glassbox-trace 0.10.0 → 0.10.1

package/README.md

@@ -181,6 +181,7 @@ Six kinds of claim (tests pass, shipped, verified, fixed, nothing changed, file
 | `low-cache-hit` | < 50% of input served from cache over ≥ 5 requests |
 | `slow-model` | ≥ 60 s for a response under 1,500 tokens (latency / rate limit / stall) |
 | `long-generation` | ≥ 60 s for a big response streamed below 15 tok/s |
+| `missing-final-usage` | subagent responses whose final usage record never reached the transcript — the cost shown is a floor, with an estimate of the gap (info; a Claude Code bug, [#93620](https://github.com/anthropics/claude-code/issues/93620), not the agent's doing) |
 | `slow-tool` | a tool call ≥ 60 s (warn at 5 min); one line per tool when repeated; blocking `TaskOutput` waits roll up per background task as info; time spent on the human (questions, permission prompts) is excluded |
 | `permission-denied` | tool calls the human rejected, a permission rule or auto mode blocked, or an interrupt stopped (info) |
 | `max-tokens`, `api-error`, `hook-error`, `compaction`, `thinking-heavy`, `subagent-share`, `long-turn` | what they say (`long-turn`: ≥ 30 tool calls after one prompt in the main conversation) |
@@ -190,7 +191,7 @@ Identical findings print once with a `×N` count in `check`. All mechanical, no
 
 ## Token and cost accounting
 
-- Usage is counted once per API request (assistant records sharing a `requestId`), not once per record. Subagent transcripts rewrite `output_tokens` while a response streams, so the largest value across the request's records is used.
+- Usage is counted once per API request (assistant records sharing a `requestId`), not once per record. Subagent transcripts rewrite `output_tokens` while a response streams, so the largest value across the request's records is used. Since Claude Code ~2.1.278 many subagent responses never get their final record written at all (every record has `stop_reason: null` and a partial count — [anthropics/claude-code#93620](https://github.com/anthropics/claude-code/issues/93620)); those output tokens are not on disk anywhere, so Glassbox counts the responses, shows the cost as **≥ $X** (a floor) and estimates the gap from the median of the session's own complete subagent tool-use responses (needs five). Compare with Claude Code's own total (`/cost`, the status line) before budgeting from a floor.
 - Context size = uncached input + cache read + cache write — the prompt the model actually saw.
 - Thinking tokens are part of output tokens; shown as a share, never added twice.
 - Cost is estimated from an editable rate card (defaults from the Claude pricing page, 2026-09-05; prefix-matched so dated model ids resolve). A `stream-json` `result` record with `total_cost_usd` overrides the estimate. Unknown models show "—" rather than a wrong number.

package/dist/glassbox.html

@@ -408,7 +408,7 @@ kbd{font:11px var(--mono);border:1px solid var(--line);border-bottom-width:2px;b
 })(typeof globalThis !== 'undefined' ? globalThis : this, function () {
   'use strict';
 
-  const VERSION = '0.10.0';
+  const VERSION = '0.10.1';
 
   // ---------------------------------------------------------------------------
   // Rate card (USD per million tokens). Prefix-matched against model ids so dated
@@ -816,7 +816,7 @@ kbd{font:11px var(--mono);border:1px solid var(--line);border-bottom-width:2px;b
           let turn = currentTurn(a.id);
           if (!turn) turn = ensureTurn(rec, 'implicit', '');
           req = {
-            id: key, agent: a.id, model: msg.model || null, stopReason: null, start: t, end: t,
+            id: key, agent: a.id, model: msg.model || null, stopReason: null, start: t, end: t, version: rec.version || null,
             usage: emptyUsage(), blocks: [], turnIndex: turn.index, contextTokens: 0, records: 0,
             isApiError: !!rec.isApiErrorMessage, uuid: rec.uuid || null,
           };
@@ -831,6 +831,7 @@ kbd{font:11px var(--mono);border:1px solid var(--line);border-bottom-width:2px;b
         maxUsage(req.usage, usageFrom(msg.usage)); // turn and agent totals are summed once, after the loop
         req.records++;
         if (msg.stop_reason) req.stopReason = msg.stop_reason;
+        if (!req.version && rec.version) req.version = rec.version;
         if (rec.isApiErrorMessage) req.isApiError = true;
         const content = Array.isArray(msg.content) ? msg.content : (msg.content != null ? [{ type: 'text', text: String(msg.content) }] : []);
         const blockIndex = typeof rec.apiBlockIndex === 'number' ? rec.apiBlockIndex : req.blocks.length;
@@ -868,6 +869,10 @@ kbd{font:11px var(--mono);border:1px solid var(--line);border-bottom-width:2px;b
       req.contextTokens = req.usage.input + req.usage.cacheRead + req.usage.cacheWrite;
       addUsage(turns[req.turnIndex].usage, req.usage);
       addUsage(agents.get(req.agent).usage, req.usage);
+      // A response whose records all have stop_reason null never got its final record written, so its
+      // output_tokens is a partial running count. Only subagent transcripts are judged: in the main
+      // transcript that shape is the live, unfinished last response (anthropics/claude-code#93620).
+      req.usageFinal = !(req.agent !== 'main' && !req.stopReason && !req.isApiError && req.model !== '<synthetic>' && !isEmptyUsage(req.usage));
     }
     for (const turn of turns) if (turn.end == null) turn.end = meta.end;
     // Response timing. Assistant record timestamps are written when a block finishes streaming,
@@ -927,12 +932,28 @@ kbd{font:11px var(--mono);border:1px solid var(--line);border-bottom-width:2px;b
       toolTimeMs: toolCalls.filter((c) => c.agent === 'main').reduce((s, c) => s + (c.durationMs || 0), 0),
       modelTimeMs: agents.get('main').modelTimeMs || 0,
       cacheHitRatio: (usage.cacheRead + usage.cacheWrite + usage.input) ? usage.cacheRead / (usage.cacheRead + usage.cacheWrite + usage.input) : null,
-      attachments,
+      attachments, usagePartial: usagePartial(requests),
     };
 
