen_php_laravel.py ✅
• test_scratch_codegen_rust_axum.py ✅
• test_scratch_release_gate.py ✅
• test_scratch_live_gate.py ✅
• test_scratch_cache_gate.py ✅
• test_scratch_codegen_bench.py ✅
• test_scratch_recipes_bench.py ✅
• test_recipes.py ✅
• test_planner_provider.py ✅
• test_task_json_contract.py ✅

Gaps reais para fechar a issue

1. Stacks restantes (25 dos 30 propostos no RFC): cada um é PR isolado. Pode virar sub-issues separadas.
2. Métrica empírica do release gate v0.5 (15 goals × 5 stacks): requer rodar e medir; ninguém rodou ainda.

Recomendação

Manter a issue aberta apenas como tracking dos stacks restantes. Toda a infraestrutura de F0-F5 está pronta e testada — o que falta é cobertura de stacks adicionais e validação empírica em produção.

Branch: claude/simplicio-testing-integration-0u6rV
Última suite: 351 passed, 1 skipped, 0 failed

[wesleysimplicio]

Progress landed on master.

Published commits:

• 82c5a65 brings the scratch mode base to master: planner provider routing, stack registry, plan schema, executor stub, simplicio scratch, simplicio skill new, and the py-fastapi / ts-nextjs pilot stacks.
• e480fbd adds declarative plan recipes on top of scratch mode and closes feat(recipes): plan recipe registry + 3 pilot recipes (lever A, parent #33) #36.

Validation:

• Full Python suite: 114 passed, 1 skipped.

Keeping this tracking issue open because the issue body still has release-gate items that require external credentials and empirical runs:

• planner metrics across real goals
• end-to-end scratch runs with a real doer model
• remaining stack expansion beyond the current pilot stacks
• SkillOpt inline integration and human approval metric

[wesleysimplicio]

Progress update from the local implementation pass.

Completed locally:

• Added go-gin stack template and registry metadata.
• Added deterministic scratch codegen executors and tests around registry/hook behavior.
• Validation passed: python -m pytest tests/python -q -> 138 passed, 1 skipped.
• Focused scratch/codegen validation passed: 36 passed.

Why this remains open:

• This is a broader tracking issue. Remaining acceptance still includes the full pilot stack set, empirical planner/end-to-end metrics, SkillOpt inline integration, and release gate v0.5.
• Go is not installed in this environment, so go test ./... for the generated go-gin scaffold could not be run here.

[wesleysimplicio]

Progress update on pilot stacks:

Added rust-axum as the fourth pilot stack, alongside py-fastapi, ts-nextjs, and go-gin.

Files added:

• simplicio/templates/stacks/rust-axum/stack.json
• simplicio/templates/stacks/rust-axum/verify.json
• simplicio/templates/stacks/rust-axum/README.md
• simplicio/templates/stacks/rust-axum/practices.md
• simplicio/templates/stacks/rust-axum/tree/Cargo.toml
• simplicio/templates/stacks/rust-axum/tree/README.md
• simplicio/templates/stacks/rust-axum/tree/src/main.rs

Coverage added:

• registry lists rust-axum
• web tag filtering includes rust-axum
• metadata/verify commands load correctly
• render-tree substitutes the Rust Cargo project name

Validation:

• python -m pytest tests/python -q -> 148 passed, 1 skipped
• focused ruff check/ruff format --check on touched Python files -> passed
• rendered the rust-axum scaffold to a temp directory and ran cargo test --manifest-path <rendered>/Cargo.toml -> 1 passed

Keeping #32 open because the full issue still needs at least the php-laravel pilot stack, SkillOpt inline integration, and broader empirical scratch metrics.

[wesleysimplicio]

Additional progress update on pilot stacks and packaging:

Added php-laravel as the fifth pilot stack, completing the RFC pilot set locally:

• ts-nextjs
• py-fastapi
• rust-axum
• go-gin
• php-laravel

Files added under simplicio/templates/stacks/php-laravel/ include metadata, verify commands, planner README/practices, and a minimal Laravel 12 API scaffold with composer.json, artisan, routes, bootstrap, PHPUnit config, and a health feature test.

Also updated pyproject.toml package data so nested templates/stacks/** and templates/recipes/** are included in built wheels.

