fix(agent-health): 日次確認の報告にツールごとの起動モデルを残す

changelog.d/3438.md

@@ -0,0 +1,2 @@
+<!-- ### Fixed -->
+- **fix(agent-health): 日次確認の報告にツールごとの起動モデルを残す** (#3438): pass の check は evidence を残さないため、どのモデルで確かめたかが報告から読めなかった。レポートの各ツールに `launchedModel`（画面・hook から読めたモデル。読めなければ「不明」。opencode / opencode-v2 は複製した `model.json` の `recent[0]` も並べる）を足し、要約の最後に「起動したモデル:」を出す。この項目の無い古いレポートは「不明」として読む。

docs/user-guide/agent-health.md

@@ -153,6 +153,28 @@ pass 38・fail 0・skip 25（検査の定義が無い 5・このツールは、
 
 `--tools` / `--only` で外したマスも表には `skip（今回の実行の対象外）` として数える（理由の行には出さない）。
 
+最後に、起動したツールごとのモデルを 1 行ずつ出す（Issue #3438。pass の check は evidence を残さないので、どのモデルで確かめたかはここで読む）。
+
+```
+起動したモデル:
+- claude: Haiku 4.5（画面）
+- antigravity: gemini-3.5-flash-low（hook）
+- opencode: Claude Sonnet 5.5（画面・設定: github-copilot/claude-sonnet-5.5）
+- opencode-v2: Mistral Large 4 Ollama Cloud（画面下端・プロバイダ名を含む・設定: ollama-cloud/mistral-large-4）
+- command-code: 不明
+```
+
+モデルは、ツール自身が見せたものだけを読む（`src/lib/agent-health/launched-model.ts`）。推測では埋めない。
+
+| 読む場所 | ツール | 括弧内 |
+|---|---|---|
+| 画面（本番の `model-info-extractor`。claude のバナー、codex の下端、antigravity のバー、command-code の `# models:` 行、opencode の `▣  Build · <モデル>` の行。opencode-v2 は `▣` の無い `Build · <モデル> · 721ms` の行） | 全ツール | `画面` |
+| hook の payload の model（claude の `SessionStart`、antigravity の `modelName` など） | hook を使うツール | `hook` |
+| opencode の入力欄の下の `Build · <モデル> <プロバイダ>`（ターン前で上の行が無いとき。ANSI を除くとモデルとプロバイダの境目が無いので、まとめて残す） | opencode・opencode-v2 | `画面下端・プロバイダ名を含む` |
+
+上の順に、最初に読めたものを使う。どれも読めなければ `不明`。`設定:` は起動前に複製した `model.json` の `recent[0]`（`<providerID>/<modelID>`）で、
+読めた値と見比べるために並べる（代わりには使わない）。version だけを読むツールは起動しないので出さない。
+
 ## レポートの形
 
 子 Issue #2879 の AI はこれだけを読む。フィールドの追加はよいが、名前の変更・削除は #2879 と #2880 を壊す。
@@ -178,6 +200,11 @@ interface AgentHealthReport {
       skipKind?: 'no-definition' | 'not-shown' | 'signed-out' | 'unsupported' | 'timeout' | 'prerequisite-failed'; // skip には必ず入る
       framePaths?: string[];        // 画面全体を保存したファイル（下の「画面の保存」）
     }>;
+    launchedModel?: {               // 起動したモデル（Issue #3438）。起動しなかったツールと #3438 より前のレポートには無い（読む側は「不明」）
+      model: string | null;         // 読めなければ null（不明）
+      source?: 'screen' | 'screen-footer' | 'hook'; // model を読んだ場所
+      seeded?: string;              // 複製した model.json の recent[0]（opencode / opencode-v2）
+    };
   }>;
   safety: {
     globalConfigRestored: Array<{

scripts/agent-health/main.ts

@@ -62,7 +62,7 @@ import {
 } from '@/lib/agent-health/types';
 import { HookListener } from './hook-listener';
 import { locateServerLog, ServerLogWatch } from './production-log';
-import { probeLimitedTool, probeTool, readVersion } from './probe-tool';
+import { probeLimitedTool, probeTool, readVersion, type ProbeOutcome } from './probe-tool';
 import { AgentHealthTmux } from './tmux-driver';
 import { LIMITED_TOOL_SPECS, TOOL_PROBE_SPECS } from './tool-table';
 
