omniharness-cli 0.1.83 → 0.1.85

package/dist/config/omniRoute.js

@@ -32,6 +32,7 @@ export class OmniRouteClient {
     metrics = {
         compression: { inputTokens: 0, compressedTokens: 0, ratio: 1, strategy: 'none', updatedAt: new Date().toISOString() },
         fallback: { attempts: 0 },
+        usage: { contextTokens: 0, tokensIn: 0, tokensOut: 0, costUsd: 0 },
         requestCount: 0,
     };
     constructor(config = {}) {
@@ -144,17 +145,27 @@ export class OmniRouteClient {
     }
     /** List every model id the gateway exposes, including `auto/*` virtual combos and individual providers. */
     async listModels(signal) {
+        return (await this.listCatalog(signal)).map((entry) => entry.id);
+    }
+    /**
+     * The catalog with the context window each entry advertises. OmniRoute
+     * states `context_length` on providers' models, on `auto/*` engines and on
+     * combos alike; `max_input_tokens` stands in when an entry carries only that.
+     */
+    async listCatalog(signal) {
         const response = await this.requestWithRetry('/v1/models', { method: 'GET', signal });
         const payload = await response.json();
         const data = this.isRecord(payload) && Array.isArray(payload.data) ? payload.data : [];
-        const ids = [];
+        const entries = [];
         for (const entry of data) {
-            if (this.isRecord(entry) && typeof entry.id === 'string' && entry.id.trim() !== '')
-                ids.push(entry.id);
+            if (!this.isRecord(entry) || typeof entry.id !== 'string' || entry.id.trim() === '')
+                continue;
+            const window = this.positive(entry.context_length) ?? this.positive(entry.max_input_tokens);
+            entries.push(window !== undefined ? { id: entry.id, contextLength: window } : { id: entry.id });
         }
-        if (ids.length === 0)
+        if (entries.length === 0)
             throw new OmniRouteError(response.status, 'invalid models response');
-        return ids;
+        return entries;
     }
     /** Retry transient responses for idempotent metadata reads without touching chat/tool requests. */
     async requestWithRetry(path, init) {
@@ -186,7 +197,19 @@ export class OmniRouteClient {
         let lineBuffer = '';
         // Partially-accumulated tool calls keyed by stream index.
         const toolStreams = new Map();
+        // OmniRoute's response headers are sent at stream start, before the
+        // latency, usage and cost are known, so on a stream they carry zeros. With
+        // OMNIROUTE_SSE_COMMENTS on, the gateway ends the stream with the same
+        // fields as `: x-omniroute-<name>=<value>` comment lines — the values that
+        // were final. They are collected here and read like a second header set.
+        const trailer = new Headers();
         const flushData = (line) => {
+            if (line.startsWith(':')) {
+                const meta = /^:\s*(x-omniroute-[a-z0-9-]+)=(.*)$/i.exec(line);
+                if (meta)
+                    trailer.set(meta[1].toLowerCase(), meta[2].trim());
+                return;
+            }
             const sep = line.indexOf(':');
             if (sep === -1)
                 return;
@@ -274,6 +297,10 @@ export class OmniRouteClient {
         const toolCalls = [...toolStreams.values()]
             .filter((entry) => entry.id !== '' && entry.name !== '')
             .map((entry) => ({ id: entry.id, type: 'function', function: { name: entry.name, arguments: entry.argsFragments.join('') } }));
+        // The trailer names the provider and model that finished the stream, so it
+        // supersedes the routing picture taken from the initial headers.
