Borrowing it
Nothing to install: this file belongs to dainsiahtill-dev/Polaris. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/dainsiahtill-dev/Polaris/main/CLAUDE.mdgit clone --depth 1 https://github.com/dainsiahtill-dev/PolarisWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/instructions/dainsiahtill-dev/polaris/claude-md)<a href="https://agentmods.dev/instructions/dainsiahtill-dev/polaris/claude-md"><img src="https://agentmods.dev/badge/instructions/dainsiahtill-dev/polaris/claude-md/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/instructions/dainsiahtill-dev/polaris/claude-md"><img src="https://agentmods.dev/badge/instructions/dainsiahtill-dev/polaris/claude-md.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.17519 | $0.17519 |
| Opus 5 | $0.08760 | $0.08760 |
| Sonnet 5 | $0.03504 | $0.03504 |
| Haiku 4.5 | $0.01752 | $0.01752 |
Grade A, and why
Polaris CLAUDE.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 10d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 724 lines — stays where its author put it; the contents beside it link to each section on GitHub.
CLAUDE.md
本文件用于指导在本仓库工作的 AI 编码代理。仅保留可执行、可验证的技术约束。
必用MCP和Skill: 充分利用codegraph MCP和superpowers,必要时需要使用Playwright来真实跑测试和审计。
-1) 强制工具链:RTK + CodeGraph + Caveman
- RTK:每条 shell 命令及每个 chained segment 都必须以
rtk开头。优先使用 RTK 原生命令;无对应命令时使用rtk proxy <command>。裸 shell 命令只允许用于明确的 RTK 故障诊断。 - CodeGraph:源码发现、架构/调用链/影响分析、代码审查及修改前上下文必须先调用
mcp__codegraph__codegraph_explore。返回源码视为已读,禁止再用rg/reader 重复读取;只有 CodeGraph 未覆盖、未索引或明确不可用时,才允许使用 RTK fallback,并记录原因。 - Caveman:默认启用
caveman的full模式,压缩过程更新与结果,删除重复叙述和无关日志,但必须保留精确技术名、错误文本、命令、验证数字、风险与未完成门禁。安全警告、不可逆操作和可能产生歧义的多步流程恢复完整表达。 - 三者必须共同使用,不能用其中一个替代另两个。标准顺序:加载适用 Skill -> CodeGraph 定位/审计 -> RTK 执行与验证 -> Caveman 压缩汇报。
0) 后端权威入口(2026-03-22)
- 对于任何
src/backend任务,必须先读src/backend/AGENTS.md。 - 统一架构执行标准入口:
src/backend/docs/AGENT_ARCHITECTURE_STANDARD.md。 - 后端强制规则:
Cell开发先复用已有 Cell 公开能力;所有新开发必须基于KernelOne底座能力与契约链路。 - 若本文件与
src/backend/AGENTS.md或src/backend/docs/AGENT_ARCHITECTURE_STANDARD.md存在冲突,以后两者为准。
0.0) 无人值守项目完成权威(强制)
- 项目完成身份必须精确绑定
(workspace, project_id, run_id, completion_contract_hash);禁止按“最新 run”、路径猜测、跨 workspace 搜索或 caller 自报字段拼接完成事实。 - Chief Engineer 只拥有整个 PM task-set 的 typed completion contract;合同必须绑定 owner task、required artifact、verifier modality、canonical argv/cwd、entrypoint(或显式 N/A)和合同 hash。Director 只能无损传递该合同及 hash,不能缩减、改写或自行声明项目完成。
- PM 声明的
project_declared_entrypoint_targets必须按 exact owner task 与其target_files交集投影到 CE authority carrier;不得依赖 CE 再次把同一路径标成semantic_role=entrypoint才保留入口。CE 可改写入口描述或建议命令,但只允许归一化到同 owner 唯一 PM entrypoint command;全局入口列表不得复制成其他 task 的 authority,缺失、跨 owner、歧义或越 scope 仍 fail-closed。 factory.verification_guard只消费 owner contract 并执行物理验证;artifact hash、依赖准备、build/test/lint、entrypoint 都必须形成按 obligation/owner/hash 绑定的 typed receipt。禁止信任 caller-supplied evidence、generic mapping receipt、TaskBoundary、stage gate、日志摘要或“磁盘看起来存在”;audit.evidence只能保存审计副本,不能制造执行权威。orchestration.workflow_runtime只拥有 durable convergence cursor:残差、依赖、attempt budget、action reservation、settlement 和终态;跨进程更新必须 CAS,effect 前必须先 reserve,崩溃重放必须复用 deterministic action id。禁止在 Factory、adapter、bench 或 Agent 中另藏重试循环/第二事实源。runtime.projection是唯一可签发completed_verified的 owner。只有 exact completion contract 与全部 owner-sealed physical receipts 一致、无 missing/failed obligation 时才能完成;VerificationGuard、Director、QA、workflow runtime 和外部 Supervisor 都不得签发最终成功。- 角色失败必须阶段局部恢复,禁止默认重跑整条
