@cat-factory/executor-harness 1.64.2 → 1.66.0

package/dist/agent-runner.js

@@ -2,7 +2,8 @@ import { spawn } from 'node:child_process';
 import { mkdir, mkdtemp, rm, writeFile } from 'node:fs/promises';
 import { tmpdir } from 'node:os';
 import { dirname, join } from 'node:path';
-import { claudeAssistantContent, claudeCallUsage, isObject, numberOf, redactBody, } from './claude-stream.js';
+import { claudeAssistantContent, isObject, numberOf, redactBody } from './claude-stream.js';
+import { createClaudeRunTelemetry, subagentDispatchId } from './claude-call-aggregator.js';
 import { createCallMetricPublisher, publishCallMetric, } from './pi.js';
 import { ProgressGuard } from './progress-guard.js';
 import { killChildProcess, spawnDetached } from './process.js';
@@ -214,30 +215,39 @@ export async function runClaudeCode(opts) {
         });
     }
     // Reconstruct the full per-call request/response bodies for telemetry from the
-    // stream. `--output-format stream-json --verbose` emits each turn as a near-verbatim
-    // Anthropic Messages envelope, so `assistant` events carry the complete response
-    // (text + tool_use blocks + usage), and `user` events carry the tool_result blocks
-    // fed back — together the growing prompt transcript. We seed it with the inputs the
-    // harness supplies (they never appear in the stream): the system + first user message
-    // when the prompt rides argv, or a single folded user turn when it doesn't — so the
-    // reconstruction never shows a system turn that was never sent. Bodies are
-    // credential-scrubbed (they can echo the leased token).
+    // stream. `--output-format stream-json --verbose` emits a near-verbatim Anthropic
+    // Messages envelope per response CONTENT BLOCK (not per call), so the aggregator below
+    // folds the envelopes sharing a `message.id` back into one call and buffers that call's
+    // `user` tool_result turns — together the growing prompt transcript, in the shape the
+    // model was actually sent. We seed it with the inputs the harness supplies (they never
+    // appear in the stream): the system + first user message when the prompt rides argv, or
+    // a single folded user turn when it doesn't — so the reconstruction never shows a system
+    // turn that was never sent. Bodies are credential-scrubbed (they can echo the leased token).
     const secrets = opts.subscriptionToken ? secretsToRedact(opts.subscriptionToken) : [];
-    const messages = folded
-        ? [{ role: 'user', content: prompt }]
-        : [
-            { role: 'system', content: opts.systemPrompt },
-            { role: 'user', content: opts.userPrompt },
-        ];
     const calls = [];
     // Streams each call as the CLI yields it, EXCEPT one whose tokens `attributeCumulativeUsage`
     // may still rewrite below (a published call must be final — see the publisher).
     const publisher = createCallMetricPublisher(calls, opts.onCallMetric);
+    // `watcherOwnsSubagents` tracks the `startSubagentWatcher` wiring below: it is started only when
+    // the CLI has an isolated config home to watch, which an `ambientAuth` run does not have. The
+    // telemetry routes the CLI's tagged subagent turns accordingly — see `createClaudeRunTelemetry`.
+    const telemetry = createClaudeRunTelemetry({
+        seed: folded
+            ? [{ role: 'user', content: prompt }]
+            : [
+                { role: 'system', content: opts.systemPrompt },
+                { role: 'user', content: opts.userPrompt },
+            ],
+        secrets,
+        watcherOwnsSubagents: !opts.ambientAuth,
+        publish: (metric) => publisher.publish(metric),
+    });
     // ADR 0026 D2.1 + ADR 0027 Defect B: surface live slice progress from the two views the run
     // produces of the SAME slicing. The parent's subagent dispatches + their terminal tool_results
-    // DO appear on this stream (only a subagent's intermediate turns don't), so `sliceTracker`
-    // knows which slices are in flight and which have returned; the parent's own plan (tracked by
-    // `planTracker` + `lastTodo`) is the only place a not-yet-dispatched slice is named at all.
+    // appear on this stream (as do the subagents' own intermediate turns, tagged with the dispatch
+    // that spawned them — see `isSubagentEvent`), so `sliceTracker` knows which slices are in flight
+    // and which have returned; the parent's own plan (tracked by `planTracker` + `lastTodo`) is the
+    // only place a not-yet-dispatched slice is named at all.
