peaks-loop 4.0.35 → 4.0.37

package/CHANGELOG.md

@@ -1,5 +1,45 @@
 # Changelog
 
+## 4.0.37 — 2026-09-10 (Windows spawn 根因 + 会话解析 + 诚实性修复)
+
+**Highlights**:
+
+1. **Windows `.cmd` spawn 根因，7 处一并修复** — `execFile`/`spawnSync` 一个 `.cmd` shim：**带 shell 时路径里的空格会被 shell 切开，不带 shell 时 Node ≥20 直接 `EINVAL`**。两条分支都不成立，所以在不含空格的路径上一切正常、在每个测试里都绿。
+   - `peaks slice check` 在**路径含空格的项目上整体失效**——这是本仓库自己所在的平台。改为 `process.execPath` + 包的 JS entry（无 shim、无 shell）。
+   - `version-precheck`：`projectRoot` 作为 **git 的参数**传入，空格路径下 git 退出 128，该层退化成 "layer skipped" —— **把一个真实存在的 tag 冲突报告成了"只是延后"**。
+   - `peaks code orchestrator-can-do`：两个探针在 Windows 上**恒返回失败**，所以 `q2SubAgentAvailable` 永远是 false、`q4ContextRatio` 永远是硬编码的 `0`（那是 fallback，不是读数）。
+   - 其余：`slice-decompose-runners`（失败被吞进 `indexed:false`，codegraph 信号静默消失）、`detect-eslint`、`coverage-c8`（`--src=`/`--reports-dir=` 携带项目路径）、若干 release 脚本。
+
+2. **会话解析根因——显式 rebind 此前只对一半 CLI 生效** — 两个函数都在回答"当前是哪个 session"，来源却不同：`getCurrentSessionId` 读 `.peaks/_runtime/session.json`，`getSessionIdCanonical` **优先 per-caller binding**。`peaks workspace init --session-id X --allow-session-rebind` 只改前者，于是 caller binding **遮蔽**了显式 rebind。实测后果：`session checkpoint` 与 `24h-mode` 把状态写进**另一个 session 的目录**，而 `peaks job` 读的是正确的那个。现在 rebind 同步更新当前 caller 的 binding——**更新而非清除**，因为清除会丢掉 per-caller 身份且没有任何代码会重建它。未 rebind 的其它 caller 文件**字节不变**，隔离不受影响。
+
+3. **job 可按 session 寻址（D6）+ `--slice-id` 不再静默 no-op（D7）** — 11 个 job 子命令补上 `--session-id`，解析顺序 `flag → PEAKS_SESSION_ID → binding`；job 不在绑定的 session 里时不再吐 `UNHANDLED_ERROR`，而是 `JOB_NOT_IN_SESSION` **指名它在哪个 session** 并给出重跑命令。`--slice-id` 现在接受 label 别名，未匹配则 `SLICE_NOT_FOUND` 并**列出全部合法 id**，且**不写入任何文件**（此前 `ok: true` 却什么都没做）。
+
+4. **`peaks web open --profile <name>`——持久登录终于有人消费** — S4 落的 `~/.peaks/web-profiles/<name>/storageState.json` 此前**没有任何动词读取**。`--profile` 只加在 `open` 上；名字在 **CLI 与 daemon 两侧**各校验一次（走线上的名字不是证据）；profile **只读**（自动浏览是否该回写用户级凭据文件是未决设计问题，本轮不替你决定）；指定了不存在的 profile **按名字失败**，而不是静默用一个未登录的浏览器。
+
+5. **诚实性修复——工具不再报告它没有的证据** — `peaks best-practice-scan` 此前把 stub 片段渲染成 8 行表格 + "方案 A ★" + 强制 ⚠️ 闸门，**全部来自合成文本**；现在合成结果直接拒绝（`ok:false`、exit 1，无推荐无表格）。`peaks slice check` 的 typecheck 阶段此前跑默认 `tsconfig.json`（272 个既有错误、**全部在 `tests/`**），**永远红**；现在 gate 于干净的 build 配置，宽口径计数对基线比较并在 detail 里如实报出。`peaks scan file-size` 不再把**生成文件**（记忆索引、lockfile）算作超长。
+
+6. **两个"不可能失败"的守卫测试** — 重建被删的 session-dir 静态扫描时发现**原版三个不变量都失效**：`src/` 正则只匹配链条**末尾**的 session id（而它要防的 bug 在**中间**，命中 0 次）；`skills/` 遍历只下一层（**完全看不到** `skills/bee/<skill>/references/`，包括它当初就是为了守的那个文件）；占位符比自己的注释窄。两侧都补了永久性的可证伪用例。`withEnv` 此前通过**在测试体内部注册 `afterEach`** 来还原环境变量——作用域错误，值会漏给下一个测试；改用 `onTestFinished`。**当前没有测试是因为这个泄漏才通过的**——真实的假通过来源，暂无受害者。
+
+7. **write gate 收紧且与 shell 解耦** — 放行范围从"**除 8 个名字外全部放行**"（那会放行顶层的 `.peaks/<change-id>/`，而 CLAUDE.md 明令禁止）收紧为**仅 `.peaks/_runtime/`**。handler 从 `node -e "<bash 专属转义的内联 JS>"` 改为调用随包发布的脚本文件，因而可像它的兄弟一样 pin PowerShell。顺带查明：**这个 hook 在 Windows 上从未真正生效过**——Claude Code 的 payload 走 stdin，而它读 argv，所以每次写入都落到 fall-through。
+
+**验证**：build clean、`tsc -p tsconfig.build.json` exit 0、宽口径 `tsconfig.json` 维持既有 142；全量 **175 files / 1679 passed（3 skipped）**；0 个 `daemon-entry` 进程残留。以上每一项都在**构建产物**（`bin/peaks.js`，非 `tsx src/`）上冒烟过，含"路径含空格的项目"这一验收场景。
+
+## 4.0.36 — 2026-09-10 (派发提示词瘦身 + 波次调度 + 编排器上下文审计)
+
+**Highlights**:
+
+1. **派发提示词瘦身（不牺牲规则）** — 每次派发的提示词里约 52% 是 CLI 注入的重复 boilerplate。本轮压缩 L1/生命周期/上下文探测段，并把 test-tool 块统一为一份（去掉 vitest/jest/mocha 举例，保留全部操作性规则，补上只存在于完整块里的 PB-5 例外规则）。7 个角色现在**字节完全一致**：rd/qa/qa-business/sc/prd **3897 → 3191 B（−18.1%）**。绑定规则 **21/21**、runner 规则 **5/5** 全角色断言。
+   - 契约指针实验**已回退**——它每次多 165 B 却不减任何内容，净亏。
+
+2. **文件重叠感知的波次调度** — 新 planner 按切面文件集计算波次：同一波内文件两两不重叠，重叠的推迟到后续波并**注明被哪个文件挡住**；附加式接入 `--from-dag`（`firstLevelWaves`），不改原调度语义。
+
+3. **编排器上下文审计 + 纪律** — 新增 `peaks code context-audit`：读会话 transcript 按 (工具, 命令) 汇总上下文消耗 Top-N。本仓库实测 **779 KB / 586 组**，且 Top-5 仅占 ~15%——**是长尾，不是单一元凶**。
+   - `--summary` 用于 memory reindex/list、doctor、request list：envelope **9.4–59 KB → 1.7–1.9 KB**（−79% ~ −97%），默认 `--json` 形状字节不变。
+   - SKILL.md 绑定规则：单次工具输出 >2 KB 禁止直接灌入编排器上下文。
+   - 子代理最终报告封顶 **≤40 行 / 2 KB**（5 个必填字段，细节写 artifact），已纳入 `BINDING_RULE_TOKENS`。
+
+**验证**：build clean、tsc clean、全量 141 files / 1270 passed（1 skipped）。
+
 ## 4.0.35 — 2026-09-10 (记忆系统 overhaul — 写入合一 / 索引重建 / 按任务调取 / 漂移健康检查 / rotate)
 
