v1.46.0 — Framework 4R que encuentra lo que importa y para
Bumpeo de minor que reescribe todo el framework de review 4R alrededor de una idea: un review tiene que encontrar defectos que importan y despues PARAR.
- Reviews que frenan a tiempo: los diffs se clasifican por complejidad — los triviales se saltan el review, los estandar reciben un lente enfocado, y los de alto riesgo llevan cobertura 4R completa con limites de costo explicitos. Los sweep budgets frenan la primera pasada tras un barrido estandar o como maximo dos por lente.
- Verificacion adversarial: los hallazgos de severidad alta pasan por una fase de refutacion antes de mover una linea — un refutador en reviews estandar, tres en full-4R (correctitud, impacto, reproducibilidad).
- Remediacion acotada: los convergence budgets cortan la remediacion en dos rondas de fix, y cada re-review recibe solo el ledger persistido y el diff del fix. Ojo con esto si tenias loops eternos de review.
Tests unitarios y E2E pasaron en Ubuntu, Arch y Fedora; GoReleaser publico archivos para macOS, Linux y Windows en AMD64 y ARM64.
v1.45.0 — Perfiles GPT-5.6 en Codex e installs mas seguros
Bumpeo de minor que trae los perfiles GPT-5.6 a Codex, endurece el contrato entre memoria y respuesta, y hace mas seguro el install/upgrade.
- Perfiles GPT-5.6 en Codex (#1067): se agregan
gpt-5.6-sol,gpt-5.6-terraygpt-5.6-lunaal catalogo, con presets Low Cost / Recommended / Powerful y defaults de reasoning-effort por lane SDD. Ojo: ahora se requiere Codex>=0.144.0antes de escribir perfiles GPT-5.6 — si falta o esta viejo, falla con un comando de upgrade en vez de escribir config invalida. - La memoria nunca reemplaza la respuesta (#1075): se refuerza el contrato de Engram — guardar memoria es bookkeeping, no una excusa para no contestarle al usuario. Los agentes tienen que componer y enviar la respuesta despues de capturar memoria, incluso en paths degradados. Se agregaron tests adversariales para que nadie invierta esto sin querer.
- Installs y upgrades mas seguros: Homebrew detecta si el binario es formula o cask y usa la estrategia correcta; instalaciones duales, binarios sombra y migracion de cask legacy fallan en cerrado con guia accionable. El opt-out de persona ahora lo respeta la TUI (#1060) y el template de OpenSpec vuelve a ser YAML valido (#1008).
No hay fallback a modelos legacy: actualiza Codex antes de correr un install o sync que apunte a Codex. Suite Go completa, unit y E2E pasaron en Ubuntu, Arch y Fedora; GoReleaser publico archivos y checksums para macOS, Linux y Windows.
v1.46.0 — 4R framework that finds what matters, then stops
Minor bump that rewrites the entire 4R review framework around one idea: a review has to find defects that matter and then STOP.
- Reviews that stop on time: diffs are classified by complexity — trivial ones skip review, standard ones get one focused lens, and high-risk ones get full 4R coverage with explicit cost limits. Sweep budgets cap the first pass at one standard sweep or at most two per lens.
- Adversarial verification: high-severity findings go through a refutation phase before a single line moves — one refuter on standard reviews, three on full-4R (correctness, impact, reproducibility).
- Bounded remediation: convergence budgets cap remediation at two fix rounds, and each re-review only gets the persisted ledger and the fix diff. Watch out for this if you had endless review loops before.
Unit and E2E tests passed on Ubuntu, Arch, and Fedora; GoReleaser published artifacts for macOS, Linux, and Windows on AMD64 and ARM64.
v1.45.0 — GPT-5.6 Codex profiles and safer installs
Minor bump that brings GPT-5.6 profiles to Codex, hardens the contract between memory and response, and makes install/upgrade safer.
- GPT-5.6 profiles in Codex (#1067): adds
gpt-5.6-sol,gpt-5.6-terra, andgpt-5.6-lunato the catalog, with Low Cost / Recommended / Powerful presets and reasoning-effort defaults per SDD lane. Heads up: Codex>=0.144.0is now required before writing GPT-5.6 profiles — if it's missing or outdated, it fails closed with an upgrade command instead of writing invalid config. - Memory never replaces the response (#1075): reinforces the Engram contract — saving memory is bookkeeping, not an excuse to skip replying to the user. Agents must compose and send the response after capturing memory, even on degraded paths. Adversarial tests were added so nobody flips this by accident.
- Safer installs and upgrades: Homebrew detects whether the binary is a formula or cask and uses the right strategy; dual installs, shadow binaries, and legacy cask migration fail closed with actionable guidance. Persona opt-out is now respected by the TUI (#1060), and the OpenSpec template is valid YAML again (#1008).
No fallback to legacy models: update Codex before running an install or sync that targets Codex. Full Go suite, unit and E2E, passed on Ubuntu, Arch, and Fedora; GoReleaser published artifacts and checksums for macOS, Linux, and Windows.
v1.46.0 — Framework 4R qui trouve l'essentiel puis s'arrête
Minor bump that rewrites the entire 4R review framework around one idea: a review has to find defects that matter and then STOP.
- Reviews that stop on time: diffs are classified by complexity — trivial ones skip review, standard ones get one focused lens, and high-risk ones get full 4R coverage with explicit cost limits. Sweep budgets cap the first pass at one standard sweep or at most two per lens.
- Adversarial verification: high-severity findings go through a refutation phase before a single line moves — one refuter on standard reviews, three on full-4R (correctness, impact, reproducibility).
- Bounded remediation: convergence budgets cap remediation at two fix rounds, and each re-review only gets the persisted ledger and the fix diff. Watch out for this if you had endless review loops before.
Unit and E2E tests passed on Ubuntu, Arch, and Fedora; GoReleaser published artifacts for macOS, Linux, and Windows on AMD64 and ARM64.
v1.45.0 — Profils GPT-5.6 pour Codex et installs plus sûrs
Minor bump that brings GPT-5.6 profiles to Codex, hardens the contract between memory and response, and makes install/upgrade safer.
- GPT-5.6 profiles in Codex (#1067): adds
gpt-5.6-sol,gpt-5.6-terra, andgpt-5.6-lunato the catalog, with Low Cost / Recommended / Powerful presets and reasoning-effort defaults per SDD lane. Heads up: Codex>=0.144.0is now required before writing GPT-5.6 profiles — if it's missing or outdated, it fails closed with an upgrade command instead of writing invalid config. - Memory never replaces the response (#1075): reinforces the Engram contract — saving memory is bookkeeping, not an excuse to skip replying to the user. Agents must compose and send the response after capturing memory, even on degraded paths. Adversarial tests were added so nobody flips this by accident.
- Safer installs and upgrades: Homebrew detects whether the binary is a formula or cask and uses the right strategy; dual installs, shadow binaries, and legacy cask migration fail closed with actionable guidance. Persona opt-out is now respected by the TUI (#1060), and the OpenSpec template is valid YAML again (#1008).
No fallback to legacy models: update Codex before running an install or sync that targets Codex. Full Go suite, unit and E2E, passed on Ubuntu, Arch, and Fedora; GoReleaser published artifacts and checksums for macOS, Linux, and Windows.