@@ -319,7 +319,7 @@ async function run(options: AgentHealthOptions, startedAt: Date): Promise<AgentH
         readPickerSettings(readTextIfPresent(pickerSettings.path), pickerSettings.format, pickerSettings.keys);
       const settingsBefore = readSettings();
 
-      let outcome: { version: string | null; checks: AgentHealthCheck[] };
+      let outcome: ProbeOutcome;
       try {
         outcome = await probeTool({
           spec,
@@ -373,7 +373,15 @@ async function run(options: AgentHealthOptions, startedAt: Date): Promise<AgentH
         ];
       }
 
-      results.push(buildToolResult({ tool, version: outcome.version, previousVersion, checks: outcome.checks }));
+      results.push(
+        buildToolResult({
+          tool,
+          version: outcome.version,
+          previousVersion,
+          checks: outcome.checks,
+          launchedModel: outcome.launchedModel,
+        })
+      );
     }
 
     // Issue #3313: the tools the probe does not launch are rows too — on the

scripts/agent-health/probe-tool.ts

@@ -24,6 +24,15 @@ import {
   expectedHookEvents,
 } from '@/lib/agent-health/hook-correlation';
 import { skipCheck } from '@/lib/agent-health/coverage';
+import {
+  laterScreenModel,
+  launchedModelLabel,
+  readHookModel,
+  readScreenModel,
+  readSeededModel,
+  resolveLaunchedModel,
+  type ScreenModelReading,
+} from '@/lib/agent-health/launched-model';
 import { firstVersionLine, paneEvidence } from '@/lib/agent-health/report';
 import { archiveCheckFrames, type FrameArchive, type JudgedFrame } from '@/lib/agent-health/frame-archive';
 import { judgeServerEvents, judgeTurnScreen, turnEndedUnauthorized } from '@/lib/agent-health/model-auth';
@@ -43,6 +52,7 @@ import {
   probeInstanceId,
   type AgentHealthCheck,
   type AgentHealthCheckId,
+  type AgentHealthLaunchedModel,
 } from '@/lib/agent-health/types';
 import type { AgentEventSource } from '@/lib/hooks/sources/types';
 import type { HookListener } from './hook-listener';
@@ -99,6 +109,8 @@ export interface ProbeContext {
 export interface ProbeOutcome {
   version: string | null;
   checks: AgentHealthCheck[];
+  /** Issue #3438: set once the tool was launched; left out when it never was. */
+  launchedModel?: AgentHealthLaunchedModel;
 }
 
 const sleep = (ms: number) => new Promise((resolve) => setTimeout(resolve, Math.max(0, ms)));
@@ -162,6 +174,14 @@ export function seedStateFiles(seeds: Array<{ from: string; to: string }>): stri
   return copied;
 }
 
+function readTextOrNull(file: string): string | null {
+  try {
+    return fs.readFileSync(file, 'utf8');
+  } catch {
+    return null;
+  }
+}
+
 /** Everything one session needs; one instance per tool. */
 class ToolSession {
   readonly name: string;
@@ -173,6 +193,8 @@ class ToolSession {
   runningTurnUnauthorized = false;
   /** Any turn did. */
   anyTurnUnauthorized = false;
+  /** The model the frames looked at so far show (Issue #3438). */
+  screenModel: ScreenModelReading | null = null;
 