 **Highlights**:

package/README-en.md

@@ -140,7 +140,7 @@ Every lane opens with **one slash command**.
 
 | | |
 | --- | --- |
-| **Latest** | [![npm](https://img.shields.io/npm/v/peaks-loop?style=for-the-badge&logo=npm&logoColor=white&color=cb3837)](https://www.npmjs.com/package/peaks-loop) — 4.0.17 (2026-08-07) |
+| **Latest** | [![npm](https://img.shields.io/npm/v/peaks-loop?style=for-the-badge&logo=npm&logoColor=white&color=cb3837)](https://www.npmjs.com/package/peaks-loop) — 4.0.37 (2026-09-10) |
 | **Domains** | Code (`peaks-code`) · Content (`peaks-content`) · Project health (`peaks-doctor`) · Issue sweep (`peaks-issue-fix-orchestrator`) · Custom SOP (`peaks-sop`) · Cross-domain primitives (`peaks-solo` dispatcher · `peaks-resume` · `peaks-status` · `peaks-test` · `peaks-slice-decompose`) |
 | **Sediment pool** | `~/.peaks/` local pool · twice-clean runs auto-promote to a bee · broken runs come back for you to redefine · the bee grows with your taste |
 | **Test suite** | 285+ cases · 4 packages (peaks-loop / peaks-loop-mut / peaks-loop-shared-channel / peaks-loop-shared) · **0 timeouts** · 14 BDD caller-binding edge cases |

package/README.md

@@ -140,7 +140,7 @@ npm i -g peaks-loop
 
 | | |
 | --- | --- |
-| **最新版本** | [![npm](https://img.shields.io/npm/v/peaks-loop?style=for-the-badge&logo=npm&logoColor=white&color=cb3837)](https://www.npmjs.com/package/peaks-loop) — 4.0.32(2026-09-08) |
+| **最新版本** | [![npm](https://img.shields.io/npm/v/peaks-loop?style=for-the-badge&logo=npm&logoColor=white&color=cb3837)](https://www.npmjs.com/package/peaks-loop) — 4.0.37(2026-09-10) |
 | **覆盖域** | 代码(`peaks-code`) · 内容(`peaks-content`) · 项目健康(`peaks-doctor`) · 批量修 issue(`peaks-issue-fix-orchestrator`) · 自定义 SOP(`peaks-sop`) · 通用原语(`peaks-solo` 分诊 / `peaks-resume` 续 / `peaks-status` 看 / `peaks-test` 测 / `peaks-slice-decompose` 切片) |
 | **沉淀池** | `~/.peaks/` 本地池 · 跑两次自动晋升成 bee · 跑翻车让你重定义 · bee 跟着你的口味长 |
 | **测试套件** | 1096 cases · 4 packages (peaks-loop 1015 / runtime 39 / mut 22 / shared-channel 20) · **CI 首次全绿**(ubuntu + windows) · 14 BDD caller-binding coverage |

