ODF docs
GitHub ↗
Documentation menu

Changelog — ODF Agent Team

18 min read

#Unreleased

#Fixed

  • Single-word skill triggers now match at a word boundary: review no longer selects the views skill via view, transaction no longer selects actions via action, and latest no longer selects testing via test.
  • The parallel branch prompt bound is 64 KB, leaving room for the canonical and domain skill bodies.
  • Linked the previously unreachable version references and replaced the OCA contributing wikilinks with real skill paths.

#Changed

  • Consolidated 41 pattern, testing, and version skills into 5 domain skills: odoo-modeling, odoo-views-owl, odoo-business-flows, odoo-testing, and odoo-versioning. Each former skill is now references/<topic>.md with its examples, and the registry drops from 87 to 51 entries.
  • Skill descriptions now lead with Use when and describe user intent instead of Entity. Trigger: ...; all 87 were rewritten.
  • Delegations also inject the body of the highest-scoring non-canonical skill (capped at 8,000 characters), so domain standards are usable without flooding the prompt.
  • Added a trigger-eval dataset (scripts/odf-skill-evals/trigger-queries.json, 57 realistic prompts) and a test that requires the expected skill in the top 3.
  • Odoo version filters were normalized to 16-19, and the five skills that referenced the nonexistent grep_app_searchGitHub now use local rg commands.
  • Skills now use the Agent Skills layout: all 87 registered skills live at skills/<name>/SKILL.md, 73 flat files were migrated, 50 version variants moved to references/versions/, and 23 long skills were split into SKILL.md plus references/examples.md.
  • Absolute install and personal paths in skill references were replaced with relative links; the OCA index and dispatcher paths now point at the flat layout.
  • Delegations inject the canonical stage skill's full SKILL.md body (frontmatter stripped, capped at 20,000 characters) alongside compact rules for matched skills, and list bundled references/, scripts/, and assets/ for on-demand reads. The parallel branch prompt bound is now 32 KB.

#1.6.2 (2026-10-26)

#Added

  • A real documentation site: every documentation link renders as a standalone page with a grouped sidebar, in-page table of contents, breadcrumbs, reading time, prev/next navigation, and the search palette. scripts/odf-docs-build.js rewrites Markdown links (registered sources → relative pages, repository files → GitHub blob URLs, external links untouched), reproduces GitHub heading anchors, and generates search-index.json, sitemap.xml, robots.txt, and .nojekyll (#186).
  • npm run docs:build / npm run docs:check, a CI step, and unit tests for the slugger and link rewriter keep the generated pages in sync (#186).
  • A Diagrams group in every docs sidebar, plus a styled 404 page (#186, #187).
  • A privacy guard test that fails when a published file gains an absolute home path, Windows home path, provider email, private-key block, known token format, URL with credentials, or literal password (#188).

#Changed

  • install.sh and the TUI installer rewrite any /home/<user>/.config/opencode path instead of one hardcoded author path, so the pack is portable without embedding a username (#188).
  • Skill frontmatter uses the public project handle; internal task logs, SDD state, and the archived verification report use generic paths (#188).
  • The tracked timesheet binary is rebuilt with go build -trimpath -ldflags="-s -w" (9.8 MB → 7.9 MB) and no longer embeds build-machine paths (#188).
  • The landing page links locally, keeps search inside the site, and documents the 1.6.2 docs and privacy work.

#Fixed

  • The safety contract shows password=<value> instead of a literal-looking password (#188).
  • Removed stale internal notes and a /tmp reference from the landing page source (#188).

#Verification boundary

  • Local validation covers 1405 Vitest tests (52 files), 331 YAML scenarios, 18 dedicated harness checks, npm run typecheck, registry validation, skill validation, npm run docs:check, bash -n install.sh, and node --check on the TUI installer.
  • A full internal crawl of the generated site resolves every link and anchor; 32 published files were re-scanned for personal data and credentials and are clean.
  • No npm package is published. Git history and the v1.6.1 tag still contain the pre-fix installer path; this release makes the newest downloadable artifact clean without rewriting history.

