v1.15.4 — Hardening de mem_save y prompt capture
Patch post Judgment Day review. Bug fixes: mem_save ahora trackea actividad contra el session_id resuelto; rutas de handler tolerantes a SessionActivity nil sin panics; fallas de prompt capture quedan best-effort y no rompen el save. Tests: regresiones para session_id explícito, nil-activity, fallas no fatales y dedupe sync. Docs: aclarado que solo mem_save y mem_save_prompt aceptan recovery override de proyecto ambiguo.
v1.15.3 — capture_prompt en mem_save
Nuevo flag mem_save.capture_prompt (default true): cuando el proceso MCP tiene contexto de prompt para el mismo proyecto/sesión, mem_save guarda el prompt junto a la observación. Pipelines automatizados pueden optar por capture_prompt: false. Es process-local — el mismo lifecycle MCP debe recibir el prompt antes de un mem_save posterior.
v1.15.2 — Recovery seguro para escrituras con proyecto ambiguo
Cuando MCP arranca desde un parent dir con múltiples repos, los agentes ahora pueden recuperar pidiendo al usuario que elija uno de los available_projects reportados. Validación estricta: solo acepta match exacto tras trim, evitando colisiones tipo foo-bar vs foo--bar.
v1.15.0 — Conflict surfacing, doctor, dashboard fix, Windows hardening
Release de hardening — la más grande de la semana. Memory conflicts: nuevo workflow de audit y semantic judge para contradicciones, comandos engram conflicts (list/show/scan/stats/replay), MCP mem_compare y cloud sync de relaciones. Doctor: engram doctor read-only, MCP mem_doctor, y engram doctor repair con flags --plan/--dry-run/--apply + backup SQLite previo. Cloud dashboard: el contador Browser ahora refleja estado actual (no totales históricos), preservando chunks históricos para audit. Windows plugin: UserPromptSubmit hardened para Git Bash/MSYS2, fork-light safe mode, fallback PowerShell para entornos enterprise. Beta: setup Docker Compose aislado para testing comunitario.
v1.14.2 → v1.14.13 — Patches de cierre del ciclo 1.14
12 patches entre 27 abr y 30 abr antes del bump a 1.15. Mayormente automáticos por github-actions[bot] — refinements del pipeline de release y hardening incremental antes del minor bump. Si te interesa el detalle por patch, revisá el changelog completo en GitHub.
v1.25.4 — Registra 4 skills faltantes en Go build
Fix puntual: las skills chained-pr, cognitive-doc-design, comment-writer y work-unit-commits existían en skills/ desde v1.25.0 pero faltaban en el Go skill registry/catalog/embedded assets, así que gentle-ai install y el TUI skill picker no las encontraban. Ahora están registradas con ID constants, catalog entries, preset registration y embedded SKILL.md.
v1.25.0 — Sustainable review workflow + rename a gentle-orchestrator
Minor bump con foco en review process. Sustainable review workflow para SDD changes, con planning de delivery strategy y review workload forecasting — los cambios grandes pueden planearse como single PR, chained PRs o phased delivery. Rename: el conductor canónico OpenCode SDD pasa de sdd-orchestrator a gentle-orchestrator, manteniendo migration support para configs legacy. PR size budget: nueva policy con override size:exception aplicable por maintainer para PRs aggregate o tracker. E2E: ahora usa shim local determinístico de Engram salvo que se opte explícitamente por live download testing.
v1.24.0 — Community plugins para OpenCode
Registro opcional de OpenCode Community Plugins para sub-agent-statusline y sdd-engram-plugin. Nuevo flow standalone de main menu para registrar community plugins sin correr el installer completo. gentle-ai escribe los plugins TUI seleccionados declarativamente a ~/.config/opencode/tui.json; OpenCode los resuelve y carga al próximo start/reload. gentle-ai update detecta versiones de plugin una vez que OpenCode las materializa bajo ~/.config/opencode/node_modules. Bonus: fix de upgrade/sync TUI pollution por backup diagnostics verbosos.
v1.24.1 → v1.24.3 + v1.25.1 → v1.25.3 — Patches de mantenimiento
6 patches entre 27 abr y 3 may, intercalados con los minor bumps 1.24.0 y 1.25.0. Mayormente automáticos por github-actions[bot]. Para detalle por patch, ver el listado completo en GitHub.
Tendencias Destacadas
Semana intensa: 17 releases en engram + 9 en gentle-ai. Cadencia automatizada vía github-actions[bot] sugiere release-please o similar.
Foco común en hardening operacional: doctor diagnostics + Windows hardening en engram, sustainable review workflow en gentle-ai.
v1.15.4 — mem_save and prompt capture hardening
Patch after Judgment Day review. Bug fixes: mem_save now tracks activity against the resolved session_id; handler paths are nil-safe for missing SessionActivity; prompt capture failures remain best-effort. Tests: regressions for explicit session_id, nil-activity, non-fatal capture, and dedupe sync. Docs: clarified that only mem_save and mem_save_prompt accept the ambiguous-project recovery override.
v1.15.3 — capture_prompt added to mem_save
New mem_save.capture_prompt boolean (defaults true): when the MCP process has prompt context for the same project/session, mem_save records that prompt alongside the observation. Automated pipelines can opt out with capture_prompt: false. Capture is process-local — the same MCP lifecycle must receive the prompt before a later mem_save.
v1.15.2 — Safe recovery for ambiguous-project writes
When MCP starts from an ambiguous parent dir with multiple repos, agents can now recover by asking the user to choose one of the reported available_projects. Strict validation: accepts only exact match after trimming, preventing collisions like foo-bar vs foo--bar.
