From 91b6f8453e13aba321b3f394f2c223f89f5b684f Mon Sep 17 00:00:00 2001 From: csh Date: Wed, 12 Aug 2026 12:40:31 +0800 Subject: [PATCH] :sparkles: feat(workflow): enforce auditable agent rules --- scripts/main_loop.py | 270 ++++++++++++++--- scripts/playbook.py | 77 +++-- templates/AGENT_RULES.template.md | 390 ++++++++++++++++--------- templates/README.md | 22 +- test/test_main_loop_cli.py | 286 ++++++++++++++++-- test/test_playbook.py | 75 +++++ test/test_template_contracts.py | 471 ++++++++++++++++++++++++------ 7 files changed, 1273 insertions(+), 318 deletions(-) diff --git a/scripts/main_loop.py b/scripts/main_loop.py index 42ec0586..b5712258 100644 --- a/scripts/main_loop.py +++ b/scripts/main_loop.py @@ -3,6 +3,7 @@ import argparse from contextlib import contextmanager from dataclasses import dataclass from datetime import datetime, timezone +import hashlib import json import os import re @@ -60,6 +61,7 @@ ALLOWED_STATUSES = { SATISFIED_STATUSES = {"resolved", "skipped"} THREAD_LOCKS: dict[str, threading.Lock] = {} THREAD_LOCKS_GUARD = threading.Lock() +SHA256_RE = re.compile(r"^[0-9a-f]{64}$") class StateError(ValueError): @@ -111,6 +113,13 @@ class Feature: return "queued" +@dataclass(frozen=True) +class EvidenceArtifact: + data: dict[str, Any] + digest: str + source: Path + + def _thread_lock(lock_path: Path) -> threading.Lock: key = str(lock_path.resolve()) with THREAD_LOCKS_GUARD: @@ -889,42 +898,121 @@ def claim_ticket( return render_claim_context(feature, ticket, metadata) -def evidence_field( - evidence: str, - key: str, - label: str, - expected: Optional[str] = None, -) -> str: - matches = re.findall( - rf"(?:^|[;,])\s*{re.escape(key)}\s*=\s*([^;,]+)", - evidence, - re.IGNORECASE, - ) - if not matches or not matches[0].strip(): - required = f"{key}={expected}" if expected is not None else key - raise StateError(f"{label} evidence must include {required}") - if len(matches) != 1: - raise StateError(f"{label} evidence must include exactly one {key}") - value = matches[0].strip() - if expected is not None and value.lower() != expected.lower(): - raise StateError(f"{label} evidence must include {key}={expected}") - return value - - -def require_review_passed(reviewed: str) -> tuple[str, str]: - for axis in ("standards", "spec"): - evidence_field(reviewed, axis, "review", "pass") +def canonical_json_bytes(value: dict[str, Any]) -> bytes: return ( - evidence_field(reviewed, "commit", "review"), - evidence_field(reviewed, "base", "review"), + json.dumps(value, ensure_ascii=True, sort_keys=True, separators=(",", ":")) + + "\n" + ).encode("utf-8") + + +def artifact_string(data: dict[str, Any], key: str, label: str) -> str: + value = data.get(key) + if not isinstance(value, str) or not value.strip(): + raise StateError(f"{label} evidence artifact field {key} must be a non-empty string") + return value.strip() + + +def artifact_pass(data: dict[str, Any], key: str, label: str) -> None: + value = artifact_string(data, key, label) + if value != "pass": + raise StateError(f"{label} evidence artifact field {key} must equal pass") + + +def load_evidence_artifact( + raw_path: str, + label: str, + expected_kind: str, +) -> EvidenceArtifact: + if not raw_path.strip(): + raise StateError(f"{label} evidence artifact is required") + source = Path(raw_path).expanduser().resolve() + if not source.is_file(): + raise StateError(f"{label} evidence artifact not found: {source}") + try: + data = json.loads(source.read_text(encoding="utf-8")) + except (UnicodeError, json.JSONDecodeError) as exc: + raise StateError(f"{label} evidence artifact must be valid UTF-8 JSON") from exc + if not isinstance(data, dict): + raise StateError(f"{label} evidence artifact must be a JSON object") + if data.get("version") != 1: + raise StateError(f"{label} evidence artifact version must equal 1") + if data.get("kind") != expected_kind: + raise StateError( + f"{label} evidence artifact kind must equal {expected_kind}" + ) + + artifact_string(data, "commit", label) + if expected_kind == "verification": + artifact_pass(data, "result", label) + artifact_string(data, "command", label) + if data.get("exit_code") != 0 or isinstance(data.get("exit_code"), bool): + raise StateError(f"{label} evidence artifact exit_code must equal 0") + output = artifact_string(data, "output", label) + output_sha256 = artifact_string(data, "output_sha256", label) + if not SHA256_RE.fullmatch(output_sha256): + raise StateError( + f"{label} evidence artifact output_sha256 must be a SHA-256 digest" + ) + if hashlib.sha256(output.encode("utf-8")).hexdigest() != output_sha256: + raise StateError(f"{label} evidence artifact output_sha256 mismatch") + else: + artifact_string(data, "base", label) + artifact_pass(data, "standards", label) + artifact_pass(data, "spec", label) + report = artifact_string(data, "report", label) + report_sha256 = artifact_string(data, "report_sha256", label) + if not SHA256_RE.fullmatch(report_sha256): + raise StateError( + f"{label} evidence artifact report_sha256 must be a SHA-256 digest" + ) + if hashlib.sha256(report.encode("utf-8")).hexdigest() != report_sha256: + raise StateError(f"{label} evidence artifact report_sha256 mismatch") + + canonical = canonical_json_bytes(data) + return EvidenceArtifact( + data=data, + digest=hashlib.sha256(canonical).hexdigest(), + source=source, ) -def require_verification_passed(evidence: str, label: str) -> str: - if not evidence.strip(): - raise StateError(f"{label} evidence is required") - evidence_field(evidence, "result", label, "pass") - return evidence_field(evidence, "commit", label) +def snapshot_evidence( + feature: Feature, + purpose: str, + artifact: EvidenceArtifact, +) -> dict[str, str]: + filename = f"{purpose}-{artifact.digest[:16]}.json" + relative_path = Path("evidence") / filename + atomic_write_text( + feature.path / relative_path, + json.dumps( + artifact.data, + ensure_ascii=True, + indent=2, + sort_keys=True, + ) + + "\n", + ) + return { + "path": relative_path.as_posix(), + "sha256": artifact.digest, + } + + +def require_review_passed(reviewed: str) -> tuple[str, str, EvidenceArtifact]: + artifact = load_evidence_artifact(reviewed, "review", "review") + return ( + artifact_string(artifact.data, "commit", "review"), + artifact_string(artifact.data, "base", "review"), + artifact, + ) + + +def require_verification_passed( + evidence: str, label: str +) -> tuple[str, EvidenceArtifact]: + artifact = load_evidence_artifact(evidence, label, "verification") + return artifact_string(artifact.data, "commit", label), artifact def normalize_commit(repo_root: Path, commit: str, label: str) -> str: @@ -1110,6 +1198,57 @@ def finish_nonresolved_ticket( return f"{result.upper()}={feature_slug}/{ticket_number}" +def release_blocked_ticket( + state_root: Path, + feature_slug: str, + ticket_number: str, + reason: str, +) -> str: + """Return a blocked ticket to the queue without its original owner. + + ``finish --result released`` is owner-scoped and ``reclaim`` only accepts + claimed tickets, so a blocked ticket whose session is gone would otherwise + have no owner left that could unblock it. Branch, worktree and uncommitted + changes are preserved exactly as ``finish --result released`` leaves them. + """ + if not reason.strip(): + raise StateError("ticket release requires a reason") + state_root = state_root.resolve() + with locked_state(state_root): + feature = load_feature(state_root, feature_slug) + ticket = feature.tickets.get(ticket_number) + if ticket is None: + raise StateError(f"{feature_slug}: ticket {ticket_number} not found") + if ticket.status != "blocked": + raise StateError( + f"ticket {feature_slug}/{ticket_number} is not blocked: {ticket.status}" + ) + metadata = dict(ticket.metadata) + previous_owner = str(metadata.get("claimed_by", "")) + timestamp = utc_now() + append_history( + metadata, + "ticket-released", + timestamp, + reason=reason.strip(), + previous_owner=previous_owner, + ) + if previous_owner: + metadata["last_owner"] = previous_owner + metadata["released_at"] = timestamp + metadata["released_reason"] = reason.strip() + for key in ( + "blocked_reason", + "blocked_at", + "claimed_by", + "claimed_at", + "heartbeat_at", + ): + metadata.pop(key, None) + update_ticket_state(ticket, status="ready-for-agent", metadata=metadata) + return f"TICKET_RELEASED={feature_slug}/{ticket_number}" + + def finish_resolved_ticket( state_root: Path, repo_root: Path, @@ -1128,8 +1267,10 @@ def finish_resolved_ticket( raise StateError("feature head is required") if not review_base: raise StateError("review base is required") - verification_commit = require_verification_passed(verified, "verification") - review_commit, reviewed_base = require_review_passed(reviewed) + verification_commit, verification_artifact = require_verification_passed( + verified, "verification" + ) + review_commit, reviewed_base, review_artifact = require_review_passed(reviewed) state_root = state_root.resolve() repo_root = resolve_repo_root(repo_root) @@ -1271,13 +1412,24 @@ def finish_resolved_ticket( if recovered_integration_commit is not None else git_output(repo_root, "rev-parse", feature_branch) ) + evidence = { + "verification": snapshot_evidence( + feature, + f"ticket-{ticket_number}-verification", + verification_artifact, + ), + "review": snapshot_evidence( + feature, + f"ticket-{ticket_number}-review", + review_artifact, + ), + } metadata.update( { "implementation_commit": implementation_commit, "ticket_head": ticket_tip, "integration_commit": integration_commit, - "verified": verified, - "reviewed": reviewed, + "evidence": evidence, "resolved_at": utc_now(), } ) @@ -1426,13 +1578,13 @@ def integrate_feature( ) -> str: if not feature_head: raise StateError("feature head is required") - feature_verification_commit = require_verification_passed( + feature_verification_commit, feature_verification_artifact = require_verification_passed( verified, "feature verification" ) - main_verification_commit = require_verification_passed( + main_verification_commit, main_verification_artifact = require_verification_passed( main_verified, "main candidate verification" ) - review_commit, reviewed_base = require_review_passed(reviewed) + review_commit, reviewed_base, review_artifact = require_review_passed(reviewed) state_root = state_root.resolve() repo_root = resolve_repo_root(repo_root) @@ -1576,14 +1728,29 @@ def integrate_feature( raise StateError(f"feature merge failed: {detail}") integration_commit = git_output(repo_root, "rev-parse", main_branch) + evidence = { + "feature_verification": snapshot_evidence( + requested, + "feature-verification", + feature_verification_artifact, + ), + "main_candidate_verification": snapshot_evidence( + requested, + "main-candidate-verification", + main_verification_artifact, + ), + "review": snapshot_evidence( + requested, + "feature-review", + review_artifact, + ), + } requested_state.update( { "feature_branch": feature_branch, "feature_head": current_feature_head, "integration_commit": integration_commit, - "verified": verified, - "main_verified": main_verified, - "reviewed": reviewed, + "evidence": evidence, "partial_authorized": bool(allow_partial), "integrated_at": utc_now(), } @@ -1659,6 +1826,12 @@ def build_parser() -> argparse.ArgumentParser: release_feature.add_argument("--state-root", default=".scratch") release_feature.add_argument("--feature", required=True) + release_ticket = subparsers.add_parser("release-ticket") + release_ticket.add_argument("--state-root", default=".scratch") + release_ticket.add_argument("--feature", required=True) + release_ticket.add_argument("--ticket", required=True) + release_ticket.add_argument("--reason", required=True) + integrate = subparsers.add_parser("integrate") integrate.add_argument("--state-root", default=".scratch") integrate.add_argument("--repo-root", default=".") @@ -1716,6 +1889,13 @@ def main(argv: list[str]) -> int: state_root, args.feature, ) + elif args.command == "release-ticket": + message = release_blocked_ticket( + state_root, + args.feature, + args.ticket, + args.reason, + ) elif args.command == "integrate": message = integrate_feature( state_root, @@ -1728,7 +1908,7 @@ def main(argv: list[str]) -> int: args.main_branch, args.allow_partial, ) - elif args.result == "resolved": + elif args.command == "finish" and args.result == "resolved": message = finish_resolved_ticket( state_root, Path(args.repo_root), @@ -1741,7 +1921,7 @@ def main(argv: list[str]) -> int: args.verified, args.reviewed, ) - else: + elif args.command == "finish": message = finish_nonresolved_ticket( state_root, args.feature, @@ -1750,6 +1930,8 @@ def main(argv: list[str]) -> int: args.result, args.reason, ) + else: + raise StateError(f"unhandled command: {args.command}") except (OSError, StateError, UnicodeError) as exc: print(f"ERROR: {exc}", file=sys.stderr) return 2 diff --git a/scripts/playbook.py b/scripts/playbook.py index 5e04c659..58ffe317 100644 --- a/scripts/playbook.py +++ b/scripts/playbook.py @@ -508,6 +508,32 @@ def extract_block_lines(text: str, start: str, end: str) -> list[str]: _AGENTS_BLOCK_START = "" _AGENTS_BLOCK_END = "" +_RULES_BLOCK_START = "" +_RULES_BLOCK_END = "" + + +def replace_marked_block( + text: str, + block: list[str], + start_marker: str, + end_marker: str, +) -> str: + """Swap the first start..end marked region for ``block``, keeping the rest.""" + updated: list[str] = [] + in_block = False + replaced = False + for line in text.splitlines(): + if not replaced and line.strip() == start_marker: + updated.extend(block) + in_block = True + replaced = True + continue + if in_block: + if line.strip() == end_marker: + in_block = False + continue + updated.append(line) + return "\n".join(updated) + "\n" def preserve_agents_subblock(block: list[str], agents_text: str) -> list[str]: @@ -559,23 +585,10 @@ def update_agents_section( agents_text = agents_path.read_text(encoding="utf-8") if start_marker in agents_text: block = preserve_agents_subblock(block, agents_text) - lines = agents_text.splitlines() - updated: list[str] = [] - in_block = False - replaced = False - for line in lines: - if not replaced and line.strip() == start_marker: - updated.extend(block) - in_block = True - replaced = True - continue - if in_block: - if line.strip() == end_marker: - in_block = False - continue - updated.append(line) agents_path.write_text( - "\n".join(updated) + "\n", encoding="utf-8", newline="\n" + replace_marked_block(agents_text, block, start_marker, end_marker), + encoding="utf-8", + newline="\n", ) log("Updated: AGENTS.md (section)") else: @@ -741,9 +754,6 @@ def sync_rules_action(config: dict, context: dict) -> int: rules_dst = project_root / "AGENT_RULES.md" force = bool(config.get("force", False)) - if rules_dst.exists() and not force: - log("AGENT_RULES.md already exists. Use force to overwrite.") - return 0 project_name = resolve_project_name(context) playbook_scripts = resolve_playbook_scripts(context) @@ -751,13 +761,36 @@ def sync_rules_action(config: dict, context: dict) -> int: date_value = config.get("date") or datetime.now().strftime("%Y-%m-%d") no_backup = bool(config.get("no_backup", False)) - backup_path(rules_dst, no_backup) text = rules_src.read_text(encoding="utf-8") text = replace_placeholders( text, project_name, date_value, playbook_scripts, playbook_root ) - rules_dst.write_text(text.rstrip("\n") + "\n", encoding="utf-8", newline="\n") - log("Synced: AGENT_RULES.md") + + if rules_dst.exists() and not force: + # The process itself is Playbook-owned, so keep it upgradable: refresh the + # marked block in place and leave anything the project added outside it. + # Files predating the markers still need force, as before. + block = extract_block_lines(text, _RULES_BLOCK_START, _RULES_BLOCK_END) + existing = rules_dst.read_text(encoding="utf-8") + if not block: + log("Skip: rules markers not found in template") + return 0 + if _RULES_BLOCK_START not in existing: + log("AGENT_RULES.md has no playbook:rules block. Use force to overwrite.") + return 0 + updated = replace_marked_block( + existing, block, _RULES_BLOCK_START, _RULES_BLOCK_END + ) + if updated == existing: + log("Unchanged: AGENT_RULES.md (section)") + else: + backup_path(rules_dst, no_backup) + rules_dst.write_text(updated, encoding="utf-8", newline="\n") + log("Updated: AGENT_RULES.md (section)") + else: + backup_path(rules_dst, no_backup) + rules_dst.write_text(text.rstrip("\n") + "\n", encoding="utf-8", newline="\n") + log("Synced: AGENT_RULES.md") local_rules = project_root / "AGENT_RULES.local.md" if not local_rules.exists(): diff --git a/templates/AGENT_RULES.template.md b/templates/AGENT_RULES.template.md index 4ba8b8d8..cdaf4e6c 100644 --- a/templates/AGENT_RULES.template.md +++ b/templates/AGENT_RULES.template.md @@ -1,5 +1,7 @@ # AGENT_RULES + + 目的：为本仓库提供稳定的 Matt Pocock 工程流程与 ticket 执行约束。 `.scratch/` 是 spec、ticket、feature 队列和执行状态的唯一事实源。 @@ -45,28 +47,75 @@ ## 任务入口 -### 可直接执行 +四个入口按成本递增排列，**从上往下取第一个满足的**。每个入口都定义了强制升级条件；条件成立 +时立即停止当前路径并转入指定入口，已产生的改动只作为事实输入，不视为已批准方案。 -只读分析、定位、审查，以及不改变行为的局部机械修改可以直接执行，不生成 -spec/tickets。仍须遵守项目规则、验证实际结果，并只处理本次相关改动。 +### 入口 1：直接执行 -### 新 feature 或设计变更 +只读分析、定位、审查，以及不改变可观察行为的机械修改（重命名、格式化、注释、导入整理、纯 +文档措辞）。不生成 spec/tickets，不建 branch。仍须遵守项目规则、验证实际结果，并只处理本次 +相关改动。 -需要需求澄清、设计取舍、跨 session 调度、并发执行、迁移或回滚时，进入完整工程 -主链。边界不清时也走该入口；执行中才发现存在未确认设计时停止实现，已有改动只作为 -事实输入，不视为已批准方案。 +**升级条件**：出现任何可观察行为变化 → 入口 2。 -### 已明确预期行为的 bug +### 入口 2：单切片改动 -已知正确行为且可以建立失败反馈回路的 bug，使用 `diagnosing-bugs` 完整执行 Phase 1-6， -由它负责反馈回路、回归测试、修复、验证和清理，不再进入 `implement` 或重复 `tdd`，也不为 -已确定的需求重新 grilling。若诊断暴露新的产品取舍、架构方向或缺失测试 seam，停止修复并 -转入 `grill-with-docs`。