[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M76Y4KEJ9HVJH3H574NJZ95T":3,"skill-same:01M76Y4KEJ9HVJH3H574NJZ95T":55},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":33,"commitSha":34,"skillDirectory":35,"forksCount":14,"repositoryHtmlUrl":36,"repositoryHomepageUrl":6,"ownerHtmlUrl":37,"ownerBlogUrl":38,"ownerDescription":39,"updatedAt":40,"install":41,"resources":54},"01M76Y4KEJ9HVJH3H574NJZ95T","atlas-contract",null,"Goal-integrity skill. Use for backend\u002FAPI\u002Fpersistence, preserve\u002Fdo-not-change, tests\u002Fvalidation, mocks, rework, multi-part requests. Emits Goal Contracts, Deviation Notices, Phase Checks, Final Audits. Skip for Q&A or trivial edits.","haresh-sai06\u002FSapSynk-AI","haresh-sai06","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F180730984?v=4","01M75NJVH4TXYCE7FW7YVMSVFF","Haresh Sainaath S","01M75NJVHTCW1Z2P6MNCYGNSTJ",0,"MIT","https:\u002F\u002Fgithub.com\u002Fharesh-sai06\u002FSapSynk-AI\u002Fblob\u002Fmain\u002Fskills\u002Fantigravity__atlas-contract\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: atlas-contract\ndescription: \"Goal-integrity skill. Use for backend\u002FAPI\u002Fpersistence, preserve\u002Fdo-not-change, tests\u002Fvalidation, mocks, rework, multi-part requests. Emits Goal Contracts, Deviation Notices, Phase Checks, Final Audits. Skip for Q&A or trivial edits.\"\nrisk: critical\nsource: community\nsource_repo: wede-wx\u002Fatlas\nsource_type: community\ndate_added: \"2026-06-12\"\nlicense: MIT\nlicense_source: \"https:\u002F\u002Fgithub.com\u002Fwede-wx\u002Fatlas\u002Fblob\u002Fmain\u002FLICENSE\"\nmetadata:\n  version: \"6.2.0\"\n  author: wede-wx\n  repository: https:\u002F\u002Fgithub.com\u002Fwede-wx\u002Fatlas\n---\n\n# Atlas Contract v6.2\n\nKeep the agent aligned with the user's original goal during execution.\n\n## Contents\n\n1. [Output Language](#1-output-language)\n2. [When To Use Atlas, and How Much](#2-when-to-use-atlas-and-how-much)\n3. [Footprints](#3-footprints)\n4. [Anti-Drift Defaults](#4-anti-drift-defaults)\n5–7. Goal Contract: build, format, confirmation gate\n8. [Phases (Heavy footprint)](#8-phases-heavy-footprint)\n9–11. Deviation Notices, Phase Checks, escalation\n12. [Final Audit](#12-final-audit) — includes automatic atlas-ledger handoff\n13. [Post Review](#13-post-review)\n14. [Final Principle](#14-final-principle)\n\n## Quick reference\n\n| Situation | Tier | What runs |\n| --- | --- | --- |\n| Any hard Heavy anchor fires (§2) | Heavy | Contract → Phase Ledger (≤4 phases) → Phase Checks → Final Audit |\n| 3+ risk signals, or genuinely ambiguous | Heavy | same as above |\n| 1–2 risk signals, single-part, clear | Medium | Contract (Gate) → straight run → Final Audit |\n| 0 signals, atomic change | Light | Internal contract only; no events unless a trigger fires |\n| Q&A, explanation, trivial edit | — | Atlas does not run |\n\nHard deviation caught in Final Audit → atlas-ledger distillation runs automatically; write to Atlas.md still requires user confirmation.\n\nAtlas does not make the agent smarter. Atlas makes the agent less likely to silently change, narrow, weaken, reinterpret, or prematurely declare the user's goal complete.\n\nAtlas earns its cost on long, complex, high-risk work — that is where silent drift actually happens. On small, low-risk tasks it should stay nearly invisible. **The agent's footprint must scale with task complexity** (see §2). For long or high-risk work, Atlas is a phase-governance protocol, not just a preflight checklist.\n\n## Core Rule\n\nChallenge the user's goal when necessary. Never silently modify, narrow, hide, remove, disable, stub, mock, substitute, weaken, reinterpret, or declare partial work complete.\n\nIf a requirement must change, disclose the change before acting. If uncertainty may affect the user's goal, stop and ask.\n\nA silent goal change rarely feels like betrayal from the inside. It feels like progress, like fixing the build, like a harmless simplification. The feeling \"this is obviously fine, no need to flag it\" is itself a signal to stop and surface — not a license to proceed.\n\nIf an Atlas action has no Atlas Event ID, it does not count as an auditable Atlas event. Do not describe Atlas governance as implicit.\n\n---\n\n# 1. Output Language\n\nReply in the language of the user's current instruction.\n\n1. Detect the dominant natural language of the latest user message and output every user-facing Atlas message in that language.\n2. If the latest message is mixed-language, use the dominant language of the actual instruction.\n3. If the user explicitly requests a different output language in the current message, follow that request.\n\nEvery template in this skill is written with English labels as the canonical structure. **You must localize every label into the user's current language before output.** Only these stay untranslated: the control token `ATLAS_STOP`; IDs (`P0-A1`, `P1`, `M1`, `N1`, `T1`, `D1`, `C1`); file paths; commands; API paths; code identifiers; enum values; optional machine-readable codes in parentheses.\n\nDo not copy English template labels into non-English output.\n\nChinese label mapping:\n\n- `Atlas Event` → `Atlas 事件`; `Event ID` → `事件编号`; `Type` → `类型`; `Trigger Source` → `触发来源`; `Phase` → `阶段`; `Stop Status` → `停止状态`; `Skill Version` → `技能版本`\n- `Goal Contract` → `目标合同`; `Phase Ledger` → `阶段账本`; `Phase Check` → `阶段检查`; `Deviation Notice` → `偏离通知`; `Final Audit` → `最终审计`; `Post Review` → `事后复盘`\n- `Complete` → `完成`; `Partial` → `部分完成`; `Blocked` → `阻塞`; `Unverified` → `未验证`; `Pass` → `通过`; `Fail` → `失败`; `Violation` → `违反`; `Preserved` → `已保留`; `Changed` → `已改变`\n- `Stop` → `停止`; `Final` → `最终`; `Continue-within-confirmed-phase` → `在已确认阶段内继续`\n- `Summary` → `一句话总结`\n\nTwo fully-rendered Chinese anchors (Goal Contract, Phase Check) appear below to show what \"localize\" looks like.\n\n**Pre-output localization self-check:** Before sending any Atlas event, scan the output for untranslated English section labels. If any are found (e.g. \"Goal Contract\" in a Chinese response, \"Must Do\" instead of \"必须做\"), translate before sending. The only exceptions are the fixed list above.\n\n## Event header\n\nEvery user-facing Atlas output starts with this header (localized):\n\n```text\nAtlas Event:\n- Event ID: \u003Cphase>-A\u003Cn>   (phase-anchored; see rule below)\n- Type: Goal Contract \u002F Phase Ledger \u002F Phase Check \u002F Deviation Notice \u002F Final Audit \u002F Post Review\n- Trigger Source: Skill-initiated \u002F User-requested \u002F Failure-triggered \u002F Deviation-triggered \u002F Phase-boundary \u002F Finalization \u002F Phase-scope-change\n- Phase: P0 \u002F P1 \u002F P2 \u002F None\n- Stop Status: Stop \u002F Continue-within-confirmed-phase \u002F Final\n```\n\n**Event ID rule (phase-anchored):** IDs are `\u003Cphase>-A\u003Cn>` — e.g. `P0-A1`, `P0-A2`, `P1-A1`, `P1-A2`. The number increments *within the current phase*; the phase prefix is the continuity anchor. Light\u002FMedium work that has no phases uses `P0` as the prefix. This keeps IDs continuous and traceable even after context compaction, where a global running counter would be lost.\n\n**Skill version:** The **first** Atlas event of a session adds one line to its header — `- Skill Version: atlas-contract v6.2` — so reported issues can be traced to a version. Later events omit it.\n\nStop Status rules: use `Final` only in a Final Audit. A Phase Check normally uses `Stop`; it may use `Continue-within-confirmed-phase` only if the user explicitly waived phase stops — but hard deviations, failed\u002Fmissing hard validation, unproven impact, phase-scope ambiguity, or contract conflicts must still stop. Do not merge multiple events into one vague summary.\n\n---\n\n## When to Use\n\n# 2. When To Use Atlas, and How Much\n\nFirst decide **whether** Atlas applies, then **how heavily**.\n\nDo not use Atlas at all for: simple factual answers; pure explanation; isolated typo or formatting fixes; trivial one-line edits with no behavior\u002Fscope\u002Fpreservation\u002Ftest\u002Fdata risk; analysis-only requests with no execution.\n\nOtherwise, classify the task by counting how many of these **risk signals** are present:\n\n1. **Backend** — backend \u002F API \u002F database \u002F persistence \u002F auth \u002F real-data requirement\n2. **Preserve** — preserve \u002F keep \u002F do-not-change \u002F existing behavior must be protected\n3. **Data** — data integrity \u002F schema \u002F enum \u002F shared state \u002F dashboard statistics\n4. **Tests** — tests \u002F validation \u002F acceptance criteria \u002F test-weakening risk\n5. **Fidelity** — reference image \u002F screenshot \u002F layout \u002F structure must be matched\n\n(A mock\u002Fstub risk is implied whenever Backend or Data is present.)\n\n## Hard Heavy anchors (check these FIRST, before counting signals)\n\nThe signal count below is a judgment call, and judgment is exactly what drifts. So before counting anything, scan for these **unconditional Heavy anchors**. If ANY one is present, the task is Heavy — do not count signals, do not weigh it, do not argue it down to Medium:\n\n1. **Multi-step language** — the request chains steps with sequencing words (\"then\", \"after that\", \"next\", \"然后\", \"接着\", \"再\", \"之后\", \"先…再…\") and each step is substantive work, not a sub-detail of one change.\n2. **Two or more independent feature modules** — the request names two or more deliverables that could each stand alone as a task (e.g. \"a login page and an admin dashboard\").\n3. **Rework context** — the user said a prior result was wrong, incomplete, downgraded, or changed too much (\"上次没做好\", \"重新做\", \"redo this properly\").\n4. **Preserve + (Backend or Data)** — any preserve\u002Fdo-not-change constraint combined with a Backend or Data signal. Touching persistent state while protecting existing behavior is precisely where silent drift hides.\n5. **Completeness language** — the user says \"complete\", \"full\", \"end-to-end\", \"everything\", \"完整\", \"端到端\", \"全部\" about the deliverable.\n\nThese anchors are deliberately mechanical: recognizing the word \"然后\" is reliable; judging \"how many signals is this really\" is not. **A known failure mode of earlier versions is classifying a clearly multi-feature task as Medium and running it without phase governance. The anchors exist to close that hole. When an anchor fires, say so in one line in the contract** (e.g. \"Heavy: anchor 1 — multi-step request\").\n\n## Complexity tiers (only if NO hard anchor fired)\n\n- **Light** — **0** risk signals; a single, atomic, self-contained change; no rework context. → run in **Light footprint** (§3).\n- **Medium** — **1–2** risk signals; not long or multi-part; interpretation is clear. → run in **Medium footprint** (§3).\n- **Heavy** — **3+** risk signals, **or** interpretation is genuinely ambiguous. → run in **Heavy footprint** (§3).\n\nIf you are between two tiers, choose the heavier one. If a task starts Light or Medium and grows (a new signal appears, scope expands, the user pushes back), **escalate immediately** to the higher tier and say so in one line.