package/bin/peaks.js

@@ -1,2 +1,72 @@
 #!/usr/bin/env node
-import '../dist/cli/index.js';
+/**
+ * peaks-loop CLI shim.
+ *
+ * This file used to be a bare static `import '../dist/cli/index.js'`. That
+ * dies with a raw `ERR_MODULE_NOT_FOUND` stack trace — before any peaks code
+ * runs — whenever the local build is stale or incomplete. The reproducible
+ * trigger: `scripts/sync-version.mjs` (run by `npm run build`, `pretest`,
+ * `prepublish`) unlinks `packages/peaks-loop-shared/dist/version.js`, so every
+ * `node bin/peaks.js <anything>` fails with an opaque Node error until the
+ * workspace packages are rebuilt.
+ *
+ * The guard below is the only behaviour added. The happy path is unchanged:
+ * the same entry module is imported with the same `process.argv`, and the
+ * CLI's own exit codes pass through untouched. Only a resolution failure
+ * whose missing specifier is OURS — an internal workspace package or any
+ * `dist/` artifact — is converted into an actionable message; every other
+ * error is rethrown unchanged (a genuinely missing third-party dependency
+ * must still surface as the raw Node error).
+ */
+import { fileURLToPath } from 'node:url';
+
+const PACKAGE_ROOT = fileURLToPath(new URL('..', import.meta.url));
+
+/** Workspace-internal packages whose absence always means "stale build". */
+const INTERNAL_PACKAGE_NAMES = ['peaks-loop-shared', 'peaks-loop-internal-runtime'];
+
+/**
+ * The specifier Node could not resolve. Prefer the structured `err.url`
+ * (a clean `file://` URL); fall back to the quoted target in the message.
+ * The importer path is deliberately never parsed — `dist/cli/index.js` is
+ * the importer for every error, so matching it would flag third-party
+ * misses as stale-build misses.
+ */
+function missingSpecifier(err) {
+  if (typeof err?.url === 'string' && err.url.length > 0) return err.url;
+  const message = typeof err?.message === 'string' ? err.message : '';
+  const pkg = /Cannot find package '([^']+)'/.exec(message);
+  if (pkg !== null) return pkg[1];
+  const mod = /Cannot find module '([^']+)'/.exec(message);
+  if (mod !== null) return mod[1];
+  return '';
+}
+
+function isInternalSpecifier(specifier) {
+  const normalized = specifier.replace(/\\/g, '/');
+  if (normalized.length === 0) return false;
+  for (const name of INTERNAL_PACKAGE_NAMES) {
+    if (
+      normalized === name ||
+      normalized.startsWith(`${name}/`) ||
+      normalized.includes(`/${name}/`) ||
+      normalized.endsWith(`/${name}`)
+    ) {
+      return true;
+    }
+  }
+  return normalized.startsWith('dist/') || /\/dist\//.test(normalized);
+}
+
+try {
+  await import('../dist/cli/index.js');
+} catch (err) {
+  const specifier = err?.code === 'ERR_MODULE_NOT_FOUND' ? missingSpecifier(err) : '';
+  if (specifier.length === 0 || !isInternalSpecifier(specifier)) throw err;
+  process.stderr.write(
+    'peaks-loop: internal module not found — the local build is stale or incomplete.\n' +
+      `  missing: ${specifier}\n` +
+      `  fix:     run \`npm run build\` in ${PACKAGE_ROOT}, then retry.\n`
+  );
+  process.exitCode = 1;
+}

package/dist/cli/cli-helpers.js

@@ -6,6 +6,13 @@ export function printResult(io, result, asJson = false) {
     }
     if (!result.ok) {
         io.stderr(`${result.code}: ${result.message}`);
+        // S1's F1, at the printer: a warning carried by a FAILED envelope is the
+        // only news some failures have — `peaks web login` reports a headed browser
+        // it could not close this way. `fail()` hard-codes `warnings: []`, so only
+        // an envelope that spreads its own warnings over a `fail()` reaches here.
+        for (const warning of result.warnings) {
+            io.stderr(`warning: ${warning}`);
+        }
         for (const action of result.nextActions) {
             io.stderr(`- ${action}`);
         }

package/dist/cli/commands/_register.js

@@ -80,6 +80,7 @@ import { registerVendorDetectCommand } from './vendor-detect.js';
 import { registerUpgradeCommands } from './upgrade-commands.js';
 import { registerUserTouchpointCommands } from './user-touchpoint-commands.js';
 import { registerVerdictAggregateCommands } from './verdict-aggregate-command.js';
+import { registerWebCommands } from './web-commands.js';
 import { registerWorkflowCommands } from './workflow-commands.js';
 import { registerWorkflowPlanCommands } from './workflow-plan-commands.js';
 import { registerWorktreeAuthCommand } from './worktree-auth-commands.js';
@@ -139,6 +140,7 @@ const REGISTRATIONS = [
     ['session-spill-demo', registerSpillDemoCommand],
     ['outer-cache-commands', registerOuterCacheCommands],
     ['vendor-detect', registerVendorDetectCommand],
+    ['web-commands', registerWebCommands],
 ];
 function dispatchRegister(register, program, io) {
     if (register.length <= 1) {

