copilot-reverse 0.6.0 → 0.8.0

package/dist/changes.js

@@ -1,4 +1,14 @@
 export const APP_CHANGES = [
+    {
+        "version": "0.7.0",
+        "date": "2026-06-29",
+        "summary": "feat(tui): `/metrics` now reports token usage (in/out) and an estimated cost per model and overall — the worker records prompt/completion tokens for every request (persisted in SQLite), and cost is a list-price estimate (Copilot is flat-fee). User messages also get a highlighted bar in the transcript so they stand out from muted system notes and assistant output."
+    },
+    {
+        "version": "0.6.0",
+        "date": "2026-06-29",
+        "summary": "feat(tui): add a `/changes` command listing the 10 most recent releases (version, date, summary) with a link to the full CHANGELOG, and refocus the startup \"what's new\" banner on important messages — it now points to `/changes` instead of advertising a bug fix, and still self-suppresses after 3 launches."
+    },
     {
         "version": "0.5.5",
         "date": "2026-06-29",
@@ -38,15 +48,5 @@ export const APP_CHANGES = [
         "version": "0.3.0",
         "date": "2026-06-26",
         "summary": "Restore `web_search` and `web_fetch` for Claude Code through the gateway: the worker now runs these tools internally against Microsoft Web IQ in a transparent agentic loop, and a new `/web-search-support` command stores the WebIQ API key."
-    },
-    {
-        "version": "0.2.1",
-        "date": "2026-06-25",
-        "summary": "Fix `/login` hanging with no output: the device-code prompt is now shown immediately while authorization is pending, instead of being buffered behind the blocking token poll."
-    },
-    {
-        "version": "0.2.0",
-        "date": "2026-06-23",
-        "summary": "Recover tool calls that some models emit as inline XML text into structured tool calls, and add changeset-driven automatic versioning + npm publish on merge to master."
     }
 ];

package/dist/core/model-canonical.js

@@ -0,0 +1,35 @@
+// Maps Copilot's model ids to the canonical ids Claude Code recognises, so its native /model picker
+// lights up with friendly names, tier grouping, and the 1M-context badge instead of bare ids.
+//
+// Copilot advertises DOTTED ids (claude-opus-4.8); Claude Code's built-in table keys on DASHED ids
+// (claude-opus-4-8) and shows the 1M badge only when the id ends with the [1m] suffix. So OUTBOUND
+// (/v1/models, ANTHROPIC_MODEL) we dash every claude id + add [1m] for the families Claude Code knows
+// to be 1M; INBOUND the proxy strips [1m] and fuzzy-maps the dashed id back to Copilot's dotted id
+// (see bestModelMatch). Non-claude ids (gpt*, o3*) have no canonical form → pass through untouched.
+export const ONE_M_SUFFIX = "[1m]";
+// Dashed canonical ids whose Claude Code table carries a 1M window — only these get the [1m] badge.
+// Everything else stays at its default window. Anchored on the probed v2.1.195 binary table.
+const ONE_M_MODELS = new Set(["claude-opus-4-6", "claude-opus-4-7", "claude-opus-4-8", "claude-sonnet-4-6"]);
+// claude-<family>-<major>-<minor> -> "Family Major.Minor" (e.g. claude-opus-4-8 -> "Opus 4.8").
+function displayName(dashed) {
+    const m = /^claude-(opus|sonnet|haiku)-(\d+)-(\d+)$/.exec(dashed);
+    if (!m)
+        return dashed;
+    const [, fam, maj, min] = m;
+    return `${fam[0].toUpperCase()}${fam.slice(1)} ${maj}.${min}`;
+}
+// Outbound: Copilot id -> the id+display Claude Code's picker understands. Claude ids are dashed to the
+// canonical form; the 1M families get the [1m] suffix so the picker shows the badge and sizes context
+// to 1M. Non-claude ids echo back as-is so they still appear, just without native metadata.
+export function toCanonical(copilotId) {
+    if (!copilotId.startsWith("claude-"))
+        return { id: copilotId, display_name: copilotId };
+    const dashed = copilotId.replace(/\./g, "-");
+    const id = ONE_M_MODELS.has(dashed) ? `${dashed}${ONE_M_SUFFIX}` : dashed;
+    return { id, display_name: displayName(dashed) };
+}
+// Inbound: drop the [1m] picker suffix. The dashed canonical id then resolves back to Copilot's
+// dotted id via the router's exact map + fuzzy fallback; nothing else to do here.
+export function stripOneM(model) {
+    return model.endsWith(ONE_M_SUFFIX) ? model.slice(0, -ONE_M_SUFFIX.length) : model;
+}

