@genesislcap/ai-assistant 15.11.0 → 15.12.0

package/dist/esm/utils/usage-rows.js

@@ -0,0 +1,90 @@
+import { vendorOfModel } from '@genesislcap/foundation-ai';
+/** Whether a message carries anything a ledger would record. */
+function hasUsage(m) {
+    return (m.cost != null ||
+        m.externalCostUsd != null ||
+        m.inputTokens != null ||
+        m.outputTokens != null ||
+        m.cacheReadTokens != null ||
+        m.cacheWriteTokens != null);
+}
+/**
+ * Project a transcript into one row per billable unit of spend.
+ *
+ * @remarks
+ * Reads the messages and returns a table; it stores nothing, mutates nothing, and
+ * adds nothing to history. Use it where an **aggregate is not enough** — writing
+ * usage rows for per-project attribution, a cost dashboard, or auditing which turns
+ * came back unpriced. For a single total, `sumUsage` is the answer and this is the
+ * wrong tool.
+ *
+ * Walks the same shape `sumUsage` does, and is guaranteed to agree with it: summing
+ * `costUsd` and `externalCostUsd` across these rows equals `sumUsage(...).costUsd`
+ * over the same input, and the four token buckets reconcile likewise. That property
+ * is the point of the function — a per-call view that quietly disagrees with the
+ * total is worse than no per-call view, because both look right in isolation.
+ *
+ * Covers the two places spend hides:
+ *
+ * - **Sub-agent conversations.** A delegating tool call carries the child's entire
+ *   conversation on `toolCall.subAgentTrace`, so a walk of top-level messages alone
+ *   misses everything a delegating agent spent. A consumer that reimplemented this
+ *   and omitted the branch reported $0.106 on a run that cost $0.234.
+ * - **Compactions.** A compaction deletes the messages it summarises, banking their
+ *   spend on the summary. Those turns produce a single `source: 'compaction'` row.
+ *
+ * Messages carrying no usage at all (a user turn, a narration) produce no row.
+ *
+ * **Do not pass the result to `sumUsage`** — the types prevent it, and the reason is
+ * that both recurse, so a pre-flattened list would be counted twice. Both functions
+ * take raw history.
+ *
+ * @beta
+ */
+export function usageRows(messages) {
+    return collectRows(messages, 0, undefined);
+}
+/** Map an already-bucketed {@link AggregateUsage} onto a row's token fields. */
+function bucketsOf(usage) {
+    return {
+        uncachedInputTokens: usage.uncachedInputTokens,
+        cacheReadTokens: usage.cacheReadTokens,
+        cacheWriteTokens: usage.cacheWriteTokens,
+        outputTokens: usage.outputTokens,
+    };
+}
+function collectRows(messages, depth, subAgentOf) {
+    var _a, _b, _c, _d, _e, _f, _g;
+    const rows = [];
+    for (const m of messages) {
+        if (hasUsage(m)) {
+            const cacheReadTokens = (_a = m.cacheReadTokens) !== null && _a !== void 0 ? _a : 0;
+            const cacheWriteTokens = (_b = m.cacheWriteTokens) !== null && _b !== void 0 ? _b : 0;
+            rows.push(Object.assign(Object.assign(Object.assign(Object.assign(Object.assign(Object.assign({ source: 'request', model: m.model, 
+                // The driver stamps `ChatMessage.provider` from the resolved provider's own
+                // status, so it is AUTHORITATIVE and covers vendors the model allowlists do
+                // not — a server-proxied `gpt-5` turn is `provider: 'openai'` on the message
+                // and unknown to `vendorOfModel`. Prefer it; fall back to resolving the model
+                // id only for messages persisted before the field existed. Still left
+                // undefined when neither knows, rather than guessed at.
+                provider: (_c = m.provider) !== null && _c !== void 0 ? _c : (m.model ? vendorOfModel(m.model) : undefined) }, (m.cost != null && { costUsd: m.cost })), (m.externalCostUsd != null && { externalCostUsd: m.externalCostUsd })), { 
+                // Uncached input is the REMAINDER, matching `sumUsage`: `inputTokens` is the
+                // whole prompt and the cache fields break it down. Clamped for the same
+                // reason — hand-edited history must not drive a total negative.
+                uncachedInputTokens: Math.max(0, ((_d = m.inputTokens) !== null && _d !== void 0 ? _d : 0) - cacheReadTokens - cacheWriteTokens), cacheReadTokens,
+                cacheWriteTokens, outputTokens: (_e = m.outputTokens) !== null && _e !== void 0 ? _e : 0 }), (m.agentName != null && { agentName: m.agentName })), { subAgentDepth: depth }), (subAgentOf != null && { subAgentOf })));
+        }
+        // Spend the compaction banked on this summary's behalf. Its buckets are already
+        // disjoint (it is an `AggregateUsage`), so they map across without the
+        // subtraction a per-message record needs.
