claude-spotter 1.5.4 → 1.5.6

package/CHANGELOG.md

@@ -1,5 +1,23 @@
 # Changelog
 
+## 1.5.6 — 2026-08-05
+
+- **Codexの動的nested MCP利用を採用として記録する。** `functions.exec`内で
+  `ALL_TOOLS.find(...name === "mcp__...")`から取得したtoolを`tools[tool.name](...)`で実行する
+  現行Codexの呼び方を認識する。単なるlookupや文字列・コメント内の記述は利用として数えない。
+- **実際の誤判定を回帰化。** `mcp__aiterm__pty_open`を実行済みなのに`usedToolIds=[]`となった形を
+  focused testとStop hookの評価DB統合テストへ固定し、`adopted`になることを確認する。
+
+## 1.5.5 — 2026-08-05
+
+- **提案時文脈をexact sessionへ戻した。** 評価記録がproject全体の最新threadを読む
+  `observer-read`を誤って使っていた経路を撤去し、既存のThroughline `auditor-context`へ一本化した。
+  Claude / CodexのUserPromptSubmit hookから提案元の`session_id`、`host`、`transcript_path`を渡し、
+  提案直前の完了turnを非採用caseへ保存できる。
+- **重複adapterを廃止。** 評価専用readerは、すでにfreshness・本文上限・exact-session照合を持つ
+  `loadAuditorContext`を再利用する。Throughline本文をauditor入力へ戻さず、監査と評価証拠の分離は維持する。
+- **検証。** 既存adapterへの引数接続とClaude / Codex hookのfocused testを通過した。
+
 ## 1.5.4 — 2026-08-05
 
 - **Throughlineを提案監査の実行条件から撤去。** Claude / CodexのUserPromptSubmitは、

package/README.ja.md

@@ -159,9 +159,9 @@ flowchart LR
 ### Throughlineの提案時評価文脈（任意）
 
 Spotterの提案AIはThroughlineを使わず、Throughlineの導入・設定・freshness・取得成否に関係なく
-UserPromptSubmitごとに監査します。Throughlineは、提案が出た時に改善分析用の別文脈を
-`observer-read`で一度だけ記録する任意経路です。取得できなくても`context_unavailable`として記録するだけで、
-監査や親への助言には影響しません。
+UserPromptSubmitごとに監査します。提案が出た時だけ、改善分析用の別文脈としてThroughline
+`auditor-context`を提案元のexact sessionとtranscriptで一度だけ呼びます。返されたfreshな直前完了turnは
+監査入力と混ぜずに保存します。取得できなくても`context_unavailable`として記録するだけで、監査や親への助言には影響しません。
 
 `spotter install`がPATH上のThroughlineを絶対パスへ解決できる場合、この評価証拠の取得経路を既定で設定します。
 既存互換の`--auditor-context`名はmarker設定に残っていますが、監査のON/OFFは制御しません。
@@ -189,7 +189,7 @@ spotter install -y --auditor-context throughline `
 ```
 
 `spotter doctor`はこの経路を`evaluation context`として表示し、command / argsや会話本文は表示しません。
-observer snapshotは端末内の評価SQLiteにだけ保存され、network送信、retry、background回収は行いません。
+評価文脈は端末内の評価SQLiteにだけ保存され、network送信、retry、background回収は行いません。
 
 ## よく使うコマンド
 
@@ -206,12 +206,16 @@ spotter db rebuild       # Claude local + Claude global DB を両方消してか
 spotter status           # 稼働中の daemon 一覧
 spotter doctor           # 環境診断 (Node / claude CLI / Codex readiness / tool-db 整合性)
 spotter diagnostics logs # daemon log から pass=false / backend latency / anomaly signal を集計
+spotter diagnostics factory
+                         # factory向けread-only診断snapshotをJSONで出力
+spotter diagnostics runtime-errors
+                         # opt-in端末内runtime error集計をread-only表示（network送信なし）
 spotter evaluation report
                          # 端末内DBからproject横断の提案率・tool採用率を集計
 spotter evaluation cases --outcome not-adopted
                          # 提案されたが同じturnで使われなかったtool itemを一覧
 spotter evaluation case <observation-id>
-                         # request、2種類の文脈、提案、利用、outcomeを確認
+                         # request、任意のThroughline snapshot、提案、利用、outcomeを確認
 spotter dashboard device --id mac --name Mac
                          # この端末の評価DBを127.0.0.1:53940で配信
 spotter dashboard hub --config dashboard-hub.json --host 172.18.0.1
@@ -236,8 +240,9 @@ spotter uninstall        # hook 登録を解除 (~/.spotter は残す)
 
