@dzhechkov/harness-core 0.8.45 → 0.8.47

package/src/workflow-run-dispatch.ts

@@ -60,7 +60,18 @@ export interface DispatchResult {
   /** null when the runtime did not report a count — NEVER 0, never estimated (ADR-004 C4). */
   tokensIn: number | null;
   tokensOut: number | null;
-  tokensSource: 'claude-envelope' | 'codex-stderr' | null;
+  tokensSource: 'claude-envelope' | 'codex-stderr' | 'codex-json' | null;
+  tokensTotal?: number | null;
+  tokensCacheRead?: number | null;
+  tokensCacheWrite?: number | null;
+  tokensReasoning?: number | null;
+  reportedTotalBasis?: 'raw-inclusive' | 'uncached-display' | 'reported-unknown' | 'output-only' | 'unknown';
+  inputCacheSemantics?: 'includes-cache-read-write' | 'excludes-cache-read-write' | 'uncached-display' | 'unknown';
+  reportedCostUsd?: number | null;
+  usageDiagnostics?: string[];
+  totalDerivation?: string;
+  modelProvenance?: string;
+  usageSource?: { schema: string; scope: string; threadId: string | null; turnId: string | null; receiptId: string | null };
   failure?: DispatchFailure;
 }
 
@@ -181,7 +192,7 @@ export { CODEX_EXEC_PROMPT_CEILING_CHARS };
 export function codexExecArgv(modelId: string, prompt: string, deliverable: Deliverable): string[] {
   const returnValue = (deliverable ?? 'return-value') !== 'file';
   const text = returnValue ? CODEX_SCOPING_PREFIX + '\n\n' + prompt : prompt;
-  return ['exec', '-m', modelId, ...(returnValue ? ['--sandbox', 'read-only'] : []), text];
+  return ['exec', '-m', modelId, ...(returnValue ? ['--sandbox', 'read-only'] : []), '--json', text];
 }
 