若 bug 来自已领取 ticket，Phase 6 后直接继续本地 ticket 执行协议的 -提交、review 和 finish 门禁。 +会改变行为、不属于入口 3 的已知 bug，且同时满足以下**全部**条件的改动走这里，不入队、不建 +ticket branch： + +- 只涉及一个模块，或只扩展一个已存在的公开接口 +- 能在已存在的测试 seam 上验证，不需要新建 seam +- 不新增公开接口、配置项、数据格式或第三方依赖 +- 不涉及数据迁移、回滚、兼容性窗口或并发协调 +- 预计一个 session 内交付，不需要跨 session 调度 + +流程：在已存在 seam 上按 `tdd` 完成一个可观察行为切片 → 运行局部验证 → 提交 → 对 +`HEAD` 运行 `code-review` 的 Standards 单轴 → 修完硬 finding 后重新提交和验证。证据 +写入 commit message 和 session 收尾，不写 `.scratch/`。 + +**升级条件**（任一成立即停止并转入口 4）：需要新 seam、跨出上述任一条边界、或发现 +未确认的设计取舍。 + +### 入口 3：已明确预期行为的 bug + +正确行为已知的 bug 走这里，**无论当下能否建立失败反馈回路**——建立反馈回路是 +`diagnosing-bugs` 的 Phase 1，不是进入它的前提。`diagnosing-bugs` 完整执行 Phase 1-6，负责反馈回路、 +复现最小化、假设、回归测试、修复、验证和清理；不另起 `tdd` 会话（Phase 5 内部已是 +test-first），也不为已确定的需求重新 grilling。 + +两种缺口按该 skill 自身的规定处理：Phase 1 建不出可失败的命令时停下来索取环境、artifact 或 +许可，不要换入口绕过；不存在正确 seam 时把"缺失 seam"本身作为发现记录，不因此中断修复， +修复落地后再建议 `improve-codebase-architecture`。诊断暴露新的产品取舍或架构方向时，先完成 +Phase 5-6 让缺陷不再复现，再把取舍带入入口 4，不要把未修完的缺陷留在原地等设计结论。 + +若 bug 来自已领取 ticket，Phase 6 后直接继续本地 ticket 执行协议的提交、review 和 finish +门禁。 + +### 入口 4：新 feature 或设计变更 + +以下任一条成立时进入完整工程主链： + +- 存在需要用户决策的产品取舍或架构方向 +- 需要新的测试 seam，或新增公开接口、配置项、数据格式 +- 跨两个以上模块边界 +- 涉及数据迁移、回滚或兼容性窗口 +- 需要跨 session 调度或多 session 并发 +- 预计无法在一个 session 内交付 + +**边界不清时先按入口 2 起步**，触到入口 2 的任一升级条件时立即转入本入口。"不确定"的正确 +处置是用最小路径试探到边界，不是预付最贵的流程。 + +### 非交互模式下的入口 4 + +主链的 grilling 收敛和 seam confirmation 需要用户在场。无人值守而判定为入口 4 时：已有已领取 +ticket 就 `finish --result blocked --reason "<待确认的具体决策>"`；尚无 ticket 则不得代替用户 +决策、也不得降级到入口 2 硬做。用 `to-questionnaire` 把待决问题写到项目根目录的 +`.scratch/questions/.md`，记录阻塞阶段、已知上下文、待决问题和恢复入口；用户回答后从 +`grill-with-docs` 恢复。不得把问题清单写入临时 worktree 或稳定知识文件。 ## 正式工程主链 -新 feature 或设计变更使用以下顺序： +入口 4 使用以下顺序： ```text setup-matt-pocock-skills @@ -80,15 +129,18 @@ setup-matt-pocock-skills ``` - 每个仓库首次使用时运行 `setup-matt-pocock-skills`，本地开发选择 local markdown tracker -- 正式仓库设计使用 `grill-with-docs` - 进入 `grill-with-docs` 或本地 ticket 执行协议前，重新读取 `memory-bank/project-brief.md`、 `memory-bank/tech-context.md` 和 `memory-bank/system-patterns.md` -- `grilling` 必须走完整 design tree，frontier 清空并经用户确认后才进入 `to-spec` -- `to-spec` 不重新进行已经完成的需求采访；但必须按其原始流程确认测试 seam。若 - grilling 尚未确认 seam，必须先向用户完成 seam confirmation -- `tdd` 不得在未经确认的 seam 上开始 +- `grilling` 必须走完整 design tree，**design frontier**（尚未定下的决策集合）清空并经用户 + 确认后才进入 `to-spec`。该词与调度语义里的 ticket frontier 无关 +- seam confirmation 的责任在 `to-spec` 与 `tdd`，不在 `grilling`——后者只收敛设计决策，不涉及 + 测试 seam。`to-spec` 必须按其原始流程与用户确认 seam 并写入 spec 的 Testing Decisions； + `tdd` 不得在未经确认的 seam 上开始 +- `to-spec` 不重新进行已经完成的需求采访 - `to-tickets` 产出可独立验证的 tracer-bullet tickets，并显式声明 `Blocked by` - 一次可以先生成多个 feature 的 spec/tickets，再按期望顺序逐个 `enqueue` +- `codebase-design` 是 seam、deep module 与依赖分类的词汇来源，供 `to-spec` 和 `tdd` 查阅， + 不作为独立会话运行 ## On-ramps 与 detours @@ -101,22 +153,24 @@ setup-matt-pocock-skills ## Phase boundaries -只在阶段边界按以下顺序判断，首个满足项生效： +只在阶段边界判断去向，阶段中途不做这个决定。按以下顺序取首个满足项： -1. 下一阶段需要当前 primary source 且 smart zone 足够时，继续当前 session +1. 下一阶段需要当前 session 作为 primary source，或剩余 smart zone 仍够下一阶段 + （约 150k tokens）时，继续当前 session 2. 当前上下文与下一阶段无关时，使用 `clear` 3. 仅在跨 harness、跨目录/仓库、交给同事或 mid-phase 分出旁支任务时使用 `handoff` 4. 任务可独立 AFK 完成时交给 subagent -5. 其余同 harness、同目录且仍需当前上下文的情况使用 `compact` +5. 其余同 harness、同目录且仍需当前上下文的情况使用 `compact`，并附上下一阶段要保留 + 什么的指令 -上下文过长不等于必须 `handoff`；`handoff` 解决的是可移植性。`compact` 是决策树的默认 -落点，但不是第一选择。`CONTEXT.md` 和 memory-bank 都不能替代阶段上下文。 +上下文过长不等于必须 `handoff`；`handoff` 解决的是可移植性。`compact` 是决策树的默认落点， +但不是第一选择——除 continue 以外的每个选项都把 primary source 换成 secondary source。 +`CONTEXT.md` 和 memory-bank 都不能替代阶段上下文。 ## 本地 Ticket 执行协议 -这是 Playbook 对 Matt 工程 skills 的调度适配层。它复用 `tdd` 和 `code-review`。 -不直接调用上游 `implement`；后者的 `code-review -> commit` 尾部顺序无法生成主循环要求的 -固定 commit 证据。 +这是 Playbook 对 Matt 工程 skills 的调度适配层，只复用 `tdd` 和 `code-review`。尾部顺序与 +上游的 `code-review -> commit` 相反：主循环要求证据绑定到固定 commit，所以先提交再 review。 对每个 claim 严格按以下顺序执行： @@ -139,12 +193,18 @@ setup-matt-pocock-skills - `.scratch/queue.md`：feature 开发与集成顺序 - `docs/agents/*.md`：tracker、领域文档布局和 skill 配置 -不得手工修改 ticket 的 `Status` 或 `main-loop:ticket-state` 区块；只通过主循环变更。 +不得手工修改 ticket 的 `Status` 或 `main-loop:ticket-state` 区块；只通过主循环变更。主循环 +没有对应命令的状态组合按"卡死与恢复"处理，仍然不手工改。 + +`.scratch/` 是否纳入版本控制由项目决定并写入 `AGENT_RULES.local.md`：纳入则 spec、ticket 和 +证据进入历史，可审计、新 clone 能接手队列，代价是状态变更产生提交噪音；排除则历史干净， +代价是状态只存在于本机磁盘、集成后审计线索消失，并发只靠"共享同一文件系统"兜住。无论哪种 +都不得把两份 `.scratch/` 当成同一队列。 ## 稳定知识维护 -只记录下一 session 仍需要的稳定知识；当前 feature、ticket、owner、heartbeat、验证和 -集成状态只由 `.scratch/` 与 `main_loop.py` 维护。 +只记录下一 session 仍需要的稳定知识；当前 feature、ticket、owner、heartbeat、验证和集成 +状态只由 `.scratch/` 与 `main_loop.py` 维护。 - 项目定位、边界、目标或成功定义长期变化时，更新 `project-brief.md` - 技术栈、工具链、环境差异或验证入口长期变化时，更新 `tech-context.md`；写入 @@ -153,14 +213,13 @@ setup-matt-pocock-skills - `CONTEXT.md` 只记录稳定领域词汇和定义；不把聊天流水、未验证猜测或短期进度写入 `CONTEXT.md` - 项目特有执行规则写入 `AGENT_RULES.local.md` -- 符合 phase-boundary 窄条件的可移植上下文由 `handoff` 写入 OS 临时目录，不写入稳定 - 知识文件；同 harness、同目录的续作优先按决策树选择 continue、clear 或 compact +- `handoff` 的产物写入 OS 临时目录，不写入稳定知识文件 - 不为普通实现选择创建 ADR；没有长期价值的信息时不更新这些文件 ## 调度语义 - feature 按 `.scratch/queue.md` 顺序调度 -- 同一 feature 中 blocker 全部满足的 tickets 构成 frontier，按稳定编号领取 +- 同一 feature 中 blocker 全部满足的 tickets 构成 ticket frontier，按稳定编号领取 - 同一 feature 的多个 frontier tickets 可在 worktree 模式并发执行 - 当前 feature 的 frontier 全被领取时返回 `BUSY`，不向后续 feature 扩张 - 只有前序 feature 无 frontier、无活动 claim 且确实 blocked 时，才可开发后序 feature @@ -171,27 +230,62 @@ setup-matt-pocock-skills - 含 `skipped` ticket 的 feature 标记为 partial，集成时必须显式授权 - stale 只由 heartbeat 时间派生，不自动转移 owner；只有 `reclaim` 可以接管 +### 读取输出而不是只看返回码 + +`BUSY`、`NOOP: ...`、`RETRY: ...` 都写到 stdout 并且**返回码为 0**；只有真正的错误才是 +`ERROR: ` 到 stderr、返回码 2。调用方必须解析 stdout，不能用返回码判断是否拿到 ticket。 + +`status` 输出 `FEATURE=`、`STATE=`、`PARTIAL=`、`FRONTIER=`、`CLAIM=`、`OWNER=`、 +`HEARTBEAT=`、`STALE=`、`ISOLATION=`、`WORKSPACE=`、`TICKET_ERROR=`、`BLOCKED=`、 +`FEATURE_BLOCKED=`、`MAIN_INTEGRATION_COMMIT=`；队列为空时输出 `NO FEATURES`。 + +`STATE=active` 不代表可推进：tickets 只剩 `resolved` 与 `blocked` 的 feature 同样报 +`STATE=active FRONTIER=-`，此时 `claim` 返回 `NOOP`。判断是否卡住要同时看 `FRONTIER=-` 和 +是否存在 `BLOCKED=` 行。 + +### 心跳的实际作用 + +heartbeat 没有强制力：过期 claim 仍保有全部权限，原 owner 可以继续 heartbeat、`finish` 和 +resolve。stale 的唯一后果是**别人获得 `reclaim` 的资格**。所以"每 10 分钟一次"是为了让接管 +判断准确，不是为了保住 claim。 + +### 卡死与恢复 + +- claim 的环境准备失败时，主循环把该 ticket 写成 `blocked`、记录原因、把 `claimed_by` 设为 + 本次 owner，并以返回码 2 退出。一次失败的 claim 会占住这张 ticket +- `finish --result released` 只允许原 `claimed_by` 解除 blocked；原 session 已丢失时用 + `release-ticket` +- `reclaim` 只接管 `claimed` 状态的 stale claim，不能用于 `blocked` ticket +- resume 与 `reclaim` 返回的 `BASE` 是 claim 当时记录的值，**不会刷新**；feature head 已推进 + 时直接拿它去 `finish` 会得到 `RETRY: feature advanced`，按重试流程取新的 `FEATURE_HEAD` + ## 执行隔离 -`claim --isolation` 支持： +`claim --isolation`：`in-place` 串行执行不建额外 worktree；`worktree` 每个 ticket 独立 +branch/worktree，适用于多 session 并发；`auto` 无其他活动 claim 时用 in-place，否则用 +worktree。计划并发时第一个 session 就必须指定 `worktree`。 -- `in-place`：串行执行，不创建额外 worktree；claim 生命周期内其他 owner 得到 `BUSY` -- `worktree`：每个 ticket 独立 branch/worktree，适用于多 session 并发 -- `auto`：无其他活动 claim 时使用 in-place，否则使用 worktree +`BUSY` 的作用域是**整个队列，不限于同一 feature**，来源有三个：请求 `in-place` 而队列里存在 +任何活动 claim；队列里存在任何 `in-place` 活动 claim（此时所有 isolation 的新 claim 都 +BUSY）；队首可调度 feature 的 frontier 已被领完。stale claim 也算活动 claim，必须显式 +`reclaim` 或 `release-ticket` 才能腾出位置。同一 owner 恢复自己的 ticket 不受这些检查影响。 -计划并发时，第一个 session 就必须指定 `worktree`。明确不并发时可使用 -`in-place` 或 `auto`，不必创建 worktree。 +in-place 要求 checkout 无非 `.scratch` 改动、HEAD 非 detached、目标 branch 未被其他 worktree +占用。主循环不得自动 stash、reset、覆盖或丢弃改动。唯一例外：重新 claim 自己此前以 in-place +释放的同一 ticket 且 workspace/branch 都匹配时，跳过 dirty 检查以保留未提交改动。 -in-place 模式要求 checkout 无非 `.scratch` 改动、HEAD 非 detached，且目标 branch -未被其他 worktree 占用。主循环不得自动 stash、reset、覆盖或丢弃改动。 +worktree 模式下若 `feature/` 正被 control checkout 占用，主循环会先确认它干净、再把 +control checkout 切到主干以释放该 branch。 -`main_loop.py` 只支持 local Markdown tracker，并要求所有 sessions 共享同一文件系统、 -control checkout 和 Git common directory。远程 tracker 可以由 Matt skills 单独使用， -但本主循环没有远程 tracker adapter，不能接入远程 claim/finish 状态。跨机器或独立 -clone 的并发不受支持，不得把两份 `.scratch/` 当成同一队列。 +`main_loop.py` 只支持 local Markdown tracker，要求所有 sessions 共享同一文件系统、control +checkout 和 Git common directory。远程 tracker 可以由 Matt skills 单独使用，但本主循环没有 +远程 tracker adapter，不能接入远程 claim/finish 状态；跨机器或独立 clone 的并发不受支持。 ## 主循环命令 +主干 branch 名不是 `main` 时，`claim` 和 `integrate` 必须显式传 `--main-branch `；其余 +子命令不接触主干，也不需要 `--repo-root`。 + ### 入队和状态 ```bash @@ -202,41 +296,42 @@ python {{PLAYBOOK_SCRIPTS}}/main_loop.py status \ --state-root .scratch ``` +`enqueue` 对已入队 feature 幂等，返回 `EXISTS=`。 + ### 领取 -`` 必须是当前 session 全局唯一且稳定的标识，例如 -`--`。不得在并行 session 间复用 `codex`、`claude` -等通用名称；同一 owner 的重复 claim 只用于原 session 恢复自己的 ticket。 +`` 必须全局唯一且在本 session 内稳定，例如 `--`。 +不得在并行 session 间复用 `codex`、`claude` 等通用名称；同一 owner 重复 claim 只用于原 +session 恢复自己的 ticket，其他 owner 不得接管，失联时用显式 reclaim。 ```bash python {{PLAYBOOK_SCRIPTS}}/main_loop.py claim \ --state-root .scratch --repo-root . \ - --owner "" --isolation in-place|worktree|auto + --owner "" --isolation in-place|worktree|auto \ + [--main-branch main] ``` -stdout 返回 `FEATURE`、`TICKET`、`CONTROL_ROOT`、绝对 `STATE_ROOT`、`WORKSPACE`、 -`BRANCH`、`BASE` 和 `ISOLATION`： - -- `CONTROL_ROOT` 是共享 Git control checkout -- `STATE_ROOT` 是唯一状态目录；它不必位于 `WORKSPACE` 内 -- `WORKSPACE` 是当前 ticket 的代码工作区 +stdout 返回 8 个赋值行：`FEATURE`、`TICKET`；`CONTROL_ROOT` 是共享 Git control checkout； +绝对 `STATE_ROOT` 是唯一状态目录，不必位于 `WORKSPACE` 内；`WORKSPACE` 是当前 ticket 的代码 +工作区；其余为 `BRANCH`、`BASE`、`ISOLATION`。也可能返回 `NO FEATURES`、 +`NOOP: no claimable tickets`、`BUSY`，或 `INTEGRATION_REQUIRED=`（队首 feature 已 +ready-to-integrate，必须先 `integrate`）。 领取成功后立即读取 `//spec.md` 和 -`//issues/-*.md`；不得在 claim 前根据 frontier 猜测本 session -将领取哪个 ticket。后续 review 也必须使用这两个已领取上下文，而不是 worktree 内相对 -`.scratch` 的偶然副本。 +`//issues/-*.md`；不得在 claim 前根据 frontier 猜测将领取哪个 +ticket。后续 review 也必须用这两个已领取上下文，而不是 worktree 内相对 `.scratch` 的偶然 +副本。 -实现、提交和 review 必须在返回的 `WORKSPACE`/`BRANCH` 中完成。heartbeat、reclaim、 -finish、status、block/release-feature 和 integrate 必须使用 -`--state-root ""`，不得在 ticket worktree 中使用相对 `.scratch`。 - -同一 owner 重复 claim 恢复原 ticket。其他 owner 不得接管；失联时使用显式 reclaim。 +实现、提交和 review 在返回的 `WORKSPACE`/`BRANCH` 中完成；其余子命令一律用 +`--state-root ""`，不得在 ticket worktree 中使用相对 `.scratch`。feature 集成 +workspace 的路径不在 claim 输出里，只写在 `//.main-loop.json` 的 +`integration_workspace`；纯 in-place 流程不创建该文件，feature 级操作在 `CONTROL_ROOT` 上做。 ### 心跳和接管 claim 的 stale 租约固定为 30 分钟，调用者不得缩短。claim 存续期间至少每 10 分钟发送一次 -heartbeat，并在预计耗时较长的验证、构建或 review 前后各发送一次。无法继续维持 heartbeat -时，使用 `finish --result released|blocked` 明确交还或阻塞 ticket。 +heartbeat，并在耗时较长的验证、构建或 review 前后各发送一次。无法继续维持时，根据是否需要 +保留阻塞原因分别执行下文完整的 `--result released` 或 `--result blocked --reason` 命令。 ```bash python {{PLAYBOOK_SCRIPTS}}/main_loop.py heartbeat \ @@ -248,35 +343,36 @@ python {{PLAYBOOK_SCRIPTS}}/main_loop.py reclaim \ --feature --ticket --owner "" ``` -reclaim 只允许接管 stale claim，并保留原 branch、worktree 和未提交改动。 +reclaim 只接管 stale 的 `claimed` ticket，要求新 owner 与原 owner 不同，保留原 branch、 +worktree 和未提交改动。被接管后原 owner 的 heartbeat 与 finish 立即报 +`owned by another session`。 ### Review 适配契约 -调用 Matt `code-review` 时不得依赖其自动搜索或交互补问，必须显式提供： +`code-review` 只产出 Standards 与 Spec 两段 findings，不给 pass/fail 判定，也不合并或 +重排 findings。下面的 `pass` 映射是 Playbook 加的一层，由调用方把 findings 归结为判定。 -- fixed point：ticket review 使用 claim 的 `BASE` 或重试返回的 `FEATURE_HEAD`；feature review - 使用最新 `main` HEAD -- ticket review 的 Spec sources：逻辑路径 `.scratch//spec.md` 与 - `.scratch//issues/-*.md`，实际从绝对 `STATE_ROOT` 解析 -- feature review 的 Spec sources：`.scratch//spec.md` 和该 feature 的全部 ticket - acceptance criteria，实际从绝对 `STATE_ROOT` 解析 +调用时不得依赖其自动搜索或交互补问，必须显式提供： -只有对应 axis 零个未解决的硬 finding 时才能记录 `pass`： +- fixed point：ticket review 用 claim 的 `BASE` 或重试返回的 `FEATURE_HEAD`；feature + review 用最新 `main` HEAD +- Spec sources：ticket review 用 `.scratch//spec.md` 与 + `.scratch//issues/-*.md`；feature review 用该 spec 加本 feature 全部 + ticket 的 acceptance criteria。两者都从绝对 `STATE_ROOT` 解析 -- `standards=pass`：没有未解决的仓库标准违规；baseline smell 属 judgement call，必须逐项记录 - 已修复或带理由接受，但不会仅因被提出就自动失败 -- `spec=pass`：没有遗漏、部分实现、错误实现或未授权范围扩张 +只有对应 axis 零个未解决的硬 finding 时才能记录 `pass`：`standards=pass` 指没有未解决的 +仓库标准违规（baseline smell 属 judgement call，须逐项记录已修复或带理由接受，不因被提出 +就自动失败）；`spec=pass` 指没有遗漏、部分实现、错误实现或未授权范围扩张。 任一来源缺失、Spec axis 被跳过、review 尚在询问输入，或仍有未解决的硬 finding 时， -不得据此填写 `standards=pass` 或 `spec=pass`。修复会改变 `HEAD`，因此必须重新运行受影响 -验证和完整双轴 review，不能沿用旧报告。 +不得据此填写 `standards=pass` 或 `spec=pass`。修复会改变 `HEAD`，必须重跑受影响验证和完整双轴 +review，不能沿用旧报告。 ### Ticket 完成或状态转换 `resolved` 前必须先提交实现，再按 Review 适配契约对 `BASE...HEAD` 运行 Matt -`code-review` 的 Standards/Spec 双轴审查。 -验证证据必须包含被验证 commit，review 证据必须包含被审查 commit 和固定 base；任意 -非空文本或裸 `pass` 不能替代结构化证据。 +`code-review` 的 Standards/Spec 双轴审查。验证证据必须包含被验证 commit，review 证据必须 +包含被审查 commit 和固定 base；任意非空文本或裸 `pass` 不能替代结构化证据。 ```bash python {{PLAYBOOK_SCRIPTS}}/main_loop.py finish \ @@ -285,48 +381,44 @@ python {{PLAYBOOK_SCRIPTS}}/main_loop.py finish \ --result resolved \ --implementation-commit \ --feature-head <已验证feature HEAD> --review-base <同一feature HEAD> \ - --verified "commit=; result=pass; <实际局部验证证据>" \ - --reviewed "commit=; base=; standards=pass; spec=pass" + --verified "" \ + --reviewed "" ``` +验证和 review 参数必须是下文定义的 UTF-8 JSON artifact 路径，不接受内联 `pass` 文本。 + 首次 review base 是 claim 返回的 `BASE`。feature HEAD 已推进时返回 `RETRY: feature advanced` 和新的 `FEATURE_HEAD`；把该 `FEATURE_HEAD` 合入 ticket branch， 重新运行受影响验证，并以它作为新的 review base 重跑双轴 review。重试时 -`--feature-head`、`--review-base` 和 review evidence 的 `base` 必须都使用该新值，验证与 -review evidence 的 `commit` 必须等于新的 ticket `HEAD`。只有成功集成到 feature branch -后 ticket 才变为 `resolved`。 +`--feature-head`、`--review-base` 和 review evidence 的 `base` 都用该新值，验证与 review +evidence 的 `commit` 必须等于新的 ticket `HEAD`。只有成功集成到 feature branch 后 ticket +才变为 `resolved`。 -其他转换： +`finish` 不检查验收勾选框；验收由 `code-review` 的 Spec axis 负责，勾选框在 resolve 前 +自行更新。 + +其他 `finish` 转换共用上面的 `--state-root`、`--repo-root`、`--feature`、`--ticket`、`--owner`： ```bash python {{PLAYBOOK_SCRIPTS}}/main_loop.py finish \ --state-root "" --repo-root "" \ --feature --ticket --owner "" \ - --result blocked|skipped --reason "<明确原因>" + --result blocked --reason "<明确原因>" python {{PLAYBOOK_SCRIPTS}}/main_loop.py finish \ --state-root "" --repo-root "" \ --feature --ticket --owner "" \ --result released + +python {{PLAYBOOK_SCRIPTS}}/main_loop.py release-ticket \ + --state-root "" --feature --ticket \ + --reason "<原 session 无法恢复的原因>" ``` -release 返回 `ready-for-agent` 并保留 workspace；blocked ticket 可由原 owner release 后恢复。 - -### Feature 集成阻塞 - -ready-to-integrate feature 若因 main 同步冲突或人工决策暂时不能集成，必须显式记录； -这样主循环才会继续为后序 feature 分配开发工作： - -```bash -python {{PLAYBOOK_SCRIPTS}}/main_loop.py block-feature \ - --state-root "" --feature --reason "<明确原因>" - -python {{PLAYBOOK_SCRIPTS}}/main_loop.py release-feature \ - --state-root "" --feature -``` - -release 后该 feature 重新成为队首 `INTEGRATION_REQUIRED`。后序 feature 即使已开发完成， -仍不得越过它集成到 main。 +`blocked` 和 `skipped` 都必须提供 `--reason`。`released` 回到 `ready-for-agent` 并保留 workspace， +只能由原 `claimed_by` 调用。`release-ticket` 用于原 session 已丢失、无人能 release 的 blocked +ticket；它不校验 owner，保留 branch、worktree 和未提交改动，返回 +`TICKET_RELEASED=/`。 ### Feature 顺序集成 @@ -336,40 +428,71 @@ feature 必须吸收最新 `main`，完成 feature 验证、main 候选验证和 python {{PLAYBOOK_SCRIPTS}}/main_loop.py integrate \ --state-root "" --repo-root "" \ --feature --feature-head <已验证feature HEAD> \ - --verified "commit=; result=pass; " \ - --main-verified "commit=; result=pass; " \ - --reviewed "commit=; base=<最新main HEAD>; standards=pass; spec=pass" + --verified "" \ + --main-verified "" \ + --reviewed "" ``` -partial feature 还需 `--allow-partial`。返回 `RETRY: feature needs main sync` 时， -先在 feature integration workspace 合并最新 main；若发生冲突，先 `block-feature`，解决后 -`release-feature`，重新验证和 review 后再调用。 +partial feature 还需 `--allow-partial`；未显式授权时报错而不是静默集成。 + +`RETRY: feature needs main sync` 表示 `main` 不是 feature head 的祖先。此时在 feature +integration workspace（纯 in-place 流程下是 `CONTROL_ROOT`）合并最新 main，重新验证和 +review 后再调用。main 合并冲突时主循环会**自动**把 feature 标记为 blocked，不需要手工 +`block-feature`；解决冲突后 `release-feature` 再重试。 + +成功后主循环会把 control checkout 切到主干，并移除干净的 ticket worktree 与 +`_integration` worktree；不干净或路径异常的保留并以 `WARNING=` 行报告。已集成的 feature +再次调用时幂等返回 `INTEGRATED=`。 + +ready-to-integrate feature 因人工决策暂时不能集成时必须显式记录，否则主循环不会为后序 +feature 继续分配开发工作： + +```bash +python {{PLAYBOOK_SCRIPTS}}/main_loop.py block-feature \ + --state-root "" --feature --reason "<明确原因>" + +python {{PLAYBOOK_SCRIPTS}}/main_loop.py release-feature \ + --state-root "" --feature +``` + +release 后该 feature 重新成为队首 `INTEGRATION_REQUIRED`；后序 feature 即使已开发完成， +仍不得越过它集成到 main。 ## Git 与证据门禁 -1. 每个 ticket 使用 `ticket//-` branch -2. 每个 feature 使用 `feature/` branch -3. 按 TDD 完成一个可观察行为切片并运行局部验证 -4. 提交全部实现，使 `HEAD` 成为可审查固定点 -5. 首次以 claim 的 `BASE` 运行 `code-review`；同步推进后的 feature 时改用新的 - `FEATURE_HEAD` -6. 修复 finding 后重新提交、验证和 review -7. `finish` 校验验证/review commit、review base 和 feature HEAD，再集成 ticket -8. `integrate` 把两级验证和最终 review 绑定到 feature HEAD，并把 review base 绑定到 - 最新 main HEAD +branch 命名：ticket 用 `ticket//-`，feature 用 `feature/`。 +执行顺序见"本地 Ticket 执行协议"，证据参数见"主循环命令"。 局部 ticket 验证、feature 验证和 main 候选验证是三个独立门禁，不能互相替代。 -证据必须来自实际 fresh run；无法运行时不得伪造 `pass`。 + +`--verified`、`--main-verified` 和 `--reviewed` 必须指向 UTF-8 JSON artifact。验证 artifact +包含 `version=1`、`kind=verification`、`commit`、`result=pass`、`command`、`exit_code=0`、 +`output` 和 `output_sha256`；review artifact 包含 `version=1`、`kind=review`、`commit`、`base`、 +`standards=pass`、`spec=pass`、`report` 和 `report_sha256`。输出和报告必须先移除 secret；摘要是 +对应 UTF-8 文本的 SHA-256。 + +门禁校验每个 commit 真实存在、artifact schema 与摘要、implementation/verification/review +commit 与目标 branch tip 一致、review base 正确、workspace 干净且在自己的 branch 上。成功后 +把规范化 artifact 快照到 `.scratch//evidence/`，并在 ticket 或 feature 状态中记录快照 +路径和 SHA-256；原始 artifact 删除后仍可审计。 + +artifact 能阻止空文本、字段缺失、摘要篡改和 Git 上下文错配，但无法证明命令真的执行过，也 +无法证明 report 来自真实 review。证据必须来自实际 fresh run，无法运行时不得伪造 `pass`； +需要机器强制时由 CI 或 pre-commit hook 生成 artifact。 ## 辅助能力 以下能力在对应阶段内按需使用，不替代上面的入口路由和证据门禁： -| Skill | 触发条件 | -| ---------------- | ----------------------------------- | -| `codebase-recon` | 架构、跨模块、重构、迁移或风险不明 | -| `brooks-audit` | 架构边界和长期维护性审查 | -| `commit-message` | 需要检查 staged diff 或生成提交信息 | +| Skill | 触发条件 | +| ------------------------------- | -------------------------------------------- | +| `codebase-recon` | 架构、跨模块、重构、迁移或风险不明 | +| `brooks-audit` | 架构边界和长期维护性审查 | +| `codebase-design` | 需要 seam、deep module、依赖分类的词汇与判据 | +| `improve-codebase-architecture` | 修复落地后暴露出的结构问题 | +| `resolving-merge-conflicts` | feature 吸收 main 或集成时出现冲突 | +| `to-questionnaire` | 需要把待决问题整理成清单交回用户 | +| `commit-message` | 需要检查 staged diff 或生成提交信息 | ## 需要确认的场景 @@ -379,19 +502,18 @@ partial feature 还需 `--allow-partial`。返回 `RETRY: feature needs main syn - 需要 spec 未授权的行为、兼容性或架构取舍 - 需要破坏性操作、覆盖他人改动或扩大任务范围 -已明确预期行为的 bug 按 `diagnosing-bugs` 路由推进；符合既有模式的小改动、测试补齐和 -已批准 ticket 范围内的实现不重复需求采访。 -无交互模式下不得替用户作设计决策；将 ticket 标记 blocked 并记录原因。 +入口 1、入口 2 范围内的改动和已批准 ticket 范围内的实现不重复需求采访。无交互模式按 +"非交互模式下的入口 4"处理，不得替用户作设计决策。 ## Session 收尾 - 运行与声明相匹配的 fresh verification -- ticket 执行必须通过 `finish` 写回；不得手工改状态 - 列出已完成、未完成、验证证据、风险和下一步 - 只提交当前 ticket/feature 相关改动，不混入其他 session 差异 -- 工作未结束时在阶段边界按 Continue -> clear -> handoff -> subagent -> compact 的顺序选择 - 去向；不要仅因上下文过长就创建 handoff +- 工作未结束时按 Phase boundaries 决策树选择去向 --- **最后更新**：{{DATE}} + + diff --git a/templates/README.md b/templates/README.md index e033cb72..d1960792 100644 --- a/templates/README.md +++ b/templates/README.md @@ -42,13 +42,15 @@ templates/ - 更新标记区块时保留区块外的项目内容 - 已有的人工入口和项目补充说明继续由项目维护 -### 2. 初始化后由项目维护 +### 2. 流程由 Playbook 维护，补充由项目维护 -- `AGENT_RULES.md`：项目采用的完整工程流程与执行规则 +- `AGENT_RULES.md`：`` 区块内的工程流程由 Playbook + 维护，重新同步时刷新；项目自己的补充写在区块外，同步不会动它 - `AGENT_RULES.local.md`：项目私有规则，Playbook 不覆盖 - `memory-bank/`：稳定项目定位、技术上下文和当前系统模式，不承载机器状态 Playbook 只处理框架提供的同名文件。项目新增的 `memory-bank/*` 不会被删除。 +没有标记区块的旧 `AGENT_RULES.md` 保持原样，不会被静默改写。 ### 3. 参考模板 @@ -72,15 +74,23 @@ heartbeat、ticket 或集成状态；机器状态只以 `.scratch/` 和 `main_lo ### AGENT_RULES.template.md -部署为项目的 `AGENT_RULES.md`，是 Matt Pocock 工程主链、ticket 调度、worktree、 -验证和集成协议的唯一流程权威。项目私有规则写入 `AGENT_RULES.local.md`；该文件由 -项目维护，Playbook 不覆盖。 +部署为项目的 `AGENT_RULES.md`，是任务入口路由、Matt Pocock 工程主链、ticket 调度、 +worktree、验证和集成协议的唯一流程权威。`` 区块内的 +内容由 Playbook 维护并在重新同步时刷新，项目补充写在区块外。项目私有规则写入 +`AGENT_RULES.local.md`；该文件由项目维护，Playbook 不覆盖。 ### AGENTS.template.md 部署为项目的 `AGENTS.md`，只提供语言规则、核心规则和工程上下文导航。 标记区块的更新约束见下文“AGENTS 模板标记”。 +## 任务入口 + +四个入口按成本递增，取第一个满足的：直接执行（零行为变更）、单切片改动（单模块、 +已有 seam、一个 session 内）、已明确预期行为的 bug（`diagnosing-bugs` Phase 1-6）、 +新 feature 或设计变更（完整主链）。边界不清时从单切片改动起步，触到升级条件再转入 +完整主链，不要预付最重的流程。判据与升级条件见 `AGENT_RULES.template.md`。 + ## 部署 ```toml @@ -101,6 +111,8 @@ local markdown tracker。 ## 正式开发流程 +只有第四个入口（新 feature 或设计变更）走这条链；前三个入口不入队、不建 ticket branch。 + ```text setup-matt-pocock-skills -> grill-with-docs diff --git a/test/test_main_loop_cli.py b/test/test_main_loop_cli.py index e9a9c075..91ea7e03 100644 --- a/test/test_main_loop_cli.py +++ b/test/test_main_loop_cli.py @@ -1,4 +1,6 @@ import importlib.util +import hashlib +import json import os import subprocess import sys @@ -52,12 +54,60 @@ def parse_assignments(output: str) -> dict[str, str]: return dict(line.split("=", 1) for line in output.splitlines() if "=" in line) -def verification_evidence(commit: str, detail: str) -> str: - return f"commit={commit}; result=pass; {detail}" +EVIDENCE_TMP = tempfile.TemporaryDirectory() +EVIDENCE_ROOT = Path(EVIDENCE_TMP.name) -def review_evidence(commit: str, base: str) -> str: - return f"commit={commit}; base={base}; standards=pass; spec=pass" +def write_evidence_artifact(data: dict[str, object]) -> str: + path = EVIDENCE_ROOT / f"evidence-{len(tuple(EVIDENCE_ROOT.iterdir())):04d}.json" + path.write_text( + json.dumps(data, ensure_ascii=False, sort_keys=True) + "\n", + encoding="utf-8", + newline="\n", + ) + return str(path) + + +def verification_evidence( + commit: str, + detail: str, + *, + overrides: dict[str, object] | None = None, +) -> str: + output = f"{detail}: passed" + data: dict[str, object] = { + "version": 1, + "kind": "verification", + "commit": commit, + "result": "pass", + "command": detail, + "exit_code": 0, + "output": output, + "output_sha256": hashlib.sha256(output.encode("utf-8")).hexdigest(), + } + data.update(overrides or {}) + return write_evidence_artifact(data) + + +def review_evidence( + commit: str, + base: str, + *, + overrides: dict[str, object] | None = None, +) -> str: + report = "Standards and spec have no unresolved hard findings" + data: dict[str, object] = { + "version": 1, + "kind": "review", + "commit": commit, + "base": base, + "standards": "pass", + "spec": "pass", + "report": report, + "report_sha256": hashlib.sha256(report.encode("utf-8")).hexdigest(), + } + data.update(overrides or {}) + return write_evidence_artifact(data) def init_repo(root: Path) -> None: @@ -131,6 +181,9 @@ class MainLoopCliTests(unittest.TestCase): "finish", "heartbeat", "reclaim", + "block-feature", + "release-feature", + "release-ticket", "integrate", ): self.assertIn(command, result.stdout) @@ -925,6 +978,154 @@ class MainLoopCliTests(unittest.TestCase): self.assertIn("FEATURE=alpha STATE=active PARTIAL=yes", final_status.stdout) self.assertIn("FRONTIER=02", final_status.stdout) + def test_release_ticket_recovers_a_blocked_ticket_without_its_owner(self): + with tempfile.TemporaryDirectory() as tmp_dir: + root = Path(tmp_dir) + init_repo(root) + issues = write_feature(root, "alpha") + write_ticket(issues, "01", "first", "First") + enqueue = run_cli( + "enqueue", "--state-root", ".scratch", "--feature", "alpha", cwd=root + ) + self.assertEqual(enqueue.returncode, 0, msg=enqueue.stderr) + claim = run_cli( + "claim", + "--state-root", + ".scratch", + "--repo-root", + ".", + "--owner", + "lost-session", + "--isolation", + "worktree", + cwd=root, + ) + self.assertEqual(claim.returncode, 0, msg=claim.stderr) + original = parse_assignments(claim.stdout) + blocked = run_cli( + "finish", + "--state-root", + ".scratch", + "--repo-root", + ".", + "--feature", + "alpha", + "--ticket", + "01", + "--owner", + "lost-session", + "--result", + "blocked", + "--reason", + "needs a product decision", + cwd=root, + ) + self.assertEqual(blocked.returncode, 0, msg=blocked.stderr) + + foreign_release = run_cli( + "finish", + "--state-root", + ".scratch", + "--repo-root", + ".", + "--feature", + "alpha", + "--ticket", + "01", + "--owner", + "new-session", + "--result", + "released", + cwd=root, + ) + foreign_reclaim = run_cli( + "reclaim", + "--state-root", + ".scratch", + "--repo-root", + ".", + "--feature", + "alpha", + "--ticket", + "01", + "--owner", + "new-session", + cwd=root, + ) + missing_reason = run_cli( + "release-ticket", + "--state-root", + ".scratch", + "--feature", + "alpha", + "--ticket", + "01", + "--reason", + " ", + cwd=root, + ) + recovered = run_cli( + "release-ticket", + "--state-root", + ".scratch", + "--feature", + "alpha", + "--ticket", + "01", + "--reason", + "original session is gone", + cwd=root, + ) + not_blocked_again = run_cli( + "release-ticket", + "--state-root", + ".scratch", + "--feature", + "alpha", + "--ticket", + "01", + "--reason", + "already released", + cwd=root, + ) + reclaimed = run_cli( + "claim", + "--state-root", + ".scratch", + "--repo-root", + ".", + "--owner", + "new-session", + "--isolation", + "worktree", + cwd=root, + ) + + self.assertEqual(foreign_release.returncode, 2) + self.assertIn("owned by another session", foreign_release.stderr) + self.assertEqual(foreign_reclaim.returncode, 2) + self.assertIn("is not claimed", foreign_reclaim.stderr) + self.assertEqual(missing_reason.returncode, 2) + self.assertIn("requires a reason", missing_reason.stderr) + self.assertEqual(recovered.returncode, 0, msg=recovered.stderr) + self.assertEqual(recovered.stdout.strip(), "TICKET_RELEASED=alpha/01") + self.assertEqual(not_blocked_again.returncode, 2) + self.assertIn("is not blocked", not_blocked_again.stderr) + self.assertEqual(reclaimed.returncode, 0, msg=reclaimed.stderr) + resumed = parse_assignments(reclaimed.stdout) + self.assertEqual(resumed["TICKET"], "01") + self.assertEqual(resumed["WORKSPACE"], original["WORKSPACE"]) + self.assertEqual(resumed["BRANCH"], original["BRANCH"]) + + ticket = MAIN_LOOP.load_feature(root / ".scratch", "alpha").tickets["01"] + self.assertEqual(ticket.metadata["last_owner"], "lost-session") + self.assertEqual( + ticket.metadata["released_reason"], "original session is gone" + ) + self.assertNotIn("blocked_reason", ticket.metadata) + events = [entry["event"] for entry in ticket.metadata["history"]] + self.assertIn("ticket-released", events) + def