 dashboardはlocal-firstで動く。各端末が自身の`~/.spotter/evaluation.db`を読み、hubは固定の
 端末・upstream対応だけを持つ。評価データをcloud DBへ複製しない。端末画面では
-`S/P/I/C/A/M`、提案率、採用率、project/tool内訳、非採用case、別々に取得した2種類の文脈を
-確認できる。health確認は端末一覧request時だけなので、端末がofflineでもbackground監視や
+対象ターン、ツール提案あり、提案ツール数、利用判定済み、実際に使用、判定不能、提案率、採用率、
+project/tool内訳、非採用case、監査対象request、任意の提案時Throughline証拠を確認できる。
+health確認は端末一覧request時だけなので、端末がofflineでもbackground監視や
 retry queueを作らず、その端末だけを切り離せる。
 
 4端末のservice、reverse tunnel、Caddy/Cloudflare構成は

package/README.md

@@ -160,9 +160,11 @@ Both share the principle of **"don't rely on the primary agent to do it itself."
 
 Spotter's proposal auditor does not use Throughline. Every UserPromptSubmit
 audit runs independently of Throughline installation, configuration, freshness, or read
-failures. When Spotter emits a proposal, the evaluation recorder may call
-Throughline `observer-read` once to save separate improvement evidence. A failed
-read is recorded as `context_unavailable` and never changes auditing or parent advice.
+failures. When Spotter emits a proposal, the evaluation recorder calls Throughline
+`auditor-context` once with the proposing host's exact session ID and transcript path.
+The bounded, fresh completed turns are saved as separate improvement evidence; they are
+never auditor input. A failed read is recorded as `context_unavailable` and never changes
+auditing or parent advice.
 
 When `spotter install` resolves Throughline on PATH to an absolute executable, it
 configures this evaluation-evidence path by default. The legacy option name
@@ -218,7 +220,7 @@ spotter evaluation report
 spotter evaluation cases --outcome not-adopted
                          # list proposed tools that were not used in the same turn
 spotter evaluation case <observation-id>
-                         # inspect request, Spotter context, Throughline snapshot, proposal, usage, and outcome
+                         # inspect request, optional Throughline snapshot, proposal, usage, and outcome
 spotter dashboard device --id mac --name Mac
                          # serve this terminal's local evaluation DB on 127.0.0.1:53940
 spotter dashboard hub --config dashboard-hub.json --host 172.18.0.1
@@ -244,8 +246,8 @@ spotter uninstall        # remove hooks from this project (leaves ~/.spotter int
 The dashboard is local-first. Every terminal reads its own `~/.spotter/evaluation.db`; the hub
 keeps only a static device-to-upstream map and does not copy evaluation data into a cloud database.
 The device view shows Japanese labels for every evaluation metric, proposal and adoption rates
-with their numerator and denominator, project/tool breakdowns, non-adopted cases, and the two
-separately captured context sources. The hub checks health only
+with their numerator and denominator, project/tool breakdowns, non-adopted cases, the request
+audited by Spotter, and optional proposal-time Throughline evidence. The hub checks health only
 when the device list is requested, so an offline terminal is isolated without a background monitor
 or retry queue.
