@prismer/runtime 2.0.7 → 2.2.55

package/README.md

@@ -12,7 +12,7 @@ The runtime opens **one WebSocket** to cloud, declares all hosted agents in a si
 ### macOS (LaunchAgent autostart)
 
 ```sh
-curl -fsSL https://prismer.cloud/install.sh | sh -s -- \
+curl -fsSL https://prod.docbrew.cn/install.sh | sh -s -- \
   --component=daemon --adapters=claude-code,codex,hermes
 ```
 
@@ -41,20 +41,26 @@ Without LaunchAgent, you own daemon lifecycle (run under tmux / launchd / system
 
 ### Container (daemon-first sandbox image)
 
-The runtime is PID 1 of every Prismer sandbox pod. `infra/sandbox-image/daemon-entrypoint.sh` writes a flat `~/.prismer/config.toml` from controller-injected env, then `exec prismer daemon start --port=7878 --foreground`. Required env:
+`tini` starts the resident Rust `sandbox-manager`. The manager exposes its
+health endpoint before asking the dependency-free image bootstrapper to install
+a verified signed bundle. That bundle contains Runtime, Cloud SDK/CLI, and AIP;
+the manager prepends its `node_modules/.bin` and then supervises
+`node <bundle>/dist/cli.js daemon start --port=7878 --foreground`. A fresh Pod
+with no verified bundle stays manager-healthy with `daemonAlive:false` and
+retries instead of falling back to image-builtin JavaScript. Required env:
 
 - `PRISMER_API_KEY` — required.
 - `PRISMER_DAEMON_ID` — required (else derived from hostname → `container:<podName>`).
-- `PRISMER_BASE_URL` (or `CLOUD_API_BASE`, `PRISMER_CLOUD_API_BASE`) — cloud URL; falls back to `https://prismer.cloud`.
+- `PRISMER_BASE_URL` (or `CLOUD_API_BASE`, `PRISMER_CLOUD_API_BASE`) — cloud URL; falls back to `https://prod.docbrew.cn`.
 - `PRISMER_HOSTED_AGENT_FILE` / `PRISMER_HOSTED_AGENT_JSON` — static-binding payload (one-time agent install before first `host.declare`).
 - `PRISMER_STATIC_BINDING_REQUIRED=true` — make missing static binding fatal.
 - `PRISMER_USER_ID`, `PRISMER_TASK_ID`, `PRISMER_CONTAINER_ID` — audit-only.
 
 The container daemon binds `:7878` (controller proxy target) instead of the desktop default `:3210`, and starts `OutboxWatcher` to upload files dropped under `/workspace/_outbox/` as sandbox-output assets.
 
-## CLI reference (16 verbs)
+## CLI reference
 
-`VERSION = '2.0.0'`. Each verb is a `commander` subcommand under `src/cli/commands/`. Output is JSON (`--json`) or human-friendly ANSI by default.
+`VERSION = '2.2.4'`. Each verb is a `commander` subcommand under `src/cli/commands/`. Output is JSON (`--json`) or human-friendly ANSI by default. Repository environment diagnostics and test orchestration are exposed by the separate APC binary (`npm run apc`), not by the daemon runtime or general SDK.
 
 ### Lifecycle
 
@@ -176,8 +182,8 @@ Pure HTTP client of `hermes gateway` (Hermes 0.10+). All calls are HTTP — no `
 
 - `port` (1..65535, default `8642`), `apiKey` (required, from `~/.hermes/.env API_SERVER_KEY`).
 - `hermesProfileName` (default = first 8 chars of `profile.id`), `autoStart` (default `false`), `startupTimeoutMs` (default 30 000).
-- `configurePrismerProvider` (default `true`) — writes `custom_providers` + `model.provider='custom:prismer'` + `mcp_servers['prismer-tasks']` block into `~/.hermes/profiles/<name>/config.yaml`. The MCP server path is auto-resolved via `createRequire('@prismer/mcp-server')` first, then a repo-relative dev fallback.
-- `installPrismerMcpServer` (default `true`), `prismerMcpServerPath` (override).
+- `configurePrismerProvider` (default `true`) — writes `custom_providers` + `model.provider='custom:prismer'` into `~/.hermes/profiles/<name>/config.yaml` (plus the `mcp_servers['prismer-tasks']` block only when `installPrismerMcpServer: true`). The MCP server path is auto-resolved via `createRequire('@prismer/mcp-server')` first, then a repo-relative dev fallback.
+- `installPrismerMcpServer` (default `false` — retired for first-party agents 2026-07-19, skill+CLI is primary; explicit third-party opt-in only), `prismerMcpServerPath` (override).
 - `model` (default `us-kimi-k2.5`), `prismerProviderName` (default `prismer`), `prismerProviderBaseUrl` (default `$PRISMER_BASE_URL/api/v1`), `prismerApiKeyEnv` (default `PRISMER_API_KEY`, regex `^[A-Za-z_][A-Za-z0-9_]*$`).
 - `mirrorNativeKanban` (default `true`) — mirrors Prismer `work_item` projections into Hermes Kanban as triage cards via `hermes -p <profile> kanban create … --triage --idempotency-key prismer:<parentTaskId>`.
 - `mirrorNativeGoals` (default `true`) — writes Prismer standing objectives into Hermes' SQLite `state_meta["goal:<sessionId>"]` directly.
@@ -196,6 +202,20 @@ Dispatch flow:
 
 `claude --print --model <m> [--system-prompt …] [--allowed-tools …] <prompt>` per dispatch. `--cwd` removed (uses `child_process.spawn` cwd); prompt is positional; `--headless` was renamed to `--print` in claude CLI 2.x. Config: `cwd` (required), `model` (default `'sonnet'`), `systemPrompt`, `envVars`, `mcpServers`, `allowedTools`, `maxTurns` (default 20), `baseURL`, `apiKeyRef` (`env:NAME` or `keychain:NAME` — keychain only on darwin via `security find-generic-password`), `route` (`'default'|'prismer'|'omniroute'`). `stdio: ['ignore', 'pipe', 'pipe']` to close child stdin (claude ≥ 2.1.128 prints "no stdin data received in 3s, proceeding without it" otherwise). `task.signal` → `child.kill('SIGTERM')`; `task.timeoutMs` → SIGTERM via `setTimeout`.
 