+        const rolled = (_f = m.compaction) === null || _f === void 0 ? void 0 : _f.rolledUpUsage;
+        if (rolled) {
+            rows.push(Object.assign(Object.assign(Object.assign(Object.assign(Object.assign({ source: 'compaction' }, (rolled.costUsd != null && { costUsd: rolled.costUsd })), bucketsOf(rolled)), (m.agentName != null && { agentName: m.agentName })), { subAgentDepth: depth }), (subAgentOf != null && { subAgentOf })));
+        }
+        for (const tc of (_g = m.toolCalls) !== null && _g !== void 0 ? _g : []) {
+            if (tc.subAgentTrace)
+                rows.push(...collectRows(tc.subAgentTrace, depth + 1, tc.id));
+        }
+    }
+    return rows;
+}

package/dist/esm/utils/usage-rows.test.js

@@ -0,0 +1,189 @@
+import { assert, createLogicSuite } from '@genesislcap/foundation-testing';
+import { sumUsage } from './sum-usage';
+import { usageRows } from './usage-rows';
+const msg = (over) => (Object.assign({ role: 'assistant', content: '' }, over));
+const rolled = (over) => (Object.assign({ costUsd: 0, uncachedInputTokens: 0, cacheReadTokens: 0, cacheWriteTokens: 0, outputTokens: 0 }, over));
+/** Every place spend can hide, in one transcript: a delegation and a compaction. */
+const richHistory = () => [
+    msg({
+        role: 'compacted-summary',
+        content: 'summary',
+        compaction: {
+            compactedCount: 4,
+            rolledUpUsage: rolled({ costUsd: 0.05, uncachedInputTokens: 400, outputTokens: 60 }),
+            // The summariser's model — NOT what the banked spend ran on.
+            model: 'claude-haiku-4-5-20251001',
+        },
+    }),
+    msg({ role: 'user', content: 'go' }),
+    msg({
+        model: 'claude-sonnet-5',
+        agentName: 'boss',
+        cost: 0.01,
+        inputTokens: 1000,
+        cacheReadTokens: 900,
+        outputTokens: 50,
+        toolCalls: [
+            {
+                id: 'tc1',
+                name: 'delegate',
+                args: {},
+                subAgentTrace: [
+                    msg({
+                        model: 'gemini-2.5-flash',
+                        agentName: 'worker',
+                        cost: 0.12,
+                        inputTokens: 500,
+                        outputTokens: 200,
+                    }),
+                ],
+            },
+        ],
+    }),
+];
+const suite = createLogicSuite('usageRows');
+suite('reconciles exactly with sumUsage over the same input', () => {
+    // THE guarantee. A per-call view that disagrees with the total is worse than none,
+    // because both look right in isolation — which is how the consumer's own copy
+    // under-reported for a month.
+    const history = richHistory();
+    const rows = usageRows(history);
+    const total = sumUsage(history);
+    const rowCost = rows.reduce((n, r) => { var _a, _b; return n + ((_a = r.costUsd) !== null && _a !== void 0 ? _a : 0) + ((_b = r.externalCostUsd) !== null && _b !== void 0 ? _b : 0); }, 0);
+    assert.ok(Math.abs(rowCost - total.costUsd) < 1e-12, `${rowCost} vs ${total.costUsd}`);
+    const bucket = (k) => rows.reduce((n, r) => n + r[k], 0);
+    assert.is(bucket('uncachedInputTokens'), total.uncachedInputTokens, 'uncached input');
+    assert.is(bucket('cacheReadTokens'), total.cacheReadTokens, 'cache read');
+    assert.is(bucket('cacheWriteTokens'), total.cacheWriteTokens, 'cache write');
+    assert.is(bucket('outputTokens'), total.outputTokens, 'output');
+});
+suite('emits a compaction row so banked spend is not lost from a per-call view', () => {
+    // The shortfall this function exists to close: the summary message carries no
+    // `cost` and no token fields of its own, so a naive per-message walk drops the
+    // spend of every turn the compaction deleted.
+    const rows = usageRows(richHistory());
+    const compactionRows = rows.filter((r) => r.source === 'compaction');
+    assert.is(compactionRows.length, 1);
+    assert.is(compactionRows[0].costUsd, 0.05);
+    assert.is(compactionRows[0].uncachedInputTokens, 400);
+    assert.is(compactionRows[0].model, undefined, 'no model — the banked spend may span several, and compaction.model is the summariser');
+});
+suite('recurses into sub-agent traces, tagging depth and the spawning call', () => {
+    const rows = usageRows(richHistory());
+    const child = rows.find((r) => r.model === 'gemini-2.5-flash');
+    assert.ok(child, 'the sub-agent turn produced a row');
+    assert.is(child.subAgentDepth, 1);
+    assert.is(child.subAgentOf, 'tc1');
+    assert.is(child.agentName, 'worker');
+    const parent = rows.find((r) => r.model === 'claude-sonnet-5');
+    assert.is(parent.subAgentDepth, 0);
+    assert.is(parent.subAgentOf, undefined);
+});
+suite('prefers the provider the driver stamped over re-deriving it from the model', () => {
+    // `ChatMessage.provider` comes from the resolved provider's own status, so it is
+    // authoritative and knows vendors the model allowlists do not. Re-deriving from the
+    // model id alone drops attribution for a server-proxied turn — exactly the rows a
+    // usage ledger cares most about getting right. PR review.
+    const rows = usageRows([
+        msg({ model: 'gpt-5', provider: 'openai', cost: 1, inputTokens: 1 }),
+        // Stamped provider disagrees with what the model id would derive: the stamp still wins.