 

package/docs/11_dashboard-operations.md

@@ -1,5 +1,8 @@
 # Spotter評価dashboard運用
 
+現行配布版: **v1.5.4**（2026-08-05）。Mac、main-server、FOX WSL2、FOX Windows nativeへ
+global install済み。v1.5.4は監査gate修正でありdashboard routing構成はv1.5.3から変更していない。
+
 ## 固定構成
 
 各端末のdevice serverはloopbackだけで待ち受ける。main-server、Mac、FOX WSL2は

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-spotter",
-  "version": "1.5.4",
+  "version": "1.5.6",
   "description": "Audit agent running alongside Claude Code that catches missed tool calls — 気づく役と実行する役の分離",
   "type": "module",
   "bin": {

package/src/cli/codex-hook-cmd.mjs

@@ -32,7 +32,7 @@ import {
 } from '../core/hook-event-log.mjs';
 import { observeRuntimeErrorIsolatedSafe } from '../core/runtime-error-store.mjs';
 import { createEvaluationStore } from '../core/evaluation-store.mjs';
-import { loadEvaluationObserverContext } from '../core/evaluation-context.mjs';
+import { loadEvaluationContext } from '../core/evaluation-context.mjs';
 import {
   canonicalizeCodexNestedMcpToolIds,
   canonicalizeCodexSkillReadToolIds,
@@ -130,7 +130,7 @@ export async function runCodexUserPromptSubmitHook({
   writeError = (text) => process.stderr.write(text),
   runtimeErrorObserver = async () => ({ collected: false, reason: 'observer_not_configured' }),
   createEvaluationStoreFn = createEvaluationStore,
-  loadEvaluationObserverContextFn = loadEvaluationObserverContext,
+  loadEvaluationContextFn = loadEvaluationContext,
   randomUUIDFn = randomUUID,
   now = () => Date.now(),
 } = {}) {
@@ -150,15 +150,16 @@ export async function runCodexUserPromptSubmitHook({
       return;
     }
     const proposals = canonicalizeProposedToolIds(proposedToolIds);
-    let observerContext = { status: 'not_requested', snapshot: null };
+    let evaluationContext = { status: 'not_requested', snapshot: null };
     let proposedAtMs = recordedAtMs;
     try {
       if (auditStatus === 'success' && proposals.resolvedToolIds.length > 0) {
         proposedAtMs = now();
-        observerContext = await loadEvaluationObserverContextFn({
+        evaluationContext = await loadEvaluationContextFn({
           projectRoot,
           host: 'codex',
           sessionId,
+          transcriptPath: input.transcript_path,
           recordedAtMs: proposedAtMs,
         });
       }
@@ -174,8 +175,8 @@ export async function runCodexUserPromptSubmitHook({
           auditStatus,
           requestText: prompt,
           auditorSeenContext: null,
-          observerContextStatus: observerContext.status,
-          observerSnapshot: observerContext.snapshot,
+          observerContextStatus: evaluationContext.status,
+          observerSnapshot: evaluationContext.snapshot,
           proposedToolIds: proposals.resolvedToolIds,
           backend,
           model,

package/src/core/evaluation-context.mjs

@@ -1,45 +1,33 @@
-import { execFile } from 'node:child_process';
-import { createHash } from 'node:crypto';
 import { realpath } from 'node:fs/promises';
 import { isAbsolute } from 'node:path';
-import { promisify } from 'node:util';
 
 import {
   DEFAULT_AUDITOR_CONTEXT_MAX_BUFFER,
   DEFAULT_AUDITOR_CONTEXT_TIMEOUT_MS,
+  loadAuditorContext,
   readProjectAuditorContextConfig,
 } from './auditor-context.mjs';
 
-export const THROUGHLINE_OBSERVER_READ_SCHEMA = 'throughline.observer_read.v1';
-export const EVALUATION_OBSERVER_CONTEXT_AVAILABLE = 'context_available';
-export const EVALUATION_OBSERVER_CONTEXT_UNAVAILABLE = 'context_unavailable';
-export const DEFAULT_EVALUATION_OBSERVER_LIMIT = 10;
-
-const execFileAsync = promisify(execFile);
-const UNAVAILABLE_STATUSES = new Set([
-  'projection_pending',
-  'ambiguous_parent',
-  'resync_required',
-  'error',
-]);
+export const EVALUATION_CONTEXT_AVAILABLE = 'context_available';
+export const EVALUATION_CONTEXT_UNAVAILABLE = 'context_unavailable';
 
 /**
- * Reads one bounded, proposal-time Throughline observer snapshot.