+The engine spawn path requires the exact Claude CLI `binaryPin` in
+`src/adapters/known-versions.ts` (currently `2.1.179`, aligned with the sandbox
+image). A mismatch raises `ADAPTER_BINARY_PIN_VIOLATION` before the SDK query
+starts. `PRISMER_ALLOW_UNPINNED=1` is an explicit unverified escape hatch.
+
+Claude Code runs with isolated `HOME` and `CLAUDE_CONFIG_DIR` rooted at
+`~/.prismer/claude-config`, preventing user plugins, MCP servers, and settings
+from entering daemon runs. Set `PRISMER_CC_NO_ISOLATION=1` only to opt back into
+the legacy user config. **Not yet verified end-to-end**: the sentinel spawn
+oracle (seed a marker in the real `~/.claude`, spawn, assert it is NOT loaded;
+negative control = `PRISMER_CC_NO_ISOLATION=1` must load it) does not exist. The
+current tests are pure-function assertions over the path resolver and never
+spawn a process, so "isolation holds at spawn time" is unproven.
+
 ### `codex`
 
 `codex exec --model <m> --cd <cwd> --sandbox <level> --ephemeral --json <prompt>`. Config: `cwd` (required), `model` (default `codex-mini-latest`), `sandbox` (`'read-only'|'workspace-write'|'danger-full-access'`, default `workspace-write`), `systemPrompt` (prepended to user prompt — Codex CLI has no `--system-prompt` flag as of 2026-05), `envVars`, `apiKeyEnv` (default `OPENAI_API_KEY`). `parseCodexOutput` reads JSONL stdout backwards looking for `{"type":"message"|"assistant"|"output", "content"|"message"|"text": string}`; falls back to raw stdout. Outputs are capped at 64 KiB with `\n…[truncated]` suffix.
@@ -317,7 +337,7 @@ Operators run `hermes` in their own Python venv. Sample `~/.hermes/profiles/<nam
 custom_providers:
   prismer:
     type: openai
-    base_url: https://prismer.cloud/api/v1
+    base_url: https://prod.docbrew.cn/api/v1
     api_key_env: PRISMER_API_KEY
 model:
   provider: custom:prismer
@@ -355,9 +375,9 @@ Bridge results are persisted on `IMTask.metadata.bridge.hermes` via the daemon's
 | Method | Path | Purpose |
 | ------ | ---- | ------- |
 | OPTIONS | `*` | CORS preflight |
-| GET | `/healthz` | `{status:'ok', daemonId, cloudBaseUrl, workspaceId, pid, startedAt, wsConnected, hostedAgents, observability}` |
+| GET | `/healthz` | `{status:'ok', daemonId, cloudBaseUrl, workspaceId, pid, startedAt, wsConnected, hostedAgents, observability}`. Additive, **present only when the runner wires the subsystem** (spread — CLI/K8s shape unchanged): `assetOutbox {pending, uploaded, deadLetter, sampledAt}` (desktop205 W13; a SAMPLE pushed in by the drop-folder tick — this handler does zero I/O and the desktop Tray polls it), `recallStats`, `cloudReachable`/`pendingWrites`. |
 | GET | `/agents` | `{agents: [{imUserId, name, adapterName}]}` |
-| GET | `/tasks/running` | `{taskIds: string[]}` |
+| GET | `/tasks/running` | `{taskIds: string[]}` + optional `tasks: [{taskId, agentName?, kind?, scopeLabel?, startedAt}]` (desktop205 W13 — descriptive rows for the desktop Tray; `taskIds` is unchanged for every existing consumer) |
 | POST | `/v1/runs` | Sandbox controller daemon-dispatch ack — `202 {runId, status:'accepted', taskId}`. Optional `payload.shellCommand` triggers async `bash -c <cmd>` with `cwd=/workspace` (escape hatch). |
 | POST | `/v1/agents/install` | **Cloud → daemon agent install RPC**: validates body via `validateInstallAgentPayload`, calls `runner.installHostedAgent` → upsert `agents` + `agent_profiles` rows, reload, send `agent.host.declare`, return `{ok, daemonId, installedAgent, hostedAgents}`. |
 | POST | `/v1/snapshot` | FS-manifest snapshot — walks `snapshotRoot` (default `/workspace`), computes per-file `{path, sha256, sizeBytes, mtime}`. Skips dotfiles + `_outbox/_uploaded`. 404 if root missing. |
@@ -378,6 +398,8 @@ The sandbox-controller's `installAgent` proxy at `POST /:id/installAgent` valida
 | `PRISMER_RUNTIME_MODE` | (empty) | `container` activates `OutboxWatcher` (alt to `PRISMER_CONTAINER_ID`) |
 | `PRISMER_DAEMON_URL` | `http://127.0.0.1:3210` | Loopback URL used by `prismer memory` and `agent doctor` |
 | `PRISMER_SHELL_ENABLED` | `false` | Override `[shell].enabled` in `config.toml` |
+| `PRISMER_ALLOW_UNPINNED` | (empty) | `=1` bypasses exact Claude CLI binary pin enforcement; behavior is unverified |
+| `PRISMER_CC_NO_ISOLATION` | (empty) | `=1` lets Claude Code inherit the user's real `HOME` / `~/.claude` instead of the isolated runtime config |
 | `LOCAL_ONLY` | (empty) | `=1` enables `--as-user` pair bypass |
