@genesislcap/ai-assistant 15.6.2 → 15.7.1

package/dist/esm/components/orchestrating-driver/orchestrating-driver.budget.test.js

@@ -0,0 +1,312 @@
+import { __awaiter } from "tslib";
+import { BudgetExhaustedError } from '@genesislcap/foundation-ai';
+import { assert, createLogicSuite } from '@genesislcap/foundation-testing';
+import { AgenticActivityBus } from '../../channel/ai-activity-bus';
+// Side-effect import — MUST precede `./orchestrating-driver` so the driver
+// subclasses jsdom's EventTarget (CustomEvent dispatch then works in node).
+import '../chat-driver/align-event-globals';
+import { OrchestratingDriver } from './orchestrating-driver';
+// GENC-1464 — the orchestrated turn's OUTCOME.
+//
+// `createDriver` returns an `OrchestratingDriver` for every agents-configured
+// host, and its `sendMessage` used to end with a hardcoded `return { reason:
+// 'done' }`. That made every typed failure unreachable for those hosts —
+// including `'budget-exhausted'`, which the assistant element latches its blocked
+// state off. A 402 therefore produced an apology over a still-enabled composer
+// for exactly the configurations this workstream exists to fix.
+//
+// The classifier compounds it: `classify()` is a SECOND retry ladder stacked on
+// the transport's, so an untyped catch re-issued the 402 `classifierRetries + 1`
+// times and then ran a fallback turn into the same wall.
+const makeRegistry = (provider) => ({
+    get: () => provider,
+    default: () => provider,
+    defaultName: () => 'test',
+    names: () => ['test'],
+    getStatus: () => __awaiter(void 0, void 0, void 0, function* () { return null; }),
+    listStatuses: () => __awaiter(void 0, void 0, void 0, function* () { return []; }),
+});
+const agent = (overrides) => (Object.assign({ description: 'test agent' }, overrides));
+/** A provider that ends every turn immediately with a plain-text reply (no tools). */
+const okProvider = () => ({
+    chat: () => __awaiter(void 0, void 0, void 0, function* () { return ({ role: 'assistant', content: 'ok' }); }),
+});
+/** A provider that counts calls and always refuses for budget. */
+const walledProvider = (vendorLabel = 'Anthropic') => {
+    let n = 0;
+    return {
+        calls: () => n,
+        provider: {
+            chat: () => __awaiter(void 0, void 0, void 0, function* () {
+                n += 1;
+                throw new BudgetExhaustedError(vendorLabel, 25, 25.4);
+            }),
+        },
+    };
+};
+const failureReasonOf = (r) => r.reason === 'done' ? r.failureReason : undefined;
+const Suite = createLogicSuite('orchestrating-driver budget wall');
+// ── 1. The showstopper: a walled AGENT turn must reach the caller ───────────
+Suite('a budget wall on the agent turn surfaces as the orchestrated result', () => __awaiter(void 0, void 0, void 0, function* () {
+    // ONE specialist and no fallback, so `classify` short-circuits without calling
+    // the provider — the wall is hit by the AGENT turn, inside a real tool loop.
+    // Against the old hardcoded return this deep-equal fails.
+    const { provider } = walledProvider();
+    const driver = new OrchestratingDriver(makeRegistry(provider), [agent({ name: 'Solo' })], {
+        sessionKey: 'budget-1',
+    });
+    const result = yield driver.sendMessage('go');
+    assert.equal(result, {
+        reason: 'done',
+        failureReason: 'budget-exhausted',
+        budget: { budgetUsd: 25, spentUsd: 25.4, vendorLabel: 'Anthropic', vendor: 'anthropic' },
+    });
+}));
+Suite("the vendor is the refusing transport's, not a hardcoded default", () => __awaiter(void 0, void 0, void 0, function* () {
+    // The label rides the contract; `vendor` is its normalised key. Without a
+    // second vendor exercised here, a constant would pass every other assertion in
+    // this file — and per-vendor blocking would wall the wrong vendor forever.
+    const { provider } = walledProvider('Gemini');
+    const driver = new OrchestratingDriver(makeRegistry(provider), [agent({ name: 'Solo' })], {
+        sessionKey: 'budget-gemini',
+    });
+    const result = yield driver.sendMessage('go');
+    assert.equal(result.reason === 'done' ? result.budget : undefined, {
+        budgetUsd: 25,
+        spentUsd: 25.4,
+        vendorLabel: 'Gemini',
+        vendor: 'gemini',
+    });
+}));
+Suite('an unrecognised vendor label carries no typed vendor at all', () => __awaiter(void 0, void 0, void 0, function* () {
