Ac Memory

1. Ensure .agents/state/INDEX.md exists; if missing, create it from docs/STATE_INDEX_TEMPLATE.md.
2. Require the caller to resolve the target binding first:

• new task: explicit TASK_ID and TASK_TITLE
• continue: explicit task-id, exact TASK_TITLE, or a caller-confirmed candidate selected after workflow-side matching

3. Ensure the selected task state file exists under .agents/state/tasks/; if missing for a caller-confirmed new workflow task, create it from docs/TASK_STATE_TEMPLATE.md.
4. Validate required header fields:

• TASK_ID
• TASK_TITLE
• TASK_TYPE
• PLAN_FROZEN
• CURRENT_ROLE
• WORKFLOW_STATUS
• STARTED_AT
• LAST_UPDATED_AT

5. Validate required sections exist:

• 目标
• 下一步 Top 3
• 阻塞项
• 关键决策索引（最近 3 条）
• 关键决策日志（全量追加，不覆盖历史）
• 验收证据（Evidence）
• 未验证清单（Gaps）
• Review Findings（闭环）
• 提交计划
• Execution Contract

6. Repair missing structure in place while preserving existing history and user-written details.
7. Sync .agents/state/INDEX.md with CURRENT_TASK plus Active / Blocked / Completed summaries, using the stable entry format task-id | TASK_TITLE | task-type | role | status.
8. Return a concise state summary such as [STATE] <task-id> | <role> | <status> | 已检查/已更新 for the caller to echo when helpful.
9. When todo items, decisions, Evidence, Gaps, Review Findings closure, or role/freeze/status change, update the relevant blocks before control returns to the caller. 9.5. For low-risk non-copy-edit work, allow multiple closely related small edits within the same execution burst to be recorded as one material update, as long as the task state still reflects:

• current Top 3 completion
• new Evidence
• current role / freeze / status
• updated LAST_UPDATED_AT In this context, same execution burst means from Contract freeze until the caller's next user-facing reply; do not carry a batched write across turns. Cap one batched write at 3 declared steps or roughly 50 net changed lines; if the burst exceeds that size, require another state update before the caller finishes. 9.6. For low-risk non-copy-edit work, append 关键决策日志 only when a real scope, wording, routing, or consistency decision was made; do not require artificial decision entries for trivial wording-only edits. 9.7. When a lightweight execution burst touches or nearly touches the batching cap, allow Evidence or Gaps to record a short threshold observation such as 阈值可接受 / 阈值偏紧 / 阈值偏松, so later workflow tuning has fresh evidence to reference.

10. For docs-only tasks, allow Checks / Evidence to record consistency review instead of pretending code builds ran.

Outputs:

• complete active task state structure
• updated .agents/state/INDEX.md
• current state summary
• repaired state ready for planning / execution / review

Hard rules:

• Treat the active task state file as the current workflow source of truth, with .agents/state/INDEX.md as the task registry.
• Do not infer a continue target from partial wording when multiple unfinished tasks could match; candidate matching may narrow choices, but the caller must still confirm the final TASK_TITLE or task-id.
• Do not create a new task file until the caller has explicitly chosen new task instead of continue.
• Do not continue planner / executor / reviewer work while required structure is missing.
• Preserve existing history in decision logs; append instead of overwriting.
• Never leave stale CURRENT_ROLE, PLAN_FROZEN, or WORKFLOW_STATUS values after a phase change.