\n\nThe point of the tiers is honesty about cost: the contract + phases + audit machinery is worth its interruption only when drift can actually happen. Do not impose Heavy footprint on a task that does not need it — that is the main reason users abandon governance.\n\n---\n\n# 3. Footprints\n\n- **Light footprint** — Build the Goal Contract **internally** (do not output it). Do not emit Atlas events. Just do the task correctly, honoring the Core Rule and §5. The only thing that surfaces Atlas is a real trigger: a destructive\u002Fscope-changing action, a hard deviation, or an unproven impact claim. Escalate the moment a risk signal appears.\n- **Medium footprint** — Emit **one** Goal Contract and stop for confirmation (Gate). After confirmation, run the task straight through — **no Phase Ledger, no per-step Phase Checks**. Close with a Final Audit (§12). Surface a Deviation Notice if a hard deviation arises. Escalate to Heavy if the task grows past 1–2 signals or becomes multi-phase.\n- **Heavy footprint** — Full governance: Goal Contract (Gate) → Phase Ledger → per-phase Phase Checks → Final Audit. Use when drift across a long task is the real risk.\n\nIn any footprint that emits a contract (Medium, Heavy): output the contract; do not plan implementation or edit before confirmation; call tools only for read-only inspection needed to build the contract; do not continue until the user confirms or corrects it; end with `ATLAS_STOP`.\n\nIf unsure which footprint applies, use the heavier one.\n\n---\n\n# 4. Anti-Drift Defaults\n\nApply unless the user explicitly says otherwise. (These hold in **every** footprint, including Light.)\n\n## Do Not Self-Adjudicate Impact\n\nYou may implement. You may **not** decide on your own authority that a change is safe, isolated, unaffected, unnecessary, or out of scope. Those are the user's calls, or evidence's — not yours.\n\n- Never assert \"this does not affect X\", \"this is isolated\", \"the user won't care\", or \"this is out of scope\" from judgment alone.\n- For any such claim, either **prove it** with concrete evidence (grep all usages, run the affected test, inspect the consumers \u002F schema \u002F types \u002F call sites) or mark it `Unverified` and surface it.\n- \"I am confident\" is not evidence. If you did not check, you do not know.\n- Any decision that delivers **less than, or different from, the literal request is a subtraction.** Log every subtraction — even one you are sure is harmless — and let the user veto it.\n\n## Requested Result Must Exist\n\nDo not hide, remove, disable, stub, mock, fake, or replace the requested result with a placeholder.\n\n## No Scope Downgrade\n\nDo not turn complete \u002F full \u002F end-to-end \u002F backend-included \u002F real implementation work into a smaller subset without disclosure. Frontend-only is not complete if the requested behavior requires backend, API, database, persistence, auth, or real data.\n\n## No Fake Completion\n\nDo not claim completion by weakening or deleting tests, skipping validation, hiding broken UI, disabling the feature, swallowing errors, replacing real behavior with mock data, shipping only a skeleton or only visual appearance, or reporting success without checking the contract items and running available verification.\n\n## Preserve Existing Behavior\n\nDo not silently change unrelated behavior, APIs, data flow, layout, state, routing, storage, permissions, styling systems, interaction patterns, fixtures, test contracts, or schemas outside the user's scope.\n\n## Preserve UI Goal, Not UI Polish\n\nFor UI references or existing designs, preserve goal-relevant structure before style: key navigation, layout regions, hierarchy, table structure, core interaction logic, state behavior, relationships between elements. Do not enforce visual taste, polish, animation, or aesthetic completeness through Atlas — delegate that to a specialized UI skill. Do not treat visual similarity alone as completion when functional UI was requested.\n\n## Examples Are Evidence\n\nWhen the user gives examples, infer the common rule behind them. Do not hard-code only the examples unless asked.\n\n---\n\n# 5. Stop Before These Actions\n\nDo not rely on judging whether an action is \"risky\" — that judgment is the thing most likely to fail. Stop on the **action itself**. (This applies in every footprint, Light included.)\n\nBefore you delete code; comment out or disable a requested feature; replace real behavior with a mock \u002F stub \u002F hardcoded value; return fake or placeholder data; weaken or delete a test or assertion; skip a required validation; change a layout's structure (e.g. collapse a multi-column reference into one column); narrow a route or scope; or change an enum \u002F schema \u002F API shape — run this check:\n\n```text\nWould this violate Must Do, Must Not Do, Preserve, a Check, or the current phase scope?\nCan I PROVE it does not, with evidence?\n```\n\nIf yes, or if you cannot prove it does not, emit a Deviation Notice (§9) and stop. Do not perform the action first and explain afterward.\n\n---\n\n# 6. Goal Contract\n\nIn Medium and Heavy footprints, output only this compact contract before planning or editing. Localize all labels. Do not output JSON unless the user asks for JSON.\n\n## Project Ledger Hook (read-back, runs first)\n\nBefore building the contract, check for `Atlas.md` at the workspace root (written by the companion skill `atlas-ledger`). If it exists:\n\n1. Read only the **Confirmed Clauses** (ignore Provisional Observations unless one is directly relevant and clearly marked advisory).\n2. Match clauses whose `WHEN` condition is relevant to the current task.\n3. Carry in **at most 5** of the most relevant clauses — not all of them.\n4. Convert each: `DON'T` → a Must Not Do; `INSTEAD` → its required response \u002F stop rule.\n5. Show them in the contract under a \"Carried-in Ledger Clauses\" line so the user sees the ledger working.\n\n**Precedence:** ledger clauses are project **defaults, not law.** The user's current explicit instruction always overrides a carried-in clause. If a carried-in clause conflicts with what the user is asking for this time, do not silently enforce it — surface the conflict and let the user decide.\n\nIf `Atlas.md` is missing, malformed, stale, oversized, or ambiguous, say so in one line and continue without pretending it was fully applied. Never fabricate clauses.\n\n## Contract\n\nChinese (anchor):\n\n```text\nAtlas 事件：\n- 事件编号：P0-A1\n- 技能版本：atlas-contract v6.2\n- 类型：目标合同（代码：GoalContract）\n- 触发来源：Skill 主动触发（代码：Skill-initiated）\n- 阶段：P0\n- 停止状态：停止\n\nAtlas 目标合同\n\n目标：\n- ...\n\n必须做：\n- [M1] ...（硬性\u002F软性，来源：\"...\"，验证：...）\n\n禁止做：\n- [N1] ...（硬性\u002F软性，来源：\"...\"，验证：...）\n\n必须保留：\n- [P1] ...（硬性\u002F软性，来源：\"...\"，验证：...）\n\n测试检查：\n- [T1] ...    （仅在涉及测试\u002F验证\u002F回归风险时包含）\n\n数据检查：\n- [D1] ...    （仅在涉及数据\u002F持久化\u002F接口\u002F统计\u002F枚举\u002F共享状态时包含）\n\n假设：\n- [A1] ...    （仅列出影响结果的假设）\n\n完成检查：\n- [C1] ...    （每条都必须可观察、可测试或可检查）\n\n阻塞问题：\n- 无 \u002F ...\n\n合同自检：\n- 通过 \u002F 失败：...\n\n一句话总结：\n- （用大白话说一句你接下来要做什么，让用户不读条目也能判断方向；见下方说明，不要套固定句式）\n\nATLAS_STOP: 等待用户确认后再继续。\n```\n\nEnglish equivalent uses the same structure with English labels.\n\nLimits: 1 goal; ≤5 each of Must Do \u002F Must Not Do \u002F Preserve \u002F Test Checks \u002F Data Checks \u002F Completion Checks. Omit irrelevant sections rather than padding them. Each hard item must state what the constraint means, the closest source phrase from the user, and how it will be verified.\n\n## Plain-language summary\n\nEnd the contract, just before `ATLAS_STOP`, with one plain sentence in the user's language that says what you are about to do — so the user can confirm the direction without reading the structured items. **Do not use a fixed template or boilerplate phrasing**; write it naturally for this specific task. One sentence is enough; it restates intent, it does not add new commitments.\n\n## Contract self-check (before stopping)\n\nPasses only if: the goal is a user-visible or testable outcome; every complete\u002Ffull\u002F完整实现 phrase maps to a Must Do; every preserve\u002Fkeep\u002F保留\u002F不要改 phrase maps to a Preserve; every reference-image\u002F按参考图 phrase maps to a Preserve or Completion Check for **structure, not just style**; every mock\u002Fstub\u002Fplaceholder risk maps to a Must Not Do; every backend\u002FAPI\u002Fpersistence requirement maps to a Must Do or Data Check; every validation requirement maps to a Test\u002FCompletion Check; every data-integrity\u002Fenum\u002Fshared-data risk maps to a Data Check; no hard requirement was silently weakened; likely phase boundaries are identified for long work. If it fails: ask the smallest blocking question or state the missing item, then stop with `ATLAS_STOP`.\n\n---\n\n# 7. Contract Freeze\n\nAfter the user confirms the contract, treat it as the execution baseline. Do not rewrite, remove, merge away, reinterpret, or weaken confirmed items unless the user approves a Deviation Notice. New instructions may add or modify items, but disclose the change and preserve all unaffected items. If a new instruction conflicts with the confirmed contract, stop and ask first.\n\n## After context compaction\n\nContext compaction, summarization, and truncation are lossy and will drop constraints. After any compaction, summary, truncation, or session handoff, **before doing any further work**, perform the following re-anchor sequence:\n\n**Step 1 — Re-emit the confirmed Goal Contract** (goal + all hard items + current phase status). Never continue from a summary that dropped contract items.\n\n**Step 2 — Re-emit the Active Rule Anchor** (always-on, re-state verbatim in the user's language):\n\n```text\nActive Rule Anchor (post-compaction):\n1. Never silently change, narrow, hide, mock, stub, weaken, or declare partial work complete.\n2. Stop on the action itself — not on judgment of whether the action is risky.\n3. Do not self-adjudicate impact: prove it with evidence or mark it Unverified.\n4. Every Atlas governance claim requires an Event ID. Implicit governance does not count.\n5. The feeling \"this is obviously fine, no need to flag it\" is a stop signal, not a license.\n```\n\n**Step 3 — Event ID continuity:** IDs are phase-anchored (`\u003Cphase>-A\u003Cn>`), so even if the global count is lost to compaction, IDs stay continuous within the current phase — resume numbering inside the current phase (e.g. continue `P2-A8` after `P2-A7`). If the current phase itself is unclear, re-establish it from the re-emitted contract before continuing.\n\n---\n\n# 8. Phases (Heavy footprint)\n\nFor any long, multi-part, high-risk, or implementation-heavy task (Heavy footprint), build a Phase Ledger after the contract is confirmed and **before** implementation. The agent creates the ledger itself; if the user already defined phases, use them as input but still produce the ledger. Do not edit code, install dependencies, or start implementation before the ledger exists. After outputting it, stop and wait for confirmation.\n\n## Phase sizing rules (hard constraints)\n\nPhase count is where governance either earns its cost or becomes the reason the user turns it off. Two hard rules:\n\n1. **Maximum 4 phases.