+    // Attribution degrades to "unattributable" rather than to a wrong vendor —
+    // which makes the element fall back to its vendor-agnostic block. Safe.
+    const { provider } = walledProvider('Acme AI');
+    const driver = new OrchestratingDriver(makeRegistry(provider), [agent({ name: 'Solo' })], {
+        sessionKey: 'budget-unknown-vendor',
+    });
+    const result = yield driver.sendMessage('go');
+    assert.equal(result.reason === 'done' ? result.budget : undefined, {
+        budgetUsd: 25,
+        spentUsd: 25.4,
+        vendorLabel: 'Acme AI',
+    });
+}));
+Suite('the wall also reaches the tool-loop-end bus detail through the orchestrator', () => __awaiter(void 0, void 0, void 0, function* () {
+    var _a;
+    const bus = new AgenticActivityBus();
+    const seen = [];
+    const stop = bus.subscribe('tool-loop-end', (d) => seen.push(d));
+    const { provider } = walledProvider();
+    const driver = new OrchestratingDriver(makeRegistry(provider), [agent({ name: 'Solo' })], {
+        sessionKey: 'budget-2',
+        activityBus: bus,
+    });
+    yield driver.sendMessage('go');
+    assert.is(seen.length, 1, 'exactly one tool-loop-end for the one agent turn');
+    assert.is((_a = seen[0]) === null || _a === void 0 ? void 0 : _a.failureReason, 'budget-exhausted');
+    stop();
+    bus.close();
+}));
+Suite('the bus detail carries the proxy figures, not just the failure reason', () => __awaiter(void 0, void 0, void 0, function* () {
+    var _a, _b, _c;
+    // The element latches off THIS event for an in-loop wall — the publish sits in
+    // sendMessage's `finally`, so it lands before the return-value seam and the
+    // idempotent latch makes the later one a no-op. If the figures rode only the
+    // return value they would be unreachable in the common case and the banner would
+    // show the generic copy in exactly the path they were added for.
+    const bus = new AgenticActivityBus();
+    const seen = [];
+    const stop = bus.subscribe('tool-loop-end', (d) => seen.push(d));
+    const { provider } = walledProvider();
+    const driver = new OrchestratingDriver(makeRegistry(provider), [agent({ name: 'Solo' })], {
+        sessionKey: 'budget-figures',
+        activityBus: bus,
+    });
+    const result = yield driver.sendMessage('go');
+    assert.equal((_a = seen[0]) === null || _a === void 0 ? void 0 : _a.budget, { budgetUsd: 25, spentUsd: 25.4, vendorLabel: 'Anthropic', vendor: 'anthropic' }, 'the event carries the same figures as the return value');
+    assert.is((_b = seen[0]) === null || _b === void 0 ? void 0 : _b.vendor, 'anthropic', 'and the top-level vendor agrees with the refusing transport, not with a stale last-resolved provider');
+    assert.equal(result.reason === 'done' ? result.budget : undefined, (_c = seen[0]) === null || _c === void 0 ? void 0 : _c.budget, 'event and return value agree — neither is the sole carrier');
+    stop();
+    bus.close();
+}));
+Suite('a clean turn publishes no budget key at all', () => __awaiter(void 0, void 0, void 0, function* () {
+    const bus = new AgenticActivityBus();
+    const seen = [];
+    const stop = bus.subscribe('tool-loop-end', (d) => seen.push(d));
+    const provider = {
+        chat: () => __awaiter(void 0, void 0, void 0, function* () { return ({ role: 'assistant', content: 'ok' }); }),
+    };
+    const driver = new OrchestratingDriver(makeRegistry(provider), [agent({ name: 'Solo' })], {
+        sessionKey: 'budget-clean',
+        activityBus: bus,
+    });
+    yield driver.sendMessage('go');
+    assert.ok(seen.every((d) => d === undefined || !('budget' in d)), 'no budget key on a non-budget turn — the historical detail shape is unchanged');
+    stop();
+    bus.close();
+}));
+// ── 2. No regression on the happy / handoff paths ───────────────────────────
+Suite('a clean turn still returns the bare legacy shape', () => __awaiter(void 0, void 0, void 0, function* () {
+    const driver = new OrchestratingDriver(makeRegistry(okProvider()), [agent({ name: 'Solo' })], {
+        sessionKey: 'budget-3',
+    });
+    const result = yield driver.sendMessage('go');
+    assert.equal(result, { reason: 'done' }, 'no failureReason on a clean turn');
+}));
+Suite('a handoff-capped turn never leaks agent-handoff to the caller', () => __awaiter(void 0, void 0, void 0, function* () {
+    // `'agent-handoff'` is an internal routing protocol. The two breaks that hold a
+    // non-'done' inner result both carry it, so the collapse rule in `sendMessage`