+        msg({
+            model: 'claude-sonnet-5',
+            provider: 'chrome',
+            cost: 1,
+            inputTokens: 1,
+        }),
+    ]);
+    assert.equal(rows.map((r) => r.provider), ['openai', 'chrome']);
+});
+suite('resolves the provider from the model, and leaves it unset when unknown', () => {
+    const rows = usageRows([
+        msg({ model: 'claude-sonnet-5', cost: 1, inputTokens: 1 }),
+        msg({ model: 'gemini-2.5-flash', cost: 1, inputTokens: 1 }),
+        msg({ model: 'gpt-5', cost: 1, inputTokens: 1 }),
+        msg({ cost: 1, inputTokens: 1 }),
+    ]);
+    assert.equal(rows.map((r) => r.provider), ['anthropic', 'gemini', undefined, undefined], 'an unrecognised or absent model leaves provider unset rather than guessing');
+});
+suite('splits the prompt into buckets the same way sumUsage does', () => {
+    const [row] = usageRows([
+        msg({
+            model: 'claude-sonnet-5',
+            inputTokens: 1000,
+            cacheReadTokens: 600,
+            cacheWriteTokens: 300,
+        }),
+    ]);
+    assert.is(row.uncachedInputTokens, 100, 'the remainder, not an addition');
+    assert.is(row.cacheReadTokens, 600);
+    assert.is(row.cacheWriteTokens, 300);
+});
+suite('clamps a prompt whose cache buckets exceed the total', () => {
+    const [row] = usageRows([msg({ inputTokens: 100, cacheReadTokens: 900 })]);
+    assert.is(row.uncachedInputTokens, 0, 'never negative');
+});
+suite('leaves costUsd undefined when nothing was reported, rather than zero', () => {
+    // What the "carried usage but no cost" audit keys off. A zero here would read as a
+    // free call and the alarm would never fire.
+    const [row] = usageRows([msg({ model: 'claude-sonnet-5', inputTokens: 500, outputTokens: 10 })]);
+    assert.is(row.costUsd, undefined);
+    assert.is(row.source, 'request');
+});
+suite('records external (non-LLM) cost and counts it in the reconciliation', () => {
+    const history = [msg({ externalCostUsd: 0.4 }), msg({ model: 'claude-sonnet-5', cost: 0.1 })];
+    const rows = usageRows(history);
+    assert.is(rows.find((r) => r.externalCostUsd != null).externalCostUsd, 0.4);
+    const rowCost = rows.reduce((n, r) => { var _a, _b; return n + ((_a = r.costUsd) !== null && _a !== void 0 ? _a : 0) + ((_b = r.externalCostUsd) !== null && _b !== void 0 ? _b : 0); }, 0);
+    assert.ok(Math.abs(rowCost - sumUsage(history).costUsd) < 1e-12);
+});
+suite('produces no row for a message carrying no usage', () => {
+    assert.equal(usageRows([
+        msg({ role: 'user', content: 'hello' }),
+        msg({ content: 'thinking out loud', category: 'reasoning' }),
+    ]), []);
+});
+suite('is empty for an empty transcript', () => {
+    assert.equal(usageRows([]), []);
+});
+suite('reconciles on a nested delegation two levels deep', () => {
+    const history = [
+        msg({
+            model: 'claude-sonnet-5',
+            cost: 0.01,
+            inputTokens: 100,
+            toolCalls: [
+                {
+                    id: 'a',
+                    name: 'delegate',
+                    args: {},
+                    subAgentTrace: [
+                        msg({
+                            model: 'claude-sonnet-5',
+                            cost: 0.02,
+                            inputTokens: 200,
+                            toolCalls: [
+                                {
+                                    id: 'b',
+                                    name: 'delegate',
+                                    args: {},
+                                    subAgentTrace: [msg({ model: 'gemini-2.5-flash', cost: 0.04, inputTokens: 400 })],
+                                },
+                            ],
+                        }),
+                    ],
+                },
+            ],
+        }),
+    ];
+    const rows = usageRows(history);
+    assert.equal(rows.map((r) => r.subAgentDepth), [0, 1, 2]);
+    const rowCost = rows.reduce((n, r) => { var _a; return n + ((_a = r.costUsd) !== null && _a !== void 0 ? _a : 0); }, 0);
+    assert.ok(Math.abs(rowCost - sumUsage(history).costUsd) < 1e-12, 'still reconciles');
+});
+suite('does not mutate the transcript', () => {
+    const history = richHistory();
+    const before = JSON.stringify(history);
+    usageRows(history);
+    assert.is(JSON.stringify(history), before, 'a projection, not a transform');
+});
+suite.run();

package/dist/tsconfig.tsbuildinfo

@@ -1 +1 @@