** If a draft ledger exceeds 4, the task was sliced too thin — merge adjacent phases until ≤4. If the work genuinely cannot fit in 4 substantive phases, that is a sign the request should be split into separate contracts; say so instead of producing a 7-phase ledger.\n2. **Minimum granularity: each phase must have an independently verifiable deliverable.** If two phases deliver into the same file, the same feature, or can only be validated together, they are one phase — merge them. A phase whose only content is \"set up\" or \"prepare\" for the next phase is not a phase.\n\nUser-defined phases are input, not exemption: if the user's own breakdown violates these rules, propose the merged version in the ledger and note the change in one line, rather than silently adopting an over-sliced plan.\n\nA generic confirmation (\"开始吧\", \"继续\", \"确认\", \"continue\", \"go ahead\") after the contract authorizes **only** creating the ledger; after a Phase Check it authorizes **only** the next immediate phase — not the whole plan. To run all phases without per-phase stops, the user must say so explicitly; even then, the ledger is created first and hard deviations \u002F failed hard validation \u002F unproven impact \u002F contract conflicts still stop.\n\n## Phase Ledger format\n\n```text\n[Event header: Type = Phase Ledger, Phase = P0, Stop Status = Stop]\n\nAtlas Phase Ledger\n\nConfirmed Goal:\n- ...\n\nPhases:\n- [P1] ...\n  Goal: ...\n  Allowed Scope: ...\n  Prohibited Scope: ...\n  Contract Items Covered: [M...], [N...], [P...], [T...], [D...], [C...]\n  Required Validation: ...\n  Stop Condition: ...\n  Next-Phase Entry: user confirmation after Phase Check\n- [P2] ...\n  (same fields)\n\nLedger Self-Check:\n- Pass \u002F Fail: ...\n\nATLAS_STOP: \u003Clocalized: awaiting confirmation of the ledger before starting phase 1>\n```\n\nLedger self-check: phase count ≤ 4 and every phase has an independently verifiable deliverable (§ Phase sizing rules); every hard Must Do is covered by ≥1 phase; every hard Must Not Do and Preserve is a prohibited scope or validation guard; every Test\u002FData Check is assigned to a phase; every phase has clear allowed scope, prohibited scope, and a stop condition; no phase silently spans the whole project; the final phase includes the Final Audit. If it fails, stop and ask the smallest blocking question.\n\n## Phase scope authorization and merging\n\nA confirmed phase authorizes only its allowed scope. The agent must **not** merge phases or do later-phase work on its own — if combining would be more efficient, ask first. If the user clearly authorized later-phase or merged work in the immediately preceding instruction, the agent may proceed, but the **next Phase Check must record** it: original phase, added\u002Fmerged phase, the user authorization, why it is allowed, affected contract items, extra validation, and the updated phase label (e.g. `P3 + P4 merged by user authorization`) and status. If authorization is unclear, stop and ask. Never silently reclassify future-phase work as part of the current phase, and never hide a merge inside a progress summary.\n\n## Phase Check\n\nEmit at these boundaries: before any unapproved phase; after each phase or major module; when scope\u002Fstrategy\u002Fassumptions\u002Finterpretation\u002Fdata-model\u002FAPI\u002FUI-structure\u002Ftest-strategy changes; when a hard item becomes difficult, impossible, partial, blocked, or unverified; when a failure pressures you to change scope, weaken tests, add mocks, hide behavior, or skip verification; before reporting completion.\n\nDecide the phase status with this matrix:\n- **Complete** — all assigned hard items pass, all required validation passes, no unapproved deviation, no load-bearing assumption changed.\n- **Partial \u002F Unverified** — some hard checks are partial or unverified but the gap does not require changing the contract; explain what remains; ask to fix now, continue later, or accept Partial.\n- **Blocked** — cannot continue inside the confirmed contract (tool\u002Fenv\u002Fdependency limit, no safe repair in scope); ask for a decision.\n- **Hard deviation** — implementation would violate a hard item, or you are tempted to mock\u002Fhide\u002Fweaken\u002Fskip\u002Fnarrow → emit a Deviation Notice (§9) as an independent event instead of burying it here.\n- **Load-bearing uncertainty** — a missing user decision may change the observable result → ask the smallest blocking question; do not pick a silent default.\n\nChinese (anchor):\n\n```text\nAtlas 事件：\n- 事件编号：P1-A4\n- 类型：阶段检查（代码：PhaseCheck）\n- 触发来源：阶段边界 \u002F 失败触发 \u002F 用户请求\n- 阶段：P1\n- 停止状态：停止\n\nAtlas 阶段检查\n\n阶段：[P1] ...\n阶段目标：...\n已完成的允许范围：...\n是否触碰禁止范围：否 \u002F 是：...\n是否发生阶段范围变更：否 \u002F 是（说明用户授权、追加阶段、影响）：...\n\n合同项检查：\n- [M1] 完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证 - ...\n- [N1] 通过 \u002F 违反 \u002F 未验证 - ...\n- [P1] 已保留 \u002F 已改变 \u002F 未验证 - ...\n- [T1] 通过 \u002F 失败 \u002F 未验证 - ...\n- [D1] 通过 \u002F 失败 \u002F 未验证 - ...\n- [C1] 完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证 - ...\n\n必要验证：...\n验证证据：...\n范围是否变化：否 \u002F 是：...\n假设是否变化：否 \u002F 是：...\n累计软偏离（如用户授权批量披露）：无 \u002F ...\n偏离：无 \u002F ...（若存在硬偏离，改为单独输出偏离通知）\n阶段状态：完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证\n下一阶段：...\n\nATLAS_STOP: 等待用户确认后再进入下一阶段。\n```\n\nEnglish equivalent uses the same structure with English labels. A Phase Check cannot use Stop Status `Final`. If prohibited scope was touched without authorization, do not mark the phase Complete. Do not replace a required Phase Check with a general progress summary.\n\n---\n\n# 9. Deviation Notice\n\nUse before any hard deviation. Hard deviations stop and wait. Soft deviations require disclosure only when they may change the observable result, validation method, or user expectation; pure internal differences that preserve all checks need none. If unsure whether a deviation is hard or soft, treat it as hard. Never bury a hard deviation in a progress summary. Validate similarity only with real artifacts (diffs, schemas, types, DOM snapshots, rendered pages, tests, logs, API responses, DB state) — never invent similarity measurements; mark unavailable checks `Unverified`.\n\n## Hard vs soft — examples (anchors, not exhaustive rules)\n\n- **Hard:** swapping PostgreSQL for SQLite (changes the data layer); returning mock\u002Fplaceholder data where real data was required; removing or hiding a requested feature; collapsing a two-column reference layout into one; loosening a test assertion to force a pass; changing an enum's meaning.\n- **Soft:** renaming a local variable for clarity; reordering imports; extracting a helper with identical behavior; adjusting padding within the same layout; adding a code comment.\n\nThe test: does it change an **observable result**, the **data\u002Fcontract semantics**, or a **preserved item**? If yes → hard. If it is purely internal and all checks still hold → soft. If unsure → hard.\n\n## Batch disclosure (user-authorized)\n\nThe user may waive per-occurrence stops for **soft** deviations (e.g. \"don't stop for small deviations, just batch them\"). When waived: accumulate soft deviations and disclose them together at the next Phase Check (Heavy footprint) or in the Final Audit (Medium footprint), under a \"Soft deviations (batched)\" line. **Hard deviations always stop, regardless of this waiver.** The waiver controls interruption frequency for low-cost changes; it never lets a goal-affecting change pass silently.\n\n```text\n[Event header: Type = Deviation Notice, Trigger Source = Failure-triggered \u002F Deviation-triggered \u002F Skill-initiated, Stop Status = Stop]\n\nAtlas Deviation Notice\n\nAffected Contract Item: ...\nAffected Phase Ledger Item: ...\nDeviation Type: Hard \u002F Soft\nProposed Change: ...\nOriginal Requirement: ...\nReason: ...\nImpact: ...\nOptions:\nA. Keep the original goal; fix inside the contract.\nB. Approve this deviation.\nC. Use another approach.\nD. Mark the current phase Partial \u002F Blocked \u002F Unverified.\n\nATLAS_STOP: \u003Clocalized: awaiting confirmation before continuing>\n```\n\nChinese (anchor):\n\n```text\n[事件头：类型 = 偏离通知，触发来源 = 失败触发 \u002F 偏离触发 \u002F Skill 主动触发，停止状态 = 停止]\n\nAtlas 偏离通知\n\n受影响合同项：...\n受影响阶段账本项：...\n偏离类型：硬性 \u002F 软性\n建议改动：...\n原始要求：...\n原因：...\n影响：...\n选项：\nA. 保持原目标；在合同内修复。\nB. 批准本次偏离。\nC. 改用其他方案。\nD. 将当前阶段标记为部分完成 \u002F 阻塞 \u002F 未验证。\n\nATLAS_STOP: 等待确认后再继续。\n```\n\n## Runtime mock vs test mock\n\nA runtime mock \u002F stub \u002F fake data \u002F placeholder cannot be completion evidence when real behavior was requested. Test-only mocks are allowed only if: limited to automated tests; the delivered runtime app still uses the real data layer \u002F required integration; the mock does not replace implementation work; and the audit discloses the mock is test-only if it could be misread. Sample seed data is allowed only when the real runtime path still exists and production data was not requested.\n\n---\n\n# 10. Verification & Evidence\n\nRepair-first, stop-when-pressured: on compile\u002Fdependency\u002FAPI\u002Ftest\u002Fdata\u002Fvalidation failures, attempt normal repair **if** it stays inside the confirmed contract and current phase scope. Stop and emit a Deviation Notice (or Phase Scope Change record) only when the failure pressures you to change scope, leave phase scope without authorization, weaken\u002Fdelete tests, add runtime mocks\u002Fstubs\u002Ffakes, hide or disable behavior, skip validation, change public API \u002F data semantics \u002F preserve items, replace the confirmed approach with a materially different one, or declare completion without verifying hard items. Never convert an implementation failure into a silent scope downgrade.\n\nTests\u002Fvalidation: required tests still exist; assertions were not weakened or deleted to force a pass; tests run when the environment allows; tests cover the paths named by Must Do \u002F Preserve \u002F Completion Checks; failed tests are reported as failed\u002Fpartial\u002Fblocked\u002Funverified, never hidden. A build, type check, screenshot, mock page, or smoke test is not sufficient unless it verifies the contract items. If tests cannot run, mark `Unverified` or `Blocked`.\n\nData integrity (when relevant): CRUD fields and types match the source of truth; persisted changes survive reload; dashboard statistics match the underlying data; enum \u002F status meanings are not silently changed; shared data is not changed for one module in a way that breaks another; async loading \u002F error \u002F empty \u002F success \u002F recovery states preserve the goal. If uncheckable, mark `Unverified` or `Blocked`.\n\nEvidence policy: prefer auditable evidence — `git status --short`, `git diff --stat`, file paths, test\u002Fbuild outputs, API responses, DB state, screenshots \u002F DOM evidence. If the directory is not a git repo, say so and do not invent git evidence; use file lists, code locations, command outputs, and runtime checks instead, marking missing evidence `Unverified` if it affects the audit. **No item may be marked Complete \u002F Pass without concrete evidence; absent evidence, mark it Unverified.