     return { version: VERSION, meta, requests, toolCalls, turns, agents: agentList, events, problems, totals };
   }
 
+  // Subagent responses with no final usage record: how many, what was seen, and an estimate of the gap from the
+  // session's own complete subagent tool_use responses (median; needs at least five to go on).
+  function usagePartial(requests) {
+    const sub = requests.filter((r) => r.agent !== 'main' && r.model !== '<synthetic>' && !r.isApiError && !isEmptyUsage(r.usage));
+    const partial = sub.filter((r) => r.usageFinal === false);
+    const out = { responses: partial.length, of: sub.length, outputSeen: partial.reduce((n, r) => n + r.usage.output, 0), versions: Array.from(new Set(partial.map((r) => r.version).filter(Boolean))).sort(), agents: new Set(partial.map((r) => r.agent)).size, estimate: null };
+    if (!partial.length) return out;
+    const basis = sub.filter((r) => r.usageFinal && r.stopReason === 'tool_use').map((r) => r.usage.output).sort((a, b) => a - b);
+    if (basis.length >= 5) {
+      const mid = basis.length >> 1;
+      const median = basis.length % 2 ? basis[mid] : (basis[mid - 1] + basis[mid]) / 2;
+      out.estimate = { basis: basis.length, median, missingOutput: partial.reduce((n, r) => n + Math.max(0, median - r.usage.output), 0) };
+    }
+    return out;
+  }
+
   function pushModel(meta, m) { if (m && !meta.models.includes(m)) meta.models.push(m); }
 
   function mkOrphanResult(tr, rec, agentId, text, turnIndex) {
@@ -1218,6 +1239,21 @@ kbd{font:11px var(--mono);border:1px solid var(--line);border-bottom-width:2px;b
       if (e.kind === 'compact') push({ id: 'compaction', severity: 'info', title: 'Context was compacted', detail: e.detail, evidence: { at: e.at } });
     }
 
