v1.0.22

v1.0.22 formalizes the memory-as-hint contract across the public OpenCode harness. Persistent memory and MCP context are now explicitly treated as hints, not sources of truth, and agents must verify them against current repository/artifact state before using them for routing, implementation, or recommendations.

Incluye:

Updated public agent and harness docs.
Checker enforcement for memory-as-hint.
Regression tests for missing memory verification wording.
Test fixture hardening to ignore local .codegraph/ and node_modules.

v1.0.21 - AHE preflight audit

v1.0.21 - AHE preflight audit

What's new

• Added an integrated AHE preflight audit step before /evolve evaluation.
• Added scripts/preflight-audit.mjs to generate preflight-audit.json per iteration.
• The preflight artifact includes:
  • contract coverage
  • runtime evidence coverage
  • doc/runtime alignment
  • drift severity
  • doc/runtime matrix
  • prioritized recommendations
  • confidence level
• Updated /evolve so preflight runs before session staging, evaluator, debugger, and evolver.
• Added audit-only as a lightweight branch for harness status/audit without invoking sidecars.
• Updated evaluator, debugger, and evolver contracts to consume preflight output as baseline evidence.
• Added checker enforcement and tests so the preflight contract cannot drift silently.

v1.0.20

v1.0.20 - Skill resolution metadata and resolver

What's new

• Added structured skill metadata for packaged OpenCode skills: phase, domains, stacks, allowed agents, surfaces, source, origin, and status.
• Added a generated skill_registry.json alongside the human-readable Markdown registry.
• Added a deterministic skill resolver that produces explainable shortlists by agent, phase, domain, and stack.
• Added a project capability snapshot script to detect stacks, domains, and work surfaces from repository structure.
• Added new packaged skills for improved orchestration and verification:
  • autonomous-loops
  • context-engineering
  • doubt-driven-development
  • iterative-retrieval
  • using-agent-skills
  • verification-loop
• Added tests for registry generation, skill resolution, and project capability snapshots.

v1.0.19

v1.0.19

Added

• Added SDD-oriented init preflight contracts for /init, including project-state detection, evidence capture, and follow-up guidance before generating project context.
• Added public-safe skill registry generation and validation so project skills can be documented consistently without leaking local user skill inventory.
• Added harness checks covering init detection rules, skill registry freshness, and agent/command contract alignment.
• Added project context template guidance for generated AGENTS.md content.

Updated

• Updated lead, developer, specifier, reviewer, researcher, designer, and scoper contracts to route skill usage and init/preflight responsibilities more explicitly.
• Updated harness documentation for init detection rules, command behavior, skill registry checks, and validation expectations.

v1.0.18

v1.0.18

Resumen

Esta release refuerza el flujo de automejora del harness alrededor de evidencia real de sesiones OpenCode y hace más auditable el comportamiento de /evolve, sin mezclar configuraciones privadas ni artefactos locales.

Cambios principales

1. /evolve ahora es OpenCode-first

• La fuente base de evidencia pasa a ser OpenCode, no sesiones de Codex.
• El flujo queda orientado a opencode.db como corpus principal.
• Los raw exports siguen existiendo, pero como fuente suplementaria.

2. Ingesta de sesiones y staging de evidencia

• Se añade un collector público para recolectar y normalizar session evidence.
• El collector genera artefactos staged para AHE:
  • execution-trees.jsonl
  • normalized-sessions.jsonl
  • session-sources.summary.json
  • cursor.json

3. Evidencia por árbol de ejecución

• La unidad principal de evidencia deja de ser la sesión plana.
• Ahora /evolve trabaja con árboles completos:
  • root session
  • subagentes enlazados por parent_id
• Esto alinea la evidencia con cómo realmente funciona OpenCode cuando delega.

4. Incremental por cursor canónico

• El collector soporta revisión incremental.
• El cursor se basa en:
  • tree_time_updated_max
  • root_session_id
• Los raw exports no avanzan el cursor canónico.

5. Contrato más claro para AHE

• Se documenta mejor la diferencia entre:
  • evidencia natural
  • prompts sintéticos/coercitivos
• Se evita usar árboles coercitivos como prueba suficiente de regresión normal de routing.
• Se endurece la distinción entre smoke rápido y suite larga del checker.