package/dist/cli/commands/best-practice-scan-command.d.ts

@@ -1,7 +1,7 @@
 /**
  * Slice 2026-08-12 best-practice-scan — CLI subcommand.
  *
- * `peaks best-practice-scan --project <path> [--lang <lang>] [--commit]`
+ * `peaks best-practice-scan --intent <goal> --project <path> [--lang <lang>] [--commit]`
  *
  * Pipeline:
  *   1. detect language via language-detector (or use --lang override)
@@ -12,6 +12,19 @@
  *       a future slice that opts the artifact into git tracking)
  *   5. print artifact path + the 8-row table to stdout
  *
+ * `--intent` is REQUIRED: it is the business goal the scan is for (Step 2.5
+ * fires from a non-empty business goal), and it is what the artifact slug is
+ * derived from. The project path is NOT an intent — falling back to it was
+ * the defect this command was fixed for (a `.`-titled "best-practice scan").
+ * Omitting `--intent` fails with `BEST_PRACTICE_SCAN_INTENT_REQUIRED`.
+ *
+ * Honesty gate: v1's Context7 / WebSearch lookups are stubs, so a scan can
+ * come back `synthetic: true` (see scan-orchestrator). A synthetic result is
+ * NOT a scan result — the command then prints a refusal (no table, no
+ * recommendation, no ⚠️ catch gate), records the skip, and exits non-zero
+ * with `BEST_PRACTICE_SCAN_SYNTHETIC_LOOKUP` so no automated caller can
+ * treat Step 2.5 as satisfied.
+ *
  * Catch-gate (spec §7): the command emits a 3-line prompt asking the
  * user to ack / pick alt / reject + reason. The gate is read from
  * stdin via `PEAKS_BEST_PRACTICE_STDIN` (test seam) or the real stdin