-{"root":["../src/chat-driver-node.ts","../src/index.ts","../src/channel/ai-activity-bus.ts","../src/channel/ai-activity-channel.ts","../src/components/flowing-waves-indicator.ts","../src/components/halo-overlay.ts","../src/components/plasma-orb-indicator.ts","../src/components/waves-indicator.ts","../src/components/activity-halo/activity-halo.ts","../src/components/agent-picker/agent-picker.constants.ts","../src/components/agent-picker/agent-picker.styles.ts","../src/components/agent-picker/agent-picker.template.ts","../src/components/agent-picker/agent-picker.ts","../src/components/agent-picker/index.ts","../src/components/ai-driver/ai-driver.ts","../src/components/ai-driver/index.ts","../src/components/chat-bubble/chat-bubble.styles.ts","../src/components/chat-bubble/chat-bubble.template.ts","../src/components/chat-bubble/chat-bubble.ts","../src/components/chat-bubble/index.ts","../src/components/chat-driver/align-event-globals.ts","../src/components/chat-driver/chat-driver.compact.test.ts","../src/components/chat-driver/chat-driver.invocation-scope.test.ts","../src/components/chat-driver/chat-driver.test.ts","../src/components/chat-driver/chat-driver.ts","../src/components/chat-driver/index.ts","../src/components/chat-interaction-wrapper/chat-interaction-wrapper.styles.ts","../src/components/chat-interaction-wrapper/chat-interaction-wrapper.template.ts","../src/components/chat-interaction-wrapper/chat-interaction-wrapper.test.ts","../src/components/chat-interaction-wrapper/chat-interaction-wrapper.ts","../src/components/chat-interaction-wrapper/index.ts","../src/components/chat-markdown/chat-markdown.ts","../src/components/chat-markdown/index.ts","../src/components/orchestrating-driver/index.ts","../src/components/orchestrating-driver/orchestrating-driver.budget.test.ts","../src/components/orchestrating-driver/orchestrating-driver.pin.test.ts","../src/components/orchestrating-driver/orchestrating-driver.ts","../src/components/popout-manager/index.ts","../src/components/popout-manager/popout-manager.ts","../src/components/settings-modal/index.ts","../src/components/settings-modal/settings-modal.styles.test.ts","../src/components/settings-modal/settings-modal.styles.ts","../src/components/settings-modal/settings-modal.template.test.ts","../src/components/settings-modal/settings-modal.template.ts","../src/config/config.ts","../src/config/define-stateful-agent.test.ts","../src/config/define-stateful-agent.ts","../src/config/fallback-agents.ts","../src/config/index.ts","../src/config/validate-providers.test.ts","../src/config/validate-providers.ts","../src/main/blocked-state.test.ts","../src/main/budget-meter.test.ts","../src/main/cost-session-banking.test.ts","../src/main/index.ts","../src/main/main.styles.test.ts","../src/main/main.styles.ts","../src/main/main.template.ts","../src/main/main.ts","../src/main/main.types.ts","../src/main/popout-interaction-gate.test.ts","../src/provider/ai-provider-switcher.ts","../src/provider/assistant-app-settings.ts","../src/state/ai-assistant-slice.test.ts","../src/state/ai-assistant-slice.ts","../src/state/debug-event-log.test.ts","../src/state/debug-event-log.ts","../src/state/driver-registry.test.ts","../src/state/driver-registry.ts","../src/state/interaction-context.test.ts","../src/state/interaction-context.ts","../src/state/session-store.ts","../src/state/persistence/build-timeline-entries.ts","../src/state/persistence/diagnostics-cursors.test.ts","../src/state/persistence/diagnostics-cursors.ts","../src/state/persistence/diagnostics.test.ts","../src/state/persistence/diagnostics.ts","../src/state/persistence/index.ts","../src/state/persistence/persister-registry.ts","../src/state/persistence/session-persistence-provider.test.ts","../src/state/persistence/session-persistence-provider.ts","../src/state/persistence/session-persistence.integration.test.ts","../src/state/persistence/session-persister.test.ts","../src/state/persistence/session-persister.ts","../src/state/persistence/session-snapshot.test.ts","../src/state/persistence/session-snapshot.ts","../src/state/persistence/stateful-restore.e2e.test.ts","../src/styles/ai-colours.ts","../src/styles/settings-section.ts","../src/suggestions/chat-suggestions.ts","../src/tags/index.ts","../src/types/ai-chat-widget.ts","../src/types/interaction-context.ts","../src/utils/animated-panel-toggle.ts","../src/utils/animation-exclusivity.test.ts","../src/utils/animation-exclusivity.ts","../src/utils/banked-usage-baselines.ts","../src/utils/collect-session-models.test.ts","../src/utils/collect-session-models.ts","../src/utils/condense-history.test.ts","../src/utils/condense-history.ts","../src/utils/cost-session-history.test.ts","../src/utils/cost-session-history.ts","../src/utils/derive-cost-session-title.test.ts","../src/utils/derive-cost-session-title.ts","../src/utils/flatten-sub-agent-messages.test.ts","../src/utils/flatten-sub-agent-messages.ts","../src/utils/format-usd.ts","../src/utils/history-transform.test.ts","../src/utils/history-transform.ts","../src/utils/index.ts","../src/utils/logger.ts","../src/utils/message-partition.test.ts","../src/utils/message-partition.ts","../src/utils/resolve-cost-history-config.test.ts","../src/utils/resolve-cost-history-config.ts","../src/utils/resolve-preference-baseline.test.ts","../src/utils/resolve-preference-baseline.ts","../src/utils/strip-agent-handlers.test.ts","../src/utils/strip-agent-handlers.ts","../src/utils/sum-costs.test.ts","../src/utils/sum-costs.ts","../src/utils/sum-tokens.test.ts","../src/utils/sum-tokens.ts","../src/utils/sum-usage.test.ts","../src/utils/sum-usage.ts","../src/utils/tool-fold.ts","../src/utils/with-timeout.ts"],"version":"5.9.2"}