#1.6.1 (2026-10-26)

#Added

  • feature-inline route with a conditional ASSESS gate (#180): the default standard-feature route runs DECIDE → BUILD → VERIFY and keeps the Intent + Expectations contract. A deterministic entry gate selects skip (terminal DECIDE before BUILD), assess (one bounded ASSESS writes the canonical decision), or escalate (the formal feature route keeps DECIDE + PLAN). ASSESS here returns strategy, scope, affected areas, risks, and unknowns — never a functional spec.
  • Bounded Context Manifest (ICE) (#181): odf_context_manifest builds one read-only, reference-only manifest at entry with redundancy, assessment_gate, and expectations_ref included in manifest_digest, and forwards it unchanged to ASSESS and every BUILD batch. Redundancy is a gate signal for feature-inline, not a blocking user question.
  • Decision-quality evaluation (#171–#173): odf-evaluation decisions reports routing quality against golden trajectories; a local decision review at archive/verify records a human label; the report adds human/judge agreement, period comparison, a review cadence, and improvement proposals. A null judge verdict is treated as unavailable, never as agreement.
  • Persisted stage handoff (#142): odf_workflow_status.next_handoff derives the next stage role's inputs from the route, selected store, and persisted artifact references so the orchestrator never rebuilds them from result prose.
  • Safe stop classification (#144): every blocked/failed envelope carries stop_disposition (kind, user_action_required, automatic_recovery, reason, next_action). Unknown reasons fail closed as unclassified; child-reported dispositions cannot weaken the deterministic class.
  • Project-scoped activation (#151, #163): .odf/project.json marks an initialized Odoo project, re-scanning the same root is idempotent and preserves initialized_at, and activation created after runtime boot prints ODF runtime reload required: instead of suggesting a reinstall.
  • Native delegation lifecycle (#139, #140, #161, #162): child sessions bind to the parent, agent, model, workspace, and prompt digest; parent cancellation aborts the child and settles the attempt as failed; odf_delegation_late_seal_prepare recovers an expired original token through a ten-minute, single-use capability; seal failures report a precise reason instead of a generic block.
  • Agent Skills validation (#182): scripts/odf-skill-validate.js (npm run validate:skills) validates frontmatter, name/directory parity, description limits, path containment, and the compact-rule contract, and runs in CI.
  • Route/stage contract matrix (#169): odf-plugin/odf-route-contract.test.ts walks every route, transition, stage/role, and artifact/gate requirement; docs/route-contract.md mirrors it.
  • Boundary hardening (#145, #153, #155, #157): evidence and local security hardening, exact cross-root resolution, blocked absolute traversal destinations, rejected prose-redirect bypasses, and recoverable aborted launches.

#Changed

  • Canonical status only (#164–#167): removed the legacy compatibility layer. odf_workflow_status no longer emits phase, artifacts, applyProgress, lastUpdated, or legacy_phase; the legacy state read path is gone, artifact-only changes report state_kind: none and stay non-resumable, and legacy launch tokens are rejected instead of upgraded. Only the route.legacy_phases alias survives, for bind comparison.
  • Centralized artifact vocabulary (#141): canonical artifact names, aliases, route completion alternatives, and OpenSpec discovery stems live in odf-artifact-contract.ts; store values and subsets live in odf-artifact-stores.js.
  • Strict VERIFY completion (#170): VERIFY completes only through a terminal verify-report.
  • Delegation API on canonical stages (#168): the public delegation contract accepts the canonical stage plus a role for the composite stages only.
  • Canonical status discovery (#174): active-change resolution reports none / unambiguous / ambiguous / incomplete and surfaces Engram errors instead of swallowing them.
  • Delegate hardening (#175, #178, #179): bounded workspace, artifact references, and task errors; blocked envelopes preserve the resolved agent, skills, profile, and Policy Gate; Policy Gate blocks use a stable reason; nested Expectations inputs are bounded; profiles are validated and labeled advisory; skill, registry, receipt, and status previews refined.
  • Skills (#182): all 87 compact_rules regenerated from each skill's own rules (150–600 characters, complete sentences, no banners, no duplicates); selection breaks score ties by registry order instead of prompt length; five OCA tool skills, karpathy-precision, and the OCA index frontmatter corrected.
  • Documentation (#146, #183): artifact store persistence clarified; the protected-main branch → PR → auto-merge workflow documented in AGENTS.md.