 /** The liveness probe: an allowlist says an id is SPELLABLE, only a probe says it ANSWERS. */
@@ -205,55 +216,130 @@ export function interpretCodexProbe(out: { stdout: string; exitCode: number | nu
  * builds/configs do print one; it is tested, not assumed. (Named consequence: codex runs report no
  * token counts today. That is the honest state, not a bug to paper over — see the manifest.)
  */
-export function extractCodexTokens(stderr: string): { tokensIn: number | null; tokensOut: number | null } {
-  const text = String(stderr ?? '');
-  const num = (raw: string | undefined): number | null => {
-    if (raw === undefined) return null;
-    const n = Number(raw.replace(/[,_\s]/g, ''));
-    return Number.isFinite(n) && n >= 0 ? n : null;
+type ObservedUsage = Pick<DispatchResult, 'tokensIn' | 'tokensOut' | 'tokensSource' | 'tokensTotal' | 'tokensCacheRead' | 'tokensCacheWrite' | 'tokensReasoning' | 'reportedTotalBasis' | 'inputCacheSemantics' | 'reportedCostUsd' | 'usageDiagnostics' | 'usageSource' | 'totalDerivation'>;
+const count = (value: unknown): number | null => typeof value === 'number' && Number.isSafeInteger(value) && value >= 0 ? value : null;
+const emptyUsage = (): ObservedUsage => ({ tokensIn: null, tokensOut: null, tokensTotal: null, tokensCacheRead: null,
+  tokensCacheWrite: null, tokensReasoning: null, tokensSource: null, reportedTotalBasis: 'unknown', inputCacheSemantics: 'unknown',
+  reportedCostUsd: null, usageDiagnostics: [] });
+
+export function extractCodexTokens(stderr: string): ObservedUsage {
+  const text = String(stderr ?? ''); const usage = emptyUsage();
+  const numeric = (raw: string | undefined): number | null => raw === undefined ? null : count(Number(raw.replace(/[,_\s]/g, '')));
+  const input = /\binput\b[^\n\d+.-]{0,20}([+-]?\d[\d,_]*(?:\.\d+)?)|\btokens?\s+in\b[^\n\d+.-]{0,10}([+-]?\d[\d,_]*(?:\.\d+)?)/i.exec(text);
+  const output = /\boutput\b[^\n\d+.-]{0,20}([+-]?\d[\d,_]*(?:\.\d+)?)|\btokens?\s+out\b[^\n\d+.-]{0,10}([+-]?\d[\d,_]*(?:\.\d+)?)/i.exec(text);
+  const total = /tokens used\s*\n\s*([+-]?\d[\d,_]*(?:\.\d+)?)|Token usage:\s*total=([+-]?\d[\d,_]*(?:\.\d+)?)/i.exec(text);
+  usage.tokensIn = numeric(input?.[1] ?? input?.[2]); usage.tokensOut = numeric(output?.[1] ?? output?.[2]);
+  usage.tokensTotal = numeric(total?.[1] ?? total?.[2]);
+  if (input && usage.tokensIn === null) usage.usageDiagnostics!.push('invalid-counter:input');
+  if (output && usage.tokensOut === null) usage.usageDiagnostics!.push('invalid-counter:output');
+  if (total && usage.tokensTotal === null) usage.usageDiagnostics!.push('invalid-counter:total');
+  const cached = /\(\+([\d,_]+) cached\)/i.exec(text);
+  if (total?.[2] && cached) usage.tokensCacheRead = numeric(cached[1]);
+  usage.reportedTotalBasis = total?.[2] ? 'uncached-display' : 'reported-unknown';
+  usage.inputCacheSemantics = total?.[2] ? 'uncached-display' : 'unknown';
+  usage.tokensSource = usage.tokensTotal != null || usage.tokensIn != null || usage.tokensOut != null ? 'codex-stderr' : null;
+  usage.totalDerivation = usage.tokensTotal === null ? 'not-recorded' : 'reported';
+  return usage;
+}
+
+export function extractClaudeUsage(stdout: string): ObservedUsage {
+  let chosen: Record<string, unknown> | null = null;
+  for (const line of [String(stdout).trim(), ...String(stdout).split(/\r?\n/)]) {
+    try { const obj = JSON.parse(line) as Record<string, unknown>; if (obj && obj['type'] === 'result') chosen = obj; } catch { /* non-envelope line */ }
+  }
+  const result = emptyUsage(); if (!chosen) return result;
+  const usage = chosen['usage']; if (!usage || typeof usage !== 'object' || Array.isArray(usage)) return result;
+  const raw = usage as Record<string, unknown>;
+  const invalid = (obj: Record<string, unknown>, key: string) => obj[key] != null && count(obj[key]) === null;
+  const field = (obj: Record<string, unknown>, key: string): number | null => {
+    if (invalid(obj, key)) result.usageDiagnostics!.push('invalid-count:' + key);
+    return count(obj[key]);
   };
-  // Two accepted spellings per half: `input tokens: N` / `input: N tokens` and `tokens in: N`.
-  const inMatch = /\binput\b[^\n\d]{0,20}([\d,_]+)|\btokens?\s+in\b[^\n\d]{0,10}([\d,_]+)/i.exec(text);
-  const outMatch = /\boutput\b[^\n\d]{0,20}([\d,_]+)|\btokens?\s+out\b[^\n\d]{0,10}([\d,_]+)/i.exec(text);
-  const tokensIn = inMatch === null ? null : num(inMatch[1] ?? inMatch[2]);
-  const tokensOut = outMatch === null ? null : num(outMatch[1] ?? outMatch[2]);
-  return { tokensIn, tokensOut };
+  result.tokensIn = field(raw, 'input_tokens'); result.tokensOut = field(raw, 'output_tokens');
+  result.tokensCacheRead = field(raw, 'cache_read_input_tokens'); result.tokensCacheWrite = field(raw, 'cache_creation_input_tokens');
+  const creation = raw['cache_creation'];
+  if (creation != null) {
+    if (typeof creation !== 'object' || Array.isArray(creation)) {
+      result.usageDiagnostics!.push('invalid-cache-creation-shape'); result.tokensCacheWrite = null;
+    } else {
+      const nested = creation as Record<string, unknown>;
+      const a = field(nested, 'ephemeral_5m_input_tokens'); const b = field(nested, 'ephemeral_1h_input_tokens');
+      const nestedInvalid = invalid(nested, 'ephemeral_5m_input_tokens') || invalid(nested, 'ephemeral_1h_input_tokens');
+      const total = a !== null && b !== null ? count(a + b) : null;
+      if (a !== null && b !== null && total === null) result.usageDiagnostics!.push('cache-creation-overflow');
+      if (nestedInvalid || (a !== null && b !== null && total === null)) result.tokensCacheWrite = null;
+      else if (raw['cache_creation_input_tokens'] == null) result.tokensCacheWrite = total;
+      else if (result.tokensCacheWrite !== null && total !== null && total !== result.tokensCacheWrite) {
+        result.usageDiagnostics!.push('cache-creation-mismatch'); result.tokensCacheWrite = null;
+      }
+    }
+  }
+  result.tokensReasoning = field(raw, 'reasoning_output_tokens');
+  if (result.tokensReasoning !== null && result.tokensOut !== null && result.tokensReasoning > result.tokensOut) {
+    result.usageDiagnostics!.push('reasoning-exceeds-output'); result.tokensReasoning = null;
+  }
+  result.tokensTotal = field(raw, 'total_tokens');
+  const components = [result.tokensIn, result.tokensCacheRead, result.tokensCacheWrite, result.tokensOut];
+  const partitionKnown = components.every((n) => n !== null);
+  const derived = partitionKnown ? count(components.reduce<number>((n, v) => n + (v ?? 0), 0)) : null;
+  if (partitionKnown && derived === null) result.usageDiagnostics!.push('total-overflow');
+  if (invalid(raw, 'total_tokens')) result.totalDerivation = 'invalid-reported-total';
+  else if (raw['total_tokens'] == null) {
+    result.tokensTotal = derived; result.totalDerivation = derived === null ? 'not-recorded' : 'disjoint-dimensions';
+  } else result.totalDerivation = 'reported';
+  if (derived !== null && result.tokensTotal !== null && result.tokensTotal !== derived) result.usageDiagnostics!.push('total-split-mismatch');
+  result.reportedCostUsd = typeof chosen['total_cost_usd'] === 'number' && Number.isFinite(chosen['total_cost_usd']) && chosen['total_cost_usd'] >= 0 ? chosen['total_cost_usd'] : null;
+  result.reportedTotalBasis = 'raw-inclusive'; result.inputCacheSemantics = 'excludes-cache-read-write'; result.tokensSource = 'claude-envelope';
+  return result;
 }
 