PM -> Chief Engineer -> Director:CE 输出/schema 失败只重试chief_engineer_review,复用已提交 PM contract;Director/QA 普通失败必须由workflow_orchestrationdurable cursor 生成 action/claim,携带完整 owner-sealed diagnostic,并通过runtime.task_runtime重开 exact numeric owner row,再执行 boundededit/repair -> affected verifier。TaskMarket 迁移未完成时不得作为同任务完成动作的第二事实源。PM 仅在 PM contract 本身无效或被显式 supersede 时重跑;局部预算耗尽输出model_ceiling/结构化 blocker,不自动升级上游。 retry_phase=planning必须优先解析到同 phase 内的last_failed_stage/current_stage;若失败点是chief_engineer_review,绝不能因 PM 与 CE 共用 planning phase 而固定选择pm_planning。只有缺少 stage-local failure evidence 时才允许使用 phase 默认阶段。mutation_bypass_blocked是 owner Director 的普通局部恢复态:只有Director + materialize_changes可保留同一 session、已提交合同、最终请求上下文、读取结果和失败 receipt,下一 turn 直接 edit/write 后只重跑受影响 verifier;禁止创建新 Director 规划会话或重启 PM/CE。该恢复必须有独立于自适应 extra-turn 的硬预算(当前最多 2 次 bypass);耗尽后输出director_quality_repair_stalled、model_ceiling、retry_scope=same_director_task_only并停止 Provider 调用。inline_patch_escape_blocked等权限/策略边界仍 fail-closed。- Verifier 失败归因必须先从生成/编译产物反投影到 authored source,再执行 task write-scope:例如 Node traceback 的
tsconfig.compilerOptions.outDir/*.js只有在显式rootDir/outDir可唯一解析且源文件真实存在时,才能归因到对应.ts/.tsx/.mts/.cts;不得授权 Director 编辑dist/等生成物,配置缺失/歧义时保持 fail-closed。语言分类必须有真实命令/异常锚点,禁止因临时路径包含pytest等子串把 Node/Go/Rust 失败误投影为 Python。归因成功后只回到 exact owner Director task,保留合同/上下文/失败 receipt,并只重跑 exact failed verifier。 - ContextOS control-plane 隔离审计必须区分弱自然语言名词与强 authority 标识:QA workspace-quality data-plane 中的通用
workspace路径字段和 verifiermetrics文案允许出现在 prompt content,但作为 message metadata / structured projection key 仍必须拒绝;workspace_root、factory_run_id、job_token、capability token 等强标识无论位于何种非当前用户指令内容都必须 fail-closed。禁止 blanket 跳过[UNTRUSTED_USER_MESSAGE];taxonomy 变更必须以原始 final provider request 快照重放并证明合法 QA 上下文 PASS、真实 authority 泄漏仍 FAIL。 - Provider structured-output 出现对象/数组被额外 JSON 字符串化、或剩余 root members 被串入首个 object field 时,必须先由
roles.kernel统一传输层执行 caller-schema-proven 归一化,禁止立刻消耗整角色重试。只允许有界 JSON container 解码与单层已知无效 escape 归一化;重建结果必须仅含 schema 声明字段、保持已结构化 sibling 完全不变并通过完整 JSON Schema,否则 fail-closed。每次适配必须写schema_normalization_applied/schema_normalization_policy,不得把语义补全伪装成传输归一化。 - Director 局部 repair 的“进展”必须同时满足:authoritative write receipt 对应负责路径的真实 fingerprint 变化、required verifier 复跑、诊断或 missing target 净减少且不引入新诊断。纯读取、同内容写入、只改变诊断签名、等量换错或错误增多一律记为 stagnation;连续 2 次 stagnation 必须停止 Provider 调用,输出
director_quality_repair_stalled+model_ceiling,并保持retry_scope=same_director_task_only、禁止回退 PM/CE。 - Director 局部 repair 的 attempt/coverage/log evidence 绝不能冒充 mutation evidence。确定性 repair 若只有 source_tool/coverage/失败或 no-op receipt、没有 authoritative write fingerprint,必须继续进入同一 owner Director 的 LLM edit;Node TAP
not ok等已有文件断言失败的最终请求必须携带单一 failure island、当前 UTF-8 目标内容,并强制产生edit_fileeffect,禁止用read_file/execute_command/自然语言响应满足修复轮。长 repair 必须续租原 TaskRuntime execution attempt,停止 heartbeat 后才 settlement;否则deo_execution_attempt_heartbeat_failed/session_lease_expired会把真实修复伪装成模型失败。 - Repair mutation receipt 必须同时具备:成功状态、mutation tool/operation、负责路径、before/after SHA-256(允许 create 的 before=