+    // is what keeps "report the turn that ended the loop" from changing the
+    // discriminant a host matches on.
+    let n = 0;
+    const nextId = () => {
+        n += 1;
+        return n;
+    };
+    const handoffProvider = {
+        chat: (_h, _u, options) => __awaiter(void 0, void 0, void 0, function* () {
+            var _a;
+            // The classifier call carries the routing tool; answer it with index 0.
+            if ((_a = options === null || options === void 0 ? void 0 : options.tools) === null || _a === void 0 ? void 0 : _a.some((t) => t.name === 'select_agent')) {
+                return {
+                    role: 'assistant',
+                    content: '',
+                    toolCalls: [{ id: `c-${nextId()}`, name: 'select_agent', args: { agent_index: 0 } }],
+                };
+            }
+            // Every agent turn asks to hand off, so the handoff cap is what ends it.
+            return {
+                role: 'assistant',
+                content: '',
+                toolCalls: [
+                    {
+                        id: `h-${nextId()}`,
+                        name: 'request_continuation',
+                        args: { summary: 's', remaining_task: 'more' },
+                    },
+                ],
+            };
+        }),
+    };
+    const driver = new OrchestratingDriver(makeRegistry(handoffProvider), [agent({ name: 'A' }), agent({ name: 'B' })], { sessionKey: 'budget-4', maxHandoffs: 1 });
+    const result = yield driver.sendMessage('go');
+    assert.is(result.reason, 'done', 'the internal handoff protocol never escapes');
+}));
+Suite('a wall on a post-handoff classify does not duplicate the user message', () => __awaiter(void 0, void 0, void 0, function* () {
+    // `classify()` runs at TWO seams. Before the first agent turn the user's
+    // message is unappended (only the optimistic `history-updated` echo exists),
+    // so the wall-catch must append it. At the END of a handoff iteration
+    // `chatDriver.sendMessage` has already appended it — passing it again
+    // produced ['user','assistant','tool','user','assistant'] with 'go' twice,
+    // persisting into the snapshot and into the history sent to the model once
+    // the budget was raised. The single-specialist narrowness guard below cannot
+    // catch this: with one specialist, classify short-circuits and the handoff
+    // seam never runs.
+    let n = 0;
+    const provider = {
+        chat: (_h, _u, options) => __awaiter(void 0, void 0, void 0, function* () {
+            var _a;
+            n += 1;
+            if ((_a = options === null || options === void 0 ? void 0 : options.tools) === null || _a === void 0 ? void 0 : _a.some((t) => t.name === 'select_agent')) {
+                // First classify routes; the post-handoff classify hits the wall.
+                if (n > 2)
+                    throw new BudgetExhaustedError('Anthropic', 25, 25.4);
+                return {
+                    role: 'assistant',
+                    content: '',
+                    toolCalls: [{ id: `c-${n}`, name: 'select_agent', args: { agent_index: 0 } }],
+                };
+            }
+            // The agent turn asks to hand off, forcing the second classify.
+            return {
+                role: 'assistant',
+                content: '',
+                toolCalls: [
+                    {
+                        id: `h-${n}`,
+                        name: 'request_continuation',
+                        args: { summary: 's', remaining_task: 'more' },
+                    },
+                ],
+            };
+        }),
+    };
+    const driver = new OrchestratingDriver(makeRegistry(provider), [agent({ name: 'A' }), agent({ name: 'B' })], { sessionKey: 'budget-4b' });
+    const result = yield driver.sendMessage('go');
+    assert.is(failureReasonOf(result), 'budget-exhausted');
+    const history = driver.getRawHistory();
+    const userMessages = history.filter((m) => m.role === 'user' && m.content === 'go');
+    assert.is(userMessages.length, 1, "the user's message appears exactly once");
+}));
+// ── 3. The classifier's own retry ladder ───────────────────────────────────
+Suite('a wall during classification costs exactly one provider call', () => __awaiter(void 0, void 0, void 0, function* () {
+    // TWO specialists and no fallback, so `classify` really runs. Before the narrow
+    // rethrow this was FOUR calls: `classifierRetries + 1` classifier attempts, then