+    // missing-final-usage: the transcript under-counts subagent output (Claude Code bug, not the agent's doing)
+    const P = trace.totals.usagePartial;
+    if (P && P.responses > 0) {
+      const c = estimateCost(trace, o.rates);
+      const est = P.estimate
+        ? ` From the median of ${P.estimate.basis} complete subagent tool-use responses in this session (${fmtInt(P.estimate.median)} output tokens each), about ${fmtInt(P.estimate.missingOutput)} output tokens${c.missing != null ? ` (~${usd(c.missing)})` : ''} are not counted.`
+        : ' Too few complete subagent responses in this session (fewer than 5) to estimate the gap.';
+      push({
+        id: 'missing-final-usage', severity: 'info',
+        title: `Cost is a floor: ${P.responses} of ${P.of} subagent responses never recorded their final output tokens`,
+        detail: `These responses were written to the subagent transcript with stop_reason null on every record and only a partial running output_tokens count (${fmtInt(P.outputSeen)} tokens across all of them), so their real output is not on disk.${est} Claude Code ${P.versions.join(', ') || '(version not recorded)'}; known transcript issue anthropics/claude-code#93620.`,
+        evidence: { requestIds: trace.requests.filter((r) => r.usageFinal === false).slice(0, 20).map((r) => r.id) }, metric: P.responses,
+      });
+    }
+
     // thinking-heavy
     const u = trace.totals.usage;
     if (u.output >= o.thinkingHeavyMinOutput && u.thinking / u.output > o.thinkingHeavyShare) push({
@@ -1301,7 +1337,17 @@ kbd{font:11px var(--mono);border:1px solid var(--line);border-bottom-width:2px;b
       byAgent[r.agent] = (byAgent[r.agent] || 0) + c;
     }
     const complete = known === trace.requests.length;
+    // Floor: some subagent responses never recorded their final output (see usagePartial). A run-reported cost
+    // is the real bill and stays as it is.
+    const P = (trace.totals && trace.totals.usagePartial) || { responses: 0, estimate: null };
+    const lowerBound = P.responses > 0 && trace.meta.reportedCost == null;
+    let missing = null;
+    if (lowerBound && P.estimate) {
+      missing = 0;
+      for (const r of trace.requests) { if (r.usageFinal !== false) continue; const rate = rateFor(r.model, rates); if (rate) missing += Math.max(0, P.estimate.median - r.usage.output) * rate.out / 1e6; }
+    }
     return {
+      lowerBound, partialResponses: P.responses, missing,
       total: known ? total : null, complete, byRequest, byModel, byAgent, unknownModels: Array.from(unknownModels),
       reported: trace.meta.reportedCost, source: trace.meta.reportedCost != null ? 'reported' : (known ? 'estimated' : 'unknown'),
     };
@@ -1329,6 +1375,7 @@ kbd{font:11px var(--mono);border:1px solid var(--line);border-bottom-width:2px;b
     'compaction': 'Context was compacted; state that matters later (decisions, file paths, test status) should be written to a file or note before it is summarised away.',
     'thinking-heavy': 'Most output tokens were thinking; fine for hard problems, but for routine tool loops a lower effort setting is cheaper and just as good.',
     'subagent-share': 'Most spend was in subagents; give them narrower briefs and ask for short structured returns so the parent pays for less.',
+    'missing-final-usage': "Not the agent's doing: Claude Code did not write these subagent responses' final usage to disk. Treat the transcript cost as a floor and check it against Claude Code's own total (/cost or the status line) before budgeting from it.",
     'long-turn': 'A very long autonomous turn; check in with a short progress summary at natural checkpoints so wasted work is caught earlier.',
     'permission-denied': 'Before a call the user may not want (installs, network, deleting, publishing), say what it will do and ask; for calls they always allow, suggest a permissions allow rule so the prompt goes away.',
     'duplicate-subagent-read': 'When several subagents need the same file, put the relevant excerpt in their briefs or have one agent summarise it once, instead of every agent reading it into its own context.',
@@ -1435,7 +1482,9 @@ kbd{font:11px var(--mono);border:1px solid var(--line);border-bottom-width:2px;b
     out.push(`# Glassbox report — ${oneLine(R(m.title || opts.title || 'session'), 90)}`, '');
     out.push(`- session: ${m.sessionId || '—'} · model: ${m.models.join(', ') || '—'}${m.start ? ' · started: ' + new Date(m.start).toISOString() : ''}`);
     out.push(`- wall ${fmtDur(T.wallMs)} (active ${fmtDur(T.activeMs)}) · ${T.turns} turn${T.turns === 1 ? '' : 's'} · ${T.requests} requests · ${T.toolCalls} tool calls (${T.toolErrors} failed${T.orphans ? ', ' + T.orphans + ' unanswered' : ''})`);
-    out.push(`- tokens: ${fmtInt(u.input + u.cacheRead + u.cacheWrite)} context served (${pctStr(T.cacheHitRatio)} cached) · peak prompt ${fmtInt(Math.max(0, ...trace.requests.map((r) => r.contextTokens)))} · ${fmtInt(u.output)} output (${fmtInt(u.thinking)} thinking) · est. cost ${usd(total)}${cost && cost.source === 'reported' ? ' (reported)' : ''}`, '');
+    out.push(`- tokens: ${fmtInt(u.input + u.cacheRead + u.cacheWrite)} context served (${pctStr(T.cacheHitRatio)} cached) · peak prompt ${fmtInt(Math.max(0, ...trace.requests.map((r) => r.contextTokens)))} · ${fmtInt(u.output)} output (${fmtInt(u.thinking)} thinking) · est. cost ${cost && cost.lowerBound ? 'at least ' : ''}${usd(total)}${cost && cost.source === 'reported' ? ' (reported)' : ''}`);
+    if (cost && cost.lowerBound) out.push(`- cost is a floor: ${cost.partialResponses} subagent response${cost.partialResponses === 1 ? '' : 's'} never recorded their final output tokens (anthropics/claude-code#93620)${cost.missing != null ? `; estimated ~${fmtInt(T.usagePartial.estimate.missingOutput)} output tokens, ~${usd(cost.missing)} more` : ''}.`);
+    out.push('');
     if (!findings.length) out.push('## Findings', '', 'Nothing flagged — a clean session, or a very short one.', '');
     else {
       out.push(`## Findings (${findings.length})`, '');
@@ -1672,8 +1721,8 @@ kbd{font:11px var(--mono);border:1px solid var(--line);border-bottom-width:2px;b
     const t = S.trace, T = t.totals, u = T.usage, cost = S.cost;
     const ctx = u.input + u.cacheRead + u.cacheWrite;
     const errs = S.findings.filter((f) => f.severity === 'error').length, warns = S.findings.filter((f) => f.severity === 'warn').length, infos = S.findings.filter((f) => f.severity === 'info').length;
-    const costV = cost.reported != null ? fmtUSD(cost.reported) : cost.total != null ? fmtUSD(cost.total) : '—';
-    const costS = cost.reported != null ? 'reported by the run' : cost.total != null ? (cost.complete ? 'estimated from rate card' : 'partial — unknown model') : (cost.unknownModels.length ? 'unknown model: set a rate' : 'no usage data');
+    const costV = cost.reported != null ? fmtUSD(cost.reported) : cost.total != null ? (cost.lowerBound ? '≥ ' : '') + fmtUSD(cost.total) : '—';
+    const costS = cost.reported != null ? 'reported by the run' : cost.total != null ? (cost.lowerBound ? `a floor — ${cost.partialResponses} subagent response${cost.partialResponses === 1 ? '' : 's'} missing final usage${cost.missing != null ? ` · ~${fmtUSD(cost.missing)} more` : ''}` : cost.complete ? 'estimated from rate card' : 'partial — unknown model') : (cost.unknownModels.length ? 'unknown model: set a rate' : 'no usage data');
     const tiles = [
       ['Wall time', fmtDur(T.wallMs), T.humanIdleMs ? `${fmtDur(T.activeMs)} active · ${fmtDur(T.humanIdleMs)} waiting on you` : (t.meta.hasTimestamps ? 'no human idle' : 'no timestamps in file')],
       ['Turns', fmtInt(T.turns), `${fmtInt(T.requests)} API requests`],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "glassbox-trace",
-  "version": "0.10.0",
+  "version": "0.10.1",
   "description": "Other tools show what happened in a Claude Code session; Glassbox tells you what went wrong. Timeline, token and cost burn, auto-diagnostics with evidence and advice, session compare, live tail, a fleet view, a secrets scanner for transcripts (fence), a CLAUDE.md compliance report (adhere), a CLI and a Claude Code Stop hook that feeds the findings back to the agent.",
   "main": "src/trace-core.js",
   "scripts": {