package/dist/supervisor/db.js

@@ -9,12 +9,16 @@ export function openDb(file) {
       exit_code INTEGER, stderr_tail TEXT NOT NULL, backoff_ms INTEGER NOT NULL, marked_unhealthy INTEGER NOT NULL DEFAULT 0);
     CREATE TABLE IF NOT EXISTS request_log (
       id INTEGER PRIMARY KEY AUTOINCREMENT, ts INTEGER NOT NULL, endpoint TEXT NOT NULL,
-      model TEXT NOT NULL, status INTEGER NOT NULL, latency_ms INTEGER NOT NULL, error TEXT);
+      model TEXT NOT NULL, status INTEGER NOT NULL, latency_ms INTEGER NOT NULL, tokens_in INTEGER, tokens_out INTEGER, error TEXT);
   `);
-    // Migrate request_log tables created before the error column existed.
+    // Migrate request_log tables created before later columns existed.
     const cols = db.prepare(`PRAGMA table_info(request_log)`).all();
     if (!cols.some((c) => c.name === "error"))
         db.exec(`ALTER TABLE request_log ADD COLUMN error TEXT`);
+    if (!cols.some((c) => c.name === "tokens_in"))
+        db.exec(`ALTER TABLE request_log ADD COLUMN tokens_in INTEGER`);
+    if (!cols.some((c) => c.name === "tokens_out"))
+        db.exec(`ALTER TABLE request_log ADD COLUMN tokens_out INTEGER`);
     return db;
 }
 export function recordRestart(db, e) {
@@ -26,10 +30,10 @@ export function listRestarts(db, limit) {
     FROM restart_events ORDER BY ts DESC LIMIT ?`).all(limit);
 }
 export function recordRequest(db, m) {
-    db.prepare(`INSERT INTO request_log (ts, endpoint, model, status, latency_ms, error) VALUES (@ts, @endpoint, @model, @status, @latencyMs, @error)`)
-        .run({ error: null, ...m });
+    db.prepare(`INSERT INTO request_log (ts, endpoint, model, status, latency_ms, tokens_in, tokens_out, error) VALUES (@ts, @endpoint, @model, @status, @latencyMs, @tokensIn, @tokensOut, @error)`)
+        .run({ tokensIn: null, tokensOut: null, error: null, ...m });
 }
 export function recentRequests(db, limit) {
-    return db.prepare(`SELECT ts, endpoint, model, status, latency_ms as latencyMs, error FROM request_log ORDER BY ts DESC LIMIT ?`).all(limit)
-        .map(({ error, ...r }) => (error == null ? r : { ...r, error }));
+    return db.prepare(`SELECT ts, endpoint, model, status, latency_ms as latencyMs, tokens_in as tokensIn, tokens_out as tokensOut, error FROM request_log ORDER BY ts DESC LIMIT ?`).all(limit)
+        .map(({ tokensIn, tokensOut, error, ...r }) => ({ ...r, ...(tokensIn != null ? { tokensIn } : {}), ...(tokensOut != null ? { tokensOut } : {}), ...(error != null ? { error } : {}) }));
 }

package/dist/supervisor/index.js

@@ -29,7 +29,7 @@ export function startSupervisor() {
         },
         onWorkerMessage: (m) => {
             if (m.type === "request-metric") {
-                const sample = { ts: Date.now(), endpoint: m.endpoint, model: m.model, status: m.status, latencyMs: m.latencyMs, error: m.error };
+                const sample = { ts: Date.now(), endpoint: m.endpoint, model: m.model, status: m.status, latencyMs: m.latencyMs, tokensIn: m.tokensIn, tokensOut: m.tokensOut, error: m.error };
                 recordRequest(db, sample);
                 bus.emit("metric", sample);
             }

