v0.5.9

Fixed

• 修复安装插件后，未进入 Harness 的普通 Claude Code 会话编辑 src/main/java、src/test/java 或 openapi 时被 pre-write 错误拦截的问题；已绑定会话的治理 gate 仍保持 fail-closed。
• pre-write 现在区分不存在、过期和损坏的 session binding，并为后两者返回 session-first 的稳定错误码与恢复动作，不再错误要求设置 harness/ACTIVE_CHANGE。
• 损坏或无法迁移的 active state.json 现在以稳定错误码 fail-closed，不再让关键 pre-write hook 以原始异常退出。

v0.5.8

Changed

• 将日常与发布质量门禁迁移到本地 quality:local，GitHub-hosted 平台与安全检查改为仅手动触发，避免 push、PR、tag 和 schedule 自动消耗 Actions 分钟。
• 本地 release 现在生成并验证制品、SBOM 与 release notes，再通过 gh 创建 GitHub Release。
• release 制品绑定到已提交且 tracked-clean 的 tag tree；远端 main/tag 逐一校验，并在部分发布失败时保留 worktree、输出可执行恢复 argv。

Fixed

• session binding 使用真实路径规范化，避免 macOS /var 与 /private/var 等价路径被误判为 worktree 冲突。

v0.5.7

chore: release 0.5.7

v0.5.6

chore: release 0.5.6

v0.5.5

chore: release 0.5.5

v0.5.4

chore: release 0.5.4

v0.5.2

Fixed

• 修复 runtime verify 仍依赖 v5 state === VALIDATED，导致 v6 verify/archive change 跳过 CompletionProof、独立 review 与 freshness gate 的问题。
• 将 task child outcome 升级为 typed exit / signal / spawn-error 合同；launch error、signal termination 与伪造 outcome 不再能够满足 RED/REPRODUCE 或发布 trusted receipt。
• 将 implement finalization 绑定到 canonical task receipt、当前 execute run spool、fresh handoff inputs 与 current task，拒绝 arbitrary/alias/traversal/symlink receipt path。
• malformed 或歧义 task child marker 现在 fail closed，不能触发 stale lock 自动恢复；PID identity 与 run ownership 继续参与恢复判断。

Security

• canonical receipt 的重试只允许原 execute run；其他 run 不能收养旧 receipt。
• 明确 Linux、macOS 与 Windows 的 task descendant 有界收敛合同，并区分 lifecycle cleanup 与 OS sandbox 保证。

v0.5.1

Added

• 为六阶段 Heavy-Skill 补齐 deterministic prepare/finalize scripts、assertions、self-check 与结构化结果持久化。
• 新增 schema-backed StageResult、ReviewResult、CompletionProof、TaskExecutionReceipt、waiver 与 Classification artifact 合同。
• 新增 canonical task-run，从冻结计划内部解析 execution strategy、阶段顺序与 exact argv，并在 git common directory 发布短期 runner authorization。
• 新增 session-first、classification authority、archive/abandon transaction、受治理写入和 installed-plugin E2E 回归覆盖。

Changed

• v6 阶段推进统一依赖 fresh artifact digests、独立 execute/check run、trusted agent binding 和 CompletionProof，不再接受 projection boolean 或 worker 自报完成。
• Clarify 收敛为 topology/frontier 模型；code-explore 与 doc-research 只返回 ResearchPacket，业务决策仍由 Main 与用户确认。
• Implement 的受治理路径只能由 canonical runner 的冻结子进程写入；直接 Write/Edit/NotebookEdit、外部 child argv、管道、重定向和命令串联均 fail-closed。
• hooks 收敛为 host boundary、机械 gate、snapshot attribution 与恢复提示；v6 correctness authority 留在 runtime result/lifecycle gate。
• Archive 改为 Verify completion 与 archive finalization 两阶段事务，并保持 v5 compatibility 只读隔离。

Fixed

• 修复 reviewer 自建 check handoff、result caller 与 exact run/role/session 未绑定，以及缺少 completed agent evidence 时仍可推进的问题。
• 修复 task receipt 与 handoff input digests、冻结 strategy/argv、时间戳和输出摘要绑定不足的问题。
• 修复 Bash 子进程写入无法归因、validation freshness 未失效，以及 stale runner marker 可重放的问题。
• 修复 archive/abandon 物理移动失败后 lifecycle 或 blocker 未按 CAS 回滚的问题。
• 修复 stale active-change 阻断 worktree、session bootstrap 顺序、Classification artifact/state authority 漂移和 v5 migration 状态映射问题。
• 修复 release allowlist、artifact validation、错误码文档和 installed-plugin Claude E2E 的可复现性。

v0.5.0

Architecture

Six-stage user-visible lifecycle: clarify → design → plan → implement → verify → archive.
Classification is internal and durable. TDD is an implementation strategy, not a lifecycle stage.
route and tdd exist only in v4/v5 compatibility readers.

Added

• State v6 (runtime/core/change-state.mjs): minimal revisioned state with
  updateChangeState() as the sole mutation primitive (lock → copy → mutate → validate →
  CAS → atomic write → event). No persisted ready/approved booleans; freshness is
  digest-derived.
• Handoff v2 (runtime/core/handoff-v2.mjs): canonical run transport under
  <git-common-dir>/enterprise-harness/runs/, safe across native worktrees.
• v5 → v6 migration (runtime/compat/v5-migrate.mjs, runtime/migrate-v5.mjs):
  explicit, confirmation-gated migration for active changes; archives stay read-only.
• Session and lock leases with heartbeat renewal and expiry-based recovery.
• Nine canonical skills: harness, explore-code, research-docs, design, plan,
  implement, review, verify, archive.
• Five capability agents: code-explore, doc-research, artifact-worker,
  implementer (worktree-isolated, sole product-code writer), reviewer.
• Quality loop: every stage/task requires execute → self-check → independent review → TECPC → fresh evidence.
• Native plugin layout regression smoke and v0.5 capability-surface smoke.

Changed

• Plugin assets moved to root skills/, agents/, hooks/scripts/.
• harness skill rewritten as the sole user-facing orchestrator for the six-stage lifecycle.
• Forked methodology skills return NEEDS_DECISION instead of calling AskUserQuestion.
• Hooks limited to host-boundary mechanics; agent lifecycle events are telemetry only.
• Specs rewritten: workflow, state-schema, agents-and-handoff, hooks, evidence.
• runtime/compat/ isolates all v4/v5 legacy agents, skills, and contract tests.

Removed

• .claude/rules/ auto-loaded controller rules (replaced by specs and skill methodology).
• Legacy .claude/skills/ and .claude/agents/ as plugin asset locations.
• runtime/hooks/ as hook script location (moved to hooks/scripts/).
• Legacy stage skills (harness-design, harness-plan, harness-tdd, harness-verify)
  from the canonical plugin surface (moved to runtime/compat/skills/).

v0.4.3

Changed

• Clarify now uses Claude Code AskUserQuestion to ask one focused question at a time, with a recommended option.
• Stage skills use compact per-step Expect / Verify checkpoints and few-shot handoff examples for weaker models.
• User workflow documentation now explains the expected evidence and verification at each stage.
• Post-write remains a thin hook shell while retaining the snapshot attribution gate and validation invalidation policy.

Fixed

• Restored SubagentStart lifecycle evidence required by agent completion gates while retaining plugin-scoped MCP evidence matchers.