package/src/cli.mjs

@@ -186,7 +186,7 @@ export function checkReport({ trace, findings, cost, claims }, opts = {}) {
   const redact = !!opts.redact;
   const legend = opts.legend || null;
   const keyed = (p) => legend ? legend.keyFor(p) : p;
-  const summary = { session: m.sessionId, title: m.title ? (redact ? '«' + m.title.length + ' chars»' : m.title) : null, model: m.models, wallMs: T.wallMs, activeMs: T.activeMs, turns: T.turns, requests: T.requests, toolCalls: T.toolCalls, toolErrors: T.toolErrors, orphans: T.orphans, usage: T.usage, cacheHitRatio: T.cacheHitRatio, cost: cost.reported != null ? cost.reported : cost.total, costSource: cost.source, start: m.start != null ? new Date(m.start).toISOString() : null, end: m.end != null ? new Date(m.end).toISOString() : null };
+  const summary = { session: m.sessionId, title: m.title ? (redact ? '«' + m.title.length + ' chars»' : m.title) : null, model: m.models, wallMs: T.wallMs, activeMs: T.activeMs, turns: T.turns, requests: T.requests, toolCalls: T.toolCalls, toolErrors: T.toolErrors, orphans: T.orphans, usage: T.usage, cacheHitRatio: T.cacheHitRatio, cost: cost.reported != null ? cost.reported : cost.total, costSource: cost.source, costLowerBound: !!cost.lowerBound, usagePartial: { responses: T.usagePartial.responses, of: T.usagePartial.of, outputSeen: T.usagePartial.outputSeen, versions: T.usagePartial.versions, estimatedMissingOutput: T.usagePartial.estimate ? T.usagePartial.estimate.missingOutput : null, estimatedMissingCost: cost.missing }, start: m.start != null ? new Date(m.start).toISOString() : null, end: m.end != null ? new Date(m.end).toISOString() : null };
   // Per-file use, keyed when a legend is given; with --redact and no legend the paths would leak, so it is omitted.
   if (legend || !redact) summary.files = core.fileStats(trace).map((r) => ({ key: legend ? legend.keyFor(r.path) : r.path, reads: r.reads, writes: r.writes, errors: r.errors, agents: r.agents, chars: r.chars }));
   if (claims) summary.claims = claims.summary; // said vs did, in one line of numbers (additive; schema stays 2)
@@ -204,7 +204,7 @@ export function checkReport({ trace, findings, cost, claims }, opts = {}) {
   const lines = [];
   lines.push(`Glassbox · ${m.sessionId ? m.sessionId.slice(0, 8) : 'session'} · ${m.models.join(', ') || 'unknown model'}`);
   lines.push(`  wall ${fmt(T.wallMs)} (active ${fmt(T.activeMs)}) · ${T.turns} turns · ${T.requests} requests · ${T.toolCalls} tool calls (${T.toolErrors} failed${T.orphans ? ', ' + T.orphans + ' unanswered' : ''})`);
-  lines.push(`  context served ${fi(T.usage.input + T.usage.cacheRead + T.usage.cacheWrite)} tokens (${T.cacheHitRatio != null ? Math.round(T.cacheHitRatio * 100) + '% cached' : 'no usage'}) · output ${fi(T.usage.output)} (${fi(T.usage.thinking)} thinking) · est. cost ${usd(summary.cost)}`);
+  lines.push(`  context served ${fi(T.usage.input + T.usage.cacheRead + T.usage.cacheWrite)} tokens (${T.cacheHitRatio != null ? Math.round(T.cacheHitRatio * 100) + '% cached' : 'no usage'}) · output ${fi(T.usage.output)} (${fi(T.usage.thinking)} thinking) · est. cost ${summary.costLowerBound ? '≥ ' : ''}${usd(summary.cost)}${summary.costLowerBound ? ' (floor)' : ''}`);
   lines.push('');
   if (!findings.length) lines.push('  no findings');
   for (const g of collapseFindings(findings)) lines.push('  ' + findingLine(g));