package/dist/tui/app.js

@@ -267,6 +267,10 @@ export function App({ registry, title, workerState = "starting", initialModel =
                         const tokens = Math.ceil(e.text.length / 4);
                         return (_jsxs(Box, { flexDirection: "column", children: [_jsxs(Text, { color: theme.accent, children: ["\u273D ", _jsxs(Text, { color: theme.muted, children: [frame, " ", loadingVerb(elapsed), "\u2026 (esc to interrupt \u00B7 ", fmtElapsed(elapsed), " \u00B7 \u2193 ", fmtTokens(tokens), " tokens \u00B7 thinking)"] })] }), e.text ? _jsx(Text, { color: color, children: e.text }) : null] }, i));
                     }
+                    // User turns get a clay-on-dark highlight bar so they stand out from muted system notes and
+                    // gray assistant output — a clear visual anchor for "this is what I said".
+                    if (e.type === "user")
+                        return _jsx(Box, { marginTop: 1, children: _jsx(Text, { backgroundColor: theme.accent, color: "black", bold: true, children: ` ${e.text.replace(/^›\s*/, "")} ` }) }, i);
                     return _jsx(Text, { color: color, children: e.text }, i);
                 }) }), body, _jsxs(Box, { flexDirection: "column", paddingX: 1, children: [_jsxs(Box, { children: [github && _jsxs(_Fragment, { children: [_jsx(Text, { color: theme.muted, children: "github " }), _jsx(Text, { color: github === "connected" ? theme.ready : theme.error, children: github === "connected" ? "✓" : "✗ /login" })] }), _jsxs(Text, { color: theme.muted, children: [github ? "  ·  " : "", "daemon "] }), _jsx(Text, { color: stateColor[state], children: state })] }), _jsxs(Box, { children: [_jsx(Text, { color: theme.muted, children: "web " }), _jsx(Text, { color: webBackend === "unavailable" ? theme.muted : theme.ready, children: webBackend === "webiq" ? "✓ webiq" : webBackend === "copilot" ? "✓ copilot" : "✗ /webiq" }), _jsx(Text, { color: theme.muted, children: "  \u00B7  " }), _jsx(ClientBadge, { name: "claude", status: status.claude }), _jsx(Text, { color: theme.muted, children: "  " }), _jsx(ClientBadge, { name: "codex", status: status.codex }), _jsx(Text, { color: theme.muted, children: "  \u00B7  /help" })] })] })] }));
 }

package/dist/tui/assistant/tools.js

@@ -31,7 +31,7 @@ export function buildActions(client) {
             const a = aggregate(await client.requests());
             if (!a.total)
                 return "no requests yet";
-            return `requests: ${a.total}, errors: ${a.errors}; ` + a.byModel.map((r) => `${r.model} n=${r.count} avg=${r.avgMs}ms`).join("; ");
+            return `requests: ${a.total}, errors: ${a.errors}, tokens: ${a.tokensIn}↑/${a.tokensOut}↓, est. cost: $${a.costUsd.toFixed(3)}; ` + a.byModel.map((r) => `${r.model} n=${r.count} avg=${r.avgMs}ms`).join("; ");
         },
     };
 }