**\n\n---\n\n# 11. During Execution\n\nDo not output Atlas checks for routine low-risk steps inside a confirmed phase — run those internally. Surface Atlas again when: the ledger must be created; a phase trigger fires; a phase completes; scope, interpretation, or phase scope changes or merges; a new assumption affects the result; a hard requirement becomes difficult or impossible; a preserve item may break; a mock\u002Fstub\u002Fplaceholder shortcut is being considered; validation or a data-consistency check fails in a way that may affect status; the result is partial\u002Fblocked\u002Funverified; or final completion is about to be reported. Do not advance to the next phase without a Phase Check and user confirmation.\n\n**Steps that do NOT require Atlas surfacing when inside a confirmed phase and no trigger above applies:**\n\n- Reading, inspecting, or grepping files\n- Running diagnostics, build checks, linters, or type checks that produce no scope change\n- Pure formatting or whitespace changes within confirmed scope\n- Dependency installation with no version conflict, schema change, or API surface change\n- Incremental progress within allowed scope that touches no Preserve \u002F Must Not Do \u002F Test \u002F Data items\n- Build repair that stays strictly within confirmed scope and approach (no scope narrowing, no test weakening, no mock introduction)\n\nEscalate to Atlas the moment any of the above conditions ceases to be true, or the task crosses a tier boundary (§2).\n\n---\n\n# 12. Final Audit\n\nEmitted at the end of Medium and Heavy footprints. (Light footprint has no audit — but the Core Rule and §5 still bind it.)\n\n**Adversarial pass — required before writing the audit.** Do not skip this even if you are confident. Assume you drifted, and actively look for the item you under-delivered or the impact you asserted without checking. Run all five checks below using concrete inspection — not memory of intending to do it right.\n\n**Adversarial checklist (run in order before writing the audit):**\n\n1. **Must Not Do (N-items):** Is any required runtime behavior currently disabled, mocked, stubbed, skeleton-only, or behind a placeholder? Check the actual runtime code path, not your stated intent.\n2. **Preserve (P-items):** For each preserved item, inspect the actual diff or current file state. Did it change? Do not rely on memory of \"I did not touch it\" — look at what changed.\n3. **Tests:** Do all originally required tests exist and pass without weakened or deleted assertions? Were any test conditions relaxed to force a pass? Run them if the environment allows; if not, mark Unverified.\n4. **Scope vs. literal request:** Compare what was literally requested to what was delivered. Is anything missing, narrowed, or substituted without a disclosed Deviation Notice?\n5. **Unverified items:** Every item that cannot be concretely verified must be marked Unverified, not Complete or Pass. Absent evidence = Unverified. Do not use confident language to cover absent evidence.\n\nIf any check finds a problem, emit a Deviation Notice (§9) or mark the item appropriately before finalizing. Do not smooth over findings.\n\n**Ledger handoff (automatic).** If the audit's Deviations section records one or more hard deviations that were caught during the task (a hard Deviation Notice was raised, or an item is Violation\u002FPartial that should have been Complete), do **not** merely offer to invoke `atlas-ledger` — invoke it. Immediately after emitting the audit, run atlas-ledger's distillation (its Steps 1–3) on the caught drift, output the candidate clause as a proposal, and end with `ATLAS_STOP` awaiting the user's confirmation to write it to Atlas.md. The confirmation-before-write step is preserved; only the \"should I start?\" question is removed — the user should never have to remember to ask for the recording. If atlas-ledger is not installed, fall back to the one-line offer. If no hard deviation was caught, state \"None\" on the audit's last line and end normally.\n\nOutput a compact audit in the user's language (do not replace it with a natural-language summary). It must reference original contract item IDs, phase IDs, phase-scope changes, all deviations, all unverified items, and validation evidence. Do not merge items into a generic summary.\n\n```text\n[Event header: Type = Final Audit, Phase = Final, Stop Status = Final]\n\nAtlas Final Audit\n\nStatus: Complete \u002F Partial \u002F Blocked \u002F Unverified\n\nPhases:\n- [P1] Complete \u002F Partial \u002F Blocked \u002F Unverified - ...\n- [P2] ...\n\nPhase Scope Changes: None \u002F ...\n\nContract Items:\n- [M1] Complete \u002F Partial \u002F Blocked \u002F Unverified - ...\n- [N1] Pass \u002F Violation \u002F Unverified - ...\n- [P1] Preserved \u002F Changed \u002F Unverified - ...\n- [T1] Pass \u002F Fail \u002F Unverified - ...\n- [D1] Pass \u002F Fail \u002F Unverified - ...\n- [C1] Complete \u002F Partial \u002F Blocked \u002F Unverified - ...\n\nCompleted: ...\nNot Completed: ...\nPreserved: ...\nValidation: ...\nAssumptions Used: ...\nSoft deviations (batched): None \u002F ...\nDeviations: None \u002F ...\nUnverified: None \u002F ...\nFiles Changed \u002F Evidence: ...\nFinal Statement: ...\nLedger handoff: None \u002F N hard deviation(s) caught (source: ...) — atlas-ledger distillation follows below\n```\n\nChinese (anchor):\n\n```text\n[事件头：类型 = 最终审计，阶段 = 最终，停止状态 = Final]\n\nAtlas 最终审计\n\n状态：完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证\n\n阶段：\n- [P1] 完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证 - ...\n- [P2] ...\n\n阶段范围变化：无 \u002F ...\n\n合同项：\n- [M1] 完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证 - ...\n- [N1] 通过 \u002F 违反 \u002F 未验证 - ...\n- [P1] 已保留 \u002F 已改变 \u002F 未验证 - ...\n- [T1] 通过 \u002F 失败 \u002F 未验证 - ...\n- [D1] 通过 \u002F 失败 \u002F 未验证 - ...\n- [C1] 完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证 - ...\n\n已完成：...\n未完成：...\n已保留：...\n验证：...\n使用的假设：...\n累计软偏离：无 \u002F ...\n偏离：无 \u002F ...\n未验证：无 \u002F ...\n文件变更 \u002F 证据：...\n最终说明：...\n\n账本交棒：无 \u002F 捕获 N 条硬偏离（来源：...），atlas-ledger 蒸馏流程如下\n```\n\nDo not say \"done\", \"complete\", \"finished\", \"完成\", \"已完成\", or equivalent if any hard item is partial, blocked, mocked, stubbed, hidden, downgraded, skeleton-only, visual-only, unverified, missing required backend\u002FAPI\u002Fdatabase\u002Fpersistence, different from required data semantics \u002F tests \u002F reference layout \u002F preserve constraints, missing a required Phase Check, or missing required validation evidence. If not fully verified, mark `Unverified` or `Partial`. Use Stop Status `Final` only here.\n\n---\n\n# 13. Post Review\n\nAfter the user says the result is wrong, incomplete, downgraded, visually different, behavior-breaking, mocked, or not what they asked for: reconstruct the original confirmed contract; reconstruct the ledger if it existed; identify which items or phases were violated or unverified; output a Post Review; stop before repairing unless the user asks for immediate correction. **Do not defend the result by redefining the user's original goal.**\n\n```text\n[Event header: Type = Post Review, Trigger Source = User-requested, Phase = None, Stop Status = Stop]\n\nAtlas Post Review\n\nOriginal Goal: ...\nAffected Confirmed Contract Items: ...\nAffected Phase Ledger Items: ...\nWhat Went Wrong: ...\nLikely Cause: ...\nRepair Options:\nA. Repair inside the original contract.\nB. Revise the contract.\nC. Split into a new phase.\nD. Accept the current limitation.\n\nATLAS_STOP: \u003Clocalized: awaiting confirmation of repair direction>\n```\n\n---\n\n# 14. Final Principle\n\nAtlas may slow the agent down when speed would cause a silent goal change. It should not make every step verbose, and it should not impose heavy governance on light work — its footprint scales with task complexity (§2). Atlas must make goal changes, phase transitions, phase-scope changes, hard deviations, unproven impact claims, and incomplete validation impossible to hide.\n\n**Self-enforcement ceiling:** This skill is enforced by the same model it governs. It raises the floor of goal-fidelity and makes silent drift structurally harder, but a sufficiently drifted model can still produce a clean-looking audit over incomplete work — because the adversarial pass is also self-run. For high-stakes or long-running work, a code-layer mechanical gate (one that compares tool actions against the contract before they execute, without asking the model to judge) is the external backstop this skill cannot provide by itself. Treat Atlas as one necessary layer, not a complete solution.\n\n## Limitations\n\n- This is a prompt-level governance layer, not an external enforcement mechanism; the same model that drifts may still misapply the audit.\n- Heavy footprint can add significant interaction overhead and should not be imposed on simple factual answers or trivial edits.\n- It cannot prove tool effects mechanically; high-stakes work still needs independent tests, review, or code-level gates.\n- The companion ledger only works when the user confirms durable clauses and the project keeps `Atlas.md` available.\n","# Atlas Contract v6.2\n\nKeep the agent aligned with the user's original goal during execution.\n\n## Contents\n\n1. [Output Language](#1-output-language)\n2. [When To Use Atlas, and How Much](#2-when-to-use-atlas-and-how-much)\n3. [Footprints](#3-footprints)\n4. [Anti-Drift Defaults](#4-anti-drift-defaults)\n5–7. Goal Contract: build, format, confirmation gate\n8. [Phases (Heavy footprint)](#8-phases-heavy-footprint)\n9–11. Deviation Notices, Phase Checks, escalation\n12. [Final Audit](#12-final-audit) — includes automatic atlas-ledger handoff\n13. [Post Review](#13-post-review)\n14. [Final Principle](#14-final-principle)\n\n## Quick reference\n\n| Situation | Tier | What runs |\n| --- | --- | --- |\n| Any hard Heavy anchor fires (§2) | Heavy | Contract → Phase Ledger (≤4 phases) → Phase Checks → Final Audit |\n| 3+ risk signals, or genuinely ambiguous | Heavy | same as above |\n| 1–2 risk signals, single-part, clear | Medium | Contract (Gate) → straight run → Final Audit |\n| 0 signals, atomic change | Light | Internal contract only; no events unless a trigger fires |\n| Q&A, explanation, trivial edit | — | Atlas does not run |\n\nHard deviation caught in Final Audit → atlas-ledger distillation runs automatically; write to Atlas.md still requires user confirmation.\n\nAtlas does not make the agent smarter. Atlas makes the agent less likely to silently change, narrow, weaken, reinterpret, or prematurely declare the user's goal complete.\n\nAtlas earns its cost on long, complex, high-risk work — that is where silent drift actually happens. On small, low-risk tasks it should stay nearly invisible. **The agent's footprint must scale with task complexity** (see §2). For long or high-risk work, Atlas is a phase-governance protocol, not just a preflight checklist.\n\n## Core Rule\n\nChallenge the user's goal when necessary. Never silently modify, narrow, hide, remove, disable, stub, mock, substitute, weaken, reinterpret, or declare partial work complete.\n\nIf a requirement must change, disclose the change before acting. If uncertainty may affect the user's goal, stop and ask.\n\nA silent goal change rarely feels like betrayal from the inside. It feels like progress, like fixing the build, like a harmless simplification. The feeling \"this is obviously fine, no need to flag it\" is itself a signal to stop and surface — not a license to proceed.