-/**
- * The claude `--output-format json` USAGE fields, pinned to the LIVE envelope shape.
- *
- * MEASURED this session (`claude -p --output-format json`, sonnet):
- * `usage.input_tokens = 2`, `usage.output_tokens = 4`, alongside `cache_read_input_tokens` and
- * `cache_creation_input_tokens`. Only the two plain counters are reported — cache tokens are a
- * SEPARATE dimension `wf-budget-1` has no field for, and silently folding them into `tokensIn`
- * would inflate every cached run's cost picture. Added to the architecture's export list; see the
- * manifest.
- */
-export function extractClaudeUsage(stdout: string): { tokensIn: number | null; tokensOut: number | null } {
-  const env = extractClaudeResult(String(stdout ?? ''));
-  if (!env.ok) return { tokensIn: null, tokensOut: null };
-  // re-scan for the envelope object itself: extractClaudeResult hands back only the text
-  const lines = String(stdout ?? '').split(/\r?\n/);
-  const candidates = [String(stdout ?? '').trim(), ...lines.map((l) => l.trim())].filter((c) => c.startsWith('{') && c.endsWith('}'));
-  let usage: Record<string, unknown> | null = null;
-  for (const c of candidates) {
-    let obj: unknown;
-    try {
-      obj = JSON.parse(c);
-    } catch {
-      continue;
+function codexStructuredOutput(stdout: string): { structured: boolean; valid: boolean; text: string; usage: ObservedUsage; reportedModel?: string | null } {
+  const lines = stdout.split(/\r?\n/).filter((line) => line.trim());
+  const events: Record<string, unknown>[] = []; let malformed = false; let recognized = false;
+  const types = new Set(['thread.started', 'turn.started', 'turn.completed', 'turn.failed', 'item.started', 'item.updated', 'item.completed', 'error']);
+  for (const line of lines) {
+    if (/"type"\s*:\s*"(?:thread\.|turn\.|item\.|error")/.test(line)) recognized = true;
+    try { const event = JSON.parse(line) as Record<string, unknown>;
+      if (event && typeof event === 'object' && typeof event['type'] === 'string' && (types.has(event['type']) || /^(thread|turn|item)\./.test(event['type']))) recognized = true;
+      events.push(event);
+    } catch { malformed = true; }
+  }
+  if (!recognized) return { structured: false, valid: true, text: stdout.trim(), usage: emptyUsage() };
+  const result = emptyUsage(); let failedTerminal = false; let text = ''; let threadId: string | null = null; const terminals: Record<string, unknown>[] = []; const reportedModels = new Set<string>();
+  for (const event of events) {
+    if (!event || typeof event !== 'object' || !types.has(String(event['type']))) { malformed = true; continue; }
+    if (typeof event['model'] === 'string' && event['model'].trim()) reportedModels.add(event['model']);
+    if (event['type'] === 'thread.started' && typeof event['thread_id'] === 'string') { if (threadId !== null && threadId !== event['thread_id']) malformed = true; threadId = event['thread_id']; }
+    if (event['type'] === 'item.completed' && event['item'] && typeof event['item'] === 'object') {
+      const item = event['item'] as Record<string, unknown>; if (item['type'] === 'agent_message' && typeof item['text'] === 'string') text = item['text'];
     }
-    if (obj === null || typeof obj !== 'object') continue;
-    const u = (obj as Record<string, unknown>)['usage'];
-    if (u !== null && typeof u === 'object' && !Array.isArray(u)) usage = u as Record<string, unknown>;
+    if (event['type'] === 'turn.completed' || (event['type'] === 'turn.failed' && event['usage'] != null)) terminals.push(event);
+    if (event['type'] === 'turn.failed' || event['type'] === 'error') failedTerminal = true;
   }
-  if (usage === null) return { tokensIn: null, tokensOut: null };
-  const pick = (k: string): number | null => {
-    const v = usage?.[k];
-    return typeof v === 'number' && Number.isFinite(v) && v >= 0 ? v : null;
-  };
-  return { tokensIn: pick('input_tokens'), tokensOut: pick('output_tokens') };
+  const distinct = new Set(terminals.map((v) => { const u = v['usage'] && typeof v['usage'] === 'object' ? v['usage'] as Record<string,unknown> : {}; return JSON.stringify([v['turn_id'] ?? null, ...['input_tokens','cached_input_tokens','cache_write_input_tokens','output_tokens','reasoning_output_tokens','total_tokens'].map((key) => u[key] ?? null)]); }));
+  if (distinct.size !== 1) { malformed = true; result.usageDiagnostics!.push(distinct.size === 0 ? 'terminal-usage-unavailable' : 'conflicting-terminal-usage'); }
+  if (distinct.size === 1) {
+    const terminal = terminals[0]!; const raw = terminal['usage'];
+    if (raw && typeof raw === 'object' && !Array.isArray(raw)) {
+      const usage = raw as Record<string, unknown>;
+      result.tokensIn = count(usage['input_tokens']); result.tokensOut = count(usage['output_tokens']);
+      result.tokensCacheRead = count(usage['cached_input_tokens']); result.tokensCacheWrite = count(usage['cache_write_input_tokens']);
+      result.tokensReasoning = count(usage['reasoning_output_tokens']);
+      const derived = result.tokensIn !== null && result.tokensOut !== null ? count(result.tokensIn + result.tokensOut) : null;
+      result.tokensTotal = count(usage['total_tokens']) ?? derived;
+      result.totalDerivation = count(usage['total_tokens']) !== null ? 'reported' : derived === null ? 'not-recorded' : 'input-plus-output';
+      if (derived !== null && result.tokensTotal !== derived) result.usageDiagnostics!.push('total-split-mismatch');
+      if (result.tokensCacheRead != null && result.tokensIn != null && result.tokensCacheRead > result.tokensIn) result.usageDiagnostics!.push('cache-exceeds-input');
+      if (result.tokensReasoning != null && result.tokensOut != null && result.tokensReasoning > result.tokensOut) result.usageDiagnostics!.push('reasoning-exceeds-output');
+      for (const [key, value] of Object.entries(usage)) if (key.endsWith('_tokens') && count(value) === null) result.usageDiagnostics!.push('invalid-counter:' + key);
+      result.tokensSource = 'codex-json'; result.reportedTotalBasis = 'raw-inclusive'; result.inputCacheSemantics = 'includes-cache-read-write';
+      result.usageSource = { schema: 'codex-exec-json', scope: 'terminal-turn', threadId, turnId: typeof terminal['turn_id'] === 'string' ? terminal['turn_id'] : null, receiptId: null };
+    } else malformed = true;
+  }
+  if (malformed) result.usageDiagnostics!.push('malformed-event-stream');
+  if (failedTerminal) result.usageDiagnostics!.push('runtime-turn-failed');
+  if (reportedModels.size > 1) result.usageDiagnostics!.push('provider-model-mismatch');
+  return { structured: true, valid: !malformed && !failedTerminal, text, usage: result, reportedModel: reportedModels.size === 1 ? [...reportedModels][0]! : null };
 }
 