test_parallel_processes_claim_distinct_frontier_tickets(self): if os.name == "nt": self.assertIsNotNone(MAIN_LOOP.msvcrt) @@ -1953,7 +2154,11 @@ class MainLoopCliTests(unittest.TestCase): "--review-base", context["BASE"], "--verified", - "result=pass; ticket tests", + verification_evidence( + implementation_commit, + "ticket tests", + overrides={"commit": ""}, + ), "--reviewed", review_evidence(implementation_commit, context["BASE"]), cwd=root, @@ -1979,9 +2184,10 @@ class MainLoopCliTests(unittest.TestCase): "--review-base", context["BASE"], "--verified", - ( - f"commit={implementation_commit}; result=fail; " - "result=pass; ticket tests" + verification_evidence( + implementation_commit, + "ticket tests", + overrides={"output_sha256": "0" * 64}, ), "--reviewed", review_evidence(implementation_commit, context["BASE"]), @@ -1990,10 +2196,13 @@ class MainLoopCliTests(unittest.TestCase): status = run_cli("status", "--state-root", ".scratch", cwd=root) self.assertEqual(finish.returncode, 2) - self.assertIn("verification evidence must include commit", finish.stderr) + self.assertIn( + "verification evidence artifact field commit must be a non-empty string", + finish.stderr, + ) self.assertEqual(conflicting.returncode, 2) self.assertIn( - "verification evidence must include exactly one result", + "verification evidence artifact output_sha256 mismatch", conflicting.stderr, ) self.assertIn("CLAIM=alpha/01", status.stdout) @@ -2349,9 +2558,10 @@ class MainLoopCliTests(unittest.TestCase): "--verified", verification_evidence(implementation_commit, "ticket tests"), "--reviewed", - ( - f"commit={implementation_commit}; base={context['BASE']}; " - "standards=pass" + review_evidence( + implementation_commit, + context["BASE"], + overrides={"spec": ""}, ), cwd=root, ) @@ -2365,12 +2575,23 @@ class MainLoopCliTests(unittest.TestCase): ) self.assertEqual(missing_ticket_verification.returncode, 2) self.assertIn( - "verification evidence is required", + "verification evidence artifact is required", missing_ticket_verification.stderr, ) self.assertEqual(incomplete_ticket_review.returncode, 2) - self.assertIn("spec=pass", incomplete_ticket_review.stderr) + self.assertIn( + "review evidence artifact field spec must be a non-empty string", + incomplete_ticket_review.stderr, + ) self.assertEqual(resolved.returncode, 0, msg=resolved.stderr) + ticket = MAIN_LOOP.load_feature(root / ".scratch", "alpha").tickets["01"] + ticket_evidence = ticket.metadata["evidence"] + self.assertEqual(set(ticket_evidence), {"verification", "review"}) + for record in ticket_evidence.values(): + self.assertRegex(record["sha256"], r"^[0-9a-f]{64}$") + self.assertTrue( + (root / ".scratch" / "alpha" / record["path"]).is_file() + ) feature_head = run_git(root, "rev-parse", "feature/alpha").stdout.strip() main_before = run_git(root, "rev-parse", "main").stdout.strip() @@ -2412,13 +2633,21 @@ class MainLoopCliTests(unittest.TestCase): "--main-verified", verification_evidence(feature_head, "main candidate tests"), "--reviewed", - f"commit={feature_head}; base={main_before}; standards=pass", + review_evidence( + feature_head, + main_before, + overrides={"spec": ""}, + ), cwd=root, ) unbound_feature_verification = run_cli( *integrate_args, "--verified", - "result=pass; feature tests", + verification_evidence( + feature_head, + "feature tests", + overrides={"commit": ""}, + ), "--main-verified", verification_evidence(feature_head, "main candidate tests"), "--reviewed", @@ -2437,19 +2666,22 @@ class MainLoopCliTests(unittest.TestCase): ) self.assertEqual(missing_feature_verification.returncode, 2) self.assertIn( - "feature verification evidence is required", + "feature verification evidence artifact is required", missing_feature_verification.stderr, ) self.assertEqual(missing_main_verification.returncode, 2) self.assertIn( - "main candidate verification evidence is required", + "main candidate verification evidence artifact is required", missing_main_verification.stderr, ) self.assertEqual(incomplete_final_review.returncode, 2) - self.assertIn("spec=pass", incomplete_final_review.stderr) + self.assertIn( + "review evidence artifact field spec must be a non-empty string", + incomplete_final_review.stderr, + ) self.assertEqual(unbound_feature_verification.returncode, 2) self.assertIn( - "feature verification evidence must include commit", + "feature verification evidence artifact field commit must be a non-empty string", unbound_feature_verification.stderr, ) self.assertEqual(wrong_final_review_base.returncode, 2) @@ -2478,6 +2710,18 @@ class MainLoopCliTests(unittest.TestCase): self.assertIn("INTEGRATED=alpha", integrated.stdout) self.assertIn("FEATURE=alpha STATE=integrated PARTIAL=no", status.stdout) self.assertTrue((root / "feature.txt").is_file()) + feature_state = MAIN_LOOP.load_feature_state( + MAIN_LOOP.load_feature(root / ".scratch", "alpha") + ) + self.assertEqual( + set(feature_state["evidence"]), + {"feature_verification", "main_candidate_verification", "review"}, + ) + for record in feature_state["evidence"].values(): + self.assertRegex(record["sha256"], r"^[0-9a-f]{64}$") + self.assertTrue( + (root / ".scratch" / "alpha" / record["path"]).is_file() + ) if __name__ == "__main__": diff --git a/test/test_playbook.py b/test/test_playbook.py index d36a9614..4d256fe7 100644 --- a/test/test_playbook.py +++ b/test/test_playbook.py @@ -400,6 +400,81 @@ no_backup = true self.assertIn("`.agents/tsl/index.md`", after) self.assertIn("`.agents/markdown/index.md`", after) + def test_resync_refreshes_the_rules_block_and_keeps_project_additions(self): + with tempfile.TemporaryDirectory() as tmp_dir: + project_root = Path(tmp_dir) / "project" + project_root.mkdir() + playbook_root = MODE_ROOTS["snapshot"] + config = write_config(project_root, "snapshot", playbook_root) + + seed = run_playbook(SCRIPT, config, project_root) + self.assertEqual( + seed.returncode, 0, msg=f"{seed.stdout}{seed.stderr}" + ) + + rules_md = project_root / "AGENT_RULES.md" + seeded = rules_md.read_text(encoding="utf-8") + self.assertIn("", seeded) + self.assertIn("", seeded) + + # A project appendix outside the block, and drift inside it. + appendix = "\n## 项目补充\n\n保留这段项目自己的说明。\n" + drifted = seeded.replace("## 任务入口", "## 任务入口（本地改过）") + appendix + rules_md.write_text(drifted, encoding="utf-8", newline="\n") + + resync = run_playbook(SCRIPT, config, project_root) + self.assertEqual( + resync.returncode, 0, msg=f"{resync.stdout}{resync.stderr}" + ) + + after = rules_md.read_text(encoding="utf-8") + self.assertIn( + "保留这段项目自己的说明。", + after, + msg="content outside the block belongs to the project", + ) + self.assertIn( + "## 任务入口\n", + after, + msg="the process itself is playbook-owned and must be refreshed", + ) + self.assertNotIn("## 任务入口（本地改过）", after) + self.assertEqual(after.count(""), 1) + + legacy = project_root / "legacy" / "AGENT_RULES.md" + legacy.parent.mkdir() + legacy.write_text( + "# AGENT_RULES\n\n没有 marker 的旧文件。\n", + encoding="utf-8", + newline="\n", + ) + legacy_config = project_root / "playbook-legacy.toml" + legacy_config.write_text( + f""" +[playbook] +project_root = "legacy" +playbook_root = "{playbook_root.as_posix()}" +install_mode = "snapshot" + +[sync_rules] +date = "2026-01-01" +no_backup = true +""".lstrip(), + encoding="utf-8", + newline="\n", + ) + legacy_run = run_playbook(SCRIPT, legacy_config, project_root) + self.assertEqual( + legacy_run.returncode, + 0, + msg=f"{legacy_run.stdout}{legacy_run.stderr}", + ) + self.assertEqual( + legacy.read_text(encoding="utf-8"), + "# AGENT_RULES\n\n没有 marker 的旧文件。\n", + msg="a file predating the markers must not be rewritten silently", + ) + if __name__ == "__main__": unittest.main() diff --git a/test/test_template_contracts.py b/test/test_template_contracts.py index a369e3ce..f0373bbd 100644 --- a/test/test_template_contracts.py +++ b/test/test_template_contracts.py @@ -1,9 +1,97 @@ +import argparse +import importlib.util +import re +import sys import unittest from pathlib import Path ROOT = Path(__file__).resolve().parents[1] TEMPLATES = ROOT / "templates" + +_MAIN_LOOP_SPEC = importlib.util.spec_from_file_location( + "playbook_main_loop_contracts", ROOT / "scripts" / "main_loop.py" +) +assert _MAIN_LOOP_SPEC and _MAIN_LOOP_SPEC.loader +MAIN_LOOP = importlib.util.module_from_spec(_MAIN_LOOP_SPEC) +sys.modules[_MAIN_LOOP_SPEC.name] = MAIN_LOOP +_MAIN_LOOP_SPEC.loader.exec_module(MAIN_LOOP) + + +def subcommand_parsers() -> dict[str, argparse.ArgumentParser]: + parser = MAIN_LOOP.build_parser() + for action in parser._subparsers._group_actions: # noqa: SLF001 + if isinstance(action, argparse._SubParsersAction): # noqa: SLF001 + return dict(action.choices) + raise AssertionError("main_loop.py exposes no subcommands") + + +def required_flags(parser: argparse.ArgumentParser) -> set[str]: + return { + option + for action in parser._actions # noqa: SLF001 + if action.required + for option in action.option_strings + if option.startswith("--") + } + + +def isolation_choices() -> set[str]: + claim = subcommand_parsers()["claim"] + for action in claim._actions: # noqa: SLF001 + if "--isolation" in action.option_strings: + return set(action.choices or ()) + raise AssertionError("claim exposes no --isolation choices") + + +def installed_skills() -> set[str]: + names: set[str] = set() + for base in (ROOT / "skills", ROOT / "skills" / "thirdparty"): + if not base.is_dir(): + continue + for path in base.iterdir(): + if path.is_dir() and (path / "SKILL.md").is_file(): + names.add(path.name) + return names + + +def rules_text() -> str: + return (TEMPLATES / "AGENT_RULES.template.md").read_text(encoding="utf-8") + + +def section(text: str, heading: str, until: str) -> str: + return text.split(heading, 1)[1].split(until, 1)[0] + + +def headings(text: str) -> list[str]: + return re.findall(r"^#{2,3} (.