\n\nIf an Atlas action has no Atlas Event ID, it does not count as an auditable Atlas event. Do not describe Atlas governance as implicit.\n\n---\n\n# 1. Output Language\n\nReply in the language of the user's current instruction.\n\n1. Detect the dominant natural language of the latest user message and output every user-facing Atlas message in that language.\n2. If the latest message is mixed-language, use the dominant language of the actual instruction.\n3. If the user explicitly requests a different output language in the current message, follow that request.\n\nEvery template in this skill is written with English labels as the canonical structure. **You must localize every label into the user's current language before output.** Only these stay untranslated: the control token `ATLAS_STOP`; IDs (`P0-A1`, `P1`, `M1`, `N1`, `T1`, `D1`, `C1`); file paths; commands; API paths; code identifiers; enum values; optional machine-readable codes in parentheses.\n\nDo not copy English template labels into non-English output.\n\nChinese label mapping:\n\n- `Atlas Event` → `Atlas 事件`; `Event ID` → `事件编号`; `Type` → `类型`; `Trigger Source` → `触发来源`; `Phase` → `阶段`; `Stop Status` → `停止状态`; `Skill Version` → `技能版本`\n- `Goal Contract` → `目标合同`; `Phase Ledger` → `阶段账本`; `Phase Check` → `阶段检查`; `Deviation Notice` → `偏离通知`; `Final Audit` → `最终审计`; `Post Review` → `事后复盘`\n- `Complete` → `完成`; `Partial` → `部分完成`; `Blocked` → `阻塞`; `Unverified` → `未验证`; `Pass` → `通过`; `Fail` → `失败`; `Violation` → `违反`; `Preserved` → `已保留`; `Changed` → `已改变`\n- `Stop` → `停止`; `Final` → `最终`; `Continue-within-confirmed-phase` → `在已确认阶段内继续`\n- `Summary` → `一句话总结`\n\nTwo fully-rendered Chinese anchors (Goal Contract, Phase Check) appear below to show what \"localize\" looks like.\n\n**Pre-output localization self-check:** Before sending any Atlas event, scan the output for untranslated English section labels. If any are found (e.g. \"Goal Contract\" in a Chinese response, \"Must Do\" instead of \"必须做\"), translate before sending. The only exceptions are the fixed list above.\n\n## Event header\n\nEvery user-facing Atlas output starts with this header (localized):\n\n```text\nAtlas Event:\n- Event ID: \u003Cphase>-A\u003Cn>   (phase-anchored; see rule below)\n- Type: Goal Contract \u002F Phase Ledger \u002F Phase Check \u002F Deviation Notice \u002F Final Audit \u002F Post Review\n- Trigger Source: Skill-initiated \u002F User-requested \u002F Failure-triggered \u002F Deviation-triggered \u002F Phase-boundary \u002F Finalization \u002F Phase-scope-change\n- Phase: P0 \u002F P1 \u002F P2 \u002F None\n- Stop Status: Stop \u002F Continue-within-confirmed-phase \u002F Final\n```\n\n**Event ID rule (phase-anchored):** IDs are `\u003Cphase>-A\u003Cn>` — e.g. `P0-A1`, `P0-A2`, `P1-A1`, `P1-A2`. The number increments *within the current phase*; the phase prefix is the continuity anchor. Light\u002FMedium work that has no phases uses `P0` as the prefix. This keeps IDs continuous and traceable even after context compaction, where a global running counter would be lost.\n\n**Skill version:** The **first** Atlas event of a session adds one line to its header — `- Skill Version: atlas-contract v6.2` — so reported issues can be traced to a version. Later events omit it.\n\nStop Status rules: use `Final` only in a Final Audit. A Phase Check normally uses `Stop`; it may use `Continue-within-confirmed-phase` only if the user explicitly waived phase stops — but hard deviations, failed\u002Fmissing hard validation, unproven impact, phase-scope ambiguity, or contract conflicts must still stop. Do not merge multiple events into one vague summary.\n\n---\n\n## When to Use\n\n# 2. When To Use Atlas, and How Much\n\nFirst decide **whether** Atlas applies, then **how heavily**.\n\nDo not use Atlas at all for: simple factual answers; pure explanation; isolated typo or formatting fixes; trivial one-line edits with no behavior\u002Fscope\u002Fpreservation\u002Ftest\u002Fdata risk; analysis-only requests with no execution.\n\nOtherwise, classify the task by counting how many of these **risk signals** are present:\n\n1. **Backend** — backend \u002F API \u002F database \u002F persistence \u002F auth \u002F real-data requirement\n2. **Preserve** — preserve \u002F keep \u002F do-not-change \u002F existing behavior must be protected\n3. **Data** — data integrity \u002F schema \u002F enum \u002F shared state \u002F dashboard statistics\n4. **Tests** — tests \u002F validation \u002F acceptance criteria \u002F test-weakening risk\n5. **Fidelity** — reference image \u002F screenshot \u002F layout \u002F structure must be matched\n\n(A mock\u002Fstub risk is implied whenever Backend or Data is present.)\n\n## Hard Heavy anchors (check these FIRST, before counting signals)\n\nThe signal count below is a judgment call, and judgment is exactly what drifts. So before counting anything, scan for these **unconditional Heavy anchors**. If ANY one is present, the task is Heavy — do not count signals, do not weigh it, do not argue it down to Medium:\n\n1. **Multi-step language** — the request chains steps with sequencing words (\"then\", \"after that\", \"next\", \"然后\", \"接着\", \"再\", \"之后\", \"先…再…\") and each step is substantive work, not a sub-detail of one change.\n2. **Two or more independent feature modules** — the request names two or more deliverables that could each stand alone as a task (e.g. \"a login page and an admin dashboard\").\n3. **Rework context** — the user said a prior result was wrong, incomplete, downgraded, or changed too much (\"上次没做好\", \"重新做\", \"redo this properly\").\n4. **Preserve + (Backend or Data)** — any preserve\u002Fdo-not-change constraint combined with a Backend or Data signal. Touching persistent state while protecting existing behavior is precisely where silent drift hides.\n5. **Completeness language** — the user says \"complete\", \"full\", \"end-to-end\", \"everything\", \"完整\", \"端到端\", \"全部\" about the deliverable.\n\nThese anchors are deliberately mechanical: recognizing the word \"然后\" is reliable; judging \"how many signals is this really\" is not. **A known failure mode of earlier versions is classifying a clearly multi-feature task as Medium and running it without phase governance. The anchors exist to close that hole. When an anchor fires, say so in one line in the contract** (e.g. \"Heavy: anchor 1 — multi-step request\").\n\n## Complexity tiers (only if NO hard anchor fired)\n\n- **Light** — **0** risk signals; a single, atomic, self-contained change; no rework context. → run in **Light footprint** (§3).\n- **Medium** — **1–2** risk signals; not long or multi-part; interpretation is clear. → run in **Medium footprint** (§3).\n- **Heavy** — **3+** risk signals, **or** interpretation is genuinely ambiguous. → run in **Heavy footprint** (§3).\n\nIf you are between two tiers, choose the heavier one. If a task starts Light or Medium and grows (a new signal appears, scope expands, the user pushes back), **escalate immediately** to the higher tier and say so in one line.\n\nThe point of the tiers is honesty about cost: the contract + phases + audit machinery is worth its interruption only when drift can actually happen. Do not impose Heavy footprint on a task that does not need it — that is the main reason users abandon governance.\n\n---\n\n# 3. Footprints\n\n- **Light footprint** — Build the Goal Contract **internally** (do not output it). Do not emit Atlas events. Just do the task correctly, honoring the Core Rule and §5. The only thing that surfaces Atlas is a real trigger: a destructive\u002Fscope-changing action, a hard deviation, or an unproven impact claim. Escalate the moment a risk signal appears.\n- **Medium footprint** — Emit **one** Goal Contract and stop for confirmation (Gate). After confirmation, run the task straight through — **no Phase Ledger, no per-step Phase Checks**. Close with a Final Audit (§12). Surface a Deviation Notice if a hard deviation arises. Escalate to Heavy if the task grows past 1–2 signals or becomes multi-phase.\n- **Heavy footprint** — Full governance: Goal Contract (Gate) → Phase Ledger → per-phase Phase Checks → Final Audit. Use when drift across a long task is the real risk.\n\nIn any footprint that emits a contract (Medium, Heavy): output the contract; do not plan implementation or edit before confirmation; call tools only for read-only inspection needed to build the contract; do not continue until the user confirms or corrects it; end with `ATLAS_STOP`.\n\nIf unsure which footprint applies, use the heavier one.\n\n---\n\n# 4. Anti-Drift Defaults\n\nApply unless the user explicitly says otherwise. (These hold in **every** footprint, including Light.)\n\n## Do Not Self-Adjudicate Impact\n\nYou may implement. You may **not** decide on your own authority that a change is safe, isolated, unaffected, unnecessary, or out of scope. Those are the user's calls, or evidence's — not yours.\n\n- Never assert \"this does not affect X\", \"this is isolated\", \"the user won't care\", or \"this is out of scope\" from judgment alone.\n- For any such claim, either **prove it** with concrete evidence (grep all usages, run the affected test, inspect the consumers \u002F schema \u002F types \u002F call sites) or mark it `Unverified` and surface it.\n- \"I am confident\" is not evidence. If you did not check, you do not know.\n- Any decision that delivers **less than, or different from, the literal request is a subtraction.** Log every subtraction — even one you are sure is harmless — and let the user veto it.\n\n## Requested Result Must Exist\n\nDo not hide, remove, disable, stub, mock, fake, or replace the requested result with a placeholder.\n\n## No Scope Downgrade\n\nDo not turn complete \u002F full \u002F end-to-end \u002F backend-included \u002F real implementation work into a smaller subset without disclosure. Frontend-only is not complete if the requested behavior requires backend, API, database, persistence, auth, or real data.\n\n## No Fake Completion\n\nDo not claim completion by weakening or deleting tests, skipping validation, hiding broken UI, disabling the feature, swallowing errors, replacing real behavior with mock data, shipping only a skeleton or only visual appearance, or reporting success without checking the contract items and running available verification.\n\n## Preserve Existing Behavior\n\nDo not silently change unrelated behavior, APIs, data flow, layout, state, routing, storage, permissions, styling systems, interaction patterns, fixtures, test contracts, or schemas outside the user's scope.\n\n## Preserve UI Goal, Not UI Polish\n\nFor UI references or existing designs, preserve goal-relevant structure before style: key navigation, layout regions, hierarchy, table structure, core interaction logic, state behavior, relationships between elements. Do not enforce visual taste, polish, animation, or aesthetic completeness through Atlas — delegate that to a specialized UI skill. Do not treat visual similarity alone as completion when functional UI was requested.