+    // a doomed agent turn.
+    const { calls, provider } = walledProvider();
+    const driver = new OrchestratingDriver(makeRegistry(provider), [agent({ name: 'A' }), agent({ name: 'B' })], { sessionKey: 'budget-5' });
+    const result = yield driver.sendMessage('go');
+    assert.is(calls(), 1, 'no ladder clears a budget wall — do not climb it');
+    assert.is(failureReasonOf(result), 'budget-exhausted');
+}));
+Suite('a classify-time wall lands on the transcript, without a routing apology', () => __awaiter(void 0, void 0, void 0, function* () {
+    // The old ordering emitted "I'm not sure how to help with that…" from the
+    // no-fallback exit and THEN routed anyway, so the user read a routing excuse
+    // the classifier never earned, immediately above the real error.
+    const { provider } = walledProvider();
+    const driver = new OrchestratingDriver(makeRegistry(provider), [agent({ name: 'A' }), agent({ name: 'B' })], { sessionKey: 'budget-6' });
+    yield driver.sendMessage('go');
+    const history = driver.getRawHistory();
+    const text = history.map((m) => m.content).join('\n');
+    assert.ok(text.includes('AI usage limit'), 'the real reason is what the user reads');
+    assert.not.ok(text.includes("I'm not sure how to help with that"));
+    // The user's own message must survive. `runOrchestratedTurn` dispatches it
+    // optimistically and leaves the real append to `chatDriver.sendMessage`, which
+    // a classify-time wall never reaches — so the driver's history held only the
+    // assistant bubble and the next render dropped what the user had typed.
+    assert.equal(history.map((m) => m.role), ['user', 'assistant'], 'the question stays above the answer');
+    assert.is(history[0].content, 'go');
+}));
+Suite('an in-loop wall does not double-append the user message', () => __awaiter(void 0, void 0, void 0, function* () {
+    // Narrowness guard for the append above: `reportBudgetExhausted` also serves
+    // the in-loop path, where `sendMessage` has ALREADY appended the user message.
+    // Only the classify seam passes one, so this path must be unchanged. ONE
+    // specialist and no fallback, so `classify` short-circuits and the wall is hit
+    // inside a real tool loop.
+    const { provider } = walledProvider();
+    const driver = new OrchestratingDriver(makeRegistry(provider), [agent({ name: 'Solo' })], {
+        sessionKey: 'budget-6b',
+    });
+    yield driver.sendMessage('go');
+    const history = driver.getRawHistory();
+    assert.is(history.filter((m) => m.role === 'user').length, 1, 'one user message, not two');
+}));
+Suite('a wall during classification never starts the fallback turn', () => __awaiter(void 0, void 0, void 0, function* () {
+    const { calls, provider } = walledProvider();
+    const driver = new OrchestratingDriver(makeRegistry(provider), [agent({ name: 'A' }), agent({ name: 'B' }), agent({ name: 'Catch-all', fallback: true })], { sessionKey: 'budget-7' });
+    const result = yield driver.sendMessage('go');
+    assert.is(calls(), 1, 'the fallback agent is not run into the same wall');
+    assert.is(failureReasonOf(result), 'budget-exhausted');
+}));
+Suite('a plain classifier failure still retries and falls back (narrowness guard)', () => __awaiter(void 0, void 0, void 0, function* () {
+    // Proves the rethrow did not make every classify failure terminal: only the
+    // budget wall short-circuits, because only the budget wall is unclearable.
+    let classifierCalls = 0;
+    const provider = {
+        chat: (_h, _u, options) => __awaiter(void 0, void 0, void 0, function* () {
+            var _a;
+            if ((_a = options === null || options === void 0 ? void 0 : options.tools) === null || _a === void 0 ? void 0 : _a.some((t) => t.name === 'select_agent')) {
+                classifierCalls += 1;
+                throw new Error('classifier is having a moment');
+            }
+            return { role: 'assistant', content: 'fallback handled it' };
+        }),
+    };
+    const driver = new OrchestratingDriver(makeRegistry(provider), [agent({ name: 'A' }), agent({ name: 'B' }), agent({ name: 'Catch-all', fallback: true })], { sessionKey: 'budget-8' });
+    const result = yield driver.sendMessage('go');
+    assert.is(classifierCalls, 3, 'classifierRetries (2) + the first attempt');
+    assert.is(result.reason, 'done');
+    assert.is(failureReasonOf(result), undefined, 'the fallback turn succeeded');
+}));
+Suite.run();