package/dist/tui/panels/metrics-agg.js

@@ -1,21 +1,47 @@
 // A request "failed" if it returned a 4xx/5xx OR carried an error message — runaway streams finish
 // 200 but tag an error (model degenerated, cut early), and those are exactly what we want to surface.
 const isError = (s) => s.status >= 400 || s.error != null;
+// Indicative $/1M-token list prices (in, out) used ONLY to estimate spend — Copilot is flat-fee, so
+// this is "what these tokens would cost at provider list price", not a real bill. Matched by substring;
+// unknown models fall back to a mid GPT-4o-class rate. Update as needed; precision isn't the point.
+const PRICING = [
+    { match: "opus", in: 15, out: 75 },
+    { match: "sonnet", in: 3, out: 15 },
+    { match: "haiku", in: 0.8, out: 4 },
+    { match: "gpt-5", in: 1.25, out: 10 },
+    { match: "gpt-4o-mini", in: 0.15, out: 0.6 },
+    { match: "gpt-4o", in: 2.5, out: 10 },
+    { match: "o1", in: 15, out: 60 },
+];
+const RATE_FALLBACK = { in: 2.5, out: 10 };
+const rate = (model) => PRICING.find((p) => model.toLowerCase().includes(p.match)) ?? RATE_FALLBACK;
+export function estimateCost(model, tokensIn, tokensOut) {
+    const r = rate(model);
+    return (tokensIn * r.in + tokensOut * r.out) / 1_000_000;
+}
 export function aggregate(samples) {
     const map = new Map();
     let errors = 0;
     for (const s of samples) {
         if (isError(s))
             errors++;
-        const m = map.get(s.model) ?? { count: 0, sum: 0 };
+        const m = map.get(s.model) ?? { count: 0, sum: 0, tin: 0, tout: 0 };
         m.count++;
         m.sum += s.latencyMs;
+        m.tin += s.tokensIn ?? 0;
+        m.tout += s.tokensOut ?? 0;
         map.set(s.model, m);
     }
+    const byModel = [...map.entries()].map(([model, v]) => ({
+        model, count: v.count, avgMs: Math.round(v.sum / v.count),
+        tokensIn: v.tin, tokensOut: v.tout, costUsd: estimateCost(model, v.tin, v.tout),
+    }));
     return {
-        total: samples.length,
-        errors,
-        byModel: [...map.entries()].map(([model, v]) => ({ model, count: v.count, avgMs: Math.round(v.sum / v.count) })),
+        total: samples.length, errors,
+        tokensIn: byModel.reduce((n, r) => n + r.tokensIn, 0),
+        tokensOut: byModel.reduce((n, r) => n + r.tokensOut, 0),
+        costUsd: byModel.reduce((n, r) => n + r.costUsd, 0),
+        byModel,
     };
 }
 // The failed requests (status >= 400 or any tagged error), newest-first, capped at `limit`. This is

package/dist/tui/setup/clients.js

@@ -1,3 +1,4 @@
+import { toCanonical } from "../../core/model-canonical.js";
 export function claudeCodeConfig(e) {
     const base = `http://${e.host}:${e.port}/anthropic`;
     return {
@@ -6,11 +7,14 @@ export function claudeCodeConfig(e) {
     };
 }
 export const ONE_M_SUFFIX = "[1m]";
-// Claude Code switches to its 1M context window only when ANTHROPIC_MODEL ends with `[1m]` — that
-// suffix is its built-in signal for a 1M model. Mirror agent-maestro: append it for models whose
-// window is in the ~1M band (800K..1.5M). Without it Claude Code assumes 200K -> "context 100%"
-// and /compact fails. The proxy strips the suffix again before forwarding to Copilot.
+// Claude Code switches to its 1M window only when ANTHROPIC_MODEL ends with `[1m]`, and only matches
+// the model to its native picker entry when the id is the DASHED canonical form it knows
+// (claude-opus-4-8, not Copilot's dotted claude-opus-4.8). Route the default model through toCanonical
+// so it's both dashed and 1M-badged for the known families; for non-claude ids keep the legacy
+// context-window suffix. The proxy strips [1m] + fuzzy-maps back to Copilot before forwarding.
 export function withClaude1mSuffix(model, contextWindow) {
+    if (model.startsWith("claude-"))
+        return toCanonical(model).id;
     return contextWindow && contextWindow > 800_000 && contextWindow < 1_500_000 && !model.endsWith(ONE_M_SUFFIX)
         ? `${model}${ONE_M_SUFFIX}`
         : model;