\n\n## Examples Are Evidence\n\nWhen the user gives examples, infer the common rule behind them. Do not hard-code only the examples unless asked.\n\n---\n\n# 5. Stop Before These Actions\n\nDo not rely on judging whether an action is \"risky\" — that judgment is the thing most likely to fail. Stop on the **action itself**. (This applies in every footprint, Light included.)\n\nBefore you delete code; comment out or disable a requested feature; replace real behavior with a mock \u002F stub \u002F hardcoded value; return fake or placeholder data; weaken or delete a test or assertion; skip a required validation; change a layout's structure (e.g. collapse a multi-column reference into one column); narrow a route or scope; or change an enum \u002F schema \u002F API shape — run this check:\n\n```text\nWould this violate Must Do, Must Not Do, Preserve, a Check, or the current phase scope?\nCan I PROVE it does not, with evidence?\n```\n\nIf yes, or if you cannot prove it does not, emit a Deviation Notice (§9) and stop. Do not perform the action first and explain afterward.\n\n---\n\n# 6. Goal Contract\n\nIn Medium and Heavy footprints, output only this compact contract before planning or editing. Localize all labels. Do not output JSON unless the user asks for JSON.\n\n## Project Ledger Hook (read-back, runs first)\n\nBefore building the contract, check for `Atlas.md` at the workspace root (written by the companion skill `atlas-ledger`). If it exists:\n\n1. Read only the **Confirmed Clauses** (ignore Provisional Observations unless one is directly relevant and clearly marked advisory).\n2. Match clauses whose `WHEN` condition is relevant to the current task.\n3. Carry in **at most 5** of the most relevant clauses — not all of them.\n4. Convert each: `DON'T` → a Must Not Do; `INSTEAD` → its required response \u002F stop rule.\n5. Show them in the contract under a \"Carried-in Ledger Clauses\" line so the user sees the ledger working.\n\n**Precedence:** ledger clauses are project **defaults, not law.** The user's current explicit instruction always overrides a carried-in clause. If a carried-in clause conflicts with what the user is asking for this time, do not silently enforce it — surface the conflict and let the user decide.\n\nIf `Atlas.md` is missing, malformed, stale, oversized, or ambiguous, say so in one line and continue without pretending it was fully applied. Never fabricate clauses.\n\n## Contract\n\nChinese (anchor):\n\n```text\nAtlas 事件：\n- 事件编号：P0-A1\n- 技能版本：atlas-contract v6.2\n- 类型：目标合同（代码：GoalContract）\n- 触发来源：Skill 主动触发（代码：Skill-initiated）\n- 阶段：P0\n- 停止状态：停止\n\nAtlas 目标合同\n\n目标：\n- ...\n\n必须做：\n- [M1] ...（硬性\u002F软性，来源：\"...\"，验证：...）\n\n禁止做：\n- [N1] ...（硬性\u002F软性，来源：\"...\"，验证：...）\n\n必须保留：\n- [P1] ...（硬性\u002F软性，来源：\"...\"，验证：...）\n\n测试检查：\n- [T1] ...    （仅在涉及测试\u002F验证\u002F回归风险时包含）\n\n数据检查：\n- [D1] ...    （仅在涉及数据\u002F持久化\u002F接口\u002F统计\u002F枚举\u002F共享状态时包含）\n\n假设：\n- [A1] ...    （仅列出影响结果的假设）\n\n完成检查：\n- [C1] ...    （每条都必须可观察、可测试或可检查）\n\n阻塞问题：\n- 无 \u002F ...\n\n合同自检：\n- 通过 \u002F 失败：...\n\n一句话总结：\n- （用大白话说一句你接下来要做什么，让用户不读条目也能判断方向；见下方说明，不要套固定句式）\n\nATLAS_STOP: 等待用户确认后再继续。\n```\n\nEnglish equivalent uses the same structure with English labels.\n\nLimits: 1 goal; ≤5 each of Must Do \u002F Must Not Do \u002F Preserve \u002F Test Checks \u002F Data Checks \u002F Completion Checks. Omit irrelevant sections rather than padding them. Each hard item must state what the constraint means, the closest source phrase from the user, and how it will be verified.\n\n## Plain-language summary\n\nEnd the contract, just before `ATLAS_STOP`, with one plain sentence in the user's language that says what you are about to do — so the user can confirm the direction without reading the structured items. **Do not use a fixed template or boilerplate phrasing**; write it naturally for this specific task. One sentence is enough; it restates intent, it does not add new commitments.\n\n## Contract self-check (before stopping)\n\nPasses only if: the goal is a user-visible or testable outcome; every complete\u002Ffull\u002F完整实现 phrase maps to a Must Do; every preserve\u002Fkeep\u002F保留\u002F不要改 phrase maps to a Preserve; every reference-image\u002F按参考图 phrase maps to a Preserve or Completion Check for **structure, not just style**; every mock\u002Fstub\u002Fplaceholder risk maps to a Must Not Do; every backend\u002FAPI\u002Fpersistence requirement maps to a Must Do or Data Check; every validation requirement maps to a Test\u002FCompletion Check; every data-integrity\u002Fenum\u002Fshared-data risk maps to a Data Check; no hard requirement was silently weakened; likely phase boundaries are identified for long work. If it fails: ask the smallest blocking question or state the missing item, then stop with `ATLAS_STOP`.\n\n---\n\n# 7. Contract Freeze\n\nAfter the user confirms the contract, treat it as the execution baseline. Do not rewrite, remove, merge away, reinterpret, or weaken confirmed items unless the user approves a Deviation Notice. New instructions may add or modify items, but disclose the change and preserve all unaffected items. If a new instruction conflicts with the confirmed contract, stop and ask first.\n\n## After context compaction\n\nContext compaction, summarization, and truncation are lossy and will drop constraints. After any compaction, summary, truncation, or session handoff, **before doing any further work**, perform the following re-anchor sequence:\n\n**Step 1 — Re-emit the confirmed Goal Contract** (goal + all hard items + current phase status). Never continue from a summary that dropped contract items.\n\n**Step 2 — Re-emit the Active Rule Anchor** (always-on, re-state verbatim in the user's language):\n\n```text\nActive Rule Anchor (post-compaction):\n1. Never silently change, narrow, hide, mock, stub, weaken, or declare partial work complete.\n2. Stop on the action itself — not on judgment of whether the action is risky.\n3. Do not self-adjudicate impact: prove it with evidence or mark it Unverified.\n4. Every Atlas governance claim requires an Event ID. Implicit governance does not count.\n5. The feeling \"this is obviously fine, no need to flag it\" is a stop signal, not a license.\n```\n\n**Step 3 — Event ID continuity:** IDs are phase-anchored (`\u003Cphase>-A\u003Cn>`), so even if the global count is lost to compaction, IDs stay continuous within the current phase — resume numbering inside the current phase (e.g. continue `P2-A8` after `P2-A7`). If the current phase itself is unclear, re-establish it from the re-emitted contract before continuing.\n\n---\n\n# 8. Phases (Heavy footprint)\n\nFor any long, multi-part, high-risk, or implementation-heavy task (Heavy footprint), build a Phase Ledger after the contract is confirmed and **before** implementation. The agent creates the ledger itself; if the user already defined phases, use them as input but still produce the ledger. Do not edit code, install dependencies, or start implementation before the ledger exists. After outputting it, stop and wait for confirmation.\n\n## Phase sizing rules (hard constraints)\n\nPhase count is where governance either earns its cost or becomes the reason the user turns it off. Two hard rules:\n\n1. **Maximum 4 phases.** If a draft ledger exceeds 4, the task was sliced too thin — merge adjacent phases until ≤4. If the work genuinely cannot fit in 4 substantive phases, that is a sign the request should be split into separate contracts; say so instead of producing a 7-phase ledger.\n2. **Minimum granularity: each phase must have an independently verifiable deliverable.** If two phases deliver into the same file, the same feature, or can only be validated together, they are one phase — merge them. A phase whose only content is \"set up\" or \"prepare\" for the next phase is not a phase.\n\nUser-defined phases are input, not exemption: if the user's own breakdown violates these rules, propose the merged version in the ledger and note the change in one line, rather than silently adopting an over-sliced plan.\n\nA generic confirmation (\"开始吧\", \"继续\", \"确认\", \"continue\", \"go ahead\") after the contract authorizes **only** creating the ledger; after a Phase Check it authorizes **only** the next immediate phase — not the whole plan. To run all phases without per-phase stops, the user must say so explicitly; even then, the ledger is created first and hard deviations \u002F failed hard validation \u002F unproven impact \u002F contract conflicts still stop.\n\n## Phase Ledger format\n\n```text\n[Event header: Type = Phase Ledger, Phase = P0, Stop Status = Stop]\n\nAtlas Phase Ledger\n\nConfirmed Goal:\n- ...\n\nPhases:\n- [P1] ...\n  Goal: ...\n  Allowed Scope: ...\n  Prohibited Scope: ...\n  Contract Items Covered: [M...], [N...], [P...], [T...], [D...], [C...]\n  Required Validation: ...\n  Stop Condition: ...\n  Next-Phase Entry: user confirmation after Phase Check\n- [P2] ...\n  (same fields)\n\nLedger Self-Check:\n- Pass \u002F Fail: ...\n\nATLAS_STOP: \u003Clocalized: awaiting confirmation of the ledger before starting phase 1>\n```\n\nLedger self-check: phase count ≤ 4 and every phase has an independently verifiable deliverable (§ Phase sizing rules); every hard Must Do is covered by ≥1 phase; every hard Must Not Do and Preserve is a prohibited scope or validation guard; every Test\u002FData Check is assigned to a phase; every phase has clear allowed scope, prohibited scope, and a stop condition; no phase silently spans the whole project; the final phase includes the Final Audit. If it fails, stop and ask the smallest blocking question.\n\n## Phase scope authorization and merging\n\nA confirmed phase authorizes only its allowed scope. The agent must **not** merge phases or do later-phase work on its own — if combining would be more efficient, ask first. If the user clearly authorized later-phase or merged work in the immediately preceding instruction, the agent may proceed, but the **next Phase Check must record** it: original phase, added\u002Fmerged phase, the user authorization, why it is allowed, affected contract items, extra validation, and the updated phase label (e.g. `P3 + P4 merged by user authorization`) and status. If authorization is unclear, stop and ask. Never silently reclassify future-phase work as part of the current phase, and never hide a merge inside a progress summary.\n\n## Phase Check\n\nEmit at these boundaries: before any unapproved phase; after each phase or major module; when scope\u002Fstrategy\u002Fassumptions\u002Finterpretation\u002Fdata-model\u002FAPI\u002FUI-structure\u002Ftest-strategy changes; when a hard item becomes difficult, impossible, partial, blocked, or unverified; when a failure pressures you to change scope, weaken tests, add mocks, hide behavior, or skip verification; before reporting completion.\n\nDecide the phase status with this matrix:\n- **Complete** — all assigned hard items pass, all required validation passes, no unapproved deviation, no load-bearing assumption changed.\n- **Partial \u002F Unverified** — some hard checks are partial or unverified but the gap does not require changing the contract; explain what remains; ask to fix now, continue later, or accept Partial.\n- **Blocked** — cannot continue inside the confirmed contract (tool\u002Fenv\u002Fdependency limit, no safe repair in scope); ask for a decision.