#Removed

  • odf_status (#165) — use odf_workflow_status.
  • The legacy dual-read compatibility layer, legacy status projection, and legacy state read path (#164, #166, #167).
  • odf_notebooklm_lookup and the notebooklm_sources registry map (#176); NotebookLM enrichment, when used, stays external to ODF.

#Fixed

  • Cancellation and aborts: parent cancellation propagates to delegated children, cancelled delegations stop cleanly, and aborted children bind with a precise seal reason (#149, #159, #161).
  • Late seal verifies the prepared prompt without an assistant result (#162), and aborted native children no longer dead-end recovery (#159).
  • Safety: absolute traversal destinations are blocked (#155), prose-redirect bypasses are rejected (#157), false cross-root blocks no longer stop valid work (#153), and evidence/local security is hardened (#145).
  • The registry fixture resolves from the repository root, so the suite no longer depends on the invocation directory (#177).

#Verification boundary

  • Local validation covers 1370 Vitest tests (50 files), 331 YAML scenarios, 18 dedicated harness checks, npm run typecheck, registry validation, and npm run validate:skills.
  • No npm package is published; the release is the tagged pack plus the GitHub release. Odoo-specific end-to-end validation and efficacy claims remain ODF 2.0 gates.

#1.6.0 (2026-10-09)

#Added

  • Diagnostics and safe recovery: bounded causal timelines, truthful canonical workflow status, fail-closed recovery guidance, and conversational exploration/continuation routing (#136).
  • Opt-in online feedback: /odf-feedback submits only an explicit 1–5 rating and aggregates with host-supplied usage after preview and confirmation; HTTPS endpoint, 1–30-day receiver-retention contract, and no prompts, free text, durable queue, or retries (#137).
  • Repeatable offline evaluation journeys for diagnostics, recovery, and conversational workflows (#136).

#Changed

  • Expanded the plugin to 31 tools and the command set to 23; documented the online feedback privacy and retention boundary.

#Verification boundary

  • CI on Node 22 passed for PR #137. Release-candidate validation passed 1,182 Vitest tests, 331 YAML scenarios, 18 harness checks, typecheck, and registry validation.
  • Online feedback is user-rated telemetry, not benchmark evidence. This release makes no efficacy claim and does not include an ODF-hosted collection service or npm publication.

#1.5.1 (2026-10-03)

#Fixed

  • Workflow persistence and recovery: Policy Gate writes fail closed, ARCHIVE retries repair partial OpenSpec/Engram writes, and replanning invalidates downstream artifacts (#104–#106).
  • Native delegation safety: restore V2 delegation, guard direct tool calls, check seal context before child launch, and bind recovery to the exact idle child; handle multiline prompts, V2 result text, and transport banners safely (#86, #89, #93, #100, #101, #107, #108).
  • Canonical workflow artifacts: terminal BUILD behavior is explicit, and nested assessment artifacts and block-YAML status maps are recognized (#87, #91, #102).
  • VERIFY evidence is command-aware; warning is terminal only for a valid VERIFY report; a BUILD start proof does not replace persisted post-child validation evidence (#109, #110, #112).
  • OpenCode V2 /odf-new normalizes sentence-final punctuation and binds workflow initialization to the active session workspace.

#Changed

  • Preflight defaults, strict input validation, and the runtime/specification contract are synchronized (#111).
  • Workflow route, recovery, archive, and metrics guidance reflects the merged behavior (#85, #94, #96, #97).
  • Lifecycle metrics correlate mixed-flow events and deduplicate mirrored stages (#98, #99).

#Verification boundary

  • CI on Node 22 passed for the implementation PRs #104 and #106–#112, whose merged suite contained 1,050 unit tests, 330 YAML scenarios, and 18 harness checks. The 1.5.1 release candidate adds one V2 /odf-new regression case; its local validation passed 1,051 unit tests, 330 YAML scenarios, and 18 harness checks.
  • Odoo-specific release validation and end-to-end telemetry remain outside this harness patch release; ODF 2.0 readiness is not claimed.