   constructor(
     private readonly ctx: ProbeContext,
@@ -221,6 +243,7 @@ class ToolSession {
   async look(): Promise<{ frame: string; clean: string; verdict: ScreenVerdict }> {
     const frame = await this.ctx.tmux.capture(this.name, this.spec.captureLines);
     this.lastFrame = frame;
+    this.screenModel = laterScreenModel(this.screenModel, readScreenModel(this.spec.cliToolId, frame));
     const result = detectSessionStatus(frame, this.spec.cliToolId);
     return {
       frame,
@@ -639,8 +662,10 @@ export async function probeTool(ctx: ProbeContext): Promise<ProbeOutcome> {
     }
   }
   const command = buildProbeLaunchCommand(spec, rendered, workDir);
+  let seededModel: string | null = null;
   for (const seeded of seedStateFiles(spec.seedFiles?.(workDir) ?? [])) {
     ctx.log(`${spec.tool}: seeded ${seeded}`);
+    if (path.basename(seeded) === 'model.json') seededModel = readSeededModel(readTextOrNull(seeded));
   }
   ctx.log(`${spec.tool}: launching — ${command}`);
 
@@ -732,6 +757,13 @@ export async function probeTool(ctx: ProbeContext): Promise<ProbeOutcome> {
     }
   }
 
+  const launchedModel = resolveLaunchedModel({
+    screen: session.screenModel,
+    hook: readHookModel(spec.cliToolId, ctx.listener.forTool(spec.tool)),
+    seeded: seededModel,
+  });
+  ctx.log(`${spec.tool}: model ${launchedModelLabel(launchedModel)}`);
+
   if (wantHooks) {
     const verdict = evaluateHookCorrelation(ctx.listener.forTool(spec.tool), {
       tool: spec.tool,
@@ -742,7 +774,7 @@ export async function probeTool(ctx: ProbeContext): Promise<ProbeOutcome> {
     session.record({ checkId: 'hook-correlation', ...verdict });
   }
 
-  return { version, checks: [...checks, ...session.checks.values()] };
+  return { version, checks: [...checks, ...session.checks.values()], launchedModel };
 }
 
 /**

src/lib/agent-health/coverage.ts

@@ -27,6 +27,7 @@ import {
   type AgentHealthSkipKind,
   type AgentHealthToolResult,
 } from './types';
+import { launchedModelLabel } from './launched-model';
 
 /** A `skip` check. Every place that skips goes through this, so the kind is never left out. */
 export function skipCheck(
@@ -152,5 +153,12 @@ export function summarizeCoverage(
     }
   }
   if (reasons.length > 0) lines.push('', '未実施の理由:', ...reasons);
+
+  // Issue #3438: a pass keeps no evidence, so this is where the model a tool
+  // ran on is told. Version-only rows are never launched.
+  const models = coverage.rows
+    .filter((row) => row.coverage === 'probed' && results.some((entry) => entry.tool === row.tool))
+    .map((row) => `- ${row.tool}: ${launchedModelLabel(results.find((entry) => entry.tool === row.tool)?.launchedModel)}`);
+  if (models.length > 0) lines.push('', '起動したモデル:', ...models);
   return lines;
 }