package/dist/esm/components/orchestrating-driver/orchestrating-driver.js

@@ -1,4 +1,5 @@
 import { __awaiter } from "tslib";
+import { BudgetExhaustedError } from '@genesislcap/foundation-ai';
 import { validateStaticAgentProviders } from '../../config/validate-providers';
 import { recordMetaEvent } from '../../state/debug-event-log';
 import { transformHistoryForAgent } from '../../utils/history-transform';
@@ -68,6 +69,17 @@ export class OrchestratingDriver extends EventTarget {
          * user stops. Reset at the start of each `sendMessage`.
          */
         this.cancelled = false;
+        /**
+         * Whether the CURRENT turn's user message has reached the inner driver's
+         * history. False through the pre-first-turn classify (where the only echo of
+         * the message is an optimistic `history-updated` dispatch), true from the
+         * moment `chatDriver.sendMessage` is entered — including through every later
+         * handoff classify. Read by the budget-wall catch in `sendMessage` to decide
+         * whether `reportBudgetExhausted` must append the message itself: appending
+         * it when already appended duplicated it; not appending it when unappended
+         * made it vanish. Reset at the top of each `runOrchestratedTurn`.
+         */
+        this.userMessageAppended = false;
         /**
          * Sticky user pick from the picker (or the host's `setAgent` API). Only
          * changes on explicit user action. Survives flow completion: when a stateful
@@ -115,6 +127,7 @@ export class OrchestratingDriver extends EventTarget {
             maxTurnSnapshots: options.maxTurnSnapshots,
             sessionKey: this.sessionKey,
             activityBus: options.activityBus,
+            budgetExhaustedMessage: options.budgetExhaustedMessage,
         });
         // Proxy events from the shared driver
         this.chatDriver.addEventListener('history-updated', (e) => {
@@ -210,8 +223,39 @@ export class OrchestratingDriver extends EventTarget {
         });
     }
     sendMessage(input, attachments) {
+        return __awaiter(this, void 0, void 0, function* () {
+            try {
+                return yield this.runOrchestratedTurn(input, attachments);
+            }
+            catch (e) {
+                // The classifier calls the provider directly, outside any ChatDriver turn,
+                // so a budget wall raised there would escape as an unhandled rejection —
+                // `FoundationAiAssistant.send()` has a `finally` but no `catch`. Land it on
+                // the driver that owns the transcript, which appends the user-facing copy
+                // and returns the same terminal result an in-loop wall would.
+                //
+                // The user's message rides along ONLY while it is genuinely unappended.
+                // `classify()` runs at two seams: before the first agent turn — where the
+                // `history-updated` dispatch in `runOrchestratedTurn` is optimistic only
+                // and the real append lives inside `chatDriver.sendMessage`, which the
+                // wall prevents from ever running — and again at the END of every handoff
+                // iteration, by which point `chatDriver.sendMessage` HAS appended it.
+                // Passing the message unconditionally duplicated it on the handoff seam:
+                // `['user','assistant','tool','user','assistant']` in the transcript, the
+                // snapshot, and the history sent to the model once the budget is raised.
+                if (e instanceof BudgetExhaustedError) {
+                    return this.chatDriver.reportBudgetExhausted(e, this.userMessageAppended
+                        ? undefined
+                        : Object.assign({ role: 'user', content: input }, (attachments ? { attachments } : {})));
+                }
+                throw e;
+            }
+        });
+    }
+    runOrchestratedTurn(input, attachments) {
         return __awaiter(this, void 0, void 0, function* () {
             this.cancelled = false;
+            this.userMessageAppended = false;
             const history = this.chatDriver.getHistory();
             // Emit the user message immediately so the UI reflects it during the classify
             // round-trip — without this the chat appears frozen until classify returns.
@@ -226,6 +270,8 @@ export class OrchestratingDriver extends EventTarget {
             let handoffs = 0;
             let handoffSummary = '';
             let remainingTask = '';
+            // The outcome of the agent turn that ENDS the loop — see the return below.
+            let lastResult;
             while (true) {
                 // Cancelled before a (next) agent turn started — e.g. during classify.
                 // The chat driver appends its own "Stopped." marker when a turn it was
@@ -243,9 +289,15 @@ export class OrchestratingDriver extends EventTarget {
                     result = yield this.chatDriver.continueFromHistory(contextPrimer);
                 }
                 else {
+                    // From here the inner driver owns the user message — it appends it as
+                    // its first act, before any provider call can throw. Set BEFORE the