file_absent),且 fingerprint 必须变化;缺路径、缺 hash、同 hash 或 heartbeat 曾失败均不得结算 completed。模块拆分后判断 scanner/tool 是否为平台默认实现,必须使用稳定 capability/包身份,禁止硬编码单一__module__精确字符串;否则_scan等内部迁移会静默降级到无 typed issues 的旧路径。 - M10 只能修复已有诊断可证明的局部结构,不得为满足 delivery depth 偷造 placeholder tests、领域 class/function 或业务语义。缺 test artifact 必须归还其 PM/CE 声明 owner;缺 export 且 workspace 无既存声明/唯一重导出/编译器 suggestion 时必须
covered_unplannable,由同一 Director task 消费结构化 discrepancy 后 edit,禁止 deterministic invent 假绿。 - JavaScript missing-export repair 只接受显式 unresolved symbol + module/path 诊断,并且只能导出 workspace 已存在声明、class method facade 或重导出已存在 binding。普通
AssertionError/strictEqual/npm test失败绝不能枚举测试 imports 后改写已导出的函数体,也不能从测试断言发明领域函数、常量或依赖;已导出 symbol 必须视为“非 export 缺陷”并保持零 mutation。DEOwrite_file/edit_file/edit_blocks的 file-edit event 必须由真实 before/after 内容计算 patch,禁止硬编码空 patch 后把真实修改投影为no_content_change。 - TaskRuntime 执行历史与 TaskBoundary 交付权威是独立轴。Factory 权威投影只能从已提交 PM stage event 的 immutable artifact binding 重验并取得 PM 合同任务,禁止从可变
tasks/plan.json/mirror 重建完成义务;PM contract IDs、TaskRuntime IDs、TaskBoundary IDs 归一化后集合必须完全一致。缺失合同任务、N/TASK-N合同 alias collision、重复 runtime identity 必须 fail-closed,settlement/verifier 辅助行与 boundary 不得新增或污染交付义务。TaskRuntime 已终态failed/cancelled时,只有 canonicalcompleted_verifiedTaskBoundary 同时具备 ledger append/content 坐标、evidence refs、零 missing/failed obligation,才可保留失败历史并授权进入 QA;pending/in_progress/blocked或仅磁盘扫描永远不得放行。 - Factory stage 失败不得仅以“可能本地修复”为由占住 workspace lease。Director/QA local rework 只有在 durable completion cursor 已提交 owner action receipt、TaskRuntime 已保存同 action/claim/full diagnostic 且 exact owner row 已安全重开时才能 defer drain;单独 TaskMarket receipt、纯 metadata、伪 action ref、辅助 task、错 run/contract/task 一律不得授权。TaskRuntime 行重开后必须由 backend lifespan-owned Factory run driver 消费;HTTP Router 只能提交到该 driver,不得自行持有不可恢复的执行 task。启动恢复必须合并
RUNNING/RECOVERINGruns 与 pending exact-row local-rework actions;仅把状态改成 pending 不能计作进展或成功。 - DEO 一旦把 member claim 为
EFFECT_STARTED,所有“物理执行前”的 policy/target drift 拒绝也必须先消费 one-use fence,再以 recovery +DEAD_LETTER形成终态证据;禁止返回 denied 后把 claimed member 留在非终态。若 fence 消费或终态追加存在歧义,必须保留 reconciliation blocker,不得 abort 相关 contingency 或伪造成功。该规则保证局部 Director repair 可失败、可重排、可重试,但不永久卡住 Factory settlement。 - 运行时观察链不得复制 Durable TaskRuntime/Provider/Tool 全量证据。事实流与持久化快照保留完整 payload;runtime.v2/NATS/状态投影只能发送有界语义摘要、不可变 fact 坐标与可解析 evidence ref,默认单事件预算 64 KiB。高频事件只能按明确 canonical kind 触发 status refresh,并采用单任务 latest-wins 合并;禁止用模糊字符串匹配、每事件/每连接同步全量重建或对子树反复 JSON 序列化,避免 observer 反向拖死执行面。
- Factory HTTP/runtime.v2 状态合同不得内嵌完整 workspace validation、repair rounds、audit bundle 或长 provider/tool 输出;单 metadata value 超过状态预算时必须投影
json_bytes + durable_evidence ref,完整证据由 audit-bundle/ContextOS/ReceiptStore 读取。QA 已有 authoritative failed verdict 时,归因必须优先保留qa_verdict_failed,不得用由该失败派生的task_runtime_not_completed遮蔽首因;这只改变归因优先级,不放宽完成门禁。 model_ceiling只能由 owner query 的最终请求快照、physical-attempt、Run Ledger settlement、provider health、repair coverage 和 residual 证据共同封存;禁止 caller 自报 attempt 数、预算或通用 JSON receipt 触发 terminal。未满足证据时必须 fail-closed 为非终态。- 外层 Supervisor 只做项目级调度(启动、读取归因、推进/暂停/告警),不得进入 Run Ledger 成功条件。Bench 仅在 unit/type/architecture gates 全绿后作为稀缺验证;失败先产出唯一 residual/module attribution,再修 owner Cell,禁止无归因反复长跑。