src/lib/agent-health/launched-model.ts

@@ -0,0 +1,163 @@
+/**
+ * Which model each probed tool was launched on (Issue #3438).
+ *
+ * A `pass` check keeps no evidence, so before this the report could not say
+ * which model a passing tool ran on — not even whether opencode-v2 picked up
+ * the user's `model.json` (#3428). The value is read only where the tool
+ * itself shows it, never guessed:
+ *
+ * - the screen, through the production reader
+ *   (`@/lib/detection/model-info-extractor`): claude's banner, codex's footer,
+ *   antigravity's bar, command-code's `# models:` row, opencode's `▣  Build ·
+ *   <model>` step row. opencode2 draws the step row without `▣`
+ *   (`Build · <model> · 721ms`), read here with the same shape.
+ * - opencode's composer bar (`┃  Build · <model> <provider>`), the last resort:
+ *   after `stripAnsi` the model and the provider are one string (see
+ *   `OPENCODE_STEP_MODEL_PATTERN`'s comment), so it is kept whole and marked
+ *   `screen-footer`.
+ * - the hooks the probe received (`NormalizedAgentEvent.model`: claude's
+ *   `SessionStart`, antigravity's `modelName`, …).
+ *
+ * What none of them shows is `null` — read as 不明. `seeded` is the setting
+ * the probe copied in (`model.json`'s `recent[0]`), kept beside the read value
+ * so the two can be compared, never used in its place.
+ */
+
+import { stripAnsi } from '@/lib/detection/ansi';
+import { extractModelInfo } from '@/lib/detection/model-info-extractor';
+import type { CLIToolType } from '@/lib/cli-tools/types';
+import { isAgentEventType } from '@/lib/hooks/agent-event-types';
+import { getAgentEventSource } from '@/lib/hooks/sources';
+import type { AgentHealthLaunchedModel, AgentHealthLaunchedModelSource } from './types';
+
+/** What a report shows for a tool whose model could not be read, or an old report without the field. */
+export const UNKNOWN_LAUNCHED_MODEL = '不明';
+
+const SOURCE_LABELS: Record<AgentHealthLaunchedModelSource, string> = {
+  screen: '画面',
+  'screen-footer': '画面下端・プロバイダ名を含む',
+  hook: 'hook',
+};
+
+const MAX_MODEL_CHARS = 64;
+
+/**
+ * opencode2's step row (`     Build · Mistral Large 4 · 721ms`): v1's
+ * `▣  Build · <model> · <duration>` without the `▣`. The duration is required,
+ * so a line that merely has two dots does not read as one.
+ */
+const OPENCODE_STEP_ROW_PATTERN = /^\s*(?:▣\s+)?[A-Za-z][A-Za-z0-9-]*\s+·\s+([^·]+?)\s+·\s+[\d.]+\s?m?s\s*$/;
+
+/** opencode's composer bar: `┃  Build · <model> <provider>[ · <variant>]`. */
+const OPENCODE_COMPOSER_BAR_PATTERN = /^\s*┃\s+[A-Za-z][A-Za-z0-9-]*\s+·\s+([^·]+?)(?:\s+·\s+\S+)?\s*$/;
+
+export interface ScreenModelReading {
+  model: string;
+  source: 'screen' | 'screen-footer';
+}
+
+function plausible(value: string): string | null {
+  const trimmed = value.trim();
+  return trimmed !== '' && trimmed.length <= MAX_MODEL_CHARS ? trimmed : null;
+}
+
+function scanFromEnd(frame: string, pattern: RegExp): string | null {
+  const lines = frame.split('\n');
+  for (let i = lines.length - 1; i >= 0; i--) {
+    const match = pattern.exec(stripAnsi(lines[i]));
+    const model = match ? plausible(match[1]) : null;
+    if (model) return model;
+  }
+  return null;
+}
+
+/** The model one frame shows, or null when it shows none. */
+export function readScreenModel(cliToolId: CLIToolType, frame: string): ScreenModelReading | null {
+  if (!frame) return null;
+  const opencode = cliToolId === 'opencode' || cliToolId === 'opencode-v2';
+  const read = extractModelInfo(opencode ? 'opencode' : cliToolId, frame).model;
+  if (read) return { model: read, source: 'screen' };
+  if (!opencode) return null;
+  const step = scanFromEnd(frame, OPENCODE_STEP_ROW_PATTERN);
+  if (step) return { model: step, source: 'screen' };
+  const bar = scanFromEnd(frame, OPENCODE_COMPOSER_BAR_PATTERN);
+  return bar ? { model: bar, source: 'screen-footer' } : null;
+}
+
+/**
+ * The reading to keep after a new frame: a step row or banner (`screen`) is
+ * never replaced by the composer bar; otherwise the newer frame wins.
+ */