+)$", text, re.MULTILINE) + + +# Tokens written in backticks that name something other than a skill. Machine +# state, CLI vocabulary and harness commands are derived from the implementation +# where possible so the allowlist stays small and reviewable. +NON_SKILL_BACKTICK_TOKENS = frozenset( + { + "base", + "claude", + "clear", + "codex", + "command", + "commit", + "compact", + "main", + "pass", + "pass-with-notes", + "passed", + "output", + "result", + "released", + "report", + "spec", + "standards", + } +) + LEGACY_FLOW_TERMS = ( "using-superpowers", "brainstorming", @@ -103,14 +191,17 @@ class TemplateContractsTests(unittest.TestCase): def test_templates_readme_separates_ownership_from_deployment_config(self): templates_readme = (TEMPLATES / "README.md").read_text(encoding="utf-8") - classification = templates_readme.split("## 模板分类", 1)[1].split( - "## 模板说明", 1 - )[0] + classification = section(templates_readme, "## 模板分类", "## 模板说明") + categories = [h for h in headings(classification) if h[:2] in ("1.", "2.", "3.")] - self.assertIn("### 1. 入口导航", classification) - self.assertIn("### 2. 初始化后由项目维护", classification) - self.assertIn("### 3. 参考模板", classification) + self.assertEqual(len(categories), 3, msg=f"unexpected categories: {categories}") self.assertIn("项目新增的 `memory-bank/*`", classification) + self.assertIn( + "playbook:rules:start/end", + classification, + msg="ownership of AGENT_RULES.md is split by the marker block; the " + "classification has to say which side the project owns", + ) self.assertNotIn("[sync_", classification) self.assertNotIn("force", classification) self.assertNotIn("no_backup", classification) @@ -151,61 +242,85 @@ class TemplateContractsTests(unittest.TestCase): self.assertFalse(TEMPLATES.joinpath("prompts").exists()) def test_agent_rules_template_defines_ticket_and_integration_contracts(self): - rules = (TEMPLATES / "AGENT_RULES.template.md").read_text(encoding="utf-8") - self.assertIn("Matt Pocock 工程流程与 ticket 执行约束", rules) + rules = rules_text() + normalized = " ".join(rules.split()) + self.assertIn("{{PLAYBOOK_ROOT}}", rules) - self.assertIn("Playbook 模板/供应商目录", rules) - self.assertIn("setup-matt-pocock-skills", rules) - self.assertIn("grill-with-docs", rules) - self.assertIn("to-spec", rules) - self.assertIn("to-tickets", rules) - self.assertIn("一次可以先生成多个 feature", rules) - self.assertIn("`.scratch/queue.md`", rules) - self.assertIn("多个 frontier tickets 可在 worktree 模式并发执行", rules) - self.assertIn("不创建额外 worktree", rules) - self.assertIn("共享同一文件系统", rules) - self.assertIn("跨机器或独立 clone", " ".join(rules.split())) - self.assertIn("只有 `reclaim` 可以接管", rules) - self.assertIn("先提交实现", rules) - self.assertIn("Standards/Spec 双轴审查", rules) - self.assertIn("三个独立门禁，不能互相替代", rules) + self.assertIn("{{PLAYBOOK_SCRIPTS}}", rules) + for heading in ( + "## 任务入口", + "## 正式工程主链", + "## On-ramps 与 detours", + "## Phase boundaries", + "## 本地 Ticket 执行协议", + "## 文档职责", + "## 调度语义", + "## 执行隔离", + "## 主循环命令", + "## Git 与证据门禁", + "## 辅助能力", + "## Session 收尾", + ): + self.assertIn(heading, rules, msg=f"missing section: {heading}") + + for invariant in ( + "`.scratch/queue.md`", + "多个 frontier tickets 可在 worktree 模式并发执行", + "只有 `reclaim` 可以接管", + "Standards/Spec 双轴审查", + "三个独立门禁，不能互相替代", + "跨机器或独立 clone", + "共享同一文件系统", + ): + self.assertIn(invariant, normalized, msg=f"missing invariant: {invariant}") + for legacy in LEGACY_FLOW_TERMS: self.assertNotIn(legacy, rules) - def test_agent_rules_routes_each_matt_on_ramp_to_its_real_destination(self): - rules = (TEMPLATES / "AGENT_RULES.template.md").read_text(encoding="utf-8") - on_ramps = rules.split("## On-ramps 与 detours", 1)[1].split( - "## Phase boundaries", 1 - )[0] + def test_agent_rules_states_whether_scratch_is_version_controlled(self): + rules = rules_text() + responsibilities = section(rules, "## 文档职责", "## 稳定知识维护") + + self.assertIn("AGENT_RULES.local.md", responsibilities) + self.assertIn( + "纳入版本控制", + responsibilities, + msg="the review contract reads spec sources out of .scratch/, so " + "whether it is tracked has to be an explicit project decision", + ) + + def test_agent_rules_routes_each_current_matt_on_ramp_to_its_destination(self): + rules = rules_text() + on_ramps = section(rules, "## On-ramps 与 detours", "## Phase boundaries") self.assertIn("`wayfinder`", on_ramps) self.assertIn("`to-spec -> to-tickets`", on_ramps) self.assertIn("`research`", on_ramps) self.assertIn("先进入 `grill-with-docs`", on_ramps) - self.assertIn("`handoff` 到独立目录运行", on_ramps) - self.assertIn("`prototype`", on_ramps) - self.assertIn("再 `handoff` 结论回原设计会话", on_ramps) + self.assertNotIn("`prototype`", on_ramps) - bug_route = rules.split("### 已明确预期行为的 bug", 1)[1].split( - "## 正式工程主链", 1 - )[0] + bug_route = section(rules, "### 入口 3", "### 入口 4") self.assertIn("`diagnosing-bugs` 完整执行 Phase 1-6", bug_route) - self.assertIn("不再进入 `implement` 或重复 `tdd`", bug_route) self.assertNotIn("to-spec", bug_route) + self.assertIn( + "improve-codebase-architecture", + bug_route, + msg="diagnosing-bugs hands off to improve-codebase-architecture after " + "the fix lands, not to a design session before it", + ) + self.assertNotIn( + "grill-with-docs", + bug_route, + msg="stopping a half-fixed defect to run a design session contradicts " + "the skill's own phase order", + ) def test_agent_rules_defines_a_local_ticket_execution_adapter(self): - rules = (TEMPLATES / "AGENT_RULES.template.md").read_text(encoding="utf-8") - main_flow = rules.split("## 正式工程主链", 1)[1].split( - "## On-ramps 与 detours", 1 - )[0] - adapter = rules.split("## 本地 Ticket 执行协议", 1)[1].split( - "## 文档职责", 1 - )[0] - normalized_adapter = " ".join(adapter.split()) + rules = rules_text() + main_flow = section(rules, "## 正式工程主链", "## On-ramps 与 detours") + adapter = section(rules, "## 本地 Ticket 执行协议", "## 文档职责") self.assertIn("本地 ticket 执行协议", main_flow) - self.assertNotIn("implement + tdd", main_flow) - self.assertIn("不直接调用上游 `implement`", normalized_adapter) ordered_steps = ( "读取已领取 ticket 的 spec", "按 `tdd`", @@ -216,12 +331,46 @@ class TemplateContractsTests(unittest.TestCase): positions = [adapter.index(step) for step in ordered_steps] self.assertEqual(positions, sorted(positions)) + def test_agent_rules_binds_state_and_evidence_to_claimed_git_context(self): + rules = rules_text() + claim = section(rules, "### 领取", "### 心跳和接管") + commands = " ".join(section(rules, "## 主循环命令", "## Git 与证据门禁").split()) + + # claim's contract is its output keys and what each one addresses. + for key in ( + "FEATURE", + "TICKET", + "CONTROL_ROOT", + "STATE_ROOT", + "WORKSPACE", + "BRANCH", + "BASE", + "ISOLATION", + ): + self.assertIn(f"`{key}`", claim, msg=f"claim output key undocumented: {key}") + self.assertNotIn( + "--state-root .scratch", + claim.split("stdout 返回", 1)[1], + msg="after a claim, state must be addressed by the absolute STATE_ROOT", + ) + + # Evidence has to be bound to the claimed commits, not to free text. + for binding in ( + "--review-base <同一feature HEAD>", + "--verified \"\"", + "--reviewed \"\"", + "--verified \"\"", + "--main-verified \"\"", + "--reviewed \"\"", + "把该 `FEATURE_HEAD` 合入 ticket branch", + ): + self.assertIn(binding, commands, msg=f"missing evidence binding: {binding}") + def test_agent_rules_defines_mechanical_review_inputs_and_pass_mapping(self): - rules = (TEMPLATES / "AGENT_RULES.template.md").read_text(encoding="utf-8") - review_contract = rules.split("### Review 适配契约", 1)[1].split( - "### Ticket 完成或状态转换", 1 - )[0] - normalized_review_contract = " ".join(review_contract.split()) + rules = rules_text() + review = " ".join( + section(rules, "### Review 适配契约", "### Ticket 完成或状态转换").split() + ) for required in ( "fixed point", @@ -230,15 +379,19 @@ class TemplateContractsTests(unittest.TestCase): "零个未解决的硬 