 | `PRISMER_AUTH_TOKEN` | (empty) | Fallback for `prismer setup --token <jwt>` |
 | `EDITOR` | `vi` | Used by `prismer profile edit` |
@@ -386,7 +408,7 @@ The sandbox-controller's `installAgent` proxy at `POST /:id/installAgent` valida
 
 ```toml
 api_key = "sk-prismer-..."
-cloud_api_base = "https://cloud.prismer.dev"
+cloud_api_base = "https://test.docbrew.cn"
 daemon_id = "..."
 
 [adapters.claude-code]
@@ -405,7 +427,7 @@ shell = "bash"
 
 ## Troubleshooting
 
-- **LaunchAgent respawn loop after install** — the plist must include `--foreground` in `ProgramArguments`. v2.0.0 installs do this automatically; if you're upgrading from 1.8.x re-run `curl -fsSL https://prismer.cloud/install.sh | sh -s -- --component=daemon` (the installer does `bootout` + `bootstrap` to reload).
+- **LaunchAgent respawn loop after install** — the plist must include `--foreground` in `ProgramArguments`. v2.0.0 installs do this automatically; if you're upgrading from 1.8.x re-run `curl -fsSL https://prod.docbrew.cn/install.sh | sh -s -- --component=daemon` (the installer does `bootout` + `bootstrap` to reload).
 - **Daemon shows "waiting for ~/.prismer/config.toml"** — that's expected behavior pre-pair. Run `prismer pair` (or `prismer setup sk-prismer-...`) and the daemon snaps to attention within 5s.
 - **Cloud unreachable** — `prismer status` will print the failing base URL. Common causes: stale `cloud_api_base` after a tenant move (run `prismer setup --force <new-key> --cloud <url>`), or the API key was rotated (mint a new one with `prismer setup --token <jwt>`).
 - **Agent shows under `__unbound__` device in workspace runtime view** — the daemon's `agent.host.declare` is not stamping `daemonId`. Confirm you're on `@prismer/runtime@2.0.0` (the `daemonId` payload field is required as of v2.0.0).

package/apc/skills/FIELD-DICTIONARY.md

@@ -0,0 +1,111 @@
+<!-- APC skill authoring 参照：命令真产物的字段字典 -->
+# APC skill 真产物字段字典
+
+> **为什么存在**：M1-d/f 探路批（env-doctor + test-runner）4 轮返工里 **2 轮死在字段名漂移**——SKILL.md 照 doc 12 的抽象词（"结构化 TierResult" / "doctor JSON"）写，agent 照着解析真产物却对不上字段名。**写任何引用这些产物的 SKILL.md 前，字段名以本文件为准**（本文件由真跑命令抓取，非文档转述；schema 变了先更新这里）。
+>
+> 抓取方式（可复现）：`npx tsx sdk/apc/bin/apc.ts env doctor` / `npx tsx scripts/test203/run.ts --tier=T0 --json`。
+
+## `apc env doctor` → stdout JSON
+
+顶层键：`envStatus` · `exitCode` · `generatedAt` · `items[]` · `summary` · `failed[]` · `undetected[]`
+
+- `exitCode`：`0`=全绿 / `78`=env_blocked（**这是权威的 env 判据，不是自己数 items**）
+- `envStatus`：`'ok'` | `'env_blocked'`
+- `summary`：`{ pass, fail, skip, total }`（`total==pass+fail+skip==items.length`，一项崩不塌全轮）
+- `failed[]`：失败项的 **id 字符串**数组（= 各 item 的 `item` 字段值）
+- `items[]` 每项：`item`（**id，如 `toolchain.node`**）· `section`（**故障域：infra/toolchain/secrets/project**，分类按这个）· `status`（`'pass'|'fail'|'skip'`）· `label` · `detail` · `fixHint` · `strength` · `durationMs`
+
+⚠️ 字段名是 **`item` / `section` / `status`**，不是 `name` / `category`（首版踩过）。
+
+## `apc test`（= run.ts）`--json` → stdout JSON
+
+顶层键：`doctor` · `envStatus` · `exitCode` · `tiers[]` · `regressions[]` · `fixed[]` · `schema` · `timestamp`
+
+- `doctor`：内嵌一份 doctor 段（同上结构）
+- `exitCode`：`0`=通过 / `1`=SUT 红（`--diff` 下=有回归）/ `78`=env_blocked
+- `regressions[]` / `fixed[]`：`--diff` 下 vs baseline 的**新增红 / 转绿**（顶层聚合，跨所有 tier）
+- `tiers[]` 每项（TierResult）：`tier` · `passed` · `failed` · `skipped` · `total` · **`failedNames`**（失败用例名数组）· `skippedNames` · `regressions` · `envStatus` · `durationMs`
+
+⚠️ 字段名是 **`failedNames`**（驼峰，非 "failed names"）；`command`/`exitCode` 在**顶层**不在每个 tier 内（首版踩过）。
+
+### ⚠️ 并行 flaky：单次 regression ≠ 真回归（2026-07-24 主会话查实）
+
+T0 全量并发跑会偶发 flaky（资源竞争/端口/超时），**每次红的集合都不同**（探路批见 5 条、复查见 1 条 `web-rpc.test.ts`、再跑见 0 条；`web-rpc` 单独跑 10/10 绿）。baseline（26 条，均既有 ACP mock 债）**未被污染**。
+
+⇒ **test-runner skill 的验收 oracle 不能把"单次 `apc test` 出现 regression"直接判 skill 失败**。判据应是：
+- regression 命中 baseline 已记的既有红 → 非新增，正常；
+- regression 是新面孔 → **单独重跑该文件**确认；单跑绿=flaky（记录不 fail skill），单跑仍红=真回归（如实 `--outcome failed` 上报，是被测栈的红不是 skill 的红）。
+- flaky 的签名 = 重跑红的集合漂移；真回归的签名 = 稳定复现同一批。
+
+## `cloud task verify-criterion <task-id> <criterion-id> --outcome <v>`
+
+- `--outcome` 合法值**恰好四个**：`passed` | `failed` | `n/a` | `waived`（`task.ts:1374-1394`）
+- 副作用 oracle：criterion 从 `pending` → 上报值，落 acceptance-view（读回确认）
+
+## `cloud skill ack <slug> --task <taskId>` 退出码
+
+`0`=回执落库（`im_task_logs.action='skill_ack'`）/ `3`=无 task 上下文（**产不出回执，非失败；不许 `||true` 吞掉，按 3 分流**）/ `4`=调用身份≠task assignee / `1`=其它。