- *
- * This is deliberately separate from auditor-context.mjs: the returned snapshot
- * is evaluation evidence and must never become auditor input.
+ * Reads the exact-session Throughline context used only as proposal-time
+ * evaluation evidence. It must never become auditor input.
  */
-export async function loadEvaluationObserverContext({
+export async function loadEvaluationContext({
   projectRoot,
   host,
   sessionId,
+  transcriptPath,
   config,
   recordedAtMs = Date.now(),
   timeoutMs = DEFAULT_AUDITOR_CONTEXT_TIMEOUT_MS,
   maxBuffer = DEFAULT_AUDITOR_CONTEXT_MAX_BUFFER,
-  execFileFn = execFileAsync,
+  execFileFn,
   realpathFn = realpath,
   readConfigFn = readProjectAuditorContextConfig,
+  loadAuditorContextFn = loadAuditorContext,
 } = {}) {
   if (!Number.isSafeInteger(recordedAtMs) || recordedAtMs < 0) {
     throw new TypeError('recordedAtMs must be a non-negative safe integer');
@@ -56,6 +44,9 @@ export async function loadEvaluationObserverContext({
   if (host === 'codex' && sessionId === 'codex:') {
     throw new TypeError('sessionId must identify a thread');
   }
+  if (typeof transcriptPath !== 'string' || transcriptPath.length === 0 || !isAbsoluteProjectPath(transcriptPath)) {
+    throw new TypeError('transcriptPath must be an absolute path');
+  }
 
   let canonicalProjectRoot;
   try {
@@ -75,110 +66,42 @@ export async function loadEvaluationObserverContext({
   if (effectiveConfig?.mode !== 'throughline') {
     return unavailableResult(recordedAtMs, 'provider_disabled');
   }
-  if (!isDirectCommandConfig(effectiveConfig)) {
-    return unavailableResult(recordedAtMs, 'config_unavailable');
-  }
-
-  let stdout;
-  try {
-    ({ stdout } = await execFileFn(effectiveConfig.command, [
-      ...effectiveConfig.args,
-      'observer-read',
-      '--project', canonicalProjectRoot,
-      '--limit', String(DEFAULT_EVALUATION_OBSERVER_LIMIT),
-      '--json',
-    ], {
-      encoding: 'utf8',
-      timeout: timeoutMs,
-      maxBuffer,
-      windowsHide: true,
-      shell: false,
-    }));
-  } catch {
-    return unavailableResult(recordedAtMs, 'observer_read_failed');
-  }
 
   let snapshot;
   try {
-    snapshot = JSON.parse(stdout);
+    snapshot = await loadAuditorContextFn({
+      config: effectiveConfig,
+      host,
+      sessionId,
+      projectRoot: canonicalProjectRoot,
+      transcriptPath,
+      timeoutMs,
+      maxBuffer,
+      ...(execFileFn === undefined ? {} : { execFileFn }),
+    });
   } catch {
-    return unavailableResult(recordedAtMs, 'observer_read_invalid');
+    return unavailableResult(recordedAtMs, 'auditor_context_failed');
   }
 
-  if (UNAVAILABLE_STATUSES.has(snapshot?.status)) {
-    return unavailableResult(recordedAtMs, `observer_${snapshot.status}`);
-  }
-  if (!isObserverSnapshot(snapshot)) {
-    return unavailableResult(recordedAtMs, 'observer_read_invalid');
-  }
-  if (snapshot.host !== null && snapshot.host !== host) {
-    return unavailableResult(recordedAtMs, 'observer_host_mismatch');
-  }
-  if (snapshot.thread_sha256 !== null && snapshot.thread_sha256 !== expectedThreadHash(host, sessionId)) {
-    return unavailableResult(recordedAtMs, 'observer_session_mismatch');
+  if (snapshot.status !== 'fresh') {
+    return unavailableResult(recordedAtMs, `auditor_context_${snapshot.status}`);
   }
   return Object.freeze({
-    status: EVALUATION_OBSERVER_CONTEXT_AVAILABLE,
+    status: EVALUATION_CONTEXT_AVAILABLE,
     recordedAtMs,
-    snapshot: freezeSnapshot(snapshot),