finding", "不得据此填写 `standards=pass` 或 `spec=pass`", ): - self.assertIn(required, normalized_review_contract) + self.assertIn(required, review) + self.assertIn( + "不给 pass/fail 判定", + review, + msg="code-review emits findings only; the pass mapping is this " + "adapter's own layer and must not be presented as the skill's verdict", + ) def test_agent_rules_reads_claimed_context_after_claim_and_defines_lease_policy(self): - rules = (TEMPLATES / "AGENT_RULES.template.md").read_text(encoding="utf-8") - startup = rules.split("## 会话启动", 1)[1].split("## 任务入口", 1)[0] - claim = rules.split("### 领取", 1)[1].split("### 心跳和接管", 1)[0] - lease = rules.split("### 心跳和接管", 1)[1].split( - "### Review 适配契约", 1 - )[0] + rules = rules_text() + startup = section(rules, "## 会话启动", "## 任务入口") + claim = section(rules, "### 领取", "### 心跳和接管") + lease = section(rules, "### 心跳和接管", "### Review 适配契约") self.assertNotIn("当前 `.scratch//spec.md`", startup) self.assertNotIn("当前 `.scratch//issues/.md`", startup) @@ -247,18 +400,75 @@ class TemplateContractsTests(unittest.TestCase): self.assertIn("每 10 分钟", lease) self.assertIn("固定为 30 分钟", lease) - def test_agent_rules_preserves_to_spec_seam_and_phase_boundary_contracts(self): - rules = (TEMPLATES / "AGENT_RULES.template.md").read_text(encoding="utf-8") - main_flow = rules.split("## 正式工程主链", 1)[1].split( - "## On-ramps 与 detours", 1 - )[0] - self.assertIn("不重新进行已经完成的需求采访", main_flow) - self.assertIn("完成 seam confirmation", main_flow) - self.assertIn("`tdd` 不得在未经确认的 seam 上开始", main_flow) + def test_agent_rules_orders_task_entries_by_cost_with_upgrade_conditions(self): + rules = rules_text() + entries = section(rules, "## 任务入口", "## 正式工程主链") + entry_headings = [h for h in headings(entries) if h.startswith("入口 ")] - phase_boundaries = rules.split("## Phase boundaries", 1)[1].split( - "## 文档职责", 1 - )[0] + self.assertEqual( + entry_headings, + [ + "入口 1：直接执行", + "入口 2：单切片改动", + "入口 3：已明确预期行为的 bug", + "入口 4：新 feature 或设计变更", + ], + msg="entries must stay ordered cheapest-first so the router can take " + "the first match", + ) + self.assertEqual( + entries.count("**升级条件**"), + 2, + msg="entry 1 and entry 2 each need an explicit upgrade trigger; " + "without one the router has no defined way out of a light path", + ) + + entry_two = section(entries, "### 入口 2", "### 入口 3") + entry_four = section(entries, "### 入口 4", "### 非交互模式下的入口 4") + self.assertIn( + "不属于入口 3", + entry_two, + msg="known bugs must reach diagnosing-bugs before the generic slice path", + ) + self.assertIn("入口 4", entry_two, msg="entry 2 must name its escalation target") + self.assertIn( + "边界不清时先按入口 2 起步", + entry_four, + msg="an uncertain boundary must start at the single-slice path, not " + "pre-pay the full chain", + ) + + def test_agent_rules_gives_unattended_sessions_a_pre_ticket_fallback(self): + rules = rules_text() + fallback = section( + rules, "### 非交互模式下的入口 4", "## 正式工程主链" + ) + + self.assertIn("--result blocked", fallback) + self.assertIn( + "to-questionnaire", + fallback, + msg="grilling needs a user, so a ticketless unattended session must " + "have a defined way to hand questions back", + ) + self.assertIn(".scratch/questions/.md", fallback) + self.assertIn("从 `grill-with-docs` 恢复", " ".join(fallback.split())) + + def test_agent_rules_keeps_seam_confirmation_with_to_spec_and_tdd(self): + rules = rules_text() + main_flow = section(rules, "## 正式工程主链", "## On-ramps 与 detours") + + self.assertIn("`tdd` 不得在未经确认的 seam 上开始", main_flow) + self.assertIn( + "seam confirmation 的责任在 `to-spec` 与 `tdd`", + main_flow, + msg="the grilling skills never mention seams, so the rules must not " + "route seam confirmation through them", + ) + + def test_agent_rules_orders_the_phase_boundary_options(self): + rules = rules_text() + phase_boundaries = section(rules, "## Phase boundaries", "## 本地 Ticket") ordered_options = ( "继续当前 session", "使用 `clear`", @@ -267,43 +477,120 @@ class TemplateContractsTests(unittest.TestCase): "使用 `compact`", ) positions = [phase_boundaries.index(option) for option in ordered_options] + self.assertEqual(positions, sorted(positions)) - self.assertIn("上下文过长不等于必须 `handoff`", phase_boundaries) + self.assertIn("150k", phase_boundaries) + self.assertIn("`handoff` 解决的是可移植性", phase_boundaries) - def test_agent_rules_binds_state_and_evidence_to_claimed_git_context(self): - rules = (TEMPLATES / "AGENT_RULES.template.md").read_text(encoding="utf-8") - execution = rules.split("### 领取", 1)[1] - after_claim_output = execution.split("stdout 返回", 1)[1] + def test_agent_rules_documents_every_main_loop_subcommand_and_required_flag(self): + rules = rules_text() + parsers = subcommand_parsers() - self.assertIn("绝对 `STATE_ROOT`", execution) - self.assertIn("`CONTROL_ROOT` 是共享 Git control checkout", execution) - self.assertIn("`WORKSPACE` 是当前 ticket 的代码工作区", execution) - self.assertNotIn("--state-root .scratch", after_claim_output) - self.assertIn("--review-base <同一feature HEAD>", execution) - self.assertIn( - 'commit=; base=; standards=pass; spec=pass', - execution, + for command, parser in parsers.items(): + self.assertIn( + f"main_loop.py {command}", + rules, + msg=f"undocumented subcommand: {command}", + ) + for flag in required_flags(parser): + self.assertIn( + flag, rules, msg=f"undocumented required flag: {command} {flag}" + ) + + documented = set(re.findall(r"main_loop\.py ([a-z][a-z-]*)", rules)) + self.assertEqual( + documented - set(parsers), + set(), + msg="the rules document subcommands the CLI does not expose", ) - self.assertIn("把该 `FEATURE_HEAD` 合入 ticket branch", execution) - self.assertIn( - "`--feature-head`、`--review-base` 和 review evidence 的 `base`", - execution, + + def test_agent_rules_documents_executable_state_transition_commands(self): + rules = rules_text() + lease = section(rules, "### 心跳和接管", "### Review 适配契约") + transitions = section( + rules, "### Ticket 完成或状态转换", "### Feature 顺序集成" ) - self.assertIn( - "base=<最新main HEAD>; standards=pass; spec=pass", - execution, + + self.assertNotIn("finish --result released|blocked", lease) + self.assertIn("--result blocked --reason", transitions) + self.assertIn("--result released", transitions) + self.assertIn("其他 `finish` 转换共用", transitions) + release_ticket = transitions.split("main_loop.py release-ticket", 1)[1] + self.assertNotIn("--repo-root", release_ticket) + self.assertNotIn("--owner", release_ticket) + + def test_agent_rules_requires_snapshotted_evidence_artifacts(self): + rules = rules_text() + gate = " ".join(section(rules, "## Git 与证据门禁", "## 辅助能力").split()) + + for required in ( + "UTF-8 JSON artifact", + "output_sha256", + "report_sha256", + ".scratch//evidence/", + "无法证明命令真的执行过", + ): + self.assertIn(required, gate) + + def test_agent_rules_only_references_installed_skills(self): + rules = rules_text() + skills = installed_skills() + machine_vocabulary = ( + set(subcommand_parsers()) + | set(MAIN_LOOP.ALLOWED_STATUSES) + | isolation_choices() + | NON_SKILL_BACKTICK_TOKENS ) + referenced = { + token + for token in re.findall(r"`([a-z][a-z0-9-]+)`", rules) + if token not in machine_vocabulary + } + + self.assertTrue(referenced, msg="expected the rules to reference skills") + self.assertEqual( + sorted(referenced - skills), + [], + msg="the rules reference skills that are not installed under skills/", + ) + + def test_agent_rules_states_what_the_evidence_gate_cannot_check(self): + rules = rules_text() + gate = section(rules, "## Git 与证据门禁", "## 辅助能力") + + self.assertIn("三个独立门禁，不能互相替代", gate) + self.assertIn( + "无法", + gate, + msg="the gate binds evidence to real commits but cannot prove a test " + "run happened; the rules must say so instead of implying enforcement", + ) + self.assertIn("--main-verified", gate) + + def test_agent_rules_documents_the_recovery_path_for_stuck_tickets(self): + rules = rules_text() + recovery = section(rules, "### 卡死与恢复", "## 执行隔离") + + self.assertIn("release-ticket", recovery) + self.assertIn( + "`reclaim` 只接管 `claimed`", + recovery, + msg="reclaim cannot rescue a blocked ticket; the rules must point at " + "the command that can", + ) + self.assertIn("BASE", recovery) def test_agent_rules_limits_main_loop_to_shared_local_markdown_state(self): rules = (TEMPLATES / "AGENT_RULES.template.md").read_text(encoding="utf-8") isolation = rules.split("## 执行隔离", 1)[1].split( "## 主循环命令", 1 )[0] + normalized = " ".join(isolation.split()) - self.assertIn("`main_loop.py` 只支持 local Markdown tracker", isolation) - self.assertIn("没有远程 tracker adapter", isolation) - self.assertIn("跨机器或独立\nclone 的并发不受支持", isolation) - self.assertNotIn("必须改用具备远程", isolation) + self.assertIn("`main_loop.py` 只支持 local Markdown tracker", normalized) + self.assertIn("没有 远程 tracker adapter", normalized) + self.assertIn("跨机器或独立 clone 的并发不受支持", normalized) + self.assertNotIn("必须改用具备远程", normalized) if __name__ == "__main__":