+        this.updateMetrics(trailer);
+        this.recordCompletion(response.headers, usage, trailer);
         return { content, model: answered, finishReason, reasoning: reasoning || undefined, toolCalls, usage, headers: response.headers, compression: this.compressionFrom(response.headers) };
     }
     async chat(model, messages, options = {}) {
@@ -294,6 +321,11 @@ export class OmniRouteClient {
         // (the combo may have routed anywhere); fall back to the requested id.
         const answered = typeof payload.model === 'string' && payload.model.trim() !== '' ? payload.model : model;
         const toolCalls = Array.isArray(message.tool_calls) ? message.tool_calls.map((call) => this.asToolCall(call)).filter((call) => call !== null) : undefined;
+        this.recordCompletion(response.headers, usage ? {
+            inputTokens: this.number(usage.prompt_tokens),
+            outputTokens: this.number(usage.completion_tokens),
+            totalTokens: this.number(usage.total_tokens),
+        } : undefined);
         return {
             content: typeof message.content === 'string' ? message.content : '',
             model: answered,
@@ -309,6 +341,50 @@ export class OmniRouteClient {
             compression: this.compressionFrom(response.headers),
         };
     }
+    /**
+     * OmniRoute's cost-telemetry set for one completion, or undefined when the
+     * headers carry no token counts. A stream's initial headers hold zeros for
+     * every field the gateway could not know yet, and zeros are treated as
+     * absent so they never overwrite a count read elsewhere.
+     */
+    usageFromHeaders(headers) {
+        const inputTokens = this.headerNumber(headers, 'x-omniroute-tokens-in') ?? 0;
+        const outputTokens = this.headerNumber(headers, 'x-omniroute-tokens-out') ?? 0;
+        if (inputTokens <= 0 && outputTokens <= 0)
+            return undefined;
+        const costUsd = this.headerNumber(headers, 'x-omniroute-response-cost');
+        const latencyMs = this.headerNumber(headers, 'x-omniroute-latency-ms');
+        return {
+            inputTokens: Math.max(0, inputTokens),
+            outputTokens: Math.max(0, outputTokens),
+            costUsd: costUsd !== undefined && costUsd > 0 ? costUsd : undefined,
+            latencyMs: latencyMs !== undefined && latencyMs > 0 ? latencyMs : undefined,
+        };
+    }
+    /**
+     * Fold one completion into the session's usage. Sources, most authoritative
+     * first: the SSE metadata trailer (final values of a stream), the response
+     * headers (final on a non-streaming reply, zeros on a stream), then the
+     * `usage` object in the body. Token counts come from one source only, so a
+     * reply that reports itself twice is still counted once.
+     */
+    recordCompletion(headers, body, trailer) {
+        const usage = this.metrics.usage;
+        const measured = (trailer && this.usageFromHeaders(trailer))
+            ?? this.usageFromHeaders(headers)
+            ?? (body && (body.inputTokens > 0 || body.outputTokens > 0) ? { inputTokens: body.inputTokens, outputTokens: body.outputTokens } : undefined);
+        if (!measured)
+            return;
+        if (measured.inputTokens > 0)
+            usage.contextTokens = measured.inputTokens;
+        usage.tokensIn += measured.inputTokens;
+        usage.tokensOut += measured.outputTokens;
+        if (measured.costUsd !== undefined)
+            usage.costUsd += measured.costUsd;
+        if (measured.latencyMs !== undefined)
+            usage.latencyMs = measured.latencyMs;
+        usage.updatedAt = new Date().toISOString();
+    }
     compressionFrom(headers) {
         const input = this.headerNumber(headers, 'x-omniroute-input-tokens');
         const compressed = this.headerNumber(headers, 'x-omniroute-compressed-tokens');
@@ -365,6 +441,9 @@ export class OmniRouteClient {
         const provider = headers.get('x-omniroute-provider') ?? decision.provider;
         if (provider)
             fallback.activeProvider = provider;
+        const model = headers.get('x-omniroute-model');
+        if (model && model.trim() !== '')
+            fallback.model = model.trim();
         if (decision.strategy)
             fallback.strategy = decision.strategy;
         if (decision.latencyMs !== undefined)
@@ -419,6 +498,9 @@ export class OmniRouteClient {
     number(value) {
         return typeof value === 'number' && Number.isFinite(value) ? value : 0;
     }
+    positive(value) {
+        return typeof value === 'number' && Number.isFinite(value) && value > 0 ? value : undefined;
+    }
     safeParse(text) {
         try {
             return JSON.parse(text);

package/dist/ui/modelWindows.js

@@ -6,6 +6,12 @@
  * a model id (or the provider name OmniRoute reports) to a token budget;
  * `contextMeter` turns "tokens in" into a fill fraction and a zone the UI
  * colours (ok / warn / danger) using the playbook's 70 / 90 thresholds.