6. Mejoras de agentes y revisión

• lead puede lanzar el collector de session evidence en el harness público.
• evaluator puede ejecutar opencode run --format json --thinking * para transcript replay.
• debugger prioriza artefactos staged y evita depender de scripting ad hoc para métricas mecánicas.

7. Checks y tests actualizados

• El checker público se alinea con el nuevo contrato de /evolve.
• Se añaden tests para el collector y para las reglas del nuevo flujo de evidencia.

Validación usada para publicar

• ./scripts/check.sh
• git diff --check

Impacto esperado

• Mejor base para automejora del harness.
• Menos falsos positivos en auditorías de routing.
• Mejor trazabilidad entre evidencia, diagnóstico y cambio propuesto.
• Contrato público más alineado con el comportamiento real de OpenCode.

v1.0.17

v1.0.17 - Phase Result Contracts

This release strengthens the OpenCode harness closeout protocol for non-trivial phase work.

Changes

• Added a shared Result Contract for specifier, developer, and reviewer phase closeout.
• Added a Verification Envelope requirement for developer implementation closeout.
• Registered the new workflow mechanisms in the harness evolution registry.
• Extended check-harness.mjs to mechanically enforce the new contracts.
• Added regression tests covering missing Result Contract and missing Verification Envelope.

v1.0.16

Summary

• Adds Task Contract and handoff_packet guidance to keep spec, implementation, and review aligned across multi-agent work.
• Adds AHE mechanism registries and router/skill benchmark scenarios for small-model drift detection.
• Extends mechanical harness checks and regression tests for the new contracts.

Validation

• node --check scripts/check-harness.mjs
• node scripts/check-harness.mjs
• node --test scripts/check-harness.test.mjs
• ./scripts/check.sh
• git diff --check

v1.0.15

v1.0.15 - Harness contract alignment checks

This release publishes the contract-alignment hardening from the local OpenCode
harness into the public starter kit, while keeping private configuration,
local MCP wiring, credentials, raw transcripts, and machine-specific paths out
of the repository.

Highlights

• Added mechanical checker coverage for /evolve, /scope, and /design
  command contracts in opencode/scripts/check-harness.mjs.
• Added cross-agent prompt checks so configured invariants from
  docs/ai/harness/agents.md must be reflected in the corresponding
  agents/*.md prompts.
• Fixed AHE JSON path validation so prose evidence that mentions markdown
  filenames, such as PRODUCT.md/DESIGN.md, is not treated as a missing local
  file path.
• Added node:test regression coverage for:
  • prose evidence containing markdown filenames;
  • inverted /evolve ordering;
  • inverted /scope ordering;
  • incomplete lead prompt invariants.
• Aligned the public developer prompt with the documented ownership rule:
  follow-up implementation adjustments return to developer.
• Added public-safe summary evidence for AHE iteration 016.

Validation

Validated locally with:

cd opencode
node --test scripts/check-harness.test.mjs
node scripts/check-harness.mjs
cd ..
./scripts/check.sh
git diff --check

v1.0.14

v1.0.14 - Semantic harness checker

This release publishes the latest AHE improvement from the local OpenCode
harness into the public starter kit, while keeping local providers, private MCP
configuration, credentials, raw transcripts, and machine-specific paths out of
the repository.

Highlights

• Replaced brittle literal prompt-text checks in opencode/scripts/check-harness.mjs
  with bounded semantic regex checks for lead-router contract prose.
• Kept structural checks exact for security-sensitive permissions and agent
  identifiers, including edit: deny, shell allowlist entries, and agent names.
• Added literal fallbacks for the most critical lead prohibition checks so the
  checker remains conservative where it matters.
• Added public-safe AHE summaries for iterations 014 and 015, covering the
  baseline diagnosis, manifest, validation, and keep decision.

v1.0.13

v1.0.13 - Tighter Lead Delegation Boundary

This release tightens the OpenCode harness contract so lead behaves more consistently as an orchestrator instead of drifting into implementation, deep research, or review work.

Changes

• Clarifies that lead must not develop, deeply investigate code, or review diffs as a substitute for specialist agents.
• Requires substantive code discovery to be delegated to researcher.
• Requires reviewable diffs, implementations, and plans to be delegated to reviewer.
• Adds a self-contained handoff contract for lead delegations.
• Extends harness checks so these delegation boundaries are enforced mechanically.

Validation

• ./scripts/check.sh
• git diff --check