package/dist/tui/slash/commands.js

@@ -24,12 +24,18 @@ export function buildRegistry(ctx, endpoint, opts = {}) {
                 return ["no request errors logged — everything's green ✓"];
             return errs.map((e) => `${new Date(e.ts).toISOString()} ${e.status} ${e.endpoint} ${e.model} — ${e.error ?? "(no message)"}`);
         } });
-    reg.add({ name: "/metrics", describe: "request metrics + recent errors", run: async (_a, c) => {
+    reg.add({ name: "/metrics", describe: "request metrics, tokens, cost + recent errors", run: async (_a, c) => {
             const reqs = await c.client.requests();
             const a = aggregate(reqs);
             if (!a.total)
                 return ["no requests yet"];
-            const lines = [`requests: ${a.total}  errors: ${a.errors}`, ...a.byModel.map((r) => `  ${r.model.padEnd(20)} n=${r.count} avg=${r.avgMs}ms`)];
+            const k = (n) => (n >= 1000 ? `${(n / 1000).toFixed(1)}k` : `${n}`);
+            const usd = (n) => `$${n < 1 ? n.toFixed(3) : n.toFixed(2)}`;
+            const lines = [
+                `requests: ${a.total}  errors: ${a.errors}  tokens: ${k(a.tokensIn)}↑ ${k(a.tokensOut)}↓  est. cost: ${usd(a.costUsd)}`,
+                ...a.byModel.map((r) => `  ${r.model.padEnd(20)} n=${r.count} avg=${r.avgMs}ms  ${k(r.tokensIn)}↑ ${k(r.tokensOut)}↓ ~${usd(r.costUsd)}`),
+                "  cost is a list-price estimate (Copilot is flat-fee)",
+            ];
             const errs = recentErrors(reqs, 5);
             if (errs.length) {
                 lines.push("recent errors:");

package/dist/version.js

@@ -1,2 +1,2 @@
 // AUTO-GENERATED by scripts/gen-version.mjs from package.json — do not edit.
-export const APP_VERSION = "0.6.0";
+export const APP_VERSION = "0.8.0";

package/dist/worker/anthropic-server.js

@@ -5,6 +5,7 @@ import { errorHint } from "./errors.js";
 import { CopilotAuthError } from "../providers/copilot/token.js";
 import { isGatewayTool } from "../core/server-tools.js";
 import { RunawayGuard } from "../core/stream-guard.js";
+import { toCanonical } from "../core/model-canonical.js";
 const frame = (event, data) => `event: ${event}\ndata: ${JSON.stringify(data)}\n\n`;
 const safeJson = (s) => { try {
     return JSON.parse(s);
@@ -22,9 +23,11 @@ const MAX_TOOL_ITERS = 5;
 const STREAM_DEADLINE_MS = 120_000;
 export function mountAnthropic(app, router, onMetric, runner) {
     // Model discovery — Anthropic list shape. Claude Desktop / Anthropic-protocol clients GET this
-    // before chatting; without it they 404 on the connection test.
+    // before chatting; without it they 404 on the connection test. Claude families are mapped to the
+    // canonical id + display Claude Code recognises (with [1m] for 1M models) so its native picker shows
+    // friendly names + the 1M badge; non-claude ids pass through. resolveModel maps them back inbound.
     app.get("/anthropic/v1/models", (_req, res) => {
-        res.json({ data: router.listModels().map((id) => ({ type: "model", id, display_name: id })), has_more: false });
+        res.json({ data: router.listModels().map((id) => ({ type: "model", ...toCanonical(id) })), has_more: false });
     });
     // Anthropic clients (Claude Code) call this to size the prompt and decide when to auto-compact.
     app.post("/anthropic/v1/messages/count_tokens", (req, res) => {
@@ -35,7 +38,7 @@ export function mountAnthropic(app, router, onMetric, runner) {
         const canon = anthropicRequestToCanonical(req.body);
         canon.model = router.resolveModel(canon.model);
         const provider = router.pick(canon.model);
-        const metric = (status, error) => onMetric({ endpoint: "/anthropic/v1/messages", model: canon.model, status, latencyMs: Date.now() - start, error });
+        const metric = (status, opts = {}) => onMetric({ endpoint: "/anthropic/v1/messages", model: canon.model, status, latencyMs: Date.now() - start, tokensIn: opts.tokensIn, tokensOut: opts.tokensOut, error: opts.error });
         try {
             if (canon.stream) {
                 res.setHeader("content-type", "text/event-stream");
@@ -157,7 +160,7 @@ export function mountAnthropic(app, router, onMetric, runner) {
                 res.write(frame("message_delta", { type: "message_delta", delta: { stop_reason: finalStop === "tool_use" ? "tool_use" : finalStop === "length" ? "max_tokens" : "end_turn" }, usage: deltaUsage }));
                 res.write(frame("message_stop", { type: "message_stop" }));
                 res.end();
-                metric(200, runaway ? `runaway stream cut (${runawayReason}) — model degenerated, ended early as max_tokens` : undefined);
+                metric(200, { tokensIn: inputTokens, tokensOut: sumCompletion, error: runaway ? `runaway stream cut (${runawayReason}) — model degenerated, ended early as max_tokens` : undefined });
             }
             else {
                 // Non-stream: same gateway loop without SSE — run gateway tools and re-complete until the
@@ -182,7 +185,7 @@ export function mountAnthropic(app, router, onMetric, runner) {
                 if (runner)
                     resp = { ...resp, content: resp.content.filter((b) => b.type !== "tool_use" || !isGatewayTool(b.name)) };
                 res.json(canonicalToAnthropicResponse(resp));
-                metric(200);
+                metric(200, { tokensIn: resp.usage?.promptTokens, tokensOut: resp.usage?.completionTokens });
             }
         }
         catch (err) {
@@ -201,7 +204,7 @@ export function mountAnthropic(app, router, onMetric, runner) {
                 res.write(frame("error", { type: "error", error: { type: errorType, message } }));
                 res.end();
             }
-            metric(status, message);
+            metric(status, { error: message });
         }
     });
 }