+{"root":["../src/chat-driver-node.ts","../src/index.ts","../src/channel/ai-activity-bus.ts","../src/channel/ai-activity-channel.ts","../src/components/flowing-waves-indicator.ts","../src/components/halo-overlay.ts","../src/components/plasma-orb-indicator.ts","../src/components/waves-indicator.ts","../src/components/activity-halo/activity-halo.ts","../src/components/agent-picker/agent-picker.constants.ts","../src/components/agent-picker/agent-picker.styles.ts","../src/components/agent-picker/agent-picker.template.ts","../src/components/agent-picker/agent-picker.ts","../src/components/agent-picker/index.ts","../src/components/ai-driver/ai-driver.ts","../src/components/ai-driver/index.ts","../src/components/chat-bubble/chat-bubble.styles.ts","../src/components/chat-bubble/chat-bubble.template.ts","../src/components/chat-bubble/chat-bubble.ts","../src/components/chat-bubble/index.ts","../src/components/chat-driver/align-event-globals.ts","../src/components/chat-driver/chat-driver.compact.test.ts","../src/components/chat-driver/chat-driver.invocation-scope.test.ts","../src/components/chat-driver/chat-driver.test.ts","../src/components/chat-driver/chat-driver.thinking-policy.test.ts","../src/components/chat-driver/chat-driver.trace-capture.test.ts","../src/components/chat-driver/chat-driver.ts","../src/components/chat-driver/index.ts","../src/components/chat-interaction-wrapper/chat-interaction-wrapper.styles.ts","../src/components/chat-interaction-wrapper/chat-interaction-wrapper.template.ts","../src/components/chat-interaction-wrapper/chat-interaction-wrapper.test.ts","../src/components/chat-interaction-wrapper/chat-interaction-wrapper.ts","../src/components/chat-interaction-wrapper/index.ts","../src/components/chat-markdown/chat-markdown.ts","../src/components/chat-markdown/index.ts","../src/components/orchestrating-driver/index.ts","../src/components/orchestrating-driver/orchestrating-driver.budget.test.ts","../src/components/orchestrating-driver/orchestrating-driver.pin.test.ts","../src/components/orchestrating-driver/orchestrating-driver.ts","../src/components/popout-manager/index.ts","../src/components/popout-manager/popout-manager.ts","../src/components/settings-modal/index.ts","../src/components/settings-modal/settings-modal.styles.test.ts","../src/components/settings-modal/settings-modal.styles.ts","../src/components/settings-modal/settings-modal.template.test.ts","../src/components/settings-modal/settings-modal.template.ts","../src/config/config.ts","../src/config/define-stateful-agent.test.ts","../src/config/define-stateful-agent.ts","../src/config/fallback-agents.ts","../src/config/index.ts","../src/config/validate-providers.test.ts","../src/config/validate-providers.ts","../src/main/blocked-state.test.ts","../src/main/budget-meter.test.ts","../src/main/cost-session-banking.test.ts","../src/main/index.ts","../src/main/main.styles.test.ts","../src/main/main.styles.ts","../src/main/main.template.ts","../src/main/main.ts","../src/main/main.types.ts","../src/main/popout-interaction-gate.test.ts","../src/provider/ai-provider-switcher.ts","../src/provider/assistant-app-settings.ts","../src/state/ai-assistant-slice.test.ts","../src/state/ai-assistant-slice.ts","../src/state/debug-event-log.test.ts","../src/state/debug-event-log.ts","../src/state/driver-registry.test.ts","../src/state/driver-registry.ts","../src/state/interaction-context.test.ts","../src/state/interaction-context.ts","../src/state/session-store.ts","../src/state/persistence/build-timeline-entries.ts","../src/state/persistence/diagnostics-cursors.test.ts","../src/state/persistence/diagnostics-cursors.ts","../src/state/persistence/diagnostics.test.ts","../src/state/persistence/diagnostics.ts","../src/state/persistence/index.ts","../src/state/persistence/persister-registry.ts","../src/state/persistence/session-persistence-provider.test.ts","../src/state/persistence/session-persistence-provider.ts","../src/state/persistence/session-persistence.integration.test.ts","../src/state/persistence/session-persister.test.ts","../src/state/persistence/session-persister.ts","../src/state/persistence/session-snapshot.test.ts","../src/state/persistence/session-snapshot.ts","../src/state/persistence/stateful-restore.e2e.test.ts","../src/styles/ai-colours.ts","../src/styles/settings-section.ts","../src/suggestions/chat-suggestions.ts","../src/tags/index.ts","../src/types/ai-chat-widget.ts","../src/types/interaction-context.ts","../src/utils/animated-panel-toggle.ts","../src/utils/animation-exclusivity.test.ts","../src