+ *
+ * The gateway's catalog is the first source: `/v1/models` states a
+ * `context_length` per entry, and `windowIndex` turns that into a lookup the
+ * meter consults before the substring table below. The table remains the
+ * answer for a model the catalog does not size, and when the catalog could
+ * not be read at all.
  */
 /** Known windows keyed by a substring of the model / provider id (longest match wins). */
 const WINDOWS = [
@@ -35,12 +41,52 @@ const WINDOWS = [
 ];
 /** Fallback window when nothing matches — conservative so the meter warns early rather than late. */
 export const DEFAULT_WINDOW = 128_000;
+/**
+ * Build the lookup from a catalog. Each entry is keyed by its full id and,
+ * when the id carries a provider prefix, by the bare model name after it —
+ * `X-OmniRoute-Model` reports the upstream name (`claude-sonnet-4-6`) while
+ * the catalog lists it under a prefix (`cc/claude-sonnet-4-6`). The first
+ * entry to claim a bare name keeps it, so a `dual`-mode mirror never
+ * contradicts its primary.
+ */
+export function windowIndex(entries) {
+    const index = new Map();
+    for (const entry of entries) {
+        const tokens = entry.contextLength;
+        if (tokens === undefined || !Number.isFinite(tokens) || tokens <= 0)
+            continue;
+        const id = entry.id.trim().toLowerCase();
+        if (id === '')
+            continue;
+        if (!index.has(id))
+            index.set(id, tokens);
+        const slash = id.indexOf('/');
+        if (slash > 0 && slash < id.length - 1) {
+            const bare = id.slice(slash + 1);
+            if (!index.has(bare))
+                index.set(bare, tokens);
+        }
+    }
+    return index;
+}
 /**
  * Resolve a context window for a model id and/or the provider OmniRoute
- * reported for the turn. Matching is case-insensitive substring; the longest
- * matching pattern wins so `gpt-4o-mini` beats `gpt-4o`.
+ * reported for the turn. A catalog `known` answers first, by the exact id and
+ * then by the bare name after a provider prefix. Otherwise matching is
+ * case-insensitive substring; the longest matching pattern wins so
+ * `gpt-4o-mini` beats `gpt-4o`.
  */
-export function windowFor(modelId, provider) {
+export function windowFor(modelId, provider, known) {
+    if (known && modelId) {
+        const id = modelId.trim().toLowerCase();
+        const exact = known.get(id);
+        if (exact !== undefined)
+            return exact;
+        const slash = id.indexOf('/');
+        const bare = slash > 0 ? known.get(id.slice(slash + 1)) : undefined;
+        if (bare !== undefined)
+            return bare;
+    }
     const haystack = `${modelId ?? ''} ${provider ?? ''}`.toLowerCase();
     let best;
     let bestLen = 0;
@@ -53,8 +99,8 @@ export function windowFor(modelId, provider) {
     return best ?? DEFAULT_WINDOW;
 }
 /** Build the meter. `used` below 0 clamps to 0; the fraction is capped at 1. */
-export function contextMeter(used, modelId, provider) {
-    const window = windowFor(modelId, provider);
+export function contextMeter(used, modelId, provider, known) {
+    const window = windowFor(modelId, provider, known);
     const safeUsed = Math.max(0, used);
     const fraction = Math.min(1, safeUsed / window);
     const zone = fraction >= 0.9 ? 'danger' : fraction >= 0.7 ? 'warn' : 'ok';