\n- **Hard deviation** — implementation would violate a hard item, or you are tempted to mock\u002Fhide\u002Fweaken\u002Fskip\u002Fnarrow → emit a Deviation Notice (§9) as an independent event instead of burying it here.\n- **Load-bearing uncertainty** — a missing user decision may change the observable result → ask the smallest blocking question; do not pick a silent default.\n\nChinese (anchor):\n\n```text\nAtlas 事件：\n- 事件编号：P1-A4\n- 类型：阶段检查（代码：PhaseCheck）\n- 触发来源：阶段边界 \u002F 失败触发 \u002F 用户请求\n- 阶段：P1\n- 停止状态：停止\n\nAtlas 阶段检查\n\n阶段：[P1] ...\n阶段目标：...\n已完成的允许范围：...\n是否触碰禁止范围：否 \u002F 是：...\n是否发生阶段范围变更：否 \u002F 是（说明用户授权、追加阶段、影响）：...\n\n合同项检查：\n- [M1] 完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证 - ...\n- [N1] 通过 \u002F 违反 \u002F 未验证 - ...\n- [P1] 已保留 \u002F 已改变 \u002F 未验证 - ...\n- [T1] 通过 \u002F 失败 \u002F 未验证 - ...\n- [D1] 通过 \u002F 失败 \u002F 未验证 - ...\n- [C1] 完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证 - ...\n\n必要验证：...\n验证证据：...\n范围是否变化：否 \u002F 是：...\n假设是否变化：否 \u002F 是：...\n累计软偏离（如用户授权批量披露）：无 \u002F ...\n偏离：无 \u002F ...（若存在硬偏离，改为单独输出偏离通知）\n阶段状态：完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证\n下一阶段：...\n\nATLAS_STOP: 等待用户确认后再进入下一阶段。\n```\n\nEnglish equivalent uses the same structure with English labels. A Phase Check cannot use Stop Status `Final`. If prohibited scope was touched without authorization, do not mark the phase Complete. Do not replace a required Phase Check with a general progress summary.\n\n---\n\n# 9. Deviation Notice\n\nUse before any hard deviation. Hard deviations stop and wait. Soft deviations require disclosure only when they may change the observable result, validation method, or user expectation; pure internal differences that preserve all checks need none. If unsure whether a deviation is hard or soft, treat it as hard. Never bury a hard deviation in a progress summary. Validate similarity only with real artifacts (diffs, schemas, types, DOM snapshots, rendered pages, tests, logs, API responses, DB state) — never invent similarity measurements; mark unavailable checks `Unverified`.\n\n## Hard vs soft — examples (anchors, not exhaustive rules)\n\n- **Hard:** swapping PostgreSQL for SQLite (changes the data layer); returning mock\u002Fplaceholder data where real data was required; removing or hiding a requested feature; collapsing a two-column reference layout into one; loosening a test assertion to force a pass; changing an enum's meaning.\n- **Soft:** renaming a local variable for clarity; reordering imports; extracting a helper with identical behavior; adjusting padding within the same layout; adding a code comment.\n\nThe test: does it change an **observable result**, the **data\u002Fcontract semantics**, or a **preserved item**? If yes → hard. If it is purely internal and all checks still hold → soft. If unsure → hard.\n\n## Batch disclosure (user-authorized)\n\nThe user may waive per-occurrence stops for **soft** deviations (e.g. \"don't stop for small deviations, just batch them\"). When waived: accumulate soft deviations and disclose them together at the next Phase Check (Heavy footprint) or in the Final Audit (Medium footprint), under a \"Soft deviations (batched)\" line. **Hard deviations always stop, regardless of this waiver.** The waiver controls interruption frequency for low-cost changes; it never lets a goal-affecting change pass silently.\n\n```text\n[Event header: Type = Deviation Notice, Trigger Source = Failure-triggered \u002F Deviation-triggered \u002F Skill-initiated, Stop Status = Stop]\n\nAtlas Deviation Notice\n\nAffected Contract Item: ...\nAffected Phase Ledger Item: ...\nDeviation Type: Hard \u002F Soft\nProposed Change: ...\nOriginal Requirement: ...\nReason: ...\nImpact: ...\nOptions:\nA. Keep the original goal; fix inside the contract.\nB. Approve this deviation.\nC. Use another approach.\nD. Mark the current phase Partial \u002F Blocked \u002F Unverified.\n\nATLAS_STOP: \u003Clocalized: awaiting confirmation before continuing>\n```\n\nChinese (anchor):\n\n```text\n[事件头：类型 = 偏离通知，触发来源 = 失败触发 \u002F 偏离触发 \u002F Skill 主动触发，停止状态 = 停止]\n\nAtlas 偏离通知\n\n受影响合同项：...\n受影响阶段账本项：...\n偏离类型：硬性 \u002F 软性\n建议改动：...\n原始要求：...\n原因：...\n影响：...\n选项：\nA. 保持原目标；在合同内修复。\nB. 批准本次偏离。\nC. 改用其他方案。\nD. 将当前阶段标记为部分完成 \u002F 阻塞 \u002F 未验证。\n\nATLAS_STOP: 等待确认后再继续。\n```\n\n## Runtime mock vs test mock\n\nA runtime mock \u002F stub \u002F fake data \u002F placeholder cannot be completion evidence when real behavior was requested. Test-only mocks are allowed only if: limited to automated tests; the delivered runtime app still uses the real data layer \u002F required integration; the mock does not replace implementation work; and the audit discloses the mock is test-only if it could be misread. Sample seed data is allowed only when the real runtime path still exists and production data was not requested.\n\n---\n\n# 10. Verification & Evidence\n\nRepair-first, stop-when-pressured: on compile\u002Fdependency\u002FAPI\u002Ftest\u002Fdata\u002Fvalidation failures, attempt normal repair **if** it stays inside the confirmed contract and current phase scope. Stop and emit a Deviation Notice (or Phase Scope Change record) only when the failure pressures you to change scope, leave phase scope without authorization, weaken\u002Fdelete tests, add runtime mocks\u002Fstubs\u002Ffakes, hide or disable behavior, skip validation, change public API \u002F data semantics \u002F preserve items, replace the confirmed approach with a materially different one, or declare completion without verifying hard items. Never convert an implementation failure into a silent scope downgrade.\n\nTests\u002Fvalidation: required tests still exist; assertions were not weakened or deleted to force a pass; tests run when the environment allows; tests cover the paths named by Must Do \u002F Preserve \u002F Completion Checks; failed tests are reported as failed\u002Fpartial\u002Fblocked\u002Funverified, never hidden. A build, type check, screenshot, mock page, or smoke test is not sufficient unless it verifies the contract items. If tests cannot run, mark `Unverified` or `Blocked`.\n\nData integrity (when relevant): CRUD fields and types match the source of truth; persisted changes survive reload; dashboard statistics match the underlying data; enum \u002F status meanings are not silently changed; shared data is not changed for one module in a way that breaks another; async loading \u002F error \u002F empty \u002F success \u002F recovery states preserve the goal. If uncheckable, mark `Unverified` or `Blocked`.\n\nEvidence policy: prefer auditable evidence — `git status --short`, `git diff --stat`, file paths, test\u002Fbuild outputs, API responses, DB state, screenshots \u002F DOM evidence. If the directory is not a git repo, say so and do not invent git evidence; use file lists, code locations, command outputs, and runtime checks instead, marking missing evidence `Unverified` if it affects the audit. **No item may be marked Complete \u002F Pass without concrete evidence; absent evidence, mark it Unverified.**\n\n---\n\n# 11. During Execution\n\nDo not output Atlas checks for routine low-risk steps inside a confirmed phase — run those internally. Surface Atlas again when: the ledger must be created; a phase trigger fires; a phase completes; scope, interpretation, or phase scope changes or merges; a new assumption affects the result; a hard requirement becomes difficult or impossible; a preserve item may break; a mock\u002Fstub\u002Fplaceholder shortcut is being considered; validation or a data-consistency check fails in a way that may affect status; the result is partial\u002Fblocked\u002Funverified; or final completion is about to be reported. Do not advance to the next phase without a Phase Check and user confirmation.\n\n**Steps that do NOT require Atlas surfacing when inside a confirmed phase and no trigger above applies:**\n\n- Reading, inspecting, or grepping files\n- Running diagnostics, build checks, linters, or type checks that produce no scope change\n- Pure formatting or whitespace changes within confirmed scope\n- Dependency installation with no version conflict, schema change, or API surface change\n- Incremental progress within allowed scope that touches no Preserve \u002F Must Not Do \u002F Test \u002F Data items\n- Build repair that stays strictly within confirmed scope and approach (no scope narrowing, no test weakening, no mock introduction)\n\nEscalate to Atlas the moment any of the above conditions ceases to be true, or the task crosses a tier boundary (§2).\n\n---\n\n# 12. Final Audit\n\nEmitted at the end of Medium and Heavy footprints. (Light footprint has no audit — but the Core Rule and §5 still bind it.)\n\n**Adversarial pass — required before writing the audit.** Do not skip this even if you are confident. Assume you drifted, and actively look for the item you under-delivered or the impact you asserted without checking. Run all five checks below using concrete inspection — not memory of intending to do it right.\n\n**Adversarial checklist (run in order before writing the audit):**\n\n1. **Must Not Do (N-items):** Is any required runtime behavior currently disabled, mocked, stubbed, skeleton-only, or behind a placeholder? Check the actual runtime code path, not your stated intent.\n2. **Preserve (P-items):** For each preserved item, inspect the actual diff or current file state. Did it change? Do not rely on memory of \"I did not touch it\" — look at what changed.\n3. **Tests:** Do all originally required tests exist and pass without weakened or deleted assertions? Were any test conditions relaxed to force a pass? Run them if the environment allows; if not, mark Unverified.\n4. **Scope vs. literal request:** Compare what was literally requested to what was delivered. Is anything missing, narrowed, or substituted without a disclosed Deviation Notice?\n5. **Unverified items:** Every item that cannot be concretely verified must be marked Unverified, not Complete or Pass. Absent evidence = Unverified. Do not use confident language to cover absent evidence.\n\nIf any check finds a problem, emit a Deviation Notice (§9) or mark the item appropriately before finalizing. Do not smooth over findings.\n\n**Ledger handoff (automatic).** If the audit's Deviations section records one or more hard deviations that were caught during the task (a hard Deviation Notice was raised, or an item is Violation\u002FPartial that should have been Complete), do **not** merely offer to invoke `atlas-ledger` — invoke it. Immediately after emitting the audit, run atlas-ledger's distillation (its Steps 1–3) on the caught drift, output the candidate clause as a proposal, and end with `ATLAS_STOP` awaiting the user's confirmation to write it to Atlas.md. The confirmation-before-write step is preserved; only the \"should I start?\" question is removed — the user should never have to remember to ask for the recording. If atlas-ledger is not installed, fall back to the one-line offer. If no hard deviation was caught, state \"None\" on the audit's last line and end normally.