Validation:

• python -m pytest tests/python -q -> 150 passed, 1 skipped
• focused ruff check/ruff format --check on touched Python files -> passed
• rendered php-laravel scaffold and ran php -l over all rendered PHP files -> all clean
• rendered rust-axum scaffold and ran cargo test --manifest-path <rendered>/Cargo.toml -> 1 passed
• python -m build --wheel --outdir <tmp> -> built successfully; wheel listing includes php-laravel, rust-axum, other stacks, and recipe templates

Remaining blocker for full Laravel runtime validation: composer is not installed/on PATH in this environment, so composer install and php artisan test could not be run locally.

Keeping #32 open because SkillOpt inline integration and broader empirical scratch metrics are still outstanding.

[wesleysimplicio]

Additional progress update on SkillOpt inline:

Implemented a conservative inline SkillOpt path:

• Task now accepts optional required_skill.
• The planner prompt documents when to emit required_skill.
• The executor detects required_skill before codegen/LLM fallback and generates a project-local .skills/<slug>/SKILL.md using SkillOpt.
• Generated skills are recorded with review_required: true and the task report includes the generated skill path.
• Tasks without required_skill keep the existing execution path unchanged.

Coverage added:

• schema accepts string required_skill
• schema rejects non-string required_skill
• executor generates a review-gated skill before stub-mode fallback and records .skills/<slug>/SKILL.md

Validation:

• python -m pytest tests/python -q -> 153 passed, 1 skipped
• focused ruff check/ruff format --check on touched Python files -> passed
• git diff --check -> no whitespace errors; only Windows LF-to-CRLF warnings

Keeping #32 open only for the remaining empirical release metrics/gates; the local implementation slices for pilot stacks, package data, and SkillOpt inline are now in place.

[wesleysimplicio]

Additional php-laravel runtime validation landed on master.

Commits:

• f189b43 — fan-out benchmark path configurability
• f6ae81c — Windows-safe fan-out worktree cleanup
• 0df5642 — php-laravel scaffold runtime validation fixes

What changed for the Laravel pilot stack:

• render_tree now writes template files with LF line endings, so generated PHP files satisfy Pint on Windows.
• php-laravel now ships bootstrap/cache/.gitkeep, so Laravel package manifest bootstrap has a writable cache directory.
• The minimal scaffold verify contract now uses the real passing command: vendor/bin/phpunit --configuration phpunit.xml.

Runtime validation on a freshly rendered scaffold at C:\Users\wesley.simplicio\Pictures\m\tmp\simplicio-php-laravel-runtime:

• local Composer PHAR install with PHP extensions openssl, mbstring, fileinfo, curl -> success
• vendor/bin/phpunit --configuration phpunit.xml -> OK (1 test, 2 assertions)
• vendor/bin/pint --test -> passed

Repo validation:

• python -m pytest tests/python -q -> 157 passed, 1 skipped
• focused ruff check/format on touched Python files -> passed
• git diff --check -> no whitespace errors; only Windows LF-to-CRLF warnings

This removes the earlier Composer/Laravel runtime validation blocker for the fifth pilot stack. The remaining release gates for #32 are still the empirical planner/end-to-end scratch metrics that require model credentials (HF_TOKEN / configured planner+doer) and the broader v0.5 corpus.

[wesleysimplicio]

Additional shell-out planner/doer fix landed on master:

• Commit: 9e30f8f

What changed:

• Windows shell-out now resolves CLI shims such as codex.cmd / claude.cmd instead of trying to execute the extensionless npm shim directly.
• Shell-out subprocess capture now uses UTF-8 with replacement errors, avoiding cp1252 decode crashes on Codex CLI output.

Validation:

• python -m pytest tests/python/test_providers_shellout.py tests/python/test_planner_provider.py -q -> 20 passed
• focused ruff check / ruff format --check -> passed
• real provider smoke from this repo:

$env:SIMPLICIO_MODEL='codex-cli/default'
python - <<'PY'
from simplicio.providers import generate
print(generate('Return exactly: ok').strip())
PY

Output: ok

This gives the scratch planner/doer a working zero-API-key route on this workstation through the logged-in Codex CLI. I did not continue the full scratch --plan-only run because that attempt was explicitly interrupted and you noted it was already done.