     //
     // The plan arrives in one of two tool vocabularies depending on the bundled CLI build:
     // `TodoWrite` (whole-list snapshots, tracked in `lastTodo`) or the incremental
@@ -295,12 +305,19 @@ export async function runClaudeCode(opts) {
     };
     const onEvent = (event, meta) => {
         const type = event.type;
+        // A subagent's turns ride the parent's stdout tagged with the dispatch that spawned them;
+        // `telemetry` routes them off the parent's chain (and decides who bills them). Progress, slice
+        // tracking, the guard and `stats` below deliberately see EVERY event: a subagent grinding on
+        // errors should trip the guard exactly as the parent would, and whether the agent acted at all
+        // does not depend on which channel billed it.
+        const dispatchId = subagentDispatchId(event);
         if (type === 'assistant' && isObject(event.message)) {
             const message = event.message;
             const content = Array.isArray(message.content) ? message.content : [];
-            const { text, reasoning, toolUses } = claudeAssistantContent(content);
+            const { text, toolUses } = claudeAssistantContent(content);
             stats.assistantChars += text.length;
             stats.toolCalls += toolUses;
+            telemetry.onAssistant(dispatchId, message);
             for (const block of content) {
                 if (!isObject(block) || block.type !== 'tool_use')
                     continue;
@@ -318,22 +335,6 @@ export async function runClaudeCode(opts) {
             sliceTracker.onAssistant(content);
             planTracker.onAssistant(content);
             emitProgress();
-            // Record this call BEFORE appending its turn: the prompt is the history that
-            // produced this response. The append-only array keeps each call's prompt a strict
-            // prefix of the next, so the backend's telemetry chain delta-compresses cleanly.
-            const u = claudeCallUsage(message.usage);
-            publisher.publish({
-                ...(typeof message.model === 'string' ? { model: message.model } : {}),
-                promptText: redactBody(JSON.stringify(messages), secrets),
-                messageCount: messages.length,
-                responseText: redactBody(text, secrets),
-                reasoningText: redactBody(reasoning, secrets),
-                inputTokens: u.inputTokens,
-                cachedInputTokens: u.cachedInputTokens,
-                outputTokens: u.outputTokens,
-                finishReason: typeof message.stop_reason === 'string' ? message.stop_reason : null,
-            });
-            messages.push({ role: 'assistant', content });
         }
         else if (type === 'user' && isObject(event.message)) {
             // tool_result blocks the harness fed back to the model — part of the next prompt.
@@ -346,7 +347,7 @@ export async function runClaudeCode(opts) {
                 // would kill nothing and only convert a clean exit into a spurious failure.
                 if (!meta?.final)
                     feedGuard(content);
-                messages.push({ role: 'tool', content });
+                telemetry.onToolResult(dispatchId, content);
             }
         }
         else if (type === 'result') {
@@ -432,6 +433,8 @@ export async function runClaudeCode(opts) {
                 ...appendArgs,
             ],
         }, prompt, { ...opts, signal: runSignal }, env, opts.subscriptionToken ? secretsToRedact(opts.subscriptionToken) : [], onEvent);
+        // The stream has ended, so the last call has no successor envelope to complete it.
+        telemetry.flush();
         return await assembleClaudeOutcome({
             summary,
             stats,
@@ -440,9 +443,17 @@ export async function runClaudeCode(opts) {
             publisher,
             usage,
             subagents,
+            expectSubagentCalls: telemetry.expectsWatcherCalls(),
+            ...(opts.log ? { log: opts.log } : {}),
         });
     }
     catch (err) {
+        // The stream ended abnormally (guard trip, watchdog kill, CLI crash). Complete the call in
+        // flight anyway, and release whatever the publisher was withholding: a killed run never
+        // returns an outcome, so the live channel is the ONLY record of what it spent, and dropping
+        // its last turn is what the streaming exists to avoid.