 // ── the adapter factories (pure over the injected ChildRunner) ───────────────────────────────────
@@ -264,6 +350,7 @@ function failed(family: BridgeFamily, model: string | null, wallMs: number, reas
     text: null,
     family,
     modelUsed: model,
+    modelProvenance: 'probed-request',
     wallMs,
     tokensIn: null,
     tokensOut: null,
@@ -314,25 +401,15 @@ export function makeCodexExecDispatcher(run: ChildRunner, opts?: { bin?: string;
         cwd: req.cwd,
         detached: true,
       });
+      const parsed = codexStructuredOutput(String(r.stdout ?? ''));
+      const usage = parsed.structured ? parsed.usage : extractCodexTokens(r.stderr);
       const cls = classifyChildRun(r);
-      if (cls.kind === 'timeout') return failed('openai', req.resolvedModelId, req.timeoutMs, 'dispatch-timeout', `the ${req.timeoutMs}ms deadline fired on step ${req.stepId}`);
-      if (cls.kind === 'dead') return failed('openai', req.resolvedModelId, clock() - t0, 'dispatch-dead', cls.detail);
-      const text = String(r.stdout ?? '').trim();
-      if (cls.exitCode !== 0 || text === '') {
-        return failed('openai', req.resolvedModelId, clock() - t0, 'dispatch-dead',
-          `codex exited ${cls.exitCode} with ${text === '' ? 'NO stdout' : String(text.length) + ' chars of stdout'} — a clean exit with nothing to read is the spawned-but-mute case, not an empty success`);
-      }
-      const tokens = extractCodexTokens(r.stderr);
-      return {
-        outcome: 'ok',
-        text,
-        family: 'openai',
-        modelUsed: req.resolvedModelId,
-        wallMs: clock() - t0,
-        tokensIn: tokens.tokensIn,
-        tokensOut: tokens.tokensOut,
-        tokensSource: tokens.tokensIn === null && tokens.tokensOut === null ? null : 'codex-stderr',
-      };
+      if (cls.kind === 'timeout') return { ...failed('openai', req.resolvedModelId, req.timeoutMs, 'dispatch-timeout', `the ${req.timeoutMs}ms deadline fired on step ${req.stepId}`), ...usage };
+      if (cls.kind === 'dead') return { ...failed('openai', req.resolvedModelId, clock() - t0, 'dispatch-dead', cls.detail), ...usage };
+      if (cls.exitCode !== 0 || !parsed.valid || parsed.text === '') return { ...failed('openai', req.resolvedModelId, clock() - t0, 'dispatch-dead', 'Codex exited nonzero, had an invalid structured event stream, or was spawned-but-mute with no final text'), ...usage };
+      return { outcome: 'ok', text: parsed.text, family: 'openai', modelUsed: parsed.reportedModel ?? req.resolvedModelId,
+        modelProvenance: parsed.reportedModel ? 'provider-reported' : 'probed-request', wallMs: clock() - t0, ...usage };
+
     },
   };
 }
@@ -380,33 +457,15 @@ export function makeClaudePDispatcher(run: ChildRunner, opts?: { bin?: string; i
         cwd: fileMode ? req.cwd : isolatedCwd(),
         detached: true,
       });
+      const usage = extractClaudeUsage(r.stdout);
       const cls = classifyChildRun(r);
-      if (cls.kind === 'timeout') return failed('claude', req.resolvedModelId, req.timeoutMs, 'dispatch-timeout', `the ${req.timeoutMs}ms deadline fired on step ${req.stepId}`);
-      if (cls.kind === 'dead') return failed('claude', req.resolvedModelId, clock() - t0, 'dispatch-dead', cls.detail);
-      // The EXIT CODE and the ENVELOPE must agree (Step-8 HIGH-10). A parseable success envelope
-      // from a process that exited nonzero is a CONTRADICTION, not a success: the runtime told us
-      // twice and the two answers differ, so believing the friendlier one is how a failed dispatch
-      // becomes a green step. The codex adapter already required exit 0; this one did not.
-      if (cls.exitCode !== 0) {
-        return failed('claude', req.resolvedModelId, clock() - t0, 'dispatch-dead',
-          `claude exited ${cls.exitCode} — a nonzero exit is a failed dispatch even when stdout carries a parseable success envelope; the two disagree and the exit code is the runtime's own verdict`);
-      }
+      if (cls.kind === 'timeout') return { ...failed('claude', req.resolvedModelId, req.timeoutMs, 'dispatch-timeout', `the ${req.timeoutMs}ms deadline fired on step ${req.stepId}`), ...usage };
+      if (cls.kind === 'dead') return { ...failed('claude', req.resolvedModelId, clock() - t0, 'dispatch-dead', cls.detail), ...usage };
+      if (cls.exitCode !== 0) return { ...failed('claude', req.resolvedModelId, clock() - t0, 'dispatch-dead', `claude exited ${cls.exitCode} — a nonzero exit is a failed dispatch even with a success envelope`), ...usage };
       const env = extractClaudeResult(r.stdout);
-      if (!env.ok) {
-        return failed('claude', req.resolvedModelId, clock() - t0, 'dispatch-dead',
-          `claude exited ${cls.exitCode} but the reply is not readable as a result envelope: ${env.detail}`);
-      }
-      const usage = extractClaudeUsage(r.stdout);
-      return {
-        outcome: 'ok',
-        text: env.text,
-        family: 'claude',
-        modelUsed: req.resolvedModelId,
-        wallMs: clock() - t0,
-        tokensIn: usage.tokensIn,
-        tokensOut: usage.tokensOut,
-        tokensSource: usage.tokensIn === null && usage.tokensOut === null ? null : 'claude-envelope',
-      };
+      if (!env.ok) return { ...failed('claude', req.resolvedModelId, clock() - t0, 'dispatch-dead', `reply is not readable as a result envelope: ${env.detail}`), ...usage };
+      return { outcome: 'ok', text: env.text, family: 'claude', modelUsed: req.resolvedModelId, modelProvenance: 'probed-request', wallMs: clock() - t0, ...usage };
+
     },
   };
 }