package/dist/cli/commands/best-practice-scan-command.js

@@ -9,6 +9,37 @@ const CATCH_GATE_PROMPT = [
     '⚠️ 任何跟你真实业务不一样,改 — LLM 推荐可能错。',
     '回应 (默认 = 接受): 接受 / 接受方案 A|接受方案 B|接受方案 C / 拒绝 + 原因'
 ].join('\n');
+const INTENT_REQUIRED_MESSAGE = 'best-practice-scan requires --intent <text>: the intent is the business goal the scan is ' +
+    'for, not the project path. Deriving it from the project path labelled a scan with a ' +
+    'non-goal, so there is no fallback.';
+const SYNTHETIC_SOURCE_MESSAGE = 'best-practice-scan: SKIPPED — the Context7 / WebSearch lookups in this build are stubs ' +
+    '(synthetic fragments), so no real documentation was consulted and there is nothing to ' +
+    'recommend. No recommendation, no comparison table and no ⚠️ catch gate were produced.';
+function bestPracticeDir(projectRoot) {
+    return join(projectRoot, 'best-practice');
+}
+function artifactPathFor(projectRoot, slugSource) {
+    const dateStr = new Date().toISOString().slice(0, 10);
+    const slug = slugSource.replace(/[^a-zA-Z0-9_-]+/g, '-').slice(0, 32) || 'intent';
+    return join(bestPracticeDir(projectRoot), `${dateStr}-${slug}.md`);
+}
+/** Refusal body for a synthetic scan. It is written to the artifact path so
+ *  the skip survives the fire-and-forget caller (`best-practice-auto-trigger`
+ *  spawns with `stdio: 'ignore'`, so stdout never reaches anyone). */
+function renderSyntheticRefusal(opts) {
+    return [
+        `# Best-Practice Scan — ${opts.intent}`,
+        '',
+        '> SKIPPED — synthetic (stub) lookup. No real documentation lookup was performed.',
+        '',
+        SYNTHETIC_SOURCE_MESSAGE,
+        '',
+        `- language: ${opts.language}`,
+        `- transport that answered: ${opts.source} (stub)`,
+        '- Step 2.5: recorded as skipped, reason "synthetic-lookup". Re-run this step once the real',
+        '  lookup lands; a gate cannot be satisfied from a fabricated scan.'
+    ].join('\n');
+}
 async function readUserInput() {
     const override = process.env.PEAKS_BEST_PRACTICE_STDIN;
     if (override !== undefined)
@@ -59,25 +90,54 @@ export function registerBestPracticeScanCommand(program, io) {
         .description('2026-08-12 best-practice-scan: language-aware + business-aware doc-fragment lookup ' +
         'via Context7 (priority 1) → WebSearch (priority 2) → empty fallback. ' +
         'Renders the 8-row comparison table from spec §5 + §6 + §7. ' +
-        'Writes the artifact to <project>/best-practice/<date>-<intent>.md (gitignored).')
+        'Writes the artifact to <project>/best-practice/<date>-<intent>.md (gitignored). ' +
+        '--intent is required; a synthetic (stub) lookup refuses instead of rendering a ' +
+        'recommendation and exits non-zero.')
+        .option('--intent <text>', 'business goal the scan is for (required)')
         .option('--project <path>', 'project root (default cwd)', process.cwd())
         .option('--lang <lang>', 'language override (skip auto-detect)')
         .option('--commit', 'reserved flag — opts the artifact into git tracking (future slice)')).action(async (opts) => {
         try {
+            const intent = (opts.intent ?? '').trim();
+            if (intent.length === 0) {
+                printResult(io, fail('best-practice.scan', 'BEST_PRACTICE_SCAN_INTENT_REQUIRED', INTENT_REQUIRED_MESSAGE, { project: opts.project }, [
+                    'Rerun with --intent "<business goal>" — the goal from the PRD artifact'
+                ]), opts.json === true);
+                process.exitCode = 1;
+                return;
+            }
             const detection = opts.lang === undefined ? detectLanguage(opts.project) : null;
             const language = opts.lang ?? detection?.language ?? 'unknown';
-            io.stdout(`[best-practice-scan] project=${opts.project} language=${language}`);
+            io.stdout(`[best-practice-scan] project=${opts.project} intent=${intent} language=${language}`);
             const scan = await scanBestPractice({
-                intent: opts.project,
+                intent,
                 language,
                 projectRoot: opts.project,
                 io
             });
+            if (scan.synthetic) {
+                const refusal = renderSyntheticRefusal({ intent, language, source: scan.source });
+                io.stdout(refusal);
+                const refusedPath = artifactPathFor(opts.project, intent);
+                mkdirSync(bestPracticeDir(opts.project), { recursive: true });
+                writeFileSync(refusedPath, refusal + '\n', 'utf8');
+                printResult(io, fail('best-practice.scan', 'BEST_PRACTICE_SCAN_SYNTHETIC_LOOKUP', SYNTHETIC_SOURCE_MESSAGE, {
+                    project: opts.project,
+                    intent,
+                    language,
+                    scanSource: scan.source,
+                    synthetic: true,
+                    step25: 'skipped-synthetic-lookup',
+                    artifactPath: refusedPath
+                }, ['Re-run once real Context7 / WebSearch wiring lands', `Skip record written to ${refusedPath}`]), opts.json === true);
+                process.exitCode = 1;
+                return;
+            }
             const recommendation = scan.results.length > 0 ? 'A' : 'B';
             const reasoning = `基于 ${scan.fragments.length} 个 ${scan.source} 文档片段;` +
                 `语言=${language};LLM 推断方案 ${recommendation} 与项目阶段最匹配。`;
             const table = formatOutputTable({
-                intent: opts.project,
+                intent,
                 language,
                 fragments: scan.results,
                 recommendation,
@@ -92,14 +152,12 @@ export function registerBestPracticeScanCommand(program, io) {
             io.stdout(CATCH_GATE_PROMPT);
             const userInput = await readUserInput();
             const outcome = parseCatchGateReply(userInput, recommendation);
-            const outDir = join(opts.project, 'best-practice');
-            mkdirSync(outDir, { recursive: true });
-            const dateStr = new Date().toISOString().slice(0, 10);
-            const slug = opts.project.replace(/[^a-zA-Z0-9_-]+/g, '-').slice(0, 32) || 'intent';
-            const artifactPath = join(outDir, `${dateStr}-${slug}.md`);
+            const artifactPath = artifactPathFor(opts.project, intent);
+            mkdirSync(bestPracticeDir(opts.project), { recursive: true });
             writeFileSync(artifactPath, table + '\n', 'utf8');
             printResult(io, ok('best-practice.scan', {
                 project: opts.project,
+                intent,
                 language,
                 scanSource: scan.source,
                 fragments: scan.results,

package/dist/cli/commands/code-runtime-commands.d.ts

@@ -1,10 +1,13 @@
 /**
  * Slice rid-024 — runtime probes: post-compact-detect / auto-compact /
- * context-now / gate-step-08 / emit-handoff.
+ * context-now / context-audit / gate-step-08 / emit-handoff.
  *
  * Extracted from code-commands.ts (rid-024 split).
- * Owns: 5 sub-commands that read or mutate runtime state.
+ * Owns: 6 sub-commands that read or mutate runtime state.
  * Owns the `readActiveSid` helper (only used by these runtime probes).
+ *
+ * Slice 2026-09-10-context-audit-and-discipline added `context-audit`
+ * (what fills the window, grouped by tool + short input key).
  */
 import type { Command } from 'commander';
 import { type ProgramIO } from '../cli-helpers.js';