+
+## 端到端达标配方（env-doctor 2026-07-24 真跑 `ok:true` 验证过的黄金样板）
+
+`cloud skill test <bundle> --agent <id> --json` 要拿 `ok:true`，skill.json 必须满足这些——**照 `sdk/apc/skills/env-doctor/skill.json` 抄**：
+
+1. **命令固化进 `sampleTasks[].prompt` 本体**：到达 agent 的是 prompt，**不是 SKILL.md**（SKILL.md 不进 dispatch 上下文）。prompt 里每条命令必须是 agent 在 cwd 真能跑的形态：
+   - **`apc` 不在 PATH** → 一律写 `npx tsx sdk/apc/bin/apc.ts <sub>`（不是裸 `apc`，不是 `npm run apc`）。
+   - 涉 `apc env up` → **强制 `--safe`**（bare `env up` 的 docker 步 `colima start` 在非交互 dispatch 里 block 数分钟 → reaper 杀）。prompt 里点明这一点。
+2. **`acceptanceCriteria` 必须机器可判 + 不可蒙混**：
+   - 用 `{label,match,type:"regex",required}` 对象，**不要**纯字符串（纯字符串走大小写敏感 substring = 要求 agent 逐字复述人话 → 假阴）。
+   - `match` 匹配 **agent 自然汇报里稳定出现的机器特征**（exit 码、`pass/fail/skip` 计数、字段值），**不是**要求贴原始 JSON。
+   - 正则要吃**多种合法排版**（inline `pass=15` / 表格 `| pass | 16 |` / dash）——否则同样正确的工作因排版一绿一红（脆性）。参考 env-doctor count 那条：`(pass|fail|skip)\w*\s*[:=|\-]*\s*\d+|\d+\s*[|:=\-]*\s*(pass|fail|skip)`。
+   - **必须自测负控**：蒙混文本（"looks fine to me"）喂 `matchAcceptanceCriteria` 要 `ok:false`；真汇报要 `ok:true`。放松到蒙混能过 = 作废。
+3. **写完立即端到端验**（daemon 活着时）：`cloud skill test <bundle> --agent <id> --json` 看 `ok:true` + `taskStatus:completed`；再造蒙混 prompt 的临时 bundle 验 `ok:false`。**这是达标定义，不是可选。**
+
+> harness 侧已修（`extractAgentReply` type 过滤 + 反向取最新 + `afterIso` 锚，只认比 prompt 发送时刻更新的 agent 回复）——authoring 不用管，但别写出依赖"首轮 poll"的假设。
+
+## 平台字段陷阱（authoring 必读）
+
+- **`compatibility` ≠ `scope`**：`compatibility` 是枚举，合法值 `claude-code/codex/opencode/openclaw/prismer-sdk`——**不接受 `hermes`**（写进去 422 `STD_COMPATIBILITY_INVALID`）。hermes 可用性由 `scope: common` 授予。
+- **⚠️ SKILL.md frontmatter 的 `scope:` 是描述性的、代码零消费**（2026-07-24 查实：`grep fm.scope` 全仓无命中）。真正决定"谁能装这个 skill"的 install-side 角色门是 **`role_scope`**（`built-in-skill.service.ts:354/467`），当前唯一约束值 = `orchestrator-only`（只装给 orchestrator）。约定（照 built-in `memory-dream`）：**两个都写** —— `scope: common|coding` 给人读，`role_scope: orchestrator-only` 仅在该 skill 真要限制为 orchestrator-only 时加。
+  - **APC coding/common skill 默认不写 `role_scope`**：common 装给所有 agent（正确）；coding 的限定靠 `platform-install` 的 agent 选择，**不是靠 frontmatter**。别以为写了 `scope: coding` 就把非 coding agent 挡在外面——那个字段没人读。若将来要"coding skill 只给 coding-role agent"的硬门，需在 `built-in-skill.service.ts` 扩 `role_scope` 值域（服务端改，不是 authoring）。
+- **skill.json 无 `scope` 字段**：真实字段 = `schemaVersion/slug/name/description/category/version/license/compatibility/runtime/sampleTasks/security/provenance`。发布可见性（private/workspace/public）由 `platform-install` 的 `publishScope` 承载，不在 skill.json。
+- **catalog slug 派生**：`platform-install` 后服务端 `toSlug(name,'community')` 把 slug 变成 `community-<name>`，与 bundle 目录名分叉——装载/查询按**返回的 slug**，不是 bundle 目录名。
+- **双身份债**：同一 API key 在 `/api/im/me`、skills-install actor、task-event actor 可能解析成**不同 imUserId**（P0 验签债 `api-guard.validateJwt` 只解码不验签的下游症状）。task assignee 必须对齐 key 的**真** identity（task-event actor 那个），不是 me-endpoint 返回的。凡 coding-scope skill 涉 task 归属都会踩。
+
+## 发版 5 verb 真跑 `--json` 字段字典（§0.25 抓取，供发版 skill authoring）
+
+`apc release <verb> --json` 输出形状（真跑抓取；退出码 green/staged/blocked ⇄ 0/3/1）：
+
+- **preflight** → `{decision, tier{tier,exitCode,envBlocked,regressions[]}, version{aligned,rootVersion,mismatches[]}, prisma{ok,stale[]}, blockers[], staged[]}`
+- **db-config-sync** → `{decision, migration{exitCode,pending[],stderr}, keyDiff:{status:'skipped',reason}|{status:'computed',missing[],extra[],common}, blockers[], staged[]}`（真跑：本地 replica 217 pending、无 dump 时 keyDiff=skipped ⬜）
+- **tag** → `{decision, tag, target, channel, tier{...}, approved, pushed, mirror, blockers[], notes[]}`（prod 拒时 `tier.exitCode=-1`，tag 算得出但 decision=blocked、pushed=false）
+- **ota-promote** → `{decision, env, deliveryMode, k8s:{invoked,exitCode,rolledOut,output}|{skipped,reason}, desktop:{status:'not-wired',missing[],deliveryMode}, blockers[], notes[]}`（`k8s.invoked=true` 靠 `[runtime-ota]` banner 判真 spawn；桌面 not-wired=⬜；deliveryMode=line1-kill1 非 drain_respawn）
+- **rollback** → `{decision, ledgerPath, plan:{ok:true,from,to,next[]}|{ok:false,reason}, approved, applied, mirrorRepointed, blockers[], notes[]}`
+
+**真环境事实**：`/VERSION=2.2.4`（tag 算 v2.2.4）；`prisma/schema.mysql.prisma` ⇄ generated client **真 drift** → preflight 真跑落 staged（非误报，与 §0.21 一致，需 `npm run prisma:generate:all`）。
+
+## 截图/资产 + 验收/观测 skill 字段（ui-canvas evidence + §0.33 S5b/S10 抓取）
+
+**资产上传（ui-canvas/report-card 用）**：