+        telemetry.flush();
+        publisher.flush();
         // A tripped no-progress guard aborted the CLI; streamCli rejects with its generic abort
         // message, so replace it with the guard's actionable diagnostic — carrying the stderr tail
         // it attached, since that is usually the only evidence of what the CLI was doing when it was
@@ -498,6 +509,10 @@ function buildClaudeEnv(opts, configHome) {
  * terminal `result` event's cumulative) covers ONLY the parent loop, and the subagent tokens live
  * exclusively in the per-session `subagents/*.jsonl` transcripts the watcher reads. Extracted from
  * {@link runClaudeCode} verbatim to keep its cyclomatic complexity down.
+ *
+ * The invariant is about the aggregate `usage` only. `calls` was NEVER disjoint from the watcher's
+ * on its own: the CLI streams a subagent's turns onto the parent's stdout as well, so the parent
+ * loop's telemetry must filter them (`subagentDispatchId`) for this concatenation to hold.
  */
 async function assembleClaudeOutcome(args) {
     const { summary, stats, stderrTail, calls, publisher, usage, subagents } = args;
@@ -513,6 +528,10 @@ async function assembleClaudeOutcome(args) {
     await subagents?.stop();
     const subUsage = subagents?.usage() ?? { inputTokens: 0, outputTokens: 0 };
     const subCalls = subagents?.calls() ?? [];
+    if (args.expectSubagentCalls && !subCalls.length) {
+        args.log?.warn('subagent turns were streamed but the transcript watcher captured no calls; their token ' +
+            'spend is missing from this run’s telemetry (check the CLI’s subagents/*.jsonl layout)');
+    }
     const mergedCalls = [...calls, ...subCalls];
     const mergedUsage = usage || subUsage.inputTokens || subUsage.outputTokens
         ? {
@@ -555,7 +574,11 @@ function claudeUsage(raw) {
 export async function runCodex(opts) {
     const stats = { toolCalls: 0, assistantChars: 0 };
     let summary = '';
-    let usage;
+    // The running CUMULATIVE total, kept in its reported (inclusive) form plus the cached share
+    // it contains. `PiRunOutcome.usage` needs the inclusive figure — it is the key-rotation
+    // weight — while the fallback call metric below needs the split, so both are derived from
+    // this one value rather than one being reconstructed from the other.
+    let cumulative;
     // Codex reads its credentials from $CODEX_HOME/auth.json with file-backed
     // storage. CRITICAL: this home must live OUTSIDE the cloned checkout (`opts.cwd`)
     // — the blueprint/requirements/conflict-resolver handlers finish with
@@ -616,7 +639,7 @@ export async function runCodex(opts) {
             opts.onProgress(progress);
         const turnUsage = codexUsage(event);
         if (turnUsage)
-            usage = turnUsage;
+            cumulative = turnUsage;
         // A `token_count` event closes a model turn: pair its per-turn usage with the
         // assistant text seen since the previous turn as one telemetry call.
         const perTurn = codexLastTurnUsage(event);
@@ -628,7 +651,8 @@ export async function runCodex(opts) {
                 responseText: redactBody(pendingText, secrets),
                 reasoningText: '',
                 inputTokens: perTurn.inputTokens,
-                cachedInputTokens: perTurn.cachedInputTokens,
+                cacheReadTokens: perTurn.cacheReadTokens,
+                cacheWriteTokens: perTurn.cacheWriteTokens,
                 outputTokens: perTurn.outputTokens,
                 finishReason: null,
             }, opts.onCallMetric);
@@ -654,19 +678,29 @@ export async function runCodex(opts) {
         }, prompt, opts, { ...opts.extraEnv, ...(codexHome ? { CODEX_HOME: codexHome } : {}) }, opts.subscriptionToken ? secretsToRedact(opts.subscriptionToken) : [], onEvent);
         // Fallback for a CLI/version that never emits per-turn `last_token_usage`: record a
         // single call from the cumulative total + final text so the run is still observable.
-        if (calls.length === 0 && (usage || summary)) {
+        // The cumulative total is inclusive of its cached share exactly as a per-turn one is, so
+        // it is split the same way rather than being filed wholesale as fresh — which would report
+        // a cache-heavy run as if nothing had been cached, the one reading this telemetry exists
+        // to rule out.