package/dist/cli/commands/code-runtime-commands.js

@@ -1,15 +1,20 @@
 /**
  * Slice rid-024 — runtime probes: post-compact-detect / auto-compact /
- * context-now / gate-step-08 / emit-handoff.
+ * context-now / context-audit / gate-step-08 / emit-handoff.
  *
  * Extracted from code-commands.ts (rid-024 split).
- * Owns: 5 sub-commands that read or mutate runtime state.
+ * Owns: 6 sub-commands that read or mutate runtime state.
  * Owns the `readActiveSid` helper (only used by these runtime probes).
+ *
+ * Slice 2026-09-10-context-audit-and-discipline added `context-audit`
+ * (what fills the window, grouped by tool + short input key).
  */
 import { addJsonOption, getErrorMessage, printResult } from '../cli-helpers.js';
 import { fail, ok } from 'peaks-loop-shared/result';
 import { detectPostCompactResume, formatPostCompactResumeLogLine } from '../../services/code/post-compact-detector.js';
 import { runAutoCompact } from '../../services/code/auto-compact-orchestrator.js';
+import { auditContext } from '../../services/context/context-audit.js';
+import { buildContextAuditHint } from '../../services/context/context-audit-hint.js';
 import { evaluateStep08, STEP_08_BACKUP_REGEX } from '../../services/code/step-08-gate.js';
 import { evaluateEmitHandoff, JOB_NOT_INITIALIZED, JOB_REMAINING_BLOCKED } from '../../services/code/emit-handoff.js';
 import { readJobShapeDecision, JobShapeDecisionError } from '../../services/code/job-shape-decision.js';
@@ -254,6 +259,57 @@ export function registerCodeRuntimeCommands(code, io) {
             process.exitCode = 1;
         }
     });
+    // Slice 2026-09-10-context-audit-and-discipline (Slice A): visibility into
+    // WHAT fills the orchestrator window. `context-now` returns a ratio; this
+    // returns the grouped byte breakdown of tool results from the live
+    // transcript. Read-only + fail-soft: an unavailable transcript reports
+    // `available:false` with a reason and NEVER sets a non-zero exit code.
+    addJsonOption(code
+        .command('context-audit')
+        .description('Slice 2026-09-10 Slice A: report what fills the current session\'s ' +
+        'context window, grouped by tool + short input key (command line / ' +
+        'path tail / pattern), sorted by bytes. Locates the CURRENT session\'s ' +
+        'IDE transcript through the active adapter\'s ' +
+        '`compact.resolveTranscriptPath` (vendor-neutral); emits total bytes, ' +
+        'entry count, and the top-N groups ' +
+        '(`{tool, key, bytes, pctOfTotal, count}`). Read-only and fail-soft — ' +
+        'a missing/oversized/corrupt transcript returns `available:false` ' +
+        'with a reason and never blocks. Never dumps tool result content.')
+        .requiredOption('--project <path>', 'target project root')
+        .option('--session-id <sid>', 'override session id (default: read from active presence)')
+        .option('--top <n>', 'number of top entries to emit (default 15, max 100)', (value) => Number(value))
+        .option('--transcript <path>', 'override the transcript jsonl path (test seam)')).action((opts) => {
+        try {
+            // `resolveOuterSessionId` checks the env signal FIRST, so the peaks
+            // session id is only a fallback lookup key — mirror context-now's
+            // 'unknown' default so an unbound presence still resolves via env.
+            const sessionId = opts.sessionId ?? readActiveSid(opts.project) ?? 'unknown';
+            const outerSessionId = resolveOuterSessionId(opts.project, sessionId);
+            const result = auditContext({
+                outerSessionId: outerSessionId ?? null,
+                ...(opts.top !== undefined ? { topN: opts.top } : {}),
+                ...(opts.transcript !== undefined ? { transcriptPath: opts.transcript } : {}),
+            });
+            // Fail-soft contract: unavailability is DATA, not an error. The exit
+            // code stays 0 so a `context-audit` call can never block a workflow.
+            const nextActions = result.available
+                ? [`${result.entryCount} tool result(s) across ${result.groupCount} group(s); showing top ${result.entries.length}.`]
+                : [`context-audit unavailable: ${result.reason ?? 'unknown'} — continue without it (read-only probe).`];
+            printResult(io, ok('code.context-audit', { ...result }, [], nextActions), opts.json);
+        }
+        catch (err) {
+            printResult(io, ok('code.context-audit', {
+                available: false,