#1.5.0 (2026-09-29)

#Added

  • Native delegation (OpenCode V2): odf_delegation_prepare + odf_delegation_seal run a phase as a host subagent child session — visible agent, description and live status — while every ODF gate stays plugin-side: bounded odf-tok-… tokens under .odf/ (prompt digest, two-hour TTL, workspace containment), child-session binding checks, composite gates and a proof-backed replay of the authoritative delegate (policy gate, validation evidence, workflow commit, attempt settlement, failure receipts). The sealed envelope matches odf_delegate for identical inputs (delegate/seal parity matrix).
  • Native parallel BUILD: odf_parallel_prepare + odf_parallel_seal launch one subagent per branch (background allowed) and run the authoritative aggregate scheduler: per-branch validation evidence, the parallel-join artifact, one BUILD commit and one aggregate receipt on failure, with per-branch task_session_id.
  • Chokepoint guard: the V2 execute.before hook blocks direct subagent/task calls from the orchestrator to registered ODF specialists without the prepared delegation marker; manual agent use from other sessions is untouched.
  • The plugin exposes 26 runtime tools.

#Changed

  • Orchestrator and phase commands document the native single and parallel delegation flows, their recovery reasons, and the fast-lane / cross-domain boundaries.
  • Native seals record one aggregate subagent metric (the replayed delegate metrics are suppressed), including join counts for parallel BUILD.

#Fixed

  • Heavy phases: VERIFY gets the same 900000 ms timeout budget as ASSESS/QA-PLAN/DESIGN, and the orchestrator documents expected duration, child-session visibility (ODF <phase> → <agent> · <change>, /sessions on the legacy path) and interruption semantics — interrupting aborts the child, settles the attempt as failed, applies the single transport relaunch, and /odf-continue resumes (#80).
  • Parallel BUILD recovery: a persisted join left running after a host restart is recovered instead of dead-ending BUILD (#74, #75).
  • Legacy status: progress-bearing artifacts (implement-progress, apply-progress, tasks, explore-progress, fix-progress) derive their state from content, so partial work renders as in-progress instead of done (#73, #75).
  • Metrics: the entry-to-final-gate funnel breaks down per stage so the dominant bottleneck is localizable (#76, #77).
  • Parallel scheduler tests use waitFor timeouts above the vitest 1 s default and no longer flake under load (#78, #79).

#Verification boundary

  • Local repository checks cover 998 unit tests (37 files), 325 YAML scenarios, and 18 harness checks; CI runs the same suite on Node 22.
  • Native delegation is covered by the prepare/seal flow tests, the delegate-vs-seal parity matrix (composite, proof-backed, inner-failed) and the parallel BUILD end-to-end tests; the real Odoo VERIFY run is recorded as release evidence.

#1.4.0 (2026-09-28)

BREAKING — OpenCode V2 only. The plugin entrypoint no longer exports the V1 server implementation, and @opencode-ai/plugin is not part of the runtime graph. Packs installed from this release require an OpenCode V2 host: a V1 host will not register the ODF plugin.

#Added

  • Native delegation visibility under OpenCode V2: child sessions carry a descriptive ODF <phase> → <agent> · <change> title, the delegated agent is surfaced through tool progress while the phase runs, and the outer delegation envelope reports task_session_id (#54, #56).
  • Source-authority not_applicable evidence: non-view DESIGN/IMPLEMENT work can return a bounded not_applicable envelope when the task names no concrete view/action XML ID or inherit_id/search_view_id relation; concrete references still require the deterministic odf-toolkit lookup (#50, #52).

#Changed

  • OpenCode V2 only: the entrypoint plugins/odf-delegation.ts now default-exports the official Plugin.define({ id, setup }) shape. The V1 server export, OdfDelegationPlugin and createODFRuntimeHooks were removed along with their V1 contract fixtures.
  • Runtime code no longer imports @opencode-ai/plugin. The shared tool helper moved to odf-plugin/odf-tool.ts (identity function + zod schema namespace), so the plugin graph carries only the host-provided @opencode/plugin plus the declared zod and yaml dependencies. @opencode-ai/plugin and @opencode-ai/sdk are now devDependencies, used for types and the V1/V2 schema-parity fixtures.