package/src/collect.mjs

@@ -101,6 +101,9 @@ export function collect(sources, opts = {}) {
     findings: sum(sessions, (s) => s.findings.length), errors: sum(sessions, (s) => s.findings.filter((f) => f.severity === 'error').length), warns: sum(sessions, (s) => s.findings.filter((f) => f.severity === 'warn').length),
     failed: sessions.filter((s) => s.failed).length,
     estimated: sessions.some((s) => s.summary.costSource !== 'reported'),
+    // Sessions whose transcript cost is a floor (subagent responses with no final usage, anthropics/claude-code#93620).
+    lowerBoundSessions: sessions.filter((s) => s.summary.costLowerBound).length,
+    partialResponses: sum(sessions, (s) => s.summary.usagePartial && s.summary.usagePartial.responses),
   };
   const top = topN(opts.top || 5).map((s) => ({ source: s.source, session: s.summary.session, cost: s.summary.cost || 0, share: cost ? (s.summary.cost || 0) / cost : 0, wallMs: s.summary.wallMs, turns: s.summary.turns, requests: s.summary.requests, findings: s.findings.length, worst: s.findings.slice().sort((a, b) => SEV[a.severity] - SEV[b.severity])[0] || null, model: s.summary.model }));
   return {
@@ -121,7 +124,8 @@ export function collectMarkdown(r) {
   L.push(`_${T.sources} source${T.sources === 1 ? '' : 's'} · ${T.sessions} session${T.sessions === 1 ? '' : 's'}${r.since ? ` since ${r.since.slice(0, 10)}` : ''} · generated ${r.generated.slice(0, 16).replace('T', ' ')} UTC by Glassbox ${r.glassbox}. Counts, keys and findings only — no session text._`, '');
   L.push(`## Totals`, '');
   L.push(`| | |`, `|---|---|`);
-  L.push(`| Cost${T.estimated ? ' (estimated from the rate card)' : ''} | **${usd(T.cost)}** |`);
+  L.push(`| Cost${T.estimated ? ' (estimated from the rate card)' : ''} | ${T.lowerBoundSessions ? 'at least ' : ''}**${usd(T.cost)}** |`);
+  if (T.lowerBoundSessions) L.push(`| Cost floor | ${T.lowerBoundSessions} session${T.lowerBoundSessions === 1 ? '' : 's'} a floor — ${fi(T.partialResponses)} subagent response${T.partialResponses === 1 ? '' : 's'} never recorded final usage (anthropics/claude-code#93620) |`);
   L.push(`| Wall / active time | ${fmt(T.wallMs)} / ${fmt(T.activeMs)} |`);
   L.push(`| Requests / tool calls / tool errors | ${fi(T.requests)} / ${fi(T.toolCalls)} / ${fi(T.toolErrors)} |`);
   L.push(`| Findings (error / warn) | ${fi(T.findings)} (${fi(T.errors)} / ${fi(T.warns)}) |`);
@@ -159,7 +163,7 @@ export function collectText(r) {
   const T = r.totals, fmt = core.fmtDur, fi = core.fmtInt;
   const L = [];
   L.push(`Glassbox fleet · ${T.sources} source${T.sources === 1 ? '' : 's'} · ${T.sessions} session${T.sessions === 1 ? '' : 's'}${r.since ? ` since ${r.since.slice(0, 10)}` : ''}`);
-  L.push(`  cost ${usd(T.cost)}${T.estimated ? ' (est.)' : ''} · wall ${fmt(T.wallMs)} (active ${fmt(T.activeMs)}) · ${fi(T.requests)} requests · ${fi(T.toolCalls)} tool calls (${fi(T.toolErrors)} failed)`);
+  L.push(`  cost ${T.lowerBoundSessions ? '≥ ' : ''}${usd(T.cost)}${T.estimated ? ' (est.)' : ''}${T.lowerBoundSessions ? ` (${T.lowerBoundSessions} session${T.lowerBoundSessions === 1 ? '' : 's'} a floor)` : ''} · wall ${fmt(T.wallMs)} (active ${fmt(T.activeMs)}) · ${fi(T.requests)} requests · ${fi(T.toolCalls)} tool calls (${fi(T.toolErrors)} failed)`);
   L.push(`  ${fi(T.findings)} findings (${fi(T.errors)} error, ${fi(T.warns)} warn) · ${fi(T.failed)} session${T.failed === 1 ? '' : 's'} failing`);
   if (T.sessions) L.push(`  ${r.concentration.halfOfSpendSessions} session${r.concentration.halfOfSpendSessions === 1 ? '' : 's'} = half the spend; top ${r.concentration.top.length} = ${pct(r.concentration.topShare)}`);
   L.push('');