package/dist/ui/terminalInterface.js

@@ -12,7 +12,7 @@ import { capabilityLine, recentRows, shortenPath, twoColumn } from './home.js';
 import { conversationWidth, overflowCount, sidebarMode, SIDEBAR_WIDTH, todoRows, usageRows, clip as clipRow } from './sidebar.js';
 import { planViewport } from './viewport.js';
 import { statusMarker, toolHead } from './toolrow.js';
-import { contextMeter, meterBar } from './modelWindows.js';
+import { contextMeter, meterBar, windowIndex } from './modelWindows.js';
 import { BEL, SYNC_QUERY, isSyncOutputReply, osc9Notify, osc52Copy, shouldNudgeOnFinish, wrapSynchronizedOutput } from './termcaps.js';
 import { KITTY_POP, KITTY_PUSH, KITTY_QUERY, isEncodedKey, isKittyQueryResponse, parseRawKey } from './keys.js';
 import { ownVersion } from '../update.js';
@@ -271,6 +271,9 @@ export function TerminalInterface({ engine }) {
     const [pickerItems, setPickerItems] = useState([]);
     const [pickerIndex, setPickerIndex] = useState(0);
     const [pickerError, setPickerError] = useState();
+    // Context windows the catalog states, keyed by model id. Empty until the
+    // catalog has been read; the meter falls back to its own table meanwhile.
+    const [windows, setWindows] = useState(() => new Map());
     const [mode, setMode] = useState(engine.state.mode);
     const [permMode, setPermMode] = useState(engine.state.permissionMode ?? 'ask');
     const [approval, setApproval] = useState(null);
@@ -310,6 +313,10 @@ export function TerminalInterface({ engine }) {
             if (alive)
                 setRecentSessions(found);
         }).catch(() => { });
+        void Promise.resolve().then(() => engine.client.listCatalog()).then((catalog) => {
+            if (alive)
+                setWindows(windowIndex(catalog));
+        }).catch(() => { });
         return () => { alive = false; };
     }, []);
     useEffect(() => {
@@ -434,7 +441,9 @@ export function TerminalInterface({ engine }) {
     const loadPicker = async () => {
         setPickerError(undefined);
         try {
-            const [accountCombos, modelIds] = await Promise.all([engine.client.listCombos(), engine.client.listModels()]);
+            const [accountCombos, catalog] = await Promise.all([engine.client.listCombos(), engine.client.listCatalog()]);
+            setWindows(windowIndex(catalog));
+            const modelIds = catalog.map((entry) => entry.id);
             const items = [];
             for (const combo of accountCombos) {
                 if (combo.name.trim() !== '' && !items.some((item) => item.id === combo.name)) {
@@ -981,9 +990,14 @@ export function TerminalInterface({ engine }) {
     const contentWidth = Math.max(20, convoWidth - 6);
     const terminalRows = rows;
     const metrics = engine.client.snapshotMetrics();
-    const meter = contextMeter(metrics.compression.inputTokens, engine.state.activeModel, metrics.fallback.activeProvider);
+    // The prompt tokens of the last completion, as the gateway counted them,
+    // are the size of the context the next turn will carry.
+    const contextTokens = metrics.usage?.contextTokens ?? 0;
+    // Sized to the model that answered, when the gateway named one: an `auto/*`
+    // engine or a combo can land anywhere, and the window is that model's.