#Fixed

  • Registry/telemetry warm-up used to live inside the retired V1 server entrypoint, so the V2 host silently lost it. It is now startOdfRuntime(), awaited by setupODFV2: metrics flushing, the skills/permissions cache refresh, unregistered-skill discovery and the learning loop all run under OpenCode V2.
  • OpenCode V2 now injects the one-shot context-pressure notice through the session.context hook, restoring parity with the retired V1 experimental.chat.system.transform seam.
  • install.sh installs dependencies before exposing the plugin entrypoint, and reminds you to restart the OpenCode service afterwards (--restart-service / ODF_RESTART_SERVICE=1 to do it immediately). A plugin that fails once stays failed inside a running service until it restarts — this was the root cause of the plugin never loading after a fresh install.
  • Three new installer tests cover the dependency/plugin ordering, the default no-op dry run and the opt-in restart flag.
  • Config-dir resolution no longer ignores XDG_CONFIG_HOME and no longer depends on the project launcher: a single resolver (scripts/lib/config-dir.js) orders ODF_CONFIG_DIR → the pack the running module ships in (when it carries odf-registry.json) → $XDG_CONFIG_HOME/opencode → ~/.config/opencode, and the five CLI copies of that logic delegate to it. A pack installed under <project>/.opencode resolves itself, so running opencode directly there reads the right registry. The README's XDG_CONFIG_HOME resolution claim holds again, the /odf-* prompts default PACK to the same order, and the YAML suite now passes with ODF_CONFIG_DIR unset instead of reading the installed registry instead of the checkout.
  • The installed pack is self-verifiable: tsconfig.pack.json ships and installs as both tsconfig.json and tsconfig.pack.json, so /odf-health --full can typecheck the installed pack (#32, #33).
  • Engram persistence is project-scoped and store-aware end to end: odf-init readback no longer depends on the working directory, the plugin resolves one project identity, and archive/metrics respect the selected store (#37, #38).
  • odf_health reports why the V2 context session is unavailable instead of a bare failure (#34), and an interrupted first odf_health stays retryable instead of dead-ending /odf-new with workflow-start-unauthorized (#35, #36).
  • odf_delegate blocks every path inside the installed ODF pack with symlink-resolved containment: the pack root, a symlinked config dir, and pack subdirectories (#42, #45).
  • The ODF Result text parser is section-authoritative: it prefers ## ODF Result, parses nested bullet fields into objects (prototype-pollution safe), coerces true/false/null, accepts the dash-less bold form, and no longer lets an unrelated fenced JSON block replace the result (#51, #53).
  • Workflow recovery: re-entering an archived stage reopens the archived state (#63), block-YAML completed_canonical_stages lists parse correctly (#65), Engram state resolution reads local state before Engram (#66), and orphan workflow locks are recovered instead of dead-ending delegation (#67).
  • Delegation cancellation bounds the child abort wait after a timeout (#64), and the TUI installer reports a degraded install honestly instead of claiming success (#68).
  • The installer only exposes the plugin entrypoint once runtime dependencies resolve (#47).

#Verification boundary

  • Local repository checks cover 970 unit tests (35 files), 325 YAML scenarios, and 18 harness checks; CI runs the same suite on Node 22.
  • This release is OpenCode V2 only (see the breaking note above); a V1 host cannot register the plugin.

#1.3.1 (2026-09-24)