package/src/trace-core.js

@@ -9,7 +9,7 @@
 })(typeof globalThis !== 'undefined' ? globalThis : this, function () {
   'use strict';
 
-  const VERSION = '0.10.0';
+  const VERSION = '0.10.1';
 
   // ---------------------------------------------------------------------------
   // Rate card (USD per million tokens). Prefix-matched against model ids so dated
@@ -417,7 +417,7 @@
           let turn = currentTurn(a.id);
           if (!turn) turn = ensureTurn(rec, 'implicit', '');
           req = {
-            id: key, agent: a.id, model: msg.model || null, stopReason: null, start: t, end: t,
+            id: key, agent: a.id, model: msg.model || null, stopReason: null, start: t, end: t, version: rec.version || null,
             usage: emptyUsage(), blocks: [], turnIndex: turn.index, contextTokens: 0, records: 0,
             isApiError: !!rec.isApiErrorMessage, uuid: rec.uuid || null,
           };
@@ -432,6 +432,7 @@
         maxUsage(req.usage, usageFrom(msg.usage)); // turn and agent totals are summed once, after the loop
         req.records++;
         if (msg.stop_reason) req.stopReason = msg.stop_reason;
+        if (!req.version && rec.version) req.version = rec.version;
         if (rec.isApiErrorMessage) req.isApiError = true;
         const content = Array.isArray(msg.content) ? msg.content : (msg.content != null ? [{ type: 'text', text: String(msg.content) }] : []);
         const blockIndex = typeof rec.apiBlockIndex === 'number' ? rec.apiBlockIndex : req.blocks.length;
@@ -469,6 +470,10 @@
       req.contextTokens = req.usage.input + req.usage.cacheRead + req.usage.cacheWrite;
       addUsage(turns[req.turnIndex].usage, req.usage);
       addUsage(agents.get(req.agent).usage, req.usage);
+      // A response whose records all have stop_reason null never got its final record written, so its
+      // output_tokens is a partial running count. Only subagent transcripts are judged: in the main
+      // transcript that shape is the live, unfinished last response (anthropics/claude-code#93620).
+      req.usageFinal = !(req.agent !== 'main' && !req.stopReason && !req.isApiError && req.model !== '<synthetic>' && !isEmptyUsage(req.usage));
     }
     for (const turn of turns) if (turn.end == null) turn.end = meta.end;
     // Response timing. Assistant record timestamps are written when a block finishes streaming,
@@ -528,12 +533,28 @@
       toolTimeMs: toolCalls.filter((c) => c.agent === 'main').reduce((s, c) => s + (c.durationMs || 0), 0),
       modelTimeMs: agents.get('main').modelTimeMs || 0,
       cacheHitRatio: (usage.cacheRead + usage.cacheWrite + usage.input) ? usage.cacheRead / (usage.cacheRead + usage.cacheWrite + usage.input) : null,
-      attachments,
+      attachments, usagePartial: usagePartial(requests),
     };
 
     return { version: VERSION, meta, requests, toolCalls, turns, agents: agentList, events, problems, totals };
   }
 