+- `POST /api/im/assets` 成功响应 asset id 在 **`data.id`**，不是 `data.assetId`（别处 `toAssetResponse` 用后者）。上传截图/产物一律 `data.assetId ?? data.id` 兜。
+- 主题态截图：Playwright `addInitScript` 预置 `localStorage['prismer-theme']=light|dark|system`（`src/contexts/theme-context.tsx` 挂 `.dark`/`.light` 到 `<html>`），必须在页面脚本前种。目标必须是 registry 校验过的 `/ui-canvas?set=&version=&journey=&node=`；稳定锚是 `[data-testid="ui-canvas-overview"]` 或 `[data-testid="ui-canvas-viewport"]`。
+
+**验收/观测（S5b/S10 用）**：
+- `cloud task verify-criterion <task-id> <criterion-id> --outcome passed|failed|n/a|waived [--evidence <ref...>可重复 --note]`；`--evidence` 落 `criterion.evidence[]`（`taskRun:`/`asset:`/`url:`），`cloud task acceptance` 回读字段名 `evidence`；**任一 required criterion=failed → overall=failed**（别软化真红）。
+- `cloud task add-criterion <taskId> --mode qualitative|quantitative|agent-self-check|manual --expectation <md> [--optional]` 返短 criterion id。criteria 存 **task metadata JSON 无专表**（仅 `im_task_criteria_templates`）。
+- `cloud task create --kind work_item` → kind 存 `im_tasks.metadata.kind`（JSON），`--priority`→`metadata.priority`，`--json` 返 `data.id`。
+- `scripts/debug/*`：`bundle.ts` 默认**写文件** `./debug-bundle-*.json`（打屏加 `--stdout`）；`--env=local` 走 dev admin 短路免凭据。
+- **harness `dispatch_not_created`（§0.29 修后的失败态）= 活 daemon 未为本 prompt 建 run（领养/路由层），非 criteria 失败**——真判据永远是 `im_task_runs` 真行,不是 skill-test output。
+
+### 结构化判据 checker（`type: "structured"` 的 `checker` 取值）
+
+| checker | 真值来源 | 备注 |
+| --- | --- | --- |
+| `cited-evidence` | **磁盘**：每个 `path:line` 读回复核（存在/行范围/非空行/锚定命中） | 层 2 |
+| `dimension-coverage` | **磁盘**：每维独占一行，答案是可复核引用或 `N/A — 理由` | 层 2；N/A 只校验长度不校验理由为真 |
+| `doc-sync-obligations` | **磁盘**：从 delta 自行重算义务集与 gap 集再比对 | 层 2 |
+| `json-claim` | **命令产物**：重解析 fenced JSON，钉 schema/算术/派生判定/baseline 重算/产物 mtime 新鲜度 | 层 3 |
+| `declared-id-readback` | **云端真 API**：`TASK:/ASSET:/CRITERION:/META:` 声明的 id 回读比对（含存储 JS 类型） | 层 3 |
+| `git-claim-readback` | **真 git 只读 plumbing**：commit 文件集/branch tip/bare 替身/`MERGE_HEAD` 仍在 | 层 3；硬拒 `.e2e-tmp/` 之外的仓库 |
+
+> **层 2 = 复核报告关于「文件」的声称；层 3 = 复核关于「运行时副作用」的声称。**
+> 共同边界：核的是**副作用存在性**不是**通道**；杀的是廉价伪造，不提供不可伪造性。

package/apc/skills/bug-reproduce/SKILL.md

@@ -0,0 +1,150 @@
+---
+name: bug-reproduce
+description: Turn a known bug into a tight, red-capable reproducer, then prove the reproducer locks that bug with a reversal negative control (reintroduce the bug → reproducer must go red again). Use at the start of a bugfix, before theorizing about the cause.
+license: MIT
+scope: coding
+compatibility:
+  - claude-code
+allowed-tools:
+  - Bash
+metadata:
+  category: diagnosis
+---
+
+# bug-reproduce
+
+把一个**已知 bug** 变成**紧、可变红**的复现器（失败测试 / 脚本），然后用**反转负控**证明这个复现器真的锁住了那个 bug——**不是恰好绿着**（`apc/05` §2 S1 · `apc/12` bug-reproduce 行；纪律移植自 built-in `diagnosing-bugs` 的 Phase 1 复现 loop）。
+
+**什么时候用**：一个 bugfix task 开工的**第一步**——在读代码建假设之前。`diagnosing-bugs` 的铁律：没有一个能对**这个** bug 变红的信号，再多盯代码也没用。本 skill 只补 `diagnosing-bugs` 缺的那一环：**负控——把 fix 撤掉/把 bug 注入回去，复现器必须变红**。
+
+铁律：一个复现器只有走完 **RED → GREEN → 反转 RED** 才算数。**只到 GREEN 不算**——一个无论 bug 在不在都绿的"复现器"是开卷考试，不是复现（`apc/12` §0 负控纪律）。
+
+## 复现器四判据（照 diagnosing-bugs Phase 1 完成标准）
+
+一个复现器必须是：
+
+| 判据 | 含义 |
+| --- | --- |
+| **red-capable** | 它驱动**真正的 bug 代码路径**、断言**用户描述的确切症状**——能对这个 bug 变红，修好后变绿。不是"没崩就算过"。 |
+| **deterministic** | 同一输入每次同一结论（非确定 bug：钉时间/种 RNG/隔离 FS，把复现率拉到可调试）。 |
+| **fast** | 秒级不是分钟级。 |
+| **agent-runnable** | 你能无人值守跑；一条命令（脚本/测试/curl）能贴出调用 + 输出。 |
+
+## Procedure
+
+### 1. 钉环境 + 建复现器（RED）
+
+在一个干净的、确定性的 workspace 里写出**最小失败测试/脚本**，然后跑它，**看它变红**：
+
+```bash
+d=.e2e-tmp/apc/bug-reproduce && rm -rf "$d" && mkdir -p "$d"    # repo-relative，已在 .gitignore
+# …在 "$d" 里写 subject + repro…
+node "$d"/repro.mjs; echo "exit=$?"     # 或 apc test --tier=<T> --json / pytest / go test …
+```
+
+**隔离场景的临时目录用 repo-relative 的 `.e2e-tmp/…`，不要用 `mktemp -d`**：复现器**本身就是证据**，判据在你跑完之后把你引用的 `path:line` **读回磁盘**——`/tmp` 从仓库根不可解析、收尾 `rm -rf` 更是把唯一产物销毁（test204「证据销毁反 pattern」的同一个坑）。本地真仓库里的 bug 则直接在真路径上做。
+
+- 复现器必须断言**用户描述的症状本身**（`clamp(-5,0,10)` 应得 0 而非 -5），不是"没抛异常"。
+- 记下**确切症状**：断言消息、错误、非零退出码。这就是 RED。
+- 真仓库里的 bug：复现命令可以是选层的 `apc test --tier=<T> --json`（读 `exitCode`：`0` 绿 · `1` SUT 红 · `78` env_blocked）。**env_blocked 不是复现——是环境红，先 `apc env doctor`**。
+
+### 2. 最小 fix（GREEN）
+
+打**最小**修复（只改让复现器变绿所需的那一点），重跑复现器，**看它变绿**（`exit=0`）。**不顺手清理别的**——本步只证明"这个 fix 让这个复现器绿"。
+
+### 3. 反转负控（承重——RED again）
+
+**这是本 skill 的命根子。** 把 fix 撤掉——把 bug 注入回去，**重跑同一条复现器命令**：
+
+```bash