#Fixed

  • Installer: curl | bash no longer dies silently — non-interactive stdin continues with an explicit notice (the TTY prompt is kept and tolerates EOF); regression tests cover the piped script and stdin at EOF (#17).
  • OpenCode V2 entry authorization: the prompt hook detects the expanded /odf-new command document, and the V2 adapter shares the entry-authorization maps between the loop guard and the registered tools, so the capability minted after odf_health reaches odf_workflow_bind (#18).
  • Stale IMPLEMENT attempts: audited odf_workflow_override action=settle-attempt appends a terminal settlement (preserving the running record and the override audit log) and requires confirm_no_active_run plus a human-approved reason; it refuses attempts still active in the runtime, unsafe ids, and already-terminal records (#9).
  • BUILD/VERIFY starts accept the completion-shape proof after an audited BUILD re-entry instead of rejecting it as a phase mismatch; odf_workflow_advance accepts an explicit null candidate_stage for initial transitions and documents the start convention (#28).
  • TUI installer parity with install.sh: copies odf-plugin/, policies/, and package.json (so npm install runs), syncs the stale-plugin cleanup list, and covers uninstall/backup; tool, command, and skill counts in the docs match the registry and are guarded by consistency tests.

#Added

  • GitHub Actions CI with test-gated auto-merge: typecheck, unit tests, YAML scenarios, registry validation, and diff check run on every PR; passing same-repo PRs squash-merge automatically (Node 22, actions v7).

#Verification boundary

  • Local repository checks cover 900 unit tests, 325 YAML scenarios, and 18 harness checks; CI runs the same suite on Node 22.
  • ODF 2.0 remains gated on representative Odoo validation, canary/cohort evidence, and end-to-end /odf-new//odf-fix telemetry; this release does not claim ODF 2.0 production readiness.

#1.3.0 (2026-09-21)

#Added

  • Fast-lane execution and V2 transport, including SDK child-session transport, shadow-route prediction, context manifests, and bounded implementation routing.
  • Migration assessment routing, expanded agent and entry contracts, human Expectations, and stronger workflow and policy enforcement.
  • Governance profiles and checks, provenance and disclosure updates, lifecycle and trace observability, installer/project-scope improvements, and resilience and ephemeral-subject validation.

#Verification boundary

  • Local repository checks cover 844 unit tests, 154 YAML scenarios, and 17 harness checks.
  • ODF 2.0 remains gated on representative Odoo validation, canary/cohort evidence, and end-to-end /odf-new//odf-fix telemetry; this release does not claim ODF 2.0 production readiness.

#1.2.1 (2026-08-23)

#Added

  • Gate de precisión: odf-toolkit lookup (IDs de vista/modelos/campos en el source local con file:line) y verify-refs (cada ref=/model= del módulo debe resolverse; exit 1 si no). Reglas "nunca inventar IDs" en odoo-sources, odf-design, odf-implement y backend engineer; el orquestador reenvía odoo_source_root.
  • README reescrito: estado honesto, matriz de dependencias/degradación, pipeline, CLIs y gate de precisión.

#Fixed

  • TUI: Enter en (Y/n) confirmaba en vez de cancelar; listener raw que se tragaba el Enter siguiente; guard no-TTY con hint de automatización.

#1.2.0 (2026-08-22)

#Added

  • execution_mode: auto (piloto automático): auto-continuación de fases en ok/warning, deteniéndose solo en gates obligatorios (health, preflight, Expectations, Policy Gate, evidencia, VERIFY, receipts)
  • CLIs deterministas: odf-project-scan (config completa del entorno Doodba: addons.yaml sources, compose, linting, git, CodeGraph, matriz de dependencias; persist verificado con checksum/diff/exit codes; --deep indexa repos activos) y odf-toolkit (context/state/result/resolve/evidence/metrics/manual-evidence/redundancy)
  • Triage mejorado (ICE): señales de riesgo desde module/domain + pii, claridad de intent con pregunta agrupada, standard-config sin hechos, known_modules desde odf-init, signals/clarity auditables
  • Redundancy pre-check en /odf-new (implementaciones existentes + odf-learned/{project} como base de rechazos)
  • Ejecución manual de tests en VERIFY (manual-evidence, executor: user-manual)
  • Completion criteria en odf-assess/design/implement/verify; smell baseline Fowler para el lens de readability; seams-first y anti-patrones de tests en implement/tdd; feedback-loop-first + redacción en odf-fix; handoff con Suggested Skills
  • Descubrimiento de artefactos OpenSpec anidados (design/design.md, qa-plan/plan.md, ...)