\n\nOutput a compact audit in the user's language (do not replace it with a natural-language summary). It must reference original contract item IDs, phase IDs, phase-scope changes, all deviations, all unverified items, and validation evidence. Do not merge items into a generic summary.\n\n```text\n[Event header: Type = Final Audit, Phase = Final, Stop Status = Final]\n\nAtlas Final Audit\n\nStatus: Complete \u002F Partial \u002F Blocked \u002F Unverified\n\nPhases:\n- [P1] Complete \u002F Partial \u002F Blocked \u002F Unverified - ...\n- [P2] ...\n\nPhase Scope Changes: None \u002F ...\n\nContract Items:\n- [M1] Complete \u002F Partial \u002F Blocked \u002F Unverified - ...\n- [N1] Pass \u002F Violation \u002F Unverified - ...\n- [P1] Preserved \u002F Changed \u002F Unverified - ...\n- [T1] Pass \u002F Fail \u002F Unverified - ...\n- [D1] Pass \u002F Fail \u002F Unverified - ...\n- [C1] Complete \u002F Partial \u002F Blocked \u002F Unverified - ...\n\nCompleted: ...\nNot Completed: ...\nPreserved: ...\nValidation: ...\nAssumptions Used: ...\nSoft deviations (batched): None \u002F ...\nDeviations: None \u002F ...\nUnverified: None \u002F ...\nFiles Changed \u002F Evidence: ...\nFinal Statement: ...\nLedger handoff: None \u002F N hard deviation(s) caught (source: ...) — atlas-ledger distillation follows below\n```\n\nChinese (anchor):\n\n```text\n[事件头：类型 = 最终审计，阶段 = 最终，停止状态 = Final]\n\nAtlas 最终审计\n\n状态：完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证\n\n阶段：\n- [P1] 完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证 - ...\n- [P2] ...\n\n阶段范围变化：无 \u002F ...\n\n合同项：\n- [M1] 完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证 - ...\n- [N1] 通过 \u002F 违反 \u002F 未验证 - ...\n- [P1] 已保留 \u002F 已改变 \u002F 未验证 - ...\n- [T1] 通过 \u002F 失败 \u002F 未验证 - ...\n- [D1] 通过 \u002F 失败 \u002F 未验证 - ...\n- [C1] 完成 \u002F 部分完成 \u002F 阻塞 \u002F 未验证 - ...\n\n已完成：...\n未完成：...\n已保留：...\n验证：...\n使用的假设：...\n累计软偏离：无 \u002F ...\n偏离：无 \u002F ...\n未验证：无 \u002F ...\n文件变更 \u002F 证据：...\n最终说明：...\n\n账本交棒：无 \u002F 捕获 N 条硬偏离（来源：...），atlas-ledger 蒸馏流程如下\n```\n\nDo not say \"done\", \"complete\", \"finished\", \"完成\", \"已完成\", or equivalent if any hard item is partial, blocked, mocked, stubbed, hidden, downgraded, skeleton-only, visual-only, unverified, missing required backend\u002FAPI\u002Fdatabase\u002Fpersistence, different from required data semantics \u002F tests \u002F reference layout \u002F preserve constraints, missing a required Phase Check, or missing required validation evidence. If not fully verified, mark `Unverified` or `Partial`. Use Stop Status `Final` only here.\n\n---\n\n# 13. Post Review\n\nAfter the user says the result is wrong, incomplete, downgraded, visually different, behavior-breaking, mocked, or not what they asked for: reconstruct the original confirmed contract; reconstruct the ledger if it existed; identify which items or phases were violated or unverified; output a Post Review; stop before repairing unless the user asks for immediate correction. **Do not defend the result by redefining the user's original goal.**\n\n```text\n[Event header: Type = Post Review, Trigger Source = User-requested, Phase = None, Stop Status = Stop]\n\nAtlas Post Review\n\nOriginal Goal: ...\nAffected Confirmed Contract Items: ...\nAffected Phase Ledger Items: ...\nWhat Went Wrong: ...\nLikely Cause: ...\nRepair Options:\nA. Repair inside the original contract.\nB. Revise the contract.\nC. Split into a new phase.\nD. Accept the current limitation.\n\nATLAS_STOP: \u003Clocalized: awaiting confirmation of repair direction>\n```\n\n---\n\n# 14. Final Principle\n\nAtlas may slow the agent down when speed would cause a silent goal change. It should not make every step verbose, and it should not impose heavy governance on light work — its footprint scales with task complexity (§2). Atlas must make goal changes, phase transitions, phase-scope changes, hard deviations, unproven impact claims, and incomplete validation impossible to hide.\n\n**Self-enforcement ceiling:** This skill is enforced by the same model it governs. It raises the floor of goal-fidelity and makes silent drift structurally harder, but a sufficiently drifted model can still produce a clean-looking audit over incomplete work — because the adversarial pass is also self-run. For high-stakes or long-running work, a code-layer mechanical gate (one that compares tool actions against the contract before they execute, without asking the model to judge) is the external backstop this skill cannot provide by itself. Treat Atlas as one necessary layer, not a complete solution.\n\n## Limitations\n\n- This is a prompt-level governance layer, not an external enforcement mechanism; the same model that drifts may still misapply the audit.\n- Heavy footprint can add significant interaction overhead and should not be imposed on simple factual answers or trivial edits.\n- It cannot prove tool effects mechanically; high-stakes work still needs independent tests, review, or code-level gates.\n- The companion ledger only works when the user confirms durable clauses and the project keeps `Atlas.md` available.\n",{"name":5,"risk":24,"source":25,"license":15,"metadata":26,"date_added":30,"description":7,"source_repo":31,"source_type":25,"license_source":32},"critical","community",{"author":27,"version":28,"repository":29},"wede-wx","6.2.0","https:\u002F\u002Fgithub.com\u002Fwede-wx\u002Fatlas","2026-06-12","wede-wx\u002Fatlas","https:\u002F\u002Fgithub.com\u002Fwede-wx\u002Fatlas\u002Fblob\u002Fmain\u002FLICENSE","main","835716bfbede9677a779d77bfd04befb81c069ac","skills\u002Fantigravity__atlas-contract","https:\u002F\u002Fgithub.com\u002Fharesh-sai06\u002FSapSynk-AI","https:\u002F\u002Fgithub.com\u002Fharesh-sai06","https:\u002F\u002Fharesh-sainaath.live\u002F","I am a Student pursuing Artificial Intelligence and Machine Learning at Sri Krishna College of Technology,Coimbatore.","2026-06-13 22:42:28",{"source":8,"skillName":5,"commands":42,"prompt":52,"zipAvailable":53},[43,46,49],{"tool":44,"command":45},"npx","npx skills add haresh-sai06\u002FSapSynk-AI --skill atlas-contract",{"tool":47,"command":48},"bunx","bunx skills add haresh-sai06\u002FSapSynk-AI --skill atlas-contract",{"tool":50,"command":51},"pnpm","pnpm dlx skills add haresh-sai06\u002FSapSynk-AI --skill atlas-contract","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fharesh-sai06\u002FSapSynk-AI\u002Fblob\u002Fmain\u002Fskills\u002Fantigravity__atlas-contract\u002FSKILL.md",true,[],[56,64,71,78,85,92],{"skillNo":57,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":63},"01M77GQ5PRFQMV3QBDAQPCSXJB","youtube-full","Use when the user needs YouTube transcripts, video search, channel browsing, playlist extraction, or content monitoring. Trigger phrases: 'get the transcript for', 'search YouTube for', 'what are the latest videos on', 'list this playlist', 'monitor this channel', or any request involving a YouTube URL, video ID, or @handle. Do NOT use for downloading video or audio files, YouTube engagement data (likes, comments), or private\u002Fage-restricted videos.","https:\u002F\u002Fgithub.com\u002Fharesh-sai06\u002FSapSynk-AI\u002Fblob\u002Fmain\u002Fskills\u002Fclaude__youtube-full-2\u002FSKILL.md",[],[],"2026-11-15 14:19:17",{"skillNo":65,"name":66,"nameZh":6,"description":67,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":68,"categories":69,"keywords":70,"publishedAt":63},"01M77GPXXAB3Y8ZNG919MD4703","x-twitter-growth","X\u002FTwitter growth engine for building audience, crafting viral content, and analyzing engagement. Use when the user wants to grow on X\u002FTwitter, write tweets or threads, analyze their X profile, research competitors on X, plan a posting strategy, or optimize engagement. Complements social-content (generic multi-platform) with X-specific depth: algorithm mechanics, thread engineering, reply strategy, profile optimization, and competitive intelligence via web search.","https:\u002F\u002Fgithub.com\u002Fharesh-sai06\u002FSapSynk-AI\u002Fblob\u002Fmain\u002Fskills\u002Fclaude__x-twitter-growth-2\u002FSKILL.md",[],[],{"skillNo":72,"name":73,"nameZh":6,"description":74,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":63},"01M77GPSS7AFYH90Y5S6XHRXZW","write-a-skill","Create new agent skills with proper structure, progressive disclosure, and bundled resources. Use when user wants to create, write, build, or author a new skill.","https:\u002F\u002Fgithub.com\u002Fharesh-sai06\u002FSapSynk-AI\u002Fblob\u002Fmain\u002Fskills\u002Fclaude__write-a-skill-2\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":63},"01M77GPKB6HJS2X5XK03649C6R","workflow-builder","Design and write deterministic multi-agent workflow scripts (.js files in .claude\u002Fworkflows\u002F) for Claude Code's Workflow tool. Use when a user wants to build, create, author, scaffold, or run a custom Claude Code workflow, orchestrate sub-agents (fan-out, pipeline, loop, judge-panel), or automate a repeatable multi-step task across fresh-context agents.","https:\u002F\u002Fgithub.com\u002Fharesh-sai06\u002FSapSynk-AI\u002Fblob\u002Fmain\u002Fskills\u002Fclaude__workflow-builder-2\u002FSKILL.md",[],[],{"skillNo":86,"name":87,"nameZh":6,"description":88,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":89,"categories":90,"keywords":91,"publishedAt":63},"01M77GP31EFXE1KS38XBZ1J9MN","webinar-marketing","When the user wants to plan, promote, run, or improve a webinar or virtual event to generate and convert demand. Use when the user mentions 'webinar,' 'virtual event,' 'online event,' 'live demo,' 'virtual summit,' 'workshop,' 'masterclass,' 'fireside chat,' 'roundtable,' 'registration funnel,' 'show-up rate,' 'attendance rate,' 'webinar promotion,' 'webinar follow-up,' or 'on-demand webinar.' Also use when they have a webinar that isn't converting — low registrations, low show-up, or attendees who don't buy — and want to diagnose and fix it. Covers the full funnel: registration, promotion, show-up, live engagement, live-to-close, and post-event nurture. Distinct from launch-strategy (full product launches) and email-sequence (lifecycle nurture) — this is the end-to-end webinar\u002Fevent motion. NOT for in-person field events logistics, and NOT for generic lifecycle email (use email-sequence).","https:\u002F\u002Fgithub.com\u002Fharesh-sai06\u002FSapSynk-AI\u002Fblob\u002Fmain\u002Fskills\u002Fclaude__webinar-marketing-2\u002FSKILL.md",[],[],{"skillNo":93,"name":94,"nameZh":6,"description":95,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":96,"categories":97,"keywords":98,"publishedAt":63},"01M77GNNA22HWXX6CEQQ3JY2YB","vpe-advisor","VP of Engineering advisory for startups: delivery throughput (DORA 4 metrics + bottleneck identification), engineering hiring funnel (sourcing → screen → onsite → offer conversion + time-to-fill + pipeline gap), engineering team structure (squad\u002Ftribe\u002Fchapter design + tech-lead manager-trigger thresholds), and production discipline (on-call, deployment cadence, postmortem culture). Use when sprint velocity is dropping, eng hiring is broken, team structure is unclear, or deciding when to add a tech-lead manager. NOT a CTO skill (which owns architecture) — VPE owns delivery operations and how the team ships.","https:\u002F\u002Fgithub.com\u002Fharesh-sai06\u002FSapSynk-AI\u002Fblob\u002Fmain\u002Fskills\u002Fclaude__vpe-advisor-2\u002FSKILL.md",[],[]]