package/src/workflow-run.ts

@@ -164,7 +164,8 @@ export interface WfRunState {
   dispatcherOverride?: boolean;
 }
 
-export interface WfBudgetRow {
+export interface WfBudgetRow extends Pick<DispatchResult, 'tokensTotal' | 'tokensCacheRead' | 'tokensCacheWrite' | 'tokensReasoning' | 'reportedTotalBasis' | 'inputCacheSemantics' | 'reportedCostUsd' | 'usageDiagnostics' | 'usageSource' | 'totalDerivation'> {
+  projectRoot?: string;
   schema: typeof WF_BUDGET_ROW_SCHEMA;
   kind: 'stage' | 'probe';
   runId: string;
@@ -175,10 +176,17 @@ export interface WfBudgetRow {
   attempt: number | null;
   family: BridgeFamily;
   model: string | null;
+  modelProvenance?: string;
+  requestedModel?: string | null;
   wallMs: number;
   tokensIn: number | null;
   tokensOut: number | null;
-  tokensSource: 'claude-envelope' | 'codex-stderr' | null;
+  tokensSource: DispatchResult['tokensSource'];
+  phase?: string | null;
+  role?: string | null;
+  tier?: string | null;
+  mode?: string | null;
+  estimate?: unknown;
   outcome: 'ok' | 'null' | 'error' | null;
   timeoutMs: number | null;
 }
@@ -239,6 +247,7 @@ export interface RunnerInputs {
   wallClockExtraMs: number | null;
   runnerVersion: string;
   cwdRoot: string;
+  projectRoot?: string;
 }
 