/utils/animation-exclusivity.ts","../src/utils/banked-usage-baselines.ts","../src/utils/collect-session-models.test.ts","../src/utils/collect-session-models.ts","../src/utils/condense-history.test.ts","../src/utils/condense-history.ts","../src/utils/cost-session-history.test.ts","../src/utils/cost-session-history.ts","../src/utils/derive-cost-session-title.test.ts","../src/utils/derive-cost-session-title.ts","../src/utils/flatten-sub-agent-messages.test.ts","../src/utils/flatten-sub-agent-messages.ts","../src/utils/format-usd.ts","../src/utils/history-transform.test.ts","../src/utils/history-transform.ts","../src/utils/index.ts","../src/utils/logger.ts","../src/utils/message-partition.test.ts","../src/utils/message-partition.ts","../src/utils/resolve-cost-history-config.test.ts","../src/utils/resolve-cost-history-config.ts","../src/utils/resolve-preference-baseline.test.ts","../src/utils/resolve-preference-baseline.ts","../src/utils/strip-agent-handlers.test.ts","../src/utils/strip-agent-handlers.ts","../src/utils/sum-costs.test.ts","../src/utils/sum-costs.ts","../src/utils/sum-tokens.test.ts","../src/utils/sum-tokens.ts","../src/utils/sum-usage.test.ts","../src/utils/sum-usage.ts","../src/utils/tool-fold.ts","../src/utils/usage-rows.test.ts","../src/utils/usage-rows.ts","../src/utils/with-timeout.ts"],"version":"5.9.2"}

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@genesislcap/ai-assistant",
   "description": "Genesis AI Assistant micro-frontend",
-  "version": "15.11.0",
+  "version": "15.12.0",
   "license": "SEE LICENSE IN license.txt",
   "main": "dist/esm/index.js",
   "types": "dist/ai-assistant.d.ts",
@@ -73,26 +73,26 @@
     }
   },
   "devDependencies": {
-    "@genesislcap/foundation-testing": "15.11.0",
-    "@genesislcap/genx": "15.11.0",
-    "@genesislcap/rollup-builder": "15.11.0",
-    "@genesislcap/ts-builder": "15.11.0",
-    "@genesislcap/uvu-playwright-builder": "15.11.0",
-    "@genesislcap/vite-builder": "15.11.0",
-    "@genesislcap/webpack-builder": "15.11.0",
+    "@genesislcap/foundation-testing": "15.12.0",
+    "@genesislcap/genx": "15.12.0",
+    "@genesislcap/rollup-builder": "15.12.0",
+    "@genesislcap/ts-builder": "15.12.0",
+    "@genesislcap/uvu-playwright-builder": "15.12.0",
+    "@genesislcap/vite-builder": "15.12.0",
+    "@genesislcap/webpack-builder": "15.12.0",
     "@types/dompurify": "^3.0.5",
     "@types/marked": "^5.0.2",
     "esbuild": "0.25.12"
   },
   "dependencies": {
-    "@genesislcap/foundation-ai": "15.11.0",
-    "@genesislcap/foundation-logger": "15.11.0",
-    "@genesislcap/foundation-notifications": "15.11.0",
-    "@genesislcap/foundation-redux": "15.11.0",
-    "@genesislcap/foundation-ui": "15.11.0",
-    "@genesislcap/foundation-utils": "15.11.0",
-    "@genesislcap/rapid-design-system": "15.11.0",
-    "@genesislcap/web-core": "15.11.0",
+    "@genesislcap/foundation-ai": "15.12.0",
+    "@genesislcap/foundation-logger": "15.12.0",
+    "@genesislcap/foundation-notifications": "15.12.0",
+    "@genesislcap/foundation-redux": "15.12.0",
+    "@genesislcap/foundation-ui": "15.12.0",
+    "@genesislcap/foundation-utils": "15.12.0",
+    "@genesislcap/rapid-design-system": "15.12.0",
+    "@genesislcap/web-core": "15.12.0",
     "dompurify": "^3.3.1",
     "marked": "^17.0.3"
   },
@@ -105,5 +105,5 @@
     "access": "public"
   },
   "customElements": "dist/custom-elements.json",
-  "gitHead": "261d900881ea1ca2b613033af497ffd449868eb0"
+  "gitHead": "06f5ab51086c6616c6b96a7a1aebccddb3af7881"
 }

package/src/chat-driver-node.ts

@@ -44,8 +44,33 @@ export {
   SUPPORTED_ANTHROPIC_MODEL_IDS,
   SUPPORTED_GEMINI_MODEL_IDS,
   isObservableAIProviderRegistry,
+  // Request pricing. Re-exported HERE, not left to a direct `@genesislcap/foundation-ai`
+  // import, for the same reason as everything above it: a headless host that adds its own
+  // foundation-ai dependency gets a second copy, and the driver's malformed/truncated
+  // handling does `instanceof` on the transports' error classes. Reaching the pricing
+  // should not cost a consumer that guarantee.
+  //
+  // These take RAW provider usage rather than a `ChatMessage` — see the foundation-ai
+  // module docs. They are for the path where nothing stamped a cost (a proxied or
+  // server-side call holding a usage block); when a message already carries `cost`, use
+  // it, because it was computed with the per-TTL cache split the message no longer has.
+  ANTHROPIC_CACHE_READ_MULTIPLIER,
+  ANTHROPIC_CACHE_WRITE_1H_MULTIPLIER,
+  ANTHROPIC_CACHE_WRITE_5M_MULTIPLIER,
+  anthropicRatesFor,
+  anthropicTokenCost,
+  GEMINI_CACHED_INPUT_MULTIPLIER,
+  GEMINI_LONG_CONTEXT_THRESHOLD,
+  geminiRatesFor,
+  geminiTokenCost,
+  // Resolves a model id to its vendor, so a caller holding only `message.model` can pick
+  // between the two cost functions above without maintaining its own model→vendor map.