+    const meter = contextMeter(contextTokens, metrics.fallback.model ?? engine.state.activeModel, metrics.fallback.activeProvider, windows);
     const meterColor = meter.zone === 'danger' ? PALETTE.error : meter.zone === 'warn' ? PALETTE.warn : PALETTE.muted;
-    const contextLabel = metrics.compression.inputTokens > 0 ? `ctx ${meterBar(meter.fraction, 8)} ${Math.round(meter.fraction * 100)}%` : '';
+    const contextLabel = contextTokens > 0 ? `ctx ${meterBar(meter.fraction, 8)} ${Math.round(meter.fraction * 100)}%` : '';
     const compression = metrics.compression.inputTokens > 0 ? `${Math.round((1 - metrics.compression.ratio) * 100)}% ${metrics.compression.strategy.toUpperCase()}` : '';
     const phase = busy && currentTool ? phaseFor(currentTool) : (busy ? 'working' : 'ready');
     const elapsedMs = busy && runStartedAt.current !== null ? Math.max(0, now - runStartedAt.current) : 0;
@@ -1009,7 +1023,7 @@ export function TerminalInterface({ engine }) {
     const liveThinkView = liveThinkLines.slice(-Math.max(2, Math.floor(liveBudget / 2)));
     const liveAnswerView = liveAnswerLines.slice(-liveBudget);
     const doneAgents = agents.filter((lane) => lane.status === 'done').length;
-    const panel = _jsx(SidebarPanel, { width: SIDEBAR_WIDTH, model: engine.state.activeModel, mode: mode, perm: permMode, workspace: engine.state.workspace.root, usage: { tokensIn: metrics.compression.inputTokens, requests: metrics.requestCount }, agents: agents, todos: taskQueue, skills: engine.skills.length, plugins: new Set(engine.skills.map((skill) => skill.source).filter((source) => source !== undefined)).size });
+    const panel = _jsx(SidebarPanel, { width: SIDEBAR_WIDTH, model: engine.state.activeModel, mode: mode, perm: permMode, workspace: engine.state.workspace.root, usage: { tokensIn: metrics.usage?.tokensIn, tokensOut: metrics.usage?.tokensOut, costUSD: metrics.usage?.costUsd, requests: metrics.requestCount }, agents: agents, todos: taskQueue, skills: engine.skills.length, plugins: new Set(engine.skills.map((skill) => skill.source).filter((source) => source !== undefined)).size });
     return _jsxs(Box, { flexDirection: "column", width: width, paddingLeft: gutter, paddingRight: gutter, children: [_jsx(Static, { items: lines, children: (line, index) => _jsx(TranscriptEntry, { line: line, width: contentWidth, fallbackModel: engine.state.activeModel }, index) }, staticKey), sideMode === 'replace' && _jsx(Box, { marginBottom: 1, children: panel }), _jsxs(Box, { flexDirection: "row", alignItems: "flex-start", children: [_jsxs(Box, { flexDirection: "column", width: sideMode === 'split' ? convoWidth : undefined, flexGrow: sideMode === 'split' ? 0 : 1, children: [view.showHero && _jsx(Hero, { width: contentWidth + 2, endpoint: engine.client.endpoint ?? 'omniroute', model: engine.state.activeModel, mode: mode, perm: permMode, workspace: engine.state.workspace.root, sessions: recentSessions, skills: engine.skills.length, plugins: new Set(engine.skills.map((skill) => skill.source).filter((source) => source !== undefined)).size, mcpTools: engine.mcpTools.length }), liveThink !== '' && _jsxs(Box, { flexDirection: "column", marginTop: 1, children: [_jsx(Text, { bold: true, color: PALETTE.warn, children: "\u00B7 thinking" }), liveThinkView.map((segments, index) => _jsx(SegmentText, { segments: segments, role: "thinking" }, index))] }), liveAnswer !== '' && _jsxs(Box, { flexDirection: "column", marginTop: 1, children: [_jsx(Text, { bold: true, color: PALETTE.accent, children: engine.state.activeModel }), liveAnswerView.map((segments, index) => _jsx(SegmentText, { segments: segments, role: "assistant" }, index))] }), (view.toolCards > 0 ? toolCards.slice(-view.toolCards) : []).map((card) => {
                                 const expanded = expandedTool === card.id;
                                 const status = card.status === 'running' ? 'running' : card.status === 'error' ? 'error' : 'done';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omniharness-cli",
-  "version": "0.1.83",
+  "version": "0.1.85",
   "description": "OmniHarness — local-first agent orchestration harness for OmniRoute.",
   "license": "MIT",
   "type": "module",