 /** Small, dependency-free 64-bit FNV — the same shape the checkpoint plane uses, kept local so this
@@ -1063,6 +1072,7 @@ export async function runWorkflow(inputs: RunnerInputs, pre: PreflightOk, deps:
       schema: WF_BUDGET_ROW_SCHEMA,
       kind: 'probe',
       runId: inputs.runId,
+      projectRoot: inputs.projectRoot ?? inputs.cwdRoot,
       dispatchSeq: null,
       stepId: null,
       itemKey: null,
@@ -1073,6 +1083,14 @@ export async function runWorkflow(inputs: RunnerInputs, pre: PreflightOk, deps:
       tokensIn: null,
       tokensOut: null,
       tokensSource: null,
+      tokensTotal: null,
+      tokensCacheRead: null,
+      tokensCacheWrite: null,
+      tokensReasoning: null,
+      reportedTotalBasis: 'unknown',
+      inputCacheSemantics: 'unknown',
+      reportedCostUsd: null,
+      usageDiagnostics: ['probe-usage-not-recorded'],
       outcome: null,
       timeoutMs: null,
     });
@@ -1500,16 +1518,34 @@ async function dispatchOnce(
       schema: WF_BUDGET_ROW_SCHEMA,
       kind: 'stage',
       runId: ctx.inputs.runId,
+      projectRoot: ctx.inputs.projectRoot ?? ctx.inputs.cwdRoot,
       dispatchSeq,
       stepId: spec.stepId,
       itemKey,
       attempt,
       family,
-      model: res.modelUsed ?? model,
+      model: res.modelUsed,
+      requestedModel: model,
+      modelProvenance: res.modelProvenance ?? (res.modelUsed === null ? 'not-recorded' : 'dispatcher-reported'),
       wallMs,
       tokensIn: res.tokensIn,
       tokensOut: res.tokensOut,
       tokensSource: res.tokensSource,
+      tokensTotal: res.tokensTotal ?? null,
+      totalDerivation: res.totalDerivation ?? 'not-recorded',
+      tokensCacheRead: res.tokensCacheRead ?? null,
+      tokensCacheWrite: res.tokensCacheWrite ?? null,
+      tokensReasoning: res.tokensReasoning ?? null,
+      reportedTotalBasis: res.reportedTotalBasis ?? 'unknown',
+      inputCacheSemantics: res.inputCacheSemantics ?? 'unknown',
+      reportedCostUsd: res.reportedCostUsd ?? null,
+      usageDiagnostics: res.usageDiagnostics ?? [],
+      ...(res.usageSource !== undefined ? { usageSource: res.usageSource } : {}),
+      phase: spec.phase,
+      role: spec.role ?? null,
+      tier: spec.tier ?? null,
+      mode: spec.mode ?? null,
+      estimate: spec.estimate ?? null,
       outcome,
       timeoutMs,
     });
@@ -1699,7 +1735,7 @@ function appendLedger(ctx: RunCtx, outcome: string): void {
     outcome,
     date: null,
   });
-  if (line !== null) ctx.deps.store.appendLedgerLine(line);
+  if (line !== null) ctx.deps.store.appendLedgerLine(JSON.stringify({ ...JSON.parse(line), summary: true, sourceProjection: 'workflow-budget', workflowRunId: ctx.inputs.runId }));
 }
 
 /** PAUSE — flush WITHOUT `run.closed` (parity with the render's top-level terminal return, which

package/dist/ledger-cost-fill.d.ts

@@ -1,58 +0,0 @@
-/**
- * Заполнение пустых полей стоимости в леджере прогонов ИЗ ЗАПИСИ ХОСТА.
- *
- * ЗАЧЕМ. Запись 3c026d5d: стоимость фичи на три четверти набита руками. ИЗМЕРЕНО 2026-09-19 по
- * живому `.dz/feature-adr/run-cost-ledger.jsonl`: 393 строки, из них 293 без числа токенов.
- * Числа при этом НЕ пропали — для прогонов Workflow они лежат в записи хоста
- * `~/.claude/projects/<проект>/<сессия>/workflows/wf_<runId>.json`, где есть `totalTokens`,
- * `durationMs` и `agentCount`. Песочница воркфлоу их видеть не может (у неё нет файловой системы
- * и нет уведомления о завершении), поэтому соединение делается СНАРУЖИ и ПОСЛЕ прогона.
- *
- * ЧЕСТНАЯ ДОСЯГАЕМОСТЬ, и она меньше, чем подразумевает запись. Из 293 пустых строк:
- *   • 75 — прогоны Workflow (`auto:true`); только их и может достать это соединение,
- *          и лишь 25 из них несут `runId`, по которому запись хоста находится СЕГОДНЯ;
- *   • 218 — ручные круги ведущего, для которых записи хоста НЕ СУЩЕСТВУЕТ ПО ПОСТРОЕНИЮ:
- *          Workflow не запускался, и стоимость этим путём не восстановима в принципе.
- * То есть механизм закрывает не «три четверти», а максимум 26% разрыва, а сегодня 9%.
- * Остальное — задача ПИСАТЕЛЯ (проставлять пригодный `runId`) и отдельный вопрос про то,
- * чем мерить ручные круги.
- *
- * ПРОВЕНАНС ОБЯЗАТЕЛЕН. Заполненное поле помечается `costSource: 'derived'`, потому что
- * выведенное число и набранное руками — разные по доверию величины, и строка, где их нельзя
- * различить, обесценивает обе.
- */
-/** Стоимость прогона, вычитанная из записи хоста. */
-export interface DerivedRunCost {
-    readonly tokens: number;
-    readonly minutes: number;
-    readonly agents: number | null;
-}
-/** Строка леджера в той части, которая нас касается. */
-export interface LedgerCostRow {
-    readonly runId?: unknown;
-    readonly tokens?: unknown;
-    readonly minutes?: unknown;
-    readonly agents?: unknown;
-    readonly costSource?: unknown;
-    readonly [k: string]: unknown;
-}
-/**
- * Прочитать стоимость из записи хоста. `null`, если запись не несёт обязательных чисел —
- * частичная запись НЕ даёт частичного заполнения: половина выведенной строки хуже пустой,
- * потому что выглядит как измерение.
- */
-export declare function readHostRunCost(record: unknown): DerivedRunCost | null;
-/**
- * Заполнить ТОЛЬКО пустые поля. Уже набранное руками число НИКОГДА не перетирается: оно могло
- * быть исправлено человеком по другим основаниям, и выведенное значение не имеет права его
- * вытеснять. Возвращает `null`, если заполнять нечего — чтобы вызывающий не переписывал файл зря.
- */
-export declare function fillLedgerCost(row: LedgerCostRow, derived: DerivedRunCost): LedgerCostRow | null;
-/** Три исхода на строку — и «нечего заполнять» отделено от «не нашли запись». */
-export type FillOutcome = 'filled' | 'nothing-to-fill' | 'no-host-record';
-/** Решение по одной строке: чем она разрешается и почему. */
-export declare function decideFill(row: LedgerCostRow, record: unknown): {
-    outcome: FillOutcome;
-    next: LedgerCostRow | null;
-};
-//# sourceMappingURL=ledger-cost-fill.d.ts.map