+                reason: `audit-failed: ${getErrorMessage(err)}`,
+                transcriptPath: null,
+                totalBytes: 0,
+                entryCount: 0,
+                groupCount: 0,
+                topN: 0,
+                entries: []
+            }, [], ['context-audit is a read-only probe; continue without it.']), opts.json);
+        }
+    });
     // v3.1.2 Step 0.8 — Mechanical PreToolUse gate.
     // Wire-installed by `peaks workspace init` (extends the existing hook
     // installer). Exit code is the load-bearing contract:
@@ -265,7 +321,8 @@ export function registerCodeRuntimeCommands(code, io) {
         'fail-closed backup regex when missing. Exit 0 = allow, exit 2 = block. ' +
         'When the decision says isJob=true AND progress.json exists, the stdout ' +
         'also carries `Next: slice #N+1 of M (<currentSlice>)` so the LLM cannot ' +
-        'wake up cold.')
+        'wake up cold. When the context ratio is ≥ 0.70 the stdout gains ONE more ' +
+        'line naming the largest context consumer (audit cached ≥ 5 min; fail-soft).')
         .requiredOption('--project <path>', 'target project root (the hook passes "." so resolveCanonicalProjectRoot promotes it to the git root)')
         .option('--session-id <sid>', 'override session id (default: read from active presence)')
         .option('--prompt <text>', 'explicit prompt text (default: read last-prompt.txt; stdin ignored)')).action((opts) => {
@@ -285,6 +342,17 @@ export function registerCodeRuntimeCommands(code, io) {
                 printResult(io, envelope, opts.json);
                 return;
             }
+            // Slice 2026-09-10-three-fixes (Slice 2): proactive context-consumer
+            // hint. Runs ONLY when the window is ≥ 0.70 full, caches the audit
+            // result for ≥ 5 min so the transcript is scanned at most once per
+            // TTL window, and is fail-soft (null → no extra line). It never
+            // changes the exit code and never blocks.
+            const hintLine = buildContextAuditHint({
+                projectRoot: opts.project,
+                sessionId,
+                outerSessionId: resolveOuterSessionId(opts.project, sessionId)
+            });
+            const hintActions = hintLine === null ? [] : [hintLine];
             const evalInput = {
                 projectRoot: opts.project,
                 sessionId
@@ -300,7 +368,7 @@ export function registerCodeRuntimeCommands(code, io) {
                     decision: verdict.decision,
                     progress: verdict.progress,
                     nextSlice: result.nextSliceLine
-                }, [], result.nextSliceLine !== null ? [result.nextSliceLine] : []);
+                }, [], [...(result.nextSliceLine !== null ? [result.nextSliceLine] : []), ...hintActions]);
                 printResult(io, envelope, opts.json);
                 return;
             }
@@ -311,7 +379,8 @@ export function registerCodeRuntimeCommands(code, io) {
                     decision: null,
                     nextSlice: null
                 }, [], [
-                    'job-shape.json says isJob=false; single-rid mode (gate allows).'
+                    'job-shape.json says isJob=false; single-rid mode (gate allows).',
+                    ...hintActions
                 ]);
                 printResult(io, envelope, opts.json);
                 return;
@@ -326,7 +395,8 @@ export function registerCodeRuntimeCommands(code, io) {
                     backupRegex: STEP_08_BACKUP_REGEX.toString()
                 }, [
                     'Run `peaks code detect-job --is-job true --rationale <text> --suggested-job-id <slug>` to record the Job-shape verdict.',
-                    'Then re-run the Bash tool call.'
+                    'Then re-run the Bash tool call.',
+                    ...hintActions
                 ]);
                 io.stderr(`${blockMessage}\n`);
                 printResult(io, envelope, opts.json);
@@ -341,7 +411,8 @@ export function registerCodeRuntimeCommands(code, io) {
                 nextSlice: null,
                 promptSource: verdict.promptSource
             }, [], [
-                'No job-shape.json AND no backup-regex match on prompt → allow (most prompts are not Job-shaped).'
+                'No job-shape.json AND no backup-regex match on prompt → allow (most prompts are not Job-shaped).',
+                ...hintActions
             ]);
             printResult(io, envelope, opts.json);
             return;