/tmp只允许放可丢弃 scratch/worktree。bench 审计、最终 provider request 快照、缺陷 manifest、handoff 与可复核测试证据必须落在~/.polaris/bench_runs、~/.polaris/audit_archives或仓内治理资产;不得把/tmp当唯一证据源。- 首个 fresh isolated
COMPLETED_VERIFIED经验基线见src/backend/docs/governance/UNATTENDED_COMPLETION_FIRST_PROOF_20260812.md。后续失败必须按final provider request -> provider response -> tool lifecycle -> effect receipt/settlement -> TaskBoundary/TaskRuntime -> Run Ledger/QA -> Factory report顺序动态定位唯一断边;未完成 owner 修复、focused tests 与平台 cascade 前禁止盲目重跑。单项目通过只允许推进下一个顺序项目,不能宣称 L1-L12 或 N-batch 封板。 - QA verdict 只能授权其精确审计的 Director delivery epoch:
(task_id, director_run_id)必须与该 task 最新 canonical TaskBoundary 完全一致。Run Ledger 必须保留历史 verdict,但只把当前 epoch 投影进effective_gates;Factory 只能消费effective_gates,rawgates仅允许无 effective 投影的 legacy 兼容。Director-only retry 使旧 QA verdict 失效,但不得重跑 PM/CE;随后只重跑 QA/受影响 verifier。 - KFS 逻辑证据路径不能靠字符串或旧绝对路径猜测。审计
runtime/qa/*、receipt 或 ContextOS 文件前必须通过当前 workspace 的 storage-root resolver/有界文件清单定位实际 system runtime root;workspace 物理目录中看不到逻辑文件不等于证据丢失。缺失结论必须附 resolved roots 与 searched paths。 - runtime.v2/ContextOS 的正式 Factory 事件必须以结构化
result.status/ok/ severity 判定成功或失败,禁止用消息子串作为更高权威。成功摘要会合法包含error_code=none;若先用/error/文本启发式分类,会把 PM/CE/Director 成功事件伪装成“异常闭环”。结构化成功必须压过文本启发式,结构化失败仍必须保持 error,并有正反回归测试。
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 10d ago First seen · 724 lines · 17,519 tokens per session scan A 373cdf1b75d5
Polaris CLAUDE.md is an instructions file published in the GitHub repository dainsiahtill-dev/Polaris (103 stars, last pushed 13d ago), licensed MIT. It adds 17,519 tokens to every session, about $0.0876 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other instructions, from other repositories
next.js AGENTS.md
AGENTS.md instructions for vercel/next.js, covering next.js development guide, codebase structure, monorepo overview, core package: packages/next and other important packages.
codex AGENTS.md
AGENTS.md instructions for openai/codex, covering rust/codex-rs, the codex-core crate, code review rules, crate api surface and model visible context.
vscode buildNext.instructions.md
Working notes and architecture documentation for the new esbuild-based build system in build/next. Use when making changes to the new build pipeline (transpile/bundle commands, NLS plugin, source-map handling, resource copying, or self-hosting watch tasks).
spec-kit AGENTS.md
AGENTS.md instructions for github/spec-kit, covering agents.md, about spec kit and specify, quickstart — add a new integration in 5 steps, integration architecture and integrationmanifest — file tracking.
vscode oss-third-party-notices.instructions.md
Instructions for microsoft/vscode, covering vs code oss third-party-notices pipeline, architecture, pipeline flow in ci, applying the notice (cutover) and fallback chain (never fail the build).
langchain AGENTS.md
AGENTS.md instructions for langchain-ai/langchain, covering global development guidelines for the langchain monorepo, corridor security analysis, project architecture and context, monorepo structure and development tools & commands.