+# 首选：直接编辑源文件，把 buggy 的那一行 body 原样写回去
+node "$d"/repro.mjs; echo "exit=$?"     # 必须再次 RED（非零退出 / AssertionError）
+```
+
+**反转手法只有一个首选：直接编辑源文件把 bug 写回去。**
+
+> ⚠️ **`git stash` / `git checkout` 在共享工作树里是有害指令，默认不许用。** 本仓库常态是**多个并发 session 共用一个 working tree**、几百个未提交改动同时在飞——`git stash` 会把**别人正在做的改动**一起卷走，`git checkout <file>` 会**丢掉**别人对同一文件的未提交编辑。这类操作不可逆地毁坏他人工作，代价远大于它省下的那一次手工编辑。
+>
+> 只有在你**独占**这个 working tree（自己的 worktree / 一次性容器）**且已确认 `git status` 干净**时，`git stash` 才是等价手法；否则一律手工反转。反转的**语义**（把 bug 放回去）才是承重的，用什么手法不是。
+
+- **复现器必须再次变红。** 这证明它**红在这个 bug 上**，不是恒绿的摆设。
+- **若撤掉 fix 后复现器仍然绿 → 复现器是假的，任务失败**（`apc/12` bug-reproduce 负控）。不许放松断言让它"看起来锁住了"。
+
+### 4. 恢复 + 收尾（**不要删证据**）
+
+把 fix 放回去，确认最后一次 **GREEN**（`exit=0`）。**复现器留在盘上别删**——判据是在你跑完之后才读盘复核你引用的行的，`rm -rf` 掉复现器 = 自毁唯一产物。
+
+### 5. 写 BugfixCard（`repro` 阶段）
+
+复现 + 反转都拿到证据后，把这一阶段落成一张 **BugfixCard**——`repro` 是 BugfixCard 的**首阶段**，段锚是 `<h2 id="stage-repro">`（`src/lib/apc-cards/card-profile.ts` 的 `BUGFIX_STAGES[0].name = 'repro'`）。
+
+> ⚠️ **段名是 `repro`，不是 `reproduce`。** `docs/apc/12` 那行写的 `stage-reproduce` 在实现里**不存在**，写成它会被 `card-stage-invalid` / `card-stage-section-missing` 挡下。以代码为准。
+
+- `card` frontmatter：`{ kind: 'bugfix', stage: 'repro', taskId: <真 task id>, schemaRev: 1 }`——`taskId` 必填，**无 task 上下文时不要伪造绑定**：如实报告"本轮没有 task，卡未写"，别硬造一张。
+- `repro` 阶段的证据要求是 `asset`：段内要有 `prismer://…/asset/<id>` 链接，挂**复现器本体 + RED 输出 + 反转 RED 输出**三份证据（先 `cloud asset upload` 拿 id）。
+- 段序检查只对**已过阶段**要求证据，当前阶段进行中不强求；但首阶段段锚 `stage-repro` 必须在场。
+
+## 输出契约（机器判据按这个复验，别自由发挥格式）
+
+旧判据全是关键词匹配（"报告里出现过 `node`/`assert`/`exit=1`/`revert` 这些词"）——**一篇一条命令都没跑过的报告照样满分**。现在多一条判据（`cited-evidence`）把复现器**读回磁盘**复核：复现器不存在、引用的行不存在、或者引用的行根本不含被测符号，当场判红。
+
+**① 复现器锚点声明**（一行；符号会被复核确实出现在该文件里）：
+
+```
+CHANGE-POINT: <被测符号> @ <复现器/被测文件的 repo-relative 路径>
+```
+
+**② RED→GREEN→反转 RED 相位表**（markdown 表，**首列是 `path:line`**）：每相位一行，首列是该相位**转在哪一行**上。判据要求 **≥3 行**带可复核引用，且 **≥2 条引用所在的那一行真的含被测符号**。
+
+```
+| path:line | phase | command | exit | observed |
+| --- | --- | --- | --- | --- |
+| .e2e-tmp/apc/bug-reproduce/repro.mjs:3 | 1 RED (buggy) | node .e2e-tmp/apc/bug-reproduce/repro.mjs | 1 | AssertionError: clamp must respect lower bound |
+| .e2e-tmp/apc/bug-reproduce/clamp.mjs:1 | 2 GREEN (fixed) | node .e2e-tmp/apc/bug-reproduce/repro.mjs | 0 | REPRO OK |
+| .e2e-tmp/apc/bug-reproduce/repro.mjs:3 | 3 REVERSAL RED | node .e2e-tmp/apc/bug-reproduce/repro.mjs | 1 | AssertionError: clamp must respect lower bound |
+```
+
+**行号照抄 `cat -n` / `rg -n` 的真实输出**，不许凭印象写。引用必须 **repo-root-relative**（绝对路径与 `/tmp/...` 一律判红——这也是上面要求复现器落 `.e2e-tmp/` 且不删的原因）。产物存**可引用的扩展名**：引用解析器只认固定后缀集（`ts/tsx/js/jsx/mjs/cjs/prisma/sql/md/json/sh/py/yml/yaml`），存成 `.txt`/`.log` 的引用不会被解析。
+
+## 产出（副作用 oracle，报告里必须给）
+
+1. **复现器命令**：你**真跑过**的那一条（`node repro.mjs` / `apc test --tier=…` / pytest…），可贴调用。
+2. **初始 RED**：buggy 代码下复现器失败（AssertionError + 非零退出码）。
+3. **fix 后 GREEN**：`exit=0`。
+4. **反转 RED**（承重）：把 bug 注入回去后复现器**再次变红**（非零退出码）——这是"锁住了"的唯一证明。
+5. **复现器本体留在盘上**（repo-relative 路径），锚点声明 + 相位表的每条引用都能读回。
+6.（有 task 上下文时）**BugfixCard `stage-repro` 段**：三份证据 asset link 挂上，卡可回读。
+
+**不许**：只报到 GREEN 就收工（没有反转 = 没证明锁住）；放松断言让恒绿脚本冒充复现器；把 `env_blocked`（exit 78）当作 bug 复现上报；凭推理宣称"应该能复现"而没真跑；**跑完 `rm -rf` 掉复现器**（证据必须留到判据读盘之后）；**在共享工作树里用 `git stash`/`git checkout` 做反转**（会卷走别人的未提交改动）。
+
+## 诚实边界
+
+- **本 skill 不保证任何 bug 都能建出紧复现器**：跨进程时序、真机才暴露、非确定性极低复现率的 bug，可能只能拉高复现率而非稳定复现（`diagnosing-bugs` Phase 1）。建不出就**显式停手说清楚**、列出试过什么、要人给可复现环境/捕获产物——**不许**无复现器就跳去建假设。
+- **反转负控证明的是"复现器红在这个 bug 上"，不是"这个 fix 是最终正解"**——正解由后续 fix + 评审 + test-runner 兜。
+- **共享工作树是常态**：本仓库经常多个 session 同时在同一个 working tree 上飞着几百个未提交改动。因此反转 fix **默认手工编辑源文件**；`git stash`/`git checkout` 只在你确认独占该工作树、且 `git status` 干净时才等价可用。这不是洁癖——stash 卷走别人的改动是不可逆的破坏。
+- **BugfixCard 的段名以实现为准（`repro`）**：`docs/apc/12` 里写的 `stage-reproduce` 是文档漂移，代码 `src/lib/apc-cards/card-profile.ts` 的 BugfixCard 首阶段叫 `repro`。无 `PRISMER_TASK_ID` 时 `card.taskId` 无法合法填写——**如实报告"卡未写"**，不要伪造一个绑定去凑一张卡。
+
+<!-- APC-ACK:v1 -->
+
+## 调用回执（APC 平台方运营纪律 · 自动注入，勿手改）
+
+执行本 skill 的**第一步**，先落一条调用回执：
+
+```bash
+cloud skill ack bug-reproduce --task "$PRISMER_TASK_ID"
+```
+
+**按退出码分流**（这条命令的退出码是承重信息，禁止用 `||` 兜底、`; true`、
+`set +e` 或重定向把它抹掉）：
+
+| exit | 含义 | 你要做的 |
+| --- | --- | --- |
+| `0` | 回执已落库（`im_task_logs.action='skill_ack'`） | 继续执行本 skill |
+| `3` | **无 task 上下文**——本次运行没有 task，产不出回执 | 继续执行本 skill；但本次运行**没有回执**，任何报告里都不得声称已 ack |
+| `4` | 你不是该 task 的 assignee，服务端拒绝 | 停下并上报：回执只能由执行该 task 的 agent 产生 |