v1.15.0 — Conflict surfacing, doctor, dashboard fix, Windows hardening
Hardening release — the biggest of the week. Memory conflicts: new audit and semantic judge workflows, engram conflicts commands (list/show/scan/stats/replay), MCP mem_compare, and cloud sync for relations. Doctor: read-only engram doctor, MCP mem_doctor, and engram doctor repair with explicit --plan/--dry-run/--apply modes + SQLite backup. Cloud dashboard: Browser counts now reflect current state (not historical totals), preserving chunks for audit. Windows plugin: hardened UserPromptSubmit for Git Bash/MSYS2, fork-light safe mode, PowerShell fallback for enterprise. Beta: isolated Docker Compose setup for community testing.
v1.14.2 → v1.14.13 — Closing patches for the 1.14 cycle
12 patches between Apr 27 and Apr 30 before the 1.15 minor bump. Mostly automated github-actions[bot] publications — release pipeline refinements and incremental hardening leading up to the minor bump. See the full changelog on GitHub for per-patch detail.
v1.25.4 — Registers 4 missing skills in Go build
Targeted fix: skills chained-pr, cognitive-doc-design, comment-writer, and work-unit-commits existed in skills/ since v1.25.0 but were missing from the Go skill registry, catalog, and embedded assets — so gentle-ai install and the TUI skill picker could not find them. Now registered with ID constants, catalog entries, preset registration, and embedded SKILL.md.
v1.25.0 — Sustainable review workflow + rename to gentle-orchestrator
Minor bump focused on review process. Sustainable review workflow for SDD changes with delivery strategy planning and review workload forecasting — large changes can be planned as single PR, chained PRs, or phased delivery. Rename: the canonical OpenCode SDD conductor moves from sdd-orchestrator to gentle-orchestrator, preserving legacy migration support. PR size budget: new policy with maintainer-applied size:exception for aggregate or tracker PRs. E2E: now uses a deterministic local Engram shim unless live download testing is explicitly enabled.
v1.24.0 — Community plugins for OpenCode
Optional registration of OpenCode Community Plugins for sub-agent-statusline and sdd-engram-plugin. New standalone main menu flow to register community plugins without running the full installer. gentle-ai writes selected TUI plugins declaratively to ~/.config/opencode/tui.json; OpenCode resolves and loads them on next start/reload. gentle-ai update detects plugin versions once OpenCode materializes them under ~/.config/opencode/node_modules. Bonus: fixed upgrade/sync TUI pollution from verbose backup diagnostics.
v1.24.1 → v1.24.3 + v1.25.1 → v1.25.3 — Maintenance patches
6 patches between Apr 27 and May 3, interleaved with the 1.24.0 and 1.25.0 minor bumps. Mostly automated github-actions[bot] publications. See the full release list on GitHub for per-patch detail.
Notable Trends
Heavy week: 17 releases on engram + 9 on gentle-ai. Automated cadence via github-actions[bot] suggests release-please or similar.
Common focus on operational hardening: doctor diagnostics + Windows hardening on engram, sustainable review workflow on gentle-ai.
v1.15.4 — Durcissement de mem_save et prompt capture
Patch après revue Judgment Day. Corrections : mem_save trace l'activité sur le session_id résolu ; les handlers tolèrent SessionActivity nil ; les échecs de prompt capture restent best-effort.
v1.15.3 — capture_prompt ajouté à mem_save
Nouveau booléen mem_save.capture_prompt (par défaut true) : si le processus MCP a un contexte de prompt pour le même projet/session, mem_save enregistre ce prompt avec l'observation.
v1.15.2 — Récupération sûre pour écritures à projet ambigu
Quand MCP démarre d'un parent dir ambigu, les agents peuvent récupérer en demandant à l'utilisateur de choisir un des available_projects rapportés.
v1.15.0 — Conflict surfacing, doctor, dashboard fix, Windows hardening
Release de durcissement — la plus grosse de la semaine. Conflits mémoire : nouveaux workflows audit et juge sémantique, commandes engram conflicts, MCP mem_compare et cloud sync. Doctor : engram doctor read-only, MCP mem_doctor, engram doctor repair avec modes --plan/--dry-run/--apply. Windows : durcissement Git Bash/MSYS2 et fallback PowerShell.
v1.14.2 → v1.14.13 — Patches de clôture du cycle 1.14
12 patches entre le 27 avril et le 30 avril avant le passage à 1.15. Principalement automatisés via github-actions[bot].
v1.25.4 — Enregistre 4 skills manquants dans le build Go
Correction ciblée : 4 skills présents dans skills/ depuis v1.25.0 mais absents du registre Go sont maintenant enregistrés.
v1.25.0 — Workflow de review durable + renommage en gentle-orchestrator
Minor bump centré sur le processus de review. Workflow durable pour les changements SDD avec planification de stratégie de livraison. Renommage : le conducteur OpenCode SDD passe de sdd-orchestrator à gentle-orchestrator, en préservant la compatibilité legacy.
v1.24.0 — Plugins communautaires pour OpenCode
Enregistrement optionnel des Community Plugins OpenCode. Nouveau flow standalone pour enregistrer des plugins sans lancer l'installeur complet.
v1.24.1 → v1.24.3 + v1.25.1 → v1.25.3 — Patches de maintenance
6 patches entre le 27 avril et le 3 mai, intercalés avec les minor bumps 1.24.0 et 1.25.0.
Tendances Notables
Semaine chargée : 17 releases sur engram et 9 sur gentle-ai. Cadence automatisée via github-actions[bot] suggère release-please ou équivalent.
Focus commun sur le durcissement opérationnel : doctor diagnostics et Windows sur engram, review workflow sur gentle-ai.