#Fixed

  • Agente DESIGN/PLAN bloqueado por design_closed: "true" (string): coerción asBoolean en el validador
  • Resolver de agentes elegía frontend por el token genérico "odoo": scoring por coincidencias + stop word
  • odf-init colgado buscando el pack: ruta determinista $ODF_CONFIG_DIR sin filesystem search
  • Scan degradado con --repo relativo (resolución contra odoo/custom/src) + guarda anti-sobrescritura
  • Persist de config truncado por el límite de ~50KB de Engram: persist compacto + readback verificado
  • Falso positivo de seguridad con el guard obligatorio de base de datos (masking de prohibiciones)
  • Hybrid BUILD/VERIFY escribía solo Engram: ahora escribe OpenSpec (autoridad) + espejo Engram
  • odf-qa hardcodeaba mem_save: ahora persiste en el store seleccionado con artifact_ref

#1.1.0 (2026-06-18)

#Added

  • Orquestador conversacional (agent/odoo_orchestrator.md) con preflight gate y máquina de estados
  • Comandos nativos /odf-new, /odf-continue, /odf-status, /odf-explore registrados en odf-registry.json
  • Plugin odf-delegation.ts invoca la API nativa task() de OpenCode con fallback determinista
  • install.sh idempotente con --yes, --dry-run, --force, backup con timestamp y soporte ODF_SOURCE_DIR
  • package.json con dependencias, scripts de test y peer dependencies del plugin SDK
  • Validación de Node.js 18+ en install.sh
  • Resolución de rutas relativas en odf-registry.json con flag use_relative_paths
  • Metadatos de paquete en odf-registry.json (name, version, repository, dependencies)
  • scripts/odf-registry-validate.js para verificar que todas las rutas registradas resuelven
  • Helpers de preflight y orchestrator en scripts/lib/ con tests unitarios
  • Parser CLI mínimo en scripts/odf-cli.js para los comandos nativos
  • Tests unitarios con Vitest (92 tests) y escenarios YAML (118 aserciones)

#Changed

  • README actualizado con instalación via ./install.sh, referencia de comandos nativos y mención del orquestador/preflight
  • Versión del proyecto a 1.1.0 (VERSION, package.json, odf-registry.json, install.sh)

#1.0.0 (2026-05-14)

#Added

  • 31 skills organizados por categoría (ODF, OCA Governance, OCA Style, Patterns)
  • 12 agentes especializados en desarrollo Odoo
  • Plugin odf-delegation.ts con auto-refresh, cache, métricas, y learning loop
  • Pipeline completo: ASSESS → QA-PLAN → DESIGN → IMPLEMENT → VERIFY
  • Skill Registry CLI (/odf-registry-refresh)
  • Multi-profile switching (/odf-profile list|switch|create|delete)
  • Backup & Rollback (/odf-backup create|list|restore)
  • Skill version tracking (/odf-skill-log)
  • Agent Observatory (/odf-metrics) con delegación metrics + learning loop
  • Strict TDD mode (/odf-tdd on|off)
  • Chained PRs para PRs >400 líneas
  • PR Size Budget check (/odf-pr-size)
  • Issue-First workflow check (/odf-issue-check)
  • Health Checks (/odf-health --quick|--full)
  • Uninstall Flow seguro (/odf-uninstall)
  • 36 test cases para resolución de skills y agentes
  • OCA commit messages con 12 tags y formato completo
  • OCA work-unit commits
  • Judgment Day adversarial review (3-pass: reviewer, maintainer, attacker)
  • Auto-descubrimiento de skills en subagentes
  • Perfiles de modelo: default (deepseek-r1 + kimi-k2.6), cheap (kimi-k2.6)

#Changed

  • Todos los skills restructurados al estándar gentle-ai (180-450 tokens)
  • Registry convertido a formato named profiles
  • Plugin de 766→1044 líneas con metrics + learning loop

#Fixed

  • Rutas relativas de skills → absolutas (34 archivos)
  • Permisos de agentes (12 agentes con acceso completo)
  • mgrep → fff/fff_grep en todos los skills

#0.1.0 (2026-04-01)

#Added

  • Initial ODF workflow with 5 phases
  • Basic OCA compliance skills
  • odf-registry.json with 22 skills
  • ODF delegation plugin