+| `1` | 其它失败（网络 / 服务端） | 重试一次；仍失败则继续执行，并在结果里显式标注「回执缺失」 |
+
+回执只证明本 skill **被调度**，不证明**执行正确**——效果证明由本 skill 自己的
+acceptanceCriteria 副作用断言承担（apc/04 §2 层 1 诚实标注）。
+
+<!-- /APC-ACK:v1 -->

package/apc/skills/bug-reproduce/skill.json

@@ -0,0 +1,96 @@
+{
+  "schemaVersion": 1,
+  "slug": "bug-reproduce",
+  "name": "Bug Reproduce",
+  "description": "Turn a known bug into a tight, red-capable reproducer (failing test / script), then PROVE the reproducer actually locks that bug with a reversal negative control: reintroduce the bug and the reproducer must go red again. A reproducer that stays green when the bug is put back is a fake and fails the task.",
+  "category": "diagnosis",
+  "version": "1.0.0",
+  "license": "MIT",
+  "compatibility": [
+    "claude-code"
+  ],
+  "runtime": {
+    "kind": "text-workflow",
+    "requires": {
+      "env": [],
+      "bins": [
+        "node",
+        "git"
+      ],
+      "capabilities": []
+    }
+  },
+  "sampleTasks": [
+    {
+      "title": "Reproduce a clamp() lower-bound bug and prove the lock with a reversal",
+      "prompt": "Run the bug-reproduce skill on this known bug. BUG: a `clamp(x, lo, hi)` helper ignores its lower bound — it returns `Math.min(x, hi)` so `clamp(-5, 0, 10)` wrongly returns -5 instead of 0. Build a tight, deterministic reproducer in a throwaway dir and PROVE it locks the bug with a reversal negative control. Run these exact steps and report the observed exit codes: (1) `d=.e2e-tmp/apc/bug-reproduce && rm -rf \"$d\" && mkdir -p \"$d\"` — a clean, deterministic workspace that is REPO-RELATIVE and SURVIVES the run. Do NOT use `mktemp -d`: the reproducer you build IS the evidence and it is re-read off disk after the run, so a temp dir you delete at the end destroys exactly the artifact the acceptance re-verifies (`.e2e-tmp/` is gitignored, so this pollutes nothing); (2) write `\"$d\"/clamp.mjs` containing the BUGGY version: `export const clamp = (x, lo, hi) => Math.min(x, hi);`; (3) write the reproducer `\"$d\"/repro.mjs`: `import assert from 'node:assert'; import { clamp } from './clamp.mjs'; assert.strictEqual(clamp(-5, 0, 10), 0, 'clamp must respect lower bound'); console.log('REPRO OK');`; (4) run the reproducer command `node \"$d\"/repro.mjs; echo \"exit=$?\"` — it MUST fail (AssertionError, non-zero exit): this is the bug reproduced RED; (5) apply the minimal fix — rewrite `\"$d\"/clamp.mjs` to `export const clamp = (x, lo, hi) => Math.max(lo, Math.min(x, hi));` — and re-run `node \"$d\"/repro.mjs; echo \"exit=$?\"`: it MUST now pass (exit 0) GREEN; (6) REVERSAL NEGATIVE CONTROL (load-bearing): revert the fix by EDITING `\"$d\"/clamp.mjs` DIRECTLY back to the buggy one-line body — do NOT use `git stash` / `git checkout` for this: you are in a SHARED working tree with other sessions' uncommitted changes, and stashing would revert THEIR work too. Re-run `node \"$d\"/repro.mjs; echo \"exit=$?\"` — it MUST go RED again (non-zero). This proves the reproducer is red-capable ON THIS bug, not accidentally green; (7) restore the fix, confirm GREEN one last time, and LEAVE `.e2e-tmp/apc/bug-reproduce/` in place — do not delete it.\n\n--- YOUR REPORT MUST USE THIS MACHINE-CHECKED OUTPUT CONTRACT (it is re-verified against the filesystem, so a reproducer you never actually wrote is caught):\n\nLine 1 — the reproducer anchor, exactly this shape (the symbol is re-read off disk and must really occur in that file):\n`CHANGE-POINT: clamp @ .e2e-tmp/apc/bug-reproduce/clamp.mjs`\n\nThen the RED→GREEN→REVERSAL phase table — one row per phase you actually ran, the FIRST CELL being the `path:line` in your reproducer/subject that the phase turns on (copy the line number from `cat -n` or `rg -n` over the files you just wrote; a missing file, an out-of-range line or a blank line FAILS the report):\n```\n| path:line | phase | command | exit | observed |\n| --- | --- | --- | --- | --- |\n| .e2e-tmp/apc/bug-reproduce/repro.mjs:3 | 1 RED (buggy) | node .e2e-tmp/apc/bug-reproduce/repro.mjs | 1 | AssertionError: clamp must respect lower bound |\n| .e2e-tmp/apc/bug-reproduce/clamp.mjs:1 | 2 GREEN (fixed) | node .e2e-tmp/apc/bug-reproduce/repro.mjs | 0 | REPRO OK |\n| .e2e-tmp/apc/bug-reproduce/repro.mjs:3 | 3 REVERSAL RED | node .e2e-tmp/apc/bug-reproduce/repro.mjs | 1 | AssertionError: clamp must respect lower bound |\n```\nYou need at least 3 phase rows carrying a verified `path:line`, and at least 2 of your citations must land on a line that LITERALLY CONTAINS `clamp` — a real line that is not a real hit in the reproducer is not evidence.