+        if (calls.length === 0 && (cumulative || summary)) {
             publishCallMetric(calls, {
                 model: opts.model,
                 promptText: redactBody(JSON.stringify(messages), secrets),
                 messageCount: messages.length,
                 responseText: redactBody(summary, secrets),
                 reasoningText: '',
-                inputTokens: usage?.inputTokens ?? 0,
-                cachedInputTokens: 0,
-                outputTokens: usage?.outputTokens ?? 0,
+                inputTokens: Math.max(0, (cumulative?.inputTokens ?? 0) - (cumulative?.cachedInputTokens ?? 0)),
+                cacheReadTokens: cumulative?.cachedInputTokens ?? 0,
+                // Codex reports no separate cache-WRITE class; 0 rather than guessed.
+                cacheWriteTokens: 0,
+                outputTokens: cumulative?.outputTokens ?? 0,
                 finishReason: null,
             }, opts.onCallMetric);
         }
+        // The outcome's usage is the key-rotation WEIGHT, so it keeps the inclusive input count.
+        const usage = cumulative
+            ? { inputTokens: cumulative.inputTokens, outputTokens: cumulative.outputTokens }
+            : undefined;
         return {
             summary,
             stats,
@@ -746,8 +780,6 @@ function codexPlanProgress(event) {
  * other shapes put it on `usage` / `info.usage` directly. We read the cumulative
  * total when present so the caller can simply overwrite (not sum) — summing
  * cumulative totals across events would multiply-count. Checked most-likely first.
- * `input_tokens` is the TOTAL prompt count (OpenAI semantics: `cached_input_tokens`
- * is a subset already inside it), so it is NOT summed with the cached share.
  */
 function codexUsage(event) {
     const info = isObject(event.info) ? event.info : undefined;
@@ -761,14 +793,21 @@ function codexUsage(event) {
     const output = numberOf(raw.output_tokens);
     if (input === 0 && output === 0)
         return undefined;
-    return { inputTokens: input, outputTokens: output };
+    return {
+        inputTokens: input,
+        cachedInputTokens: numberOf(raw.cached_input_tokens),
+        outputTokens: output,
+    };
 }
 /**
  * Per-TURN Codex token usage off a `token_count` event's `info.last_token_usage` (the
  * delta for the turn just completed, as opposed to `codexUsage`'s cumulative total).
- * `input_tokens` is the total prompt count for the turn and already INCLUDES the cached
- * share (OpenAI semantics), so `cachedInputTokens` is surfaced as the subset it is —
- * NOT added on top (adding it would double-count every cached token).
+ *
+ * OpenAI semantics: `input_tokens` is the turn's WHOLE prompt count and already INCLUDES
+ * the cached share, so the fresh figure is the difference. Clamped at 0 because the two
+ * counts come off the same event and a vendor inconsistency must not mint a negative token
+ * count. Codex reports no separate cache-WRITE class, so that class is 0 here rather than
+ * guessed.
  */
 function codexLastTurnUsage(event) {
     const info = isObject(event.info) ? event.info : undefined;
@@ -780,7 +819,12 @@ function codexLastTurnUsage(event) {
     const output = numberOf(raw.output_tokens);
     if (input === 0 && output === 0)
         return undefined;
-    return { inputTokens: input, cachedInputTokens: cached, outputTokens: output };
+    return {
+        inputTokens: Math.max(0, input - cached),
+        cacheReadTokens: cached,
+        cacheWriteTokens: 0,
+        outputTokens: output,
+    };
 }
 /** Dispatch to the configured subscription harness runner. */
 export function runSubscriptionHarness(harness, opts) {

package/dist/claude-call-aggregator.js

@@ -0,0 +1,232 @@
+import { claudeAssistantContent, claudeCallUsage, isObject, redactBody } from './claude-stream.js';
+/**
+ * Assemble per-call telemetry out of Claude Code's per-block stream envelopes.
+ *
+ * `onCallStart` fires when a call's FIRST envelope arrives, which is the moment the caller must
+ * snapshot the prompt: the history at that point is what produced the response. `onCall` fires
+ * once the call is complete (a different `message.id` began, or the stream ended).
+ *
+ * Usage is merged as the MAXIMUM of each bucket across the call's envelopes rather than the last
+ * one seen. The envelopes carry a snapshot of the same call's usage, and which of them holds the
+ * final output count is a CLI detail we should not depend on; a max is right whether the value is
+ * repeated verbatim or grows.