package/dist/cli/commands/core/doctor-command.d.ts

@@ -1,5 +1,13 @@
 import type { Command } from 'commander';
 import { type ProgramIO } from '../../cli-helpers.js';
+/**
+ * Slice 2026-09-10-context-audit-and-discipline (Slice B): bounded view of
+ * the doctor envelope. The `checks` array (one object per check, each with a
+ * full message) and the stale-binding instance list collapse to counts +
+ * names-of-first-N; the summary counters and log section are scalars and are
+ * kept verbatim. ≤ 2 KB by construction.
+ */
+export declare function buildDoctorSummary(data: Record<string, unknown>): Record<string, unknown>;
 export type BindingSource = 'canonical' | 'legacy';
 export type DoctorLogsSection = {
     logDir: string;

package/dist/cli/commands/core/doctor-command.js

@@ -4,6 +4,7 @@ import { findProjectRoot } from '../../../services/config/config-safety.js';
 import { loadSkillRegistry } from '../../../services/skills/skill-registry.js';
 import { planStatusLineInstall } from '../../../services/skills/statusline-settings-service.js';
 import { addJsonOption, printResult } from '../../cli-helpers.js';
+import { boundedNames, fitSummaryToBytes } from '../../../services/context/summary-view.js';
 import { fail, ok } from 'peaks-loop-shared/result';
 // slice-3b Option C: the doctor subpackage owns the check pipeline but
 // does NOT import cross-domain utils from the main package (avoids
@@ -38,6 +39,43 @@ function statusLineAlreadyInstalledForScope(scope, projectRoot) {
         return false;
     }
 }
+/**
+ * Slice 2026-09-10-context-audit-and-discipline (Slice B): bounded view of
+ * the doctor envelope. The `checks` array (one object per check, each with a
+ * full message) and the stale-binding instance list collapse to counts +
+ * names-of-first-N; the summary counters and log section are scalars and are
+ * kept verbatim. ≤ 2 KB by construction.
+ */
+export function buildDoctorSummary(data) {
+    const checks = Array.isArray(data.checks) ? data.checks : [];
+    const label = (c) => {
+        const id = typeof c.id === 'string' ? c.id : 'unknown';
+        if (c.ok === true)
+            return `ok ${id}`;
+        return `${c.severity === 'warning' ? 'warn' : 'FAIL'} ${id}`;
+    };
+    const failed = checks.filter((c) => c.ok === false);
+    const stale = typeof data.staleBinding === 'object' && data.staleBinding !== null
+        ? data.staleBinding
+        : null;
+    const view = {
+        view: 'summary',
+        summary: data.summary,
+        checks: boundedNames(checks.map(label)),
+        failed: boundedNames(failed.map((c) => `${typeof c.id === 'string' ? c.id : 'unknown'}: ${typeof c.message === 'string' ? c.message : ''}`)),
+    };
+    if (stale !== null) {
+        view.staleBinding = {
+            ttlMs: stale.ttlMs,
+            staleCount: stale.staleCount,
+            droppedCount: stale.droppedCount,
+            droppedSids: boundedNames(Array.isArray(stale.droppedSids) ? stale.droppedSids.map((s) => String(s)) : []),
+        };
+    }
+    if (data.logs !== undefined)
+        view.logs = data.logs;
+    return fitSummaryToBytes(view);
+}
 function doctorStatusLineInstalledProbe() {
     const projectRoot = findProjectRoot(process.cwd());
     // Check both scopes: a user may have installed the statusLine globally, which
@@ -136,7 +174,8 @@ export function registerDoctorCommand(program, io) {
         // scan) addressable for non-current projects. The doctor was
         // hardcoded to findProjectRoot(process.cwd()) which is the
         // wrong default for users inspecting a sibling project.
-        .option('--project <path>', 'target project root (defaults to git root or cwd)')).action(async (options) => {
+        .option('--project <path>', 'target project root (defaults to git root or cwd)')
+        .option('--summary', 'JSON envelope only: emit check counts + names-of-first-N (≤ 2 KB) instead of the full checks/stale-binding arrays; the default envelope is unchanged')).action(async (options) => {
         // v2.18.2 cycle 2 (Q2 arbitration): --rebuild-binding and
         // --cleanup-stale BOTH mutate the binding file. Running them
         // together is ambiguous (rebuild rewrites callerIds; cleanup
@@ -209,9 +248,12 @@ export function registerDoctorCommand(program, io) {
             droppedCount: droppedStale.length,
             droppedSids: droppedStale
         };
-        const data = logsSection === null
+        const fullData = logsSection === null
             ? { ...report, staleBinding: staleBindingSection }
             : { ...report, logs: logsSection, staleBinding: staleBindingSection };
+        // Slice B: `--summary` is opt-in and affects the JSON envelope only (the
+        // human-readable path below already prints one line per check).
+        const data = options.summary === true ? buildDoctorSummary(fullData) : fullData;
         // Slice 2026-08-05-statusline-sid-only-marker-and-multi-binary-drift-guard
         // repair cycle: `report.summary.ok` already factors in the
         // severity-aware aggregation in `buildReport` (warnings do NOT

package/dist/cli/commands/core/memory-command.js

@@ -53,6 +53,7 @@ export function registerMemoryCommand(program, io) {
         .option('--kind <kind>', `filter by memory kind (one of: ${KIND_HELP})`)
         .option('--pick', 'spawn fzf for interactive multi-select (requires fzf >= 0.38); writes picked.json')
         .option('--fzf-bin <path>', 'override fzf binary path (default: fzf on PATH)', 'fzf')
+        .option('--summary', 'emit counts + names-of-first-N only (≤ 2 KB) instead of the full entry array; the default envelope is unchanged')
         .option('--project <path>', 'target project root (defaults to git root or cwd)')).action((options) => {
         void import('../memory-commands.js').then(({ runMemoryList }) => {
             void runMemoryList(io, {
@@ -60,6 +61,7 @@ export function registerMemoryCommand(program, io) {
                 ...(options.pick === true ? { pick: true } : {}),
                 ...(options.fzfBin ? { fzfBin: options.fzfBin } : {}),
                 ...(options.project !== undefined ? { project: options.project } : {}),
+                ...(options.summary === true ? { summary: true } : {}),
                 ...(options.json !== undefined ? { json: options.json } : {}),
             });
         }).catch((error) => {
@@ -79,12 +81,14 @@ export function registerMemoryCommand(program, io) {
         .description('Rebuild .peaks/memory/index.json from every memory file on disk and regenerate MEMORY.md; reports unclassified files and orphans both ways. Dry-run by default; pass --apply to write.')
         .option('--project <path>', 'target project root (defaults to git root or cwd)')
         .option('--dry-run', 'report drift without writing (default)')
-        .option('--apply', 'rebuild index.json and regenerate MEMORY.md')).action((options) => {
+        .option('--apply', 'rebuild index.json and regenerate MEMORY.md')
+        .option('--summary', 'emit drift counts + names-of-first-N only (≤ 2 KB) instead of the full arrays; the default envelope is unchanged')).action((options) => {
         void import('../memory-commands.js').then(({ runMemoryReindex }) => {
             void runMemoryReindex(io, {
                 ...(options.project !== undefined ? { project: options.project } : {}),
                 ...(options.dryRun === true ? { dryRun: true } : {}),
                 ...(options.apply === true ? { apply: true } : {}),
+                ...(options.summary === true ? { summary: true } : {}),
                 ...(options.json !== undefined ? { json: options.json } : {}),
             });
         }).catch((error) => {