package/dist/ledger-cost-fill.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"ledger-cost-fill.d.ts","sourceRoot":"","sources":["../src/ledger-cost-fill.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAEH,qDAAqD;AACrD,MAAM,WAAW,cAAc;IAC7B,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;CAChC;AAED,wDAAwD;AACxD,MAAM,WAAW,aAAa;IAC5B,QAAQ,CAAC,KAAK,CAAC,EAAE,OAAO,CAAC;IACzB,QAAQ,CAAC,MAAM,CAAC,EAAE,OAAO,CAAC;IAC1B,QAAQ,CAAC,OAAO,CAAC,EAAE,OAAO,CAAC;IAC3B,QAAQ,CAAC,MAAM,CAAC,EAAE,OAAO,CAAC;IAC1B,QAAQ,CAAC,UAAU,CAAC,EAAE,OAAO,CAAC;IAC9B,QAAQ,EAAE,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC;CAC/B;AAID;;;;GAIG;AACH,wBAAgB,eAAe,CAAC,MAAM,EAAE,OAAO,GAAG,cAAc,GAAG,IAAI,CAUtE;AAED;;;;GAIG;AACH,wBAAgB,cAAc,CAAC,GAAG,EAAE,aAAa,EAAE,OAAO,EAAE,cAAc,GAAG,aAAa,GAAG,IAAI,CAUhG;AAED,iFAAiF;AACjF,MAAM,MAAM,WAAW,GAAG,QAAQ,GAAG,iBAAiB,GAAG,gBAAgB,CAAC;AAE1E,6DAA6D;AAC7D,wBAAgB,UAAU,CAAC,GAAG,EAAE,aAAa,EAAE,MAAM,EAAE,OAAO,GAAG;IAAE,OAAO,EAAE,WAAW,CAAC;IAAC,IAAI,EAAE,aAAa,GAAG,IAAI,CAAA;CAAE,CAKpH"}

package/dist/ledger-cost-fill.js

@@ -1,78 +0,0 @@
-/**
- * Заполнение пустых полей стоимости в леджере прогонов ИЗ ЗАПИСИ ХОСТА.
- *
- * ЗАЧЕМ. Запись 3c026d5d: стоимость фичи на три четверти набита руками. ИЗМЕРЕНО 2026-09-19 по
- * живому `.dz/feature-adr/run-cost-ledger.jsonl`: 393 строки, из них 293 без числа токенов.
- * Числа при этом НЕ пропали — для прогонов Workflow они лежат в записи хоста
- * `~/.claude/projects/<проект>/<сессия>/workflows/wf_<runId>.json`, где есть `totalTokens`,
- * `durationMs` и `agentCount`. Песочница воркфлоу их видеть не может (у неё нет файловой системы
- * и нет уведомления о завершении), поэтому соединение делается СНАРУЖИ и ПОСЛЕ прогона.
- *
- * ЧЕСТНАЯ ДОСЯГАЕМОСТЬ, и она меньше, чем подразумевает запись. Из 293 пустых строк:
- *   • 75 — прогоны Workflow (`auto:true`); только их и может достать это соединение,
- *          и лишь 25 из них несут `runId`, по которому запись хоста находится СЕГОДНЯ;
- *   • 218 — ручные круги ведущего, для которых записи хоста НЕ СУЩЕСТВУЕТ ПО ПОСТРОЕНИЮ:
- *          Workflow не запускался, и стоимость этим путём не восстановима в принципе.
- * То есть механизм закрывает не «три четверти», а максимум 26% разрыва, а сегодня 9%.
- * Остальное — задача ПИСАТЕЛЯ (проставлять пригодный `runId`) и отдельный вопрос про то,
- * чем мерить ручные круги.
- *
- * ПРОВЕНАНС ОБЯЗАТЕЛЕН. Заполненное поле помечается `costSource: 'derived'`, потому что
- * выведенное число и набранное руками — разные по доверию величины, и строка, где их нельзя
- * различить, обесценивает обе.
- */
-const isFiniteNumber = (v) => typeof v === 'number' && Number.isFinite(v);
-/**
- * Прочитать стоимость из записи хоста. `null`, если запись не несёт обязательных чисел —
- * частичная запись НЕ даёт частичного заполнения: половина выведенной строки хуже пустой,
- * потому что выглядит как измерение.
- */
-export function readHostRunCost(record) {
-    if (record === null || typeof record !== 'object')
-        return null;
-    const r = record;
-    if (!isFiniteNumber(r['totalTokens']) || !isFiniteNumber(r['durationMs']))
-        return null;
-    if (r['totalTokens'] < 0 || r['durationMs'] < 0)
-        return null;
-    return {
-        tokens: r['totalTokens'],
-        minutes: Math.round(r['durationMs'] / 60_000),
-        agents: isFiniteNumber(r['agentCount']) ? r['agentCount'] : null,
-    };
-}
-/**
- * Заполнить ТОЛЬКО пустые поля. Уже набранное руками число НИКОГДА не перетирается: оно могло
- * быть исправлено человеком по другим основаниям, и выведенное значение не имеет права его
- * вытеснять. Возвращает `null`, если заполнять нечего — чтобы вызывающий не переписывал файл зря.
- */
-export function fillLedgerCost(row, derived) {
-    const next = { ...row };
-    let filled = false;
-    if (!isFiniteNumber(row['tokens'])) {
-        next['tokens'] = derived.tokens;
-        filled = true;
-    }
-    if (!isFiniteNumber(row['minutes'])) {
-        next['minutes'] = derived.minutes;
-        filled = true;
-    }
-    if (!isFiniteNumber(row['agents']) && derived.agents !== null) {
-        next['agents'] = derived.agents;
-        filled = true;
-    }
-    if (!filled)
-        return null;
-    // Провенанс ставится ТОЛЬКО когда что-то действительно заполнено.
-    next['costSource'] = 'derived';
-    return next;
-}
-/** Решение по одной строке: чем она разрешается и почему. */
-export function decideFill(row, record) {
-    const derived = record === undefined ? null : readHostRunCost(record);
-    if (derived === null)
-        return { outcome: 'no-host-record', next: null };
-    const next = fillLedgerCost(row, derived);
-    return next === null ? { outcome: 'nothing-to-fill', next: null } : { outcome: 'filled', next };
-}
-//# sourceMappingURL=ledger-cost-fill.js.map