+ *
+ * An envelope with no `message.id` cannot be attributed, so it is treated as a call of its own —
+ * the pre-aggregation behaviour, kept so a CLI build (or a transcript) that omits the id degrades
+ * to over-counting rather than to silently merging unrelated calls.
+ */
+export function createClaudeCallAggregator(handlers) {
+    let pending;
+    let anonymous = 0;
+    const complete = () => {
+        if (!pending)
+            return;
+        const { id: _id, ...call } = pending;
+        pending = undefined;
+        handlers.onCall(call);
+    };
+    return {
+        onAssistant(message) {
+            // `#anon-<n>` cannot collide with a real id (the API mints `msg_…`), so an envelope
+            // with no id keeps its own call rather than merging into whatever came before it.
+            const id = typeof message.id === 'string' && message.id ? message.id : `#anon-${anonymous++}`;
+            if (pending && pending.id !== id)
+                complete();
+            const content = Array.isArray(message.content) ? message.content : [];
+            const { text, reasoning, toolUses } = claudeAssistantContent(content);
+            const usage = claudeCallUsage(message.usage);
+            const stopReason = typeof message.stop_reason === 'string' ? message.stop_reason : null;
+            const model = typeof message.model === 'string' ? message.model : undefined;
+            if (!pending) {
+                pending = {
+                    id,
+                    content: [],
+                    text: '',
+                    reasoning: '',
+                    stopReason: null,
+                    inputTokens: 0,
+                    cacheReadTokens: 0,
+                    cacheWriteTokens: 0,
+                    outputTokens: 0,
+                    toolResults: [],
+                    toolUses: 0,
+                };
+                handlers.onCallStart?.();
+            }
+            pending.content.push(...content);
+            pending.text += text;
+            pending.reasoning += reasoning;
+            pending.toolUses += toolUses;
+            pending.inputTokens = Math.max(pending.inputTokens, usage.inputTokens);
+            pending.cacheReadTokens = Math.max(pending.cacheReadTokens, usage.cacheReadTokens);
+            pending.cacheWriteTokens = Math.max(pending.cacheWriteTokens, usage.cacheWriteTokens);
+            pending.outputTokens = Math.max(pending.outputTokens, usage.outputTokens);
+            // A block-split response reports its stop reason on the envelope that carries the end of the
+            // message; earlier ones report none. Keep the first non-null rather than the last seen.
+            if (stopReason && !pending.stopReason)
+                pending.stopReason = stopReason;
+            if (model && !pending.model)
+                pending.model = model;
+        },
+        onToolResult(content) {
+            // Results can only belong to the tool_use blocks of the call in flight. Before the first
+            // assistant envelope there is nothing they could attach to.
+            if (pending)
+                pending.toolResults.push(content);
+        },
+        flush: complete,
+    };
+}
+/**
+ * Assemble ONE conversation's per-call telemetry from the CLI stream: the growing request
+ * transcript and the per-call token/body metrics.
+ *
+ * Owns the transcript because the two are one concern — a call's `promptText` is the transcript as
+ * of that call, and its turns may only be appended once the call that produced them is complete.
+ * `seed` is what the harness supplied and the stream therefore never shows (the system + first user
+ * message, or the single folded user turn), so the reconstruction never claims a system turn that
+ * was not sent. A subagent's conversation seeds EMPTY — its prompt was minted by the CLI and never
+ * crosses this stream — which is also why its first call carries `messageCount: 0` (the backend's
+ * `latestChainTip` skips those on purpose: there is no re-sendable chain to delta against).
+ * Bodies are credential-scrubbed; they can echo the leased token.
+ *
+ * Deliberately does NOT touch {@link PiRunStats}: the run's tool/output counters describe whether
+ * the agent ACTED at all (`agentNeverActed`), which is true of a subagent's turns whichever channel
+ * ends up owning their telemetry rows. The caller accumulates them off the raw stream instead.
+ */
+export function createClaudeStreamTelemetry(opts) {
+    const messages = [...opts.seed];
+    let callPrompt = '';
+    let callMessageCount = 0;
+    // The aggregator IS the surface: the transcript and metric work happens in its callbacks, so
+    // there is nothing to wrap it in.