\n\nHARD RULE: do not guess line numbers. Every `path:line` anywhere in your report is re-read off disk AFTER the run, so the files must still be there. Citations MUST be repo-root-relative (`.e2e-tmp/apc/bug-reproduce/repro.mjs:3`, never an absolute path and never a `/tmp/...` path, which cannot be resolved from the repo root).\n\nReport: the exact reproducer command you ran, the initial RED (exit code + the assertion that failed), the GREEN after the fix (exit 0), and the reversal RED after reintroducing the bug (exit code) — that reversal is the whole point: it is how you know the reproducer locks the bug.",
+      "expectedArtifacts": [
+        "the reproducer left on disk at `.e2e-tmp/apc/bug-reproduce/` (repo-relative, survives the run) + a `CHANGE-POINT: clamp @ .e2e-tmp/apc/bug-reproduce/clamp.mjs` declaration",
+        "RED→GREEN→REVERSAL phase table: ≥3 rows whose first cell is a `path:line` that resolves on disk, ≥2 landing on a line containing `clamp`",
+        "the exact reproducer command (a `node repro.mjs` invocation) that was actually run",
+        "initial RED: the reproducer failed on the buggy code (AssertionError, non-zero exit)",
+        "GREEN after the minimal fix (exit 0)",
+        "reversal negative control: reintroducing the bug makes the reproducer RED again (non-zero exit) — proves the lock"
+      ],
+      "acceptanceCriteria": [
+        {
+          "label": "the reproducer REALLY EXISTS on disk — anchored on `clamp @ .e2e-tmp/apc/bug-reproduce/clamp.mjs`, every path:line re-read off disk, ≥2 citations landing on a line that actually contains the subject symbol, ≥3 phase rows (RED/GREEN/REVERSAL) carrying a verified citation",
+          "type": "structured",
+          "checker": "cited-evidence",
+          "args": {
+            "minRows": 3,
+            "minAnchoredRows": 2,
+            "anchorMinLength": 4,
+            "expectedChangePoint": {
+              "symbol": "clamp",
+              "path": ".e2e-tmp/apc/bug-reproduce/clamp.mjs"
+            }
+          },
+          "match": "<structured:cited-evidence>",
+          "required": true
+        },
+        {
+          "label": "names a concrete runnable reproducer command (node/test invocation)",
+          "match": "\\b(node|npx|tsx|npm|pnpm|yarn|pytest|go\\s+test|cargo\\s+test|apc\\s+test)\\b",
+          "type": "regex",
+          "required": true
+        },
+        {
+          "label": "initial RED: reproducer fails on the bug (assertion / non-zero exit)",
+          "match": "(assert\\w*|assertionerror|❌|✗|fail\\w*|exit(\\s*code)?[^0-9]{0,6}[1-9])",
+          "type": "regex",
+          "required": true
+        },
+        {
+          "label": "GREEN after the fix (pass / exit 0)",
+          "match": "(pass\\w*|✓|✔|success\\w*|repro\\s*ok|green\\b|exit(\\s*code)?[^0-9]{0,6}0)\\b",
+          "type": "regex",
+          "required": true
+        },
+        {
+          "label": "REVERSAL negative control: reintroducing the bug makes the reproducer RED again (the lock)",
+          "match": "((revert|reversal|revers\\w*|reintroduc|re-?inject|re-?insert|put(ting)?\\s+(the\\s+)?bug\\s+back|remov\\w*\\s+the\\s+fix|back\\s?out|stash|undo\\s+the\\s+fix|buggy\\s*(version|body|one-?liner|代码|码|恢复|复原)|反转|复原|放回|重新(注入|引入)|注入回|撤(掉|回|销)\\s*(了\\s*)?(修复|fix))[\\s\\S]{0,220}(fail\\w*|red\\b|❌|✗|assert\\w*|non-?zero|exit(\\s*code)?[^0-9]{0,6}[1-9]))|((fail\\w*|red\\b|❌|✗|assert\\w*|non-?zero|exit(\\s*code)?[^0-9]{0,6}[1-9])[\\s\\S]{0,220}(revert|reversal|revers\\w*|reintroduc|re-?inject|re-?insert|put(ting)?\\s+(the\\s+)?bug\\s+back|remov\\w*\\s+the\\s+fix|back\\s?out|stash|undo\\s+the\\s+fix|buggy\\s*(version|body|one-?liner|代码|码|恢复|复原)|反转|复原|放回|重新(注入|引入)|注入回|撤(掉|回|销)\\s*(了\\s*)?(修复|fix)))",
+          "type": "regex",
+          "flags": "is",
+          "required": true
+        }
+      ]
+    }
+  ],
+  "security": {
+    "dataAccess": [
+      "local-filesystem"
+    ],
+    "humanApprovalRequiredFor": []
+  },
+  "provenance": {
+    "sourceKind": "inline-spec",
+    "sourceRefs": [
+      "docs/apc/12-skill-acceptance-tasks.md",
+      "docs/apc/05-devchain-gaps-and-skills.md",
+      "sdk/prismer-cloud/built-in-skills/diagnosing-bugs/SKILL.md"
+    ],
+    "authoredBy": "prismer-platform",
+    "authoredAt": "2026-07-24T00:00:00Z"
+  }
+}