+export function laterScreenModel(
+  previous: ScreenModelReading | null,
+  next: ScreenModelReading | null
+): ScreenModelReading | null {
+  if (!next) return previous;
+  if (previous?.source === 'screen' && next.source === 'screen-footer') return previous;
+  return next;
+}
+
+/** The first model a hook payload names, in arrival order (claude's `SessionStart` comes first). */
+export function readHookModel(
+  cliToolId: CLIToolType,
+  deliveries: ReadonlyArray<{ kind: string; event: string | null; body: Record<string, unknown> | null }>
+): string | null {
+  for (const delivery of deliveries) {
+    if (delivery.kind !== 'agent-event' || delivery.body === null) continue;
+    try {
+      const normalized = getAgentEventSource(cliToolId).normalizeEvent({
+        payload: delivery.body,
+        event: isAgentEventType(delivery.event) ? delivery.event : null,
+      });
+      const model = normalized?.model ? plausible(normalized.model) : null;
+      if (model) return model;
+    } catch {
+      // A payload the source cannot read names no model.
+    }
+  }
+  return null;
+}
+
+/** `<providerID>/<modelID>` of `model.json`'s `recent[0]`, or null. */
+export function readSeededModel(text: string | null): string | null {
+  if (text === null) return null;
+  try {
+    const parsed = JSON.parse(text) as { recent?: unknown };
+    if (!Array.isArray(parsed?.recent)) return null;
+    const first = parsed.recent[0] as { providerID?: unknown; modelID?: unknown } | undefined;
+    if (typeof first?.modelID !== 'string' || first.modelID === '') return null;
+    return typeof first.providerID === 'string' && first.providerID !== ''
+      ? `${first.providerID}/${first.modelID}`
+      : first.modelID;
+  } catch {
+    return null;
+  }
+}
+
+/** The screen's step row or banner, then the hooks, then the composer bar; null when none. */
+export function resolveLaunchedModel(input: {
+  screen: ScreenModelReading | null;
+  hook: string | null;
+  seeded?: string | null;
+}): AgentHealthLaunchedModel {
+  const seeded = input.seeded ? { seeded: input.seeded } : {};
+  if (input.screen?.source === 'screen') return { model: input.screen.model, source: 'screen', ...seeded };
+  if (input.hook) return { model: input.hook, source: 'hook', ...seeded };
+  if (input.screen) return { model: input.screen.model, source: input.screen.source, ...seeded };
+  return { model: null, ...seeded };
+}
+
+/**
+ * One line for the summary: `Haiku 4.5（画面）`, `不明`, `不明（設定: github-copilot/…）`.
+ * A report written before #3438 has no `launchedModel` and reads 不明.
+ */
+export function launchedModelLabel(launched: AgentHealthLaunchedModel | null | undefined): string {
+  const model = typeof launched?.model === 'string' && launched.model !== '' ? launched.model : null;
+  const source = launched?.source && SOURCE_LABELS[launched.source] ? SOURCE_LABELS[launched.source] : null;
+  const notes = [
+    ...(model && source ? [source] : []),
+    ...(typeof launched?.seeded === 'string' && launched.seeded !== '' ? [`設定: ${launched.seeded}`] : []),
+  ];
+  const head = model ?? UNKNOWN_LAUNCHED_MODEL;
+  return notes.length > 0 ? `${head}（${notes.join('・')}）` : head;
+}

src/lib/agent-health/report.ts

@@ -12,6 +12,7 @@ import {
   EVIDENCE_PANE_LINES,
   MAX_EVIDENCE_CHARS,
   type AgentHealthCheck,
+  type AgentHealthLaunchedModel,
   type AgentHealthReport,
   type AgentHealthState,
   type AgentHealthReportTool,
@@ -68,6 +69,8 @@ export function buildToolResult(input: {
   version: string | null;
   previousVersion: string | null;
   checks: readonly AgentHealthCheck[];
+  /** Issue #3438: left out when the tool was not launched. */
+  launchedModel?: AgentHealthLaunchedModel;
 }): AgentHealthToolResult {
   return {
     tool: input.tool,
@@ -78,6 +81,7 @@ export function buildToolResult(input: {
       ...check,
       ...(check.evidence !== undefined ? { evidence: truncateEvidence(check.evidence) } : {}),
     })),
+    ...(input.launchedModel ? { launchedModel: input.launchedModel } : {}),
   };
 }