+    return createClaudeCallAggregator({
+        // Snapshotted when a call's FIRST envelope arrives: the history at that moment is what
+        // produced the response, and later envelopes of the same call must not see the turns it
+        // went on to add.
+        onCallStart: () => {
+            callPrompt = redactBody(JSON.stringify(messages), opts.secrets);
+            callMessageCount = messages.length;
+        },
+        onCall: (call) => {
+            opts.publish({
+                ...(call.model ? { model: call.model } : {}),
+                promptText: callPrompt,
+                messageCount: callMessageCount,
+                responseText: redactBody(call.text, opts.secrets),
+                reasoningText: redactBody(call.reasoning, opts.secrets),
+                inputTokens: call.inputTokens,
+                cacheReadTokens: call.cacheReadTokens,
+                cacheWriteTokens: call.cacheWriteTokens,
+                outputTokens: call.outputTokens,
+                finishReason: call.stopReason,
+            });
+            // Appended only now, so each call's prompt stays a strict prefix of the next and the
+            // backend's telemetry chain delta-compresses cleanly.
+            messages.push({ role: 'assistant', content: call.content });
+            for (const result of call.toolResults)
+                messages.push({ role: 'tool', content: result });
+        },
+    });
+}
+/**
+ * The dispatch (`Agent`/`Task` tool_use) id a stream envelope is tagged with, or `undefined` for a
+ * parent-loop turn.
+ *
+ * Claude Code streams the turns of the subagents it dispatches onto the parent's stdout, tagged
+ * with the tool_use id that spawned them. Those same turns are also written to the per-session
+ * `subagents/*.jsonl` transcripts the watcher reads, so recording both channels counted every
+ * subagent call twice — and splicing them into the parent's message reconstruction produced a
+ * `promptText` chain that interleaves several conversations and therefore matches no real request.
+ *
+ * The id is what makes the fallback below possible: concurrent subagents interleave on one stdout,
+ * so it is the ONLY thing separating their conversations.
+ */
+export function subagentDispatchId(event) {
+    if (!isObject(event))
+        return undefined;
+    const id = event.parent_tool_use_id;
+    return typeof id === 'string' && id ? id : undefined;
+}
+/** Whether a stream envelope describes a SUBAGENT's turn rather than the parent loop's. */
+export function isSubagentEvent(event) {
+    return subagentDispatchId(event) !== undefined;
+}
+/**
+ * Per-call telemetry for the subagents whose turns ride the parent's stdout — the FALLBACK channel,
+ * used only when no `subagents/*.jsonl` watcher will run (see `startSubagentWatcher`, which is
+ * wired only when the CLI has an isolated config home; an `ambientAuth` run has none).
+ *
+ * Without this, filtering tagged events out of the parent's telemetry leaves a subagent-heavy run
+ * with its spend recorded by NEITHER channel — an under-count, which reads as a cheap run and is
+ * the worse failure direction than the double-count the filter exists to fix.
+ *
+ * Each dispatch id gets its OWN transcript, because concurrent subagents interleave arbitrarily on
+ * one stream: folding them into a single chain is exactly the defect this whole module removes,
+ * one level down.
+ */
+function createSubagentStreamTelemetry(opts) {
+    const perDispatch = new Map();
+    const forDispatch = (dispatchId) => {
+        let telemetry = perDispatch.get(dispatchId);
+        if (!telemetry) {
+            // Seeded EMPTY: the CLI minted this subagent's prompt and it never crossed the stream.
+            telemetry = createClaudeStreamTelemetry({
+                seed: [],
+                secrets: opts.secrets,
+                publish: opts.publish,
+            });
+            perDispatch.set(dispatchId, telemetry);
+        }
+        return telemetry;
+    };
+    return {
+        onAssistant: (dispatchId, message) => forDispatch(dispatchId).onAssistant(message),
+        // Only against a dispatch already seen: a result for a subagent whose assistant turns never
+        // reached us has no conversation to attach to, and minting one would publish a call that is
+        // all tool output and no request.