+                    // await so a wall raised anywhere inside the turn (or on a later
+                    // handoff classify) sees the append as done.
+                    this.userMessageAppended = true;
                     // oxlint-disable-next-line no-await-in-loop
                     result = yield this.chatDriver.sendMessage(input, attachments);
                 }
+                lastResult = result;
                 // Release check: a stateful agent called `releaseAgent` from a terminal
                 // tool handler. Fire onDeactivate, clear the pin, drop the user back to
                 // classifier-mode. The LLM has already emitted its final wrap-up message
@@ -273,7 +325,20 @@ export class OrchestratingDriver extends EventTarget {
                 // oxlint-disable-next-line no-await-in-loop
                 currentAgent = yield this.classify(remainingTask, updatedHistory);
             }
-            return { reason: 'done' };
+            // The orchestrated turn reports the outcome of the agent turn that ENDED it.
+            // The hardcoded `{ reason: 'done' }` that used to live here made every
+            // typed failure unreachable for an agents-configured host — including
+            // `'budget-exhausted'`, which the element latches its blocked state off.
+            //
+            // "Last wins" cannot hide an earlier failure behind a later success: a
+            // failing inner turn always ends the loop, because the loop only continues
+            // on `reason === 'agent-handoff'`. The two breaks that hold a non-'done'
+            // result (the handoff cap, and the pinned/fallback break) both carry
+            // `'agent-handoff'`, which is an internal routing protocol that must never
+            // escape to the host — the collapse below strips it. A cancel before the
+            // first turn leaves `lastResult` undefined; the release break returns the
+            // inner turn's real result, which is correct, because that turn ran.
+            return (lastResult === null || lastResult === void 0 ? void 0 : lastResult.reason) === 'done' ? lastResult : { reason: 'done' };
         });
     }
     continueFromHistory(transientPrimer) {
@@ -514,6 +579,11 @@ export class OrchestratingDriver extends EventTarget {
                     },
                 },
             };
+            // True only when the classifier RAN and answered "no match" (index -1). The
+            // no-fallback apology below is gated on it: an exhausted-retries exit never
+            // earned that apology, and emitting it there put a routing excuse ahead of
+            // the real error the routed turn is about to produce.
+            let routedNoMatch = false;
             for (let attempt = 0; attempt <= this.classifierRetries; attempt += 1) {
                 try {
                     const options = {
@@ -535,9 +605,18 @@ export class OrchestratingDriver extends EventTarget {
                         return this.specialists[index];
                     }
                     // index === -1 — fall through to fallback
+                    routedNoMatch = true;
                     break;
                 }
                 catch (e) {
+                    // Terminal: no ladder clears a budget wall. Short-circuit the
+                    // classifier's own retries the same way 402 short-circuits the
+                    // transport's — this loop is a SECOND retry ladder stacked on that one,
+                    // so without this a single wall costs `classifierRetries + 1` doomed
+                    // classifier calls and then a doomed fallback turn on top. Deliberately
+                    // narrow: a generic classifier failure must still retry and fall back.
+                    if (e instanceof BudgetExhaustedError)
+                        throw e;
                     logger.warn(`OrchestratingDriver: classifier attempt ${attempt + 1} failed:`, e);
                     if (attempt === this.classifierRetries) {
                         logger.error('OrchestratingDriver: classifier failed after all retries, using fallback');
@@ -546,9 +625,15 @@ export class OrchestratingDriver extends EventTarget {
             }
             if (this.fallback)
                 return this.fallback;
-            // No fallback configured — emit an inline response and return first specialist as no-op
-            const specialistNames = this.specialists.map((s) => s.name).join(', ');
-            this.appendInlineMessage(`I'm not sure how to help with that. I can assist with: ${specialistNames}.`);
+            // No fallback configured. Route to the first specialist as a no-op — but
+            // only APOLOGISE when the classifier actually said "none of these". On an
+            // exhausted-retries exit the classifier never expressed an opinion, so the
+            // apology would be a routing excuse the user reads immediately above the
+            // real error that turn is about to produce.
+            if (routedNoMatch) {
+                const specialistNames = this.specialists.map((s) => s.name).join(', ');
+                this.appendInlineMessage(`I'm not sure how to help with that. I can assist with: ${specialistNames}.`);
+            }
             return this.specialists[0];
         });
     }