+    snapshot,
   });
 }
 
 function unavailableResult(recordedAtMs, reason) {
   return Object.freeze({
-    status: EVALUATION_OBSERVER_CONTEXT_UNAVAILABLE,
+    status: EVALUATION_CONTEXT_UNAVAILABLE,
     recordedAtMs,
     reason,
     snapshot: null,
   });
 }
 
-function isDirectCommandConfig(config) {
-  return config && config.mode === 'throughline' &&
-    isAbsoluteProjectPath(config.command) &&
-    !/\.(?:cmd|bat)$/i.test(config.command) &&
-    Array.isArray(config.args) &&
-    config.args.every((arg) => typeof arg === 'string' && arg.length > 0);
-}
-
 function isAbsoluteProjectPath(value) {
   return isAbsolute(value) || /^[A-Za-z]:[\\/]/.test(value);
 }
-
-function isObserverSnapshot(value) {
-  if (!isRecord(value) || value.schema !== THROUGHLINE_OBSERVER_READ_SCHEMA || value.status !== 'snapshot') return false;
-  if (!Array.isArray(value.turns) || value.turns.length > DEFAULT_EVALUATION_OBSERVER_LIMIT) return false;
-  if (typeof value.historyTruncated !== 'boolean' || !isNullableString(value.afterCursor) || !isNullableString(value.throughCursor)) return false;
-  if (!isRecord(value.page) || typeof value.page.complete !== 'boolean' || !isNullableString(value.page.nextToken)) return false;
-  if (value.host === null) return value.thread_sha256 === null && value.turns.length === 0;
-  if ((value.host !== 'claude' && value.host !== 'codex') || !/^[a-f0-9]{64}$/.test(value.thread_sha256)) return false;
-  return value.turns.every((turn) => isRecord(turn) &&
-    turn.host === value.host && turn.thread_sha256 === value.thread_sha256);
-}
-
-// Throughline observer-read hashes the Claude session ID directly. Codex hook
-// session IDs may already carry Throughline's `codex:` namespace, while the
-// observer feed hashes the underlying Codex thread ID.
-function expectedThreadHash(host, sessionId) {
-  const threadId = host === 'codex' && sessionId.startsWith('codex:')
-    ? sessionId.slice('codex:'.length)
-    : sessionId;
-  if (threadId.length === 0) throw new TypeError('sessionId must identify a thread');
-  return createHash('sha256').update(threadId, 'utf8').digest('hex');
-}
-
-function freezeSnapshot(snapshot) {
-  return Object.freeze({
-    ...snapshot,
-    turns: Object.freeze(snapshot.turns.map((turn) => Object.freeze({ ...turn }))),
-    page: Object.freeze({ ...snapshot.page }),
-  });
-}
-
-function isNullableString(value) {
-  return value === null || typeof value === 'string';
-}
-
-function isRecord(value) {
-  return value !== null && typeof value === 'object' && !Array.isArray(value);
-}

package/src/core/evaluation-tool-id.mjs

@@ -167,8 +167,26 @@ function codexShellInputs(toolInput) {
 function nestedMcpToolIds(toolInput) {
   if (typeof toolInput !== 'string') return [];
   const executable = maskJavaScriptLiteralsAndComments(toolInput);
-  const matches = executable.matchAll(/\btools\.(mcp__[A-Za-z0-9_-]+__[A-Za-z0-9_.:/-]+)\s*\(/gu);
-  return [...new Set([...matches].map((match) => match[1]).filter(validCatalogId))];
+  const adopted = new Set(
+    [...executable.matchAll(/\btools\.(mcp__[A-Za-z0-9_-]+__[A-Za-z0-9_.:/-]+)\s*\(/gu)]
+      .map((match) => match[1])
+      .filter(validCatalogId),
+  );
+
+  // Codex may resolve a lazily exposed MCP tool by its exact name, then invoke it through the
+  // returned metadata object: `const tool = ALL_TOOLS.find(x => x.name === "mcp__...");
+  // await tools[tool.name](...)`. Count only bindings that are followed by that executable call.