+        onToolResult: (dispatchId, content) => perDispatch.get(dispatchId)?.onToolResult(content),
+        flush: () => {
+            for (const telemetry of perDispatch.values())
+                telemetry.flush();
+        },
+    };
+}
+/**
+ * Assemble a run's per-call telemetry, routing each envelope to the conversation it belongs to.
+ *
+ * The routing is the whole point. A subagent's turns ride the parent's stdout tagged with the
+ * dispatch that spawned them, and they must never join the PARENT's chain — that splice produced a
+ * `promptText` interleaving several conversations, matching no request that was ever sent.
+ *
+ * Who RECORDS them is a separate question, decided once per run rather than per event:
+ * `watcherOwnsSubagents` says a `subagents/*.jsonl` watcher will run, and it is the better source
+ * (it reads the settled transcript, so its usage and stop reason are final). With no watcher — an
+ * `ambientAuth` run has no isolated config home to watch — the tagged turns are recorded here
+ * instead, on per-dispatch transcripts of their own. Dropping them in that case would leave the run
+ * billed by neither channel, and an under-count reads as a cheap run rather than as an error.
+ */
+export function createClaudeRunTelemetry(opts) {
+    const parent = createClaudeStreamTelemetry(opts);
+    const subagents = opts.watcherOwnsSubagents ? undefined : createSubagentStreamTelemetry(opts);
+    let sawSubagentTurn = false;
+    return {
+        onAssistant(dispatchId, message) {
+            if (!dispatchId)
+                return parent.onAssistant(message);
+            sawSubagentTurn = true;
+            subagents?.onAssistant(dispatchId, message);
+        },
+        onToolResult(dispatchId, content) {
+            if (!dispatchId)
+                return parent.onToolResult(content);
+            sawSubagentTurn = true;
+            subagents?.onToolResult(dispatchId, content);
+        },
+        flush() {
+            parent.flush();
+            subagents?.flush();
+        },
+        expectsWatcherCalls: () => opts.watcherOwnsSubagents && sawSubagentTurn,
+    };
+}

package/dist/claude-stream.js

@@ -51,16 +51,22 @@ export function claudeAssistantContent(content) {
 }
 /**
  * Per-CALL token usage off a Claude `assistant` message's `usage` (this turn only, not
- * the cumulative `result` total). `inputTokens` counts every billed input bucket (fresh
- * + both cache buckets); `cachedInputTokens` is the cache share, surfaced separately.
+ * the cumulative `result` total).
+ *
+ * Anthropic reports all three input classes SEPARATELY and `input_tokens` is already
+ * exclusive of both caches, so the three fields here are orthogonal and additive:
+ * total input = `inputTokens + cacheReadTokens + cacheWriteTokens`. Do NOT re-lump the
+ * reads and the writes — a cache write costs 1.25–2× base input while a read costs ~0.1×,
+ * so a turn that keeps invalidating the prefix and one that rides a warm cache are
+ * indistinguishable once they are summed.
  */
 export function claudeCallUsage(raw) {
     if (!isObject(raw))
-        return { inputTokens: 0, cachedInputTokens: 0, outputTokens: 0 };
-    const cached = numberOf(raw.cache_read_input_tokens) + numberOf(raw.cache_creation_input_tokens);
+        return { inputTokens: 0, cacheReadTokens: 0, cacheWriteTokens: 0, outputTokens: 0 };
     return {
-        inputTokens: numberOf(raw.input_tokens) + cached,
-        cachedInputTokens: cached,
+        inputTokens: numberOf(raw.input_tokens),
+        cacheReadTokens: numberOf(raw.cache_read_input_tokens),
+        cacheWriteTokens: numberOf(raw.cache_creation_input_tokens),
         outputTokens: numberOf(raw.output_tokens),
     };
 }

package/dist/inline.js

@@ -48,7 +48,7 @@ export async function handleInline(job, opts) {
         return {
             text: outcome.summary,
             finishReason: deriveFinishReason(outcome.callMetrics),
-            ...(outcome.usage ? { usage: outcome.usage } : {}),
+            ...(outcome.usage ? { usage: inlineUsage(outcome.usage, outcome.callMetrics) } : {}),
             ...(outcome.callMetrics ? { callMetrics: outcome.callMetrics } : {}),
         };
     }
@@ -56,3 +56,31 @@ export async function handleInline(job, opts) {
         await rm(cwd, { recursive: true, force: true }).catch(() => { });
     }
 }
+/**
+ * Split the run's coarse usage into the three orthogonal input classes an {@link InlineResult}
+ * carries. `outcome.usage` is the ROTATION-window weight — every billed input bucket summed —
+ * so the split has to come from the per-call metrics, the only channel that kept the classes
+ * apart. Fresh input is likewise taken from the calls rather than derived by subtraction, so a
+ * CLI whose per-call and cumulative counts disagree can never produce a negative class.