package/dist/esm/components/settings-modal/settings-modal.styles.js

@@ -304,9 +304,9 @@ export const settingsModalStyles = css `
   }
 
   /*
-   * Deliberately quieter than the live block: no pulsing badge and no accent, so the
-   * eye still lands on what is being built now. It leads only because it is the wider
-   * figure, not because it is the more urgent one.
+   * Deliberately quieter than the live block: no pulsing badge and no accent, so the eye lands
+   * on what is being built now — which is why that block leads the tab and this one follows,
+   * under the history heading whose rows it totals.
    */
   .settings-lifetime-usage-top {
     display: flex;
@@ -489,7 +489,14 @@ export const settingsModalStyles = css `
 
   .settings-token-bucket {
     display: flex;
-    min-width: 0;
+
+    /*
+     * min-content (not 0) so a bucket cannot be shrunk narrower than its own unbreakable label.
+     * With the label set to nowrap, a zero minimum would overflow it into the next column instead
+     * of wrapping — so the container's flex-wrap takes over and moves a bucket to a second row,
+     * which is the degradation we want when the modal is dragged narrow.
+     */
+    min-width: min-content;
     flex-direction: column;
     gap: 2px;
   }
@@ -500,6 +507,15 @@ export const settingsModalStyles = css `
     text-transform: uppercase;
     letter-spacing: 0.06em;
     color: var(--neutral-foreground-hint);
+
+    /*
+     * One line, always. The rate marker is a separate inline box after collapsible whitespace, so
+     * a wrappable label gives it a break opportunity of its own — in a narrow column it orphaned
+     * onto line 2, which also pushed that bucket's value a line below the other three and broke
+     * the row of figures. nowrap removes every break opportunity in the label, the space before
+     * the marker included, so no glue character is needed.
+     */
+    white-space: nowrap;
   }
 
   .settings-token-bucket-value {
@@ -509,13 +525,24 @@ export const settingsModalStyles = css `
   }
 