+  // Subagent responses with no final usage record: how many, what was seen, and an estimate of the gap from the
+  // session's own complete subagent tool_use responses (median; needs at least five to go on).
+  function usagePartial(requests) {
+    const sub = requests.filter((r) => r.agent !== 'main' && r.model !== '<synthetic>' && !r.isApiError && !isEmptyUsage(r.usage));
+    const partial = sub.filter((r) => r.usageFinal === false);
+    const out = { responses: partial.length, of: sub.length, outputSeen: partial.reduce((n, r) => n + r.usage.output, 0), versions: Array.from(new Set(partial.map((r) => r.version).filter(Boolean))).sort(), agents: new Set(partial.map((r) => r.agent)).size, estimate: null };
+    if (!partial.length) return out;
+    const basis = sub.filter((r) => r.usageFinal && r.stopReason === 'tool_use').map((r) => r.usage.output).sort((a, b) => a - b);
+    if (basis.length >= 5) {
+      const mid = basis.length >> 1;
+      const median = basis.length % 2 ? basis[mid] : (basis[mid - 1] + basis[mid]) / 2;
+      out.estimate = { basis: basis.length, median, missingOutput: partial.reduce((n, r) => n + Math.max(0, median - r.usage.output), 0) };
+    }
+    return out;
+  }
+
   function pushModel(meta, m) { if (m && !meta.models.includes(m)) meta.models.push(m); }
 
   function mkOrphanResult(tr, rec, agentId, text, turnIndex) {
@@ -819,6 +840,21 @@
       if (e.kind === 'compact') push({ id: 'compaction', severity: 'info', title: 'Context was compacted', detail: e.detail, evidence: { at: e.at } });
     }
 
+    // missing-final-usage: the transcript under-counts subagent output (Claude Code bug, not the agent's doing)
+    const P = trace.totals.usagePartial;
+    if (P && P.responses > 0) {
+      const c = estimateCost(trace, o.rates);
+      const est = P.estimate
+        ? ` From the median of ${P.estimate.basis} complete subagent tool-use responses in this session (${fmtInt(P.estimate.median)} output tokens each), about ${fmtInt(P.estimate.missingOutput)} output tokens${c.missing != null ? ` (~${usd(c.missing)})` : ''} are not counted.`
+        : ' Too few complete subagent responses in this session (fewer than 5) to estimate the gap.';
+      push({
+        id: 'missing-final-usage', severity: 'info',
+        title: `Cost is a floor: ${P.responses} of ${P.of} subagent responses never recorded their final output tokens`,
+        detail: `These responses were written to the subagent transcript with stop_reason null on every record and only a partial running output_tokens count (${fmtInt(P.outputSeen)} tokens across all of them), so their real output is not on disk.${est} Claude Code ${P.versions.join(', ') || '(version not recorded)'}; known transcript issue anthropics/claude-code#93620.`,
+        evidence: { requestIds: trace.requests.filter((r) => r.usageFinal === false).slice(0, 20).map((r) => r.id) }, metric: P.responses,
+      });
+    }
+
     // thinking-heavy
     const u = trace.totals.usage;
     if (u.output >= o.thinkingHeavyMinOutput && u.thinking / u.output > o.thinkingHeavyShare) push({
@@ -902,7 +938,17 @@
       byAgent[r.agent] = (byAgent[r.agent] || 0) + c;
     }
     const complete = known === trace.requests.length;
+    // Floor: some subagent responses never recorded their final output (see usagePartial). A run-reported cost
+    // is the real bill and stays as it is.
+    const P = (trace.totals && trace.totals.usagePartial) || { responses: 0, estimate: null };
+    const lowerBound = P.responses > 0 && trace.meta.reportedCost == null;
+    let missing = null;
+    if (lowerBound && P.estimate) {
+      missing = 0;
+      for (const r of trace.requests) { if (r.usageFinal !== false) continue; const rate = rateFor(r.model, rates); if (rate) missing += Math.max(0, P.estimate.median - r.usage.output) * rate.out / 1e6; }
+    }
     return {
+      lowerBound, partialResponses: P.responses, missing,
       total: known ? total : null, complete, byRequest, byModel, byAgent, unknownModels: Array.from(unknownModels),
       reported: trace.meta.reportedCost, source: trace.meta.reportedCost != null ? 'reported' : (known ? 'estimated' : 'unknown'),
     };
@@ -930,6 +976,7 @@
     'compaction': 'Context was compacted; state that matters later (decisions, file paths, test status) should be written to a file or note before it is summarised away.',
     'thinking-heavy': 'Most output tokens were thinking; fine for hard problems, but for routine tool loops a lower effort setting is cheaper and just as good.',
     'subagent-share': 'Most spend was in subagents; give them narrower briefs and ask for short structured returns so the parent pays for less.',
+    'missing-final-usage': "Not the agent's doing: Claude Code did not write these subagent responses' final usage to disk. Treat the transcript cost as a floor and check it against Claude Code's own total (/cost or the status line) before budgeting from it.",
     'long-turn': 'A very long autonomous turn; check in with a short progress summary at natural checkpoints so wasted work is caught earlier.',
     'permission-denied': 'Before a call the user may not want (installs, network, deleting, publishing), say what it will do and ask; for calls they always allow, suggest a permissions allow rule so the prompt goes away.',
     'duplicate-subagent-read': 'When several subagents need the same file, put the relevant excerpt in their briefs or have one agent summarise it once, instead of every agent reading it into its own context.',
@@ -1036,7 +1083,9 @@
     out.push(`# Glassbox report — ${oneLine(R(m.title || opts.title || 'session'), 90)}`, '');
     out.push(`- session: ${m.sessionId || '—'} · model: ${m.models.join(', ') || '—'}${m.start ? ' · started: ' + new Date(m.start).toISOString() : ''}`);
     out.push(`- wall ${fmtDur(T.wallMs)} (active ${fmtDur(T.activeMs)}) · ${T.turns} turn${T.turns === 1 ? '' : 's'} · ${T.requests} requests · ${T.toolCalls} tool calls (${T.toolErrors} failed${T.orphans ? ', ' + T.orphans + ' unanswered' : ''})`);
-    out.push(`- tokens: ${fmtInt(u.input + u.cacheRead + u.cacheWrite)} context served (${pctStr(T.cacheHitRatio)} cached) · peak prompt ${fmtInt(Math.max(0, ...trace.requests.map((r) => r.contextTokens)))} · ${fmtInt(u.output)} output (${fmtInt(u.thinking)} thinking) · est. cost ${usd(total)}${cost && cost.source === 'reported' ? ' (reported)' : ''}`, '');
+    out.push(`- tokens: ${fmtInt(u.input + u.cacheRead + u.cacheWrite)} context served (${pctStr(T.cacheHitRatio)} cached) · peak prompt ${fmtInt(Math.max(0, ...trace.requests.map((r) => r.contextTokens)))} · ${fmtInt(u.output)} output (${fmtInt(u.thinking)} thinking) · est. cost ${cost && cost.lowerBound ? 'at least ' : ''}${usd(total)}${cost && cost.source === 'reported' ? ' (reported)' : ''}`);
+    if (cost && cost.lowerBound) out.push(`- cost is a floor: ${cost.partialResponses} subagent response${cost.partialResponses === 1 ? '' : 's'} never recorded their final output tokens (anthropics/claude-code#93620)${cost.missing != null ? `; estimated ~${fmtInt(T.usagePartial.estimate.missingOutput)} output tokens, ~${usd(cost.missing)} more` : ''}.`);
+    out.push('');
     if (!findings.length) out.push('## Findings', '', 'Nothing flagged — a clean session, or a very short one.', '');
     else {
       out.push(`## Findings (${findings.length})`, '');