+ *
+ * With no per-call telemetry (an older CLI build that streams nothing) the coarse total is
+ * reported as fresh with both cache classes 0. That is the honest reading: nothing is KNOWN to
+ * have been cached, and inventing a split would be worse than admitting the channel is silent.
+ */
+function inlineUsage(usage, calls) {
+    if (!calls?.length) {
+        return {
+            inputTokens: usage.inputTokens,
+            cacheReadTokens: 0,
+            cacheWriteTokens: 0,
+            outputTokens: usage.outputTokens,
+        };
+    }
+    const sum = (pick) => calls.reduce((total, call) => total + pick(call), 0);
+    return {
+        inputTokens: sum((call) => call.inputTokens),
+        cacheReadTokens: sum((call) => call.cacheReadTokens),
+        cacheWriteTokens: sum((call) => call.cacheWriteTokens),
+        outputTokens: usage.outputTokens,
+    };
+}

package/dist/subagents.js

@@ -137,7 +137,13 @@ export function startSubagentWatcher(root, opts) {
             return;
         const message = event.message;
         const u = claudeCallUsage(message.usage);
-        if (u.inputTokens === 0 && u.outputTokens === 0)
+        // Every input class counts towards "did this turn report usage at all": a turn riding a
+        // warm cache legitimately reports 0 fresh input, and skipping it would drop precisely the
+        // cache-heavy calls this telemetry exists to weigh.
+        if (u.inputTokens === 0 &&
+            u.cacheReadTokens === 0 &&
+            u.cacheWriteTokens === 0 &&
+            u.outputTokens === 0)
             return;
         const content = Array.isArray(message.content) ? message.content : [];
         const { text, reasoning } = claudeAssistantContent(content);
@@ -154,11 +160,16 @@ export function startSubagentWatcher(root, opts) {
             responseText: redactBody(text, secrets),
             reasoningText: redactBody(reasoning, secrets),
             inputTokens: u.inputTokens,
-            cachedInputTokens: u.cachedInputTokens,
+            cacheReadTokens: u.cacheReadTokens,
+            cacheWriteTokens: u.cacheWriteTokens,
             outputTokens: u.outputTokens,
             finishReason: typeof message.stop_reason === 'string' ? message.stop_reason : null,
         }, opts.onCallMetric);
-        usage.inputTokens += u.inputTokens;
+        // The run-level `usage` is the COARSE rotation-window weight, which counts every billed
+        // input bucket — unlike the per-call metric above, whose `inputTokens` is fresh-only. Sum
+        // all three classes back together here or a cache-heavy subagent looks nearly free to the
+        // rotation.
+        usage.inputTokens += u.inputTokens + u.cacheReadTokens + u.cacheWriteTokens;
         usage.outputTokens += u.outputTokens;
     };
     const NEWLINE = 0x0a;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cat-factory/executor-harness",
-  "version": "1.64.2",
+  "version": "1.66.0",
   "description": "Container payload: a thin TypeScript wrapper that runs the Pi coding agent against a cloned repo and opens a PR. Runs in the Cloudflare Container (and, in local native mode, as a host process); carries no secrets.",
   "repository": {
     "type": "git",
@@ -26,9 +26,9 @@
     "hono": "^4.12.32",
     "typescript": "7.0.2",
     "vitest": "^4.1.10",
-    "@cat-factory/kernel": "0.171.0",
-    "@cat-factory/server": "0.162.0",
-    "@cat-factory/spend": "0.12.100"
+    "@cat-factory/kernel": "0.175.0",
+    "@cat-factory/server": "0.165.0",
+    "@cat-factory/spend": "0.12.104"
   },
   "scripts": {
     "build": "tsc -p tsconfig.json",