repo-evidence-first

name	repo-evidence-first
description	在非平凡规划、方案、编辑、审查或完成声明前，先用仓库证据锚点确认项目事实；当本地约定、peer pattern、测试、配置或现有实现会影响方案/implementation 时必须先取证；无证据的项目判断必须标为 Hypothesis、Open Question、Unverified、委托调查或 blocked。

Repo Evidence First

这个技能用于防止 ROSE 凭经验猜项目事实。凡是要声称“这个仓库通常怎样做”“某文件是权威”“某测试覆盖行为”“某任务已经完成”“某引用是 stale/current”等，都必须先拿到仓库证据。

什么时候使用

用于非平凡的：

• 规划、设计、任务拆分
• 编辑前的目标定位
• code review / test review / security review
• completion claim：complete、fixed、passing、verified、ready
• 需要判断 active/current/stale/archived/generated 的证据
• 需要项目约定、peer pattern、上下游、测试覆盖或验证路径
• research-first planning gate 中需要本地仓库事实支撑方案：现有实现、docs、tests、configs、schemas、generated/source-of-truth、peer pattern、验证命令

Direct allowlist 小改可以跳过完整 evidence pack，但必须说明为什么满足 direct 条件且 subagent 不会节省上下文。

工作流

1. 命名当前 contract：用户目标、scope、acceptance、stop conditions。
2. 查本地规则：AGENTS.md、agents/rose.md、active OpenSpec、skills、README/docs。
3. 建立 evidence pack，不要把 unsupported claim 写成 fact。
4. 遇到 broad/noisy search 时，派发最轻量 subagent。
5. 把冲突、stale、missing、generated、archived 证据显式标出。
6. 给出下一步：edit、delegate、ask_user、blocked、verify。

可选证据源：当非平凡代码任务需要文件/符号/调用方/被调用方/peer pattern/测试/影响面定位，且 CodeGraph 可用且有帮助时，ROSE 可把 CodeGraph 作为可选 locality provider 分配给合格 lane owner。CodeGraph 结果必须压缩成 evidence anchors；不可当作 proof；不可替代 edit/review/test/doc lane 的最终读文件、读 diff、读测试或读文档；不可用时回退到普通 search/read，只有 materially affects confidence 时标 Unverified。

🔴 CHECKPOINT: 在进入编辑、审查结论或 completion claim 前，必须能列出至少 1 个有效证据锚点，或把状态降级为 PARTIAL / NOT_FOUND / CONFLICTING / BLOCKED。没有 anchors，就不能说 “confirmed / verified / done / repository pattern”。

Evidence pack 字段

REPO EVIDENCE STATUS: GROUNDED | PARTIAL | NOT_FOUND | CONFLICTING | BLOCKED
CONFIDENCE: HIGH | MED | LOW | VERY LOW | UNKNOWN

Active contract:
- ...

Local rules inspected:
- path:line - rule

Project facts:
- path:line-or-symbol - fact - freshness - confidence

Existing patterns:
- path:line-or-symbol - pattern
- N/A if none found

Counter-evidence / stale evidence:
- path:line-or-symbol - stale/missing/archived/generated/conflicting signal
- N/A if none found

Verification path:
- command / inspection / test - why

Unknowns:
- Open Question / Unverified item

Next action:
- edit | delegate | ask_user | blocked | verify

Claim classification

Also follow the global Evidence-Driven Claim Hygiene rule from AGENTS.md: every claim in output needs a provenance tag or downgrade; non-claim headings/labels are formatting, not claims. Repository facts without anchors become Hypothesis, Open Question, Unverified, delegated evidence work, or blocked items.

For internal evidence packs and ROSE/subagent handoffs, use English claim tags and CONFIDENCE: HIGH | MED | LOW | VERY LOW | UNKNOWN. For user-facing summaries, keep prose in the user's input language and localize tags/confidence labels when a mapping exists; do not remove Unverified or localized equivalents without new evidence.

• Grounded Fact: supported by file paths with lines/symbols, command output summary, test result, spec/task/protocol section, explicit user instruction, or reconciled subagent evidence anchor.
• Hypothesis: plausible but not proven by current repo evidence.
• Open Question: needs user/product/architecture decision.
• Unverified: evidence may exist but was not inspected or command could not run.
• Blocked: evidence conflicts or scope/approval is missing.

Unsupported project facts must become Hypothesis, delegated evidence work, user questions, or blocked items.

Routing

• Local code/config/tests/schemas/symbols/call chains: code-scout
• Local docs/workflow/OpenSpec/skills/rules: doc-researcher
• External official/current behavior: web-researcher
• Test coverage or verification strategy: test-engineer
• Secrets, auth, permissions, tool policy, install/hooks, trust model: security-auditor

Planning gate ownership:

• This skill owns local repository evidence only. It should supply current repo facts, existing patterns, constraints, verification paths, and unknowns for the方案.
• Official/API docs, fast-changing provider behavior, SDK/framework docs, and changelog-sensitive facts route to source-driven-development or external/current evidence lanes.
• Industry/GitHub/mature project prior art routes to mature-project-pattern-research.
• User-facing方案 discussion, no-objection/approval, waiver, and write-back readiness route to requirements-grilling or the active lifecycle workflow.
• Test-document artifacts route to test-document-generator.

When local evidence is part of a research-first planning gate, return facts in a form that can be merged with official-doc and prior-art lanes: official facts N/A, local facts, applicable/rejected local patterns, assumptions, Unverified gaps, and verification commands. Do not proceed from local evidence directly to implementation when the gate also requires user acceptance of the方案.

Use the lightest specialist that can return compact anchors. Do not paste raw grep dumps or long logs into MainAgent context.

Runtime-valid delegation only:

Need	If specialist is exposed and allowed	If not exposed/allowed
Broad local code evidence	Delegate to code-scout or equivalent read-only scout; optionally include CodeGraph locality evidence when available/useful	Search locally with available tools, or mark Unverified
Docs/OpenSpec/workflow evidence	Delegate to exposed docs scout	Read local docs yourself, or ask caller to dispatch
External/current behavior	Delegate/fetch only if web access is allowed	Mark external behavior Unverified; do not infer
Test/security review	Delegate only when caller/runtime permits	Provide escalation request; do not fabricate specialist verdict

🛑 STOP / Blacklist

• Do not claim a repo fact without path:line, symbol, command output, explicit user instruction, or reconciled anchored subagent evidence.
• Do not name unavailable agents as if they ran; say “escalation needed” or Unverified.
• Do not convert stale docs, generated files, partial logs, or memory into current fact without freshness checks.
• Do not proceed from CONFLICTING evidence to edit/ship without resolving or reporting the conflict.

输出给用户

简短说明：

• 已确认的项目事实
• 证据锚点
• 冲突或 stale 证据
• 下一步动作
• 未验证项