+  const bindingPattern = /\b(?:const|let|var)\s+([A-Za-z_$][\w$]*)\s*=\s*ALL_TOOLS\.find\s*\(\s*([A-Za-z_$][\w$]*)\s*=>\s*\2\.name\s*={2,3}/gu;
+  for (const binding of executable.matchAll(bindingPattern)) {
+    let valueStart = binding.index + binding[0].length;
+    while (/\s/u.test(toolInput[valueStart] ?? '')) valueStart += 1;
+    const parsedName = parseJavaScriptStringAt(toolInput, valueStart);
+    if (!parsedName || !MCP_ID_PATTERN.test(parsedName.value)) continue;
+    const variable = escapedRegExp(binding[1]);
+    const callPattern = new RegExp(String.raw`\btools\s*\[\s*${variable}\.name\s*\]\s*\(`, 'u');
+    if (callPattern.test(executable.slice(parsedName.end))) adopted.add(parsedName.value);
+  }
+  return [...adopted];
 }
 
 function escapedRegExp(value) {

package/src/hooks/user-prompt.mjs

@@ -27,7 +27,7 @@ import { discardLegacyPending } from './pending-context.mjs';
 import { projectBackendFailure, projectParentAdvice, projectToolIds } from './parent-output-projector.mjs';
 import { randomUUID } from 'node:crypto';
 import { createEvaluationStore } from '../core/evaluation-store.mjs';
-import { loadEvaluationObserverContext } from '../core/evaluation-context.mjs';
+import { loadEvaluationContext } from '../core/evaluation-context.mjs';
 import { version } from '../version.mjs';
 
 const TIMEOUT_MS = 50_000;
@@ -39,7 +39,7 @@ export async function runUserPrompt({
   discardLegacyPendingFn = discardLegacyPending,
   recordHookEventFn = recordClaudeHookEvent,
   createEvaluationStoreFn = createEvaluationStore,
-  loadEvaluationObserverContextFn = loadEvaluationObserverContext,
+  loadEvaluationContextFn = loadEvaluationContext,
   randomUUIDFn = randomUUID,
   spotterVersion = version,
   now = Date.now,
@@ -65,11 +65,12 @@ export async function runUserPrompt({
         ? projectToolIds(Array.isArray(result?.missing_tools) ? result.missing_tools.map((entry) => entry?.name) : [])
         : [];
       const proposalRecordedAtMs = proposedToolIds.length > 0 ? now() : null;
-      const observer = proposedToolIds.length > 0
-        ? await loadEvaluationObserverContextFn({
+      const evaluationContext = proposedToolIds.length > 0
+        ? await loadEvaluationContextFn({
           projectRoot,
           host: 'claude',
           sessionId,
+          transcriptPath: input.transcript_path,
           recordedAtMs: proposalRecordedAtMs,
         })
         : { status: 'not_requested', snapshot: null };
@@ -87,8 +88,8 @@ export async function runUserPrompt({
           auditStatus,
           requestText: prompt,
           auditorSeenContext: null,
-          observerContextStatus: observer.status,
-          observerSnapshot: observer.snapshot,
+          observerContextStatus: evaluationContext.status,
+          observerSnapshot: evaluationContext.snapshot,
           proposedToolIds,
           backend: result?.evaluation_meta?.backend ?? null,
           model: result?.evaluation_meta?.model ?? null,