package/dist/ledger-cost-fill.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"ledger-cost-fill.js","sourceRoot":"","sources":["../src/ledger-cost-fill.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAmBH,MAAM,cAAc,GAAG,CAAC,CAAU,EAAe,EAAE,CAAC,OAAO,CAAC,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC;AAEhG;;;;GAIG;AACH,MAAM,UAAU,eAAe,CAAC,MAAe;IAC7C,IAAI,MAAM,KAAK,IAAI,IAAI,OAAO,MAAM,KAAK,QAAQ;QAAE,OAAO,IAAI,CAAC;IAC/D,MAAM,CAAC,GAAG,MAAiC,CAAC;IAC5C,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC;QAAE,OAAO,IAAI,CAAC;IACvF,IAAI,CAAC,CAAC,aAAa,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,YAAY,CAAC,GAAG,CAAC;QAAE,OAAO,IAAI,CAAC;IAC7D,OAAO;QACL,MAAM,EAAE,CAAC,CAAC,aAAa,CAAC;QACxB,OAAO,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,YAAY,CAAC,GAAG,MAAM,CAAC;QAC7C,MAAM,EAAE,cAAc,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,IAAI;KACjE,CAAC;AACJ,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,cAAc,CAAC,GAAkB,EAAE,OAAuB;IACxE,MAAM,IAAI,GAA4B,EAAE,GAAG,GAAG,EAAE,CAAC;IACjD,IAAI,MAAM,GAAG,KAAK,CAAC;IACnB,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC,EAAE,CAAC;QAAC,IAAI,CAAC,QAAQ,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC;QAAC,MAAM,GAAG,IAAI,CAAC;IAAC,CAAC;IACvF,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC,EAAE,CAAC;QAAC,IAAI,CAAC,SAAS,CAAC,GAAG,OAAO,CAAC,OAAO,CAAC;QAAC,MAAM,GAAG,IAAI,CAAC;IAAC,CAAC;IAC1F,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC,IAAI,OAAO,CAAC,MAAM,KAAK,IAAI,EAAE,CAAC;QAAC,IAAI,CAAC,QAAQ,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC;QAAC,MAAM,GAAG,IAAI,CAAC;IAAC,CAAC;IAClH,IAAI,CAAC,MAAM;QAAE,OAAO,IAAI,CAAC;IACzB,kEAAkE;IAClE,IAAI,CAAC,YAAY,CAAC,GAAG,SAAS,CAAC;IAC/B,OAAO,IAAqB,CAAC;AAC/B,CAAC;AAKD,6DAA6D;AAC7D,MAAM,UAAU,UAAU,CAAC,GAAkB,EAAE,MAAe;IAC5D,MAAM,OAAO,GAAG,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,eAAe,CAAC,MAAM,CAAC,CAAC;IACtE,IAAI,OAAO,KAAK,IAAI;QAAE,OAAO,EAAE,OAAO,EAAE,gBAAgB,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC;IACvE,MAAM,IAAI,GAAG,cAAc,CAAC,GAAG,EAAE,OAAO,CAAC,CAAC;IAC1C,OAAO,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,iBAAiB,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC;AAClG,CAAC"}