+  vendorOfModel,
 } from '@genesislcap/foundation-ai';
 export type {
+  // Re-exported alongside `sumUsage` below — a consumer that can call the function but cannot NAME
+  // its return type has to restate the shape by hand, which is how the buckets get mis-summed.
+  AggregateUsage,
   AIProvider,
   AIProviderRegistry,
   AnthropicModelId,
@@ -53,12 +78,45 @@ export type {
   ChatFallback,
   ChatMessage,
   ChatRequestOptions,
+  ChatThinkingPolicy,
   ChatToolChoice,
   ChatToolDefinition,
   ChatToolHandlers,
   GeminiModelId,
+  // Pricing input/output shapes, alongside the functions re-exported above — a caller that
+  // can invoke them but cannot name their argument or result has to restate the shape by
+  // hand, which is how the token buckets get mixed up between the two providers.
+  AnthropicUsageRecord,
+  GeminiUsageRecord,
+  TokenCost,
+  TokenCostBreakdown,
+  TokenRates,
 } from '@genesislcap/foundation-ai';
 
+// Usage/cost summation. A headless consumer attributing spend to a run has no other route to it:
+// the transports' `getLifetimeCost()` is per-instance and lifetime-scoped, so a shared registry
+// makes it unattributable, and re-deriving cost from token counts is wrong in the one direction
+// that matters (`inputTokens` is total prompt size INCLUDING the cached part, so re-pricing charges
+// every cached token at full rate). `sumUsage` reads the per-request cost the transports already
+// stamped, and recurses into `toolCalls[].subAgentTrace` — where the entire spend of a delegating
+// agent lives — and into `compaction.rolledUpUsage`, which is spend whose messages no longer exist.
+// Both are invisible when omitted: the total is merely too low, which reads as a cheap run.
+//
+// `addUsage`/`emptyUsage` are the composition primitives for a conversation that CONTINUES across
+// rounds — see the guidance on `sumUsage` itself. `totalTokens` exists because the four buckets of
+// an `AggregateUsage` are disjoint and safe to add, which the per-message fields are NOT.
+export { addUsage, emptyUsage, sumUsage, totalTokens } from './utils/sum-usage';
+
+// Per-call projection, for the cases an aggregate cannot serve: usage rows for
+// per-project attribution, a cost dashboard, or auditing which turns came back
+// unpriced. `usageRows` is guaranteed to reconcile with `sumUsage` — including spend a
+// compaction banked, which has no message of its own and which a naive per-message walk
+// therefore drops while the total keeps it. Both take RAW history; `UsageRow` is
+// deliberately not a `ChatMessage`, so passing rows back into `sumUsage` (which recurses
+// itself, and would double-count) does not compile.
+export { usageRows } from './utils/usage-rows';
+export type { UsageRow } from './utils/usage-rows';
+
 // Debug-log harvesting (GENC-1461 unified diagnostics). A headless consumer collates its driver's
 // log — `buildTimelineEntries({ turnSnapshots: driver.getTurnSnapshots(), messages:
 // driver.getHistory(), metaEvents: getMetaEvents(sessionKey) })` → `DiagnosticEntry[]` — ships it to

package/src/components/chat-driver/chat-driver.thinking-policy.test.ts

@@ -0,0 +1,185 @@
+import type {
+  AIProvider,
+  AIProviderRegistry,
+  ChatMessage,
+  ChatRequestOptions,
+  ChatThinkingPolicy,
+} from '@genesislcap/foundation-ai';
+import { assert, createLogicSuite } from '@genesislcap/foundation-testing';
+import type { AgentConfig } from '../../config/config';
+// Side-effect import — MUST come before `./chat-driver` so the driver subclasses
+// jsdom's EventTarget rather than Node's native one. Mirrors chat-driver.test.ts.
+import './align-event-globals';
+import { ChatDriver } from './chat-driver';
+
+// ---------------------------------------------------------------------------
+// Per-turn `thinkingPolicy` resolution.
+//
+// The cost case for this option is the SHAPE of a tool loop, not a single call: the
+// opening turn is a real decision and worth reasoning over, while the iterations that
+// follow mostly pick the next tool from a narrow set — and on a thinking model each of
+// those bills reasoning at the full output rate. That saving only exists if the
+// resolver runs per iteration, so these tests assert the sequence across a loop, not
+// just that one value arrives.
+//
+// The other half is the undefined case. Every agent written before this option leaves
+// it unset, and a resolver may answer `undefined` on any given turn; both must reach
+// the transport as `undefined` so the model keeps its own default. `undefined` reaching
+// the wire as `'off'` would be a silent capability regression, and as `'auto'` a silent
+// bill increase — so it is asserted explicitly rather than assumed.
+// ---------------------------------------------------------------------------
+
+/** Captures the policy seen on each model call; calls one tool, then finishes. */
+const capturingProvider = (): AIProvider & { seen: Array<ChatThinkingPolicy | undefined> } => {
+  const seen: Array<ChatThinkingPolicy | undefined> = [];
+  let turns = 0;
+  return {
+    seen,
+    chat: async (
+      _history: ChatMessage[],
+      _userMessage: string,
+      options?: ChatRequestOptions,
+    ): Promise<ChatMessage> => {
+      seen.push(options?.thinkingPolicy);
+      turns += 1;
+      // Two tool calls, so the loop runs three model calls in total — enough for a
+      // per-turn resolver to say something different on the later ones.