   /*
-   * Cache reads bill at a fraction of the uncached input rate, and in an agentic
-   * session they are usually the largest bucket by a wide margin. Muted so the biggest
-   * number on the row doesn't read as the dominant cost driver when it is the cheapest.
+   * The price-direction marker on a cache bucket's label — down for reads (a fraction of the
+   * uncached input rate), up for writes (a premium over it). Three tiers, so a bucket with no
+   * marker is the reference rate rather than an unlabelled fourth case.
+   *
+   * Carried by the label rather than the value, which is what this used to do: cache reads are
+   * usually the largest bucket by a wide margin in an agentic session, so the value was muted to
+   * stop the biggest number reading as the dominant cost driver. That priced the treatment to the
+   * wrong thing — a dimmed 0 (every Gemini cache write) looked disabled next to a normal 0, and a
+   * rate is a property of the bucket, not of this turn's figure.
+   *
+   * Sized down and nudged up rather than tinted: the label is already the hint colour, so a
+   * dimmer marker would vanish and a brighter one would out-shout the word it qualifies.
    */
-  .settings-token-bucket.is-discounted .settings-token-bucket-value {
-    font-weight: 500;
-    color: var(--neutral-foreground-hint);
+  .settings-token-bucket-rate {
+    font-size: 8px;
+    line-height: 1;
+    vertical-align: 0.15em;
+    letter-spacing: 0;
   }
 
   /*
@@ -537,9 +564,15 @@ export const settingsModalStyles = css `
     margin-bottom: 0;
   }
 
-  /* Equal columns, so the four buckets line up as a grid rather than bunching left. */
+  /*
+   * Basis auto, not 0: a zero basis splits the row into four equal columns, so CACHE WRITE is
+   * squeezed exactly as hard as INPUT despite being twice the label. Sizing from content gives the
+   * long labels the room their text needs and lets the short ones concede it, which is what makes
+   * white-space: nowrap above safe — every column starts at least as wide as its own content,
+   * then shares the slack. Still grows and shrinks, so the four stay flush edge to edge.
+   */
   .settings-usage-row .settings-token-bucket {
-    flex: 1 1 0;
+    flex: 1 1 auto;
   }
 
   .settings-model-tags {
@@ -579,6 +612,16 @@ export const settingsModalStyles = css `
     margin: calc(var(--design-unit) * 4px) 0 calc(var(--design-unit) * 2px);
   }
 
+  /*
+   * The heading labels the lifetime card as well as the rows, so the card gives up its own top
+   * margin and the gap becomes the heading's alone. Adjacent siblings collapse to the LARGER of
+   * the two margins, so leaving both would space the caps 3 units off the block they label while
+   * the rows below sit at 2 — reading as though the heading belonged to neither.
+   */
+  .settings-cost-history-title + .settings-lifetime-usage {
+    margin-top: 0;
+  }
+
   .settings-cost-history {
     display: flex;
     flex-direction: column;
@@ -615,14 +658,16 @@ export const settingsModalStyles = css `
   }
 
   /*
-   * Both of these carry a ?hidden binding (showTokens), and both declare display: flex — which
-   * silently beat the UA [hidden] { display: none } rule, because author-origin declarations
-   * outrank user-agent origin whatever the specificity. Without this, showTokens: false still
-   * rendered the full breakdown in the lifetime block, the live block and every history row.
-   * Sibling elements that hide correctly do so only by declaring no display at all. Same
-   * treatment as .settings-modal-tab-panel[hidden] near the top of this file.
+   * All of these carry a ?hidden binding, and all declare display: flex — which silently beat the
+   * UA [hidden] { display: none } rule, because author-origin declarations outrank user-agent
+   * origin whatever the specificity. Without this, showTokens: false still rendered the full
+   * breakdown in the lifetime block, the live block and every history row, and an empty row list
+   * still occupied its padding under the lifetime total. Sibling elements that hide correctly do
+   * so only by declaring no display at all. Same treatment as .settings-modal-tab-panel[hidden]
+   * near the top of this file.
    */
   .settings-usage-row-tokens[hidden],
+  .settings-cost-history[hidden],
   .settings-cost-history-row-content[hidden] {
     display: none !important;
   }