package/src/viewer.html

@@ -475,8 +475,8 @@ kbd{font:11px var(--mono);border:1px solid var(--line);border-bottom-width:2px;b
     const t = S.trace, T = t.totals, u = T.usage, cost = S.cost;
     const ctx = u.input + u.cacheRead + u.cacheWrite;
     const errs = S.findings.filter((f) => f.severity === 'error').length, warns = S.findings.filter((f) => f.severity === 'warn').length, infos = S.findings.filter((f) => f.severity === 'info').length;
-    const costV = cost.reported != null ? fmtUSD(cost.reported) : cost.total != null ? fmtUSD(cost.total) : '—';
-    const costS = cost.reported != null ? 'reported by the run' : cost.total != null ? (cost.complete ? 'estimated from rate card' : 'partial — unknown model') : (cost.unknownModels.length ? 'unknown model: set a rate' : 'no usage data');
+    const costV = cost.reported != null ? fmtUSD(cost.reported) : cost.total != null ? (cost.lowerBound ? '≥ ' : '') + fmtUSD(cost.total) : '—';
+    const costS = cost.reported != null ? 'reported by the run' : cost.total != null ? (cost.lowerBound ? `a floor — ${cost.partialResponses} subagent response${cost.partialResponses === 1 ? '' : 's'} missing final usage${cost.missing != null ? ` · ~${fmtUSD(cost.missing)} more` : ''}` : cost.complete ? 'estimated from rate card' : 'partial — unknown model') : (cost.unknownModels.length ? 'unknown model: set a rate' : 'no usage data');
     const tiles = [
       ['Wall time', fmtDur(T.wallMs), T.humanIdleMs ? `${fmtDur(T.activeMs)} active · ${fmtDur(T.humanIdleMs)} waiting on you` : (t.meta.hasTimestamps ? 'no human idle' : 'no timestamps in file')],
       ['Turns', fmtInt(T.turns), `${fmtInt(T.requests)} API requests`],