+      if (turns <= 2) {
+        return {
+          role: 'assistant',
+          content: '',
+          toolCalls: [{ id: `t${turns}`, name: 'step', args: {} }],
+        };
+      }
+      return { role: 'assistant', content: 'done' };
+    },
+  };
+};
+
+const makeRegistry = (provider: AIProvider): AIProviderRegistry => ({
+  get: () => provider,
+  default: () => provider,
+  defaultName: () => 'test',
+  names: () => ['test'],
+  getStatus: async () => null,
+  listStatuses: async () => [],
+});
+
+const agent = (overrides: Partial<AgentConfig>): AgentConfig =>
+  ({
+    name: 'worker',
+    description: 'test agent',
+    toolDefinitions: [
+      { name: 'step', description: 'step', parameters: { type: 'object', properties: {} } },
+    ],
+    toolHandlers: { step: async () => 'stepped' },
+    ...overrides,
+  }) as AgentConfig;
+
+/** Run one user turn through a driver carrying `config`, and return the policies seen. */
+const policiesFor = async (
+  config: Partial<AgentConfig>,
+): Promise<Array<ChatThinkingPolicy | undefined>> => {
+  const provider = capturingProvider();
+  const driver = new ChatDriver(makeRegistry(provider), {
+    maxToolIterations: 10,
+    maxFoldOperations: 5,
+    sessionKey: '',
+  });
+  driver.applyAgent(agent(config));
+  await driver.sendMessage('go');
+  return provider.seen;
+};
+
+const suite = createLogicSuite('ChatDriver thinkingPolicy');
+
+suite('leaves the policy undefined when the agent does not set one', async () => {
+  // The compatibility case: every existing agent. Undefined must reach the transport
+  // as undefined so each model keeps its own default posture — not silently coerced
+  // to 'off' (a capability regression) or 'auto' (a bill increase).
+  const seen = await policiesFor({});
+  assert.ok(seen.length >= 3, `expected a multi-call loop, got ${seen.length}`);
+  assert.equal(
+    seen.filter((p) => p !== undefined),
+    [],
+    'no turn invents a policy',
+  );
+});
+
+suite('applies a static policy to every turn of the loop', async () => {
+  const seen = await policiesFor({ thinkingPolicy: 'off' });
+  assert.ok(seen.length >= 3);
+  assert.equal([...new Set(seen)], ['off'], 'a static value is not just a first-turn setting');
+});
+
+suite('pins the resolved policy for the whole tool loop', async () => {
+  // A tool-use loop is ONE assistant turn and Anthropic requires a single thinking mode
+  // for its duration. Toggling part-way does not error — the API silently disables
+  // thinking for that request and strips blocks that would leave the turn structure
+  // invalid, so an 'auto' -> 'off' switch loses the continuity the opening call
+  // established and 'off' -> 'auto' never delivers the reasoning asked for. It also
+  // invalidates the prompt cache, costing more than the reasoning it meant to save.
+  //
+  // So a resolver that changes its mind mid-loop must NOT be honoured mid-loop. This
+  // asserts the opposite of what it looks like it should: the later values are ignored.
+  let call = 0;
+  const seen = await policiesFor({
+    thinkingPolicy: () => {
+      call += 1;
+      return call === 1 ? 'auto' : 'off';
+    },
+  });
+  assert.ok(seen.length >= 3, `expected a multi-call loop, got ${seen.length}`);
+  assert.equal(
+    [...new Set(seen)],
+    ['auto'],
+    'the first call decides; later resolutions do not take effect until the next user turn',
+  );
+});
+
+suite('re-resolves on the next user turn', async () => {
+  // The flip side: pinning is per turn, not for the driver's lifetime, so a state change
+  // between turns still lands.
+  const provider = capturingProvider();
+  const driver = new ChatDriver(makeRegistry(provider), {
+    maxToolIterations: 10,
+    maxFoldOperations: 5,
+    sessionKey: '',
+  });
+  let turn = 0;
+  driver.applyAgent(
+    agent({
+      thinkingPolicy: () => (turn === 0 ? 'auto' : 'off'),
+    }),
+  );
+  await driver.sendMessage('go');
+  const firstTurn = [...provider.seen];
+  turn = 1;
+  await driver.sendMessage('again');
+  const secondTurn = provider.seen.slice(firstTurn.length);
+
+  assert.equal([...new Set(firstTurn)], ['auto'], 'turn one holds its posture');
+  assert.equal([...new Set(secondTurn)], ['off'], 'turn two picks up the new one');
+});
+
+suite('passes undefined through when the resolver declines to choose', async () => {
+  // A resolver may answer on some turns and not others. `undefined` is its third answer —
+  // "leave this model alone" — and must not be normalised into a value. Asserted across
+  // the whole loop because the first call's answer is the one that gets pinned.
+  const seen = await policiesFor({ thinkingPolicy: () => undefined });
+  assert.ok(seen.length >= 3);
+  assert.equal(
+    seen.filter((p) => p !== undefined),
+    [],
+    'declining is not the same as choosing',
+  );
+});
+
+suite('awaits an async resolver', async () => {
+  const seen = await policiesFor({
+    thinkingPolicy: async () => 'off' as ChatThinkingPolicy,
+  });
+  assert.equal([...new Set(seen)], ['off'], 'a promise is resolved, not passed through');
+});
+
+suite.run();