package/dist/worker/openai-server.js

@@ -18,7 +18,7 @@ export function mountOpenAI(app, router, onMetric) {
         const canon = openaiRequestToCanonical(req.body);
         canon.model = router.resolveModel(canon.model);
         const provider = router.pick(canon.model);
-        const metric = (status, error) => onMetric({ endpoint: "/openai/chat/completions", model: canon.model, status, latencyMs: Date.now() - start, error });
+        const metric = (status, opts = {}) => onMetric({ endpoint: "/openai/chat/completions", model: canon.model, status, latencyMs: Date.now() - start, tokensIn: opts.tokensIn, tokensOut: opts.tokensOut, error: opts.error });
         try {
             if (canon.stream) {
                 res.setHeader("content-type", "text/event-stream");
@@ -27,8 +27,11 @@ export function mountOpenAI(app, router, onMetric) {
                 const guard = new RunawayGuard();
                 const deadline = start + STREAM_DEADLINE_MS;
                 let runawayReason = "";
+                let usage;
                 for await (const chunk of provider.stream(canon)) {
                     res.write(canonicalChunkToOpenAISSE(chunk, id, canon.model));
+                    if (chunk.done)
+                        usage = chunk.usage;
                     // Backstop covers tool-call streams too: a model can loop on tool calls forever, which
                     // never feeds the text guard — the wall clock cuts those cleanly instead of freezing.
                     if (chunk.kind === "text" && guard.push(chunk.delta)) {
@@ -41,11 +44,12 @@ export function mountOpenAI(app, router, onMetric) {
                     }
                 }
                 res.end();
-                metric(200, runawayReason ? `runaway stream cut (${runawayReason}) — model degenerated, ended early` : undefined);
+                metric(200, { tokensIn: usage?.promptTokens, tokensOut: usage?.completionTokens, error: runawayReason ? `runaway stream cut (${runawayReason}) — model degenerated, ended early` : undefined });
             }
             else {
-                res.json(canonicalToOpenAIResponse(await provider.complete(canon)));
-                metric(200);
+                const resp = await provider.complete(canon);
+                res.json(canonicalToOpenAIResponse(resp));
+                metric(200, { tokensIn: resp.usage?.promptTokens, tokensOut: resp.usage?.completionTokens });
             }
         }
         catch (err) {
@@ -62,7 +66,7 @@ export function mountOpenAI(app, router, onMetric) {
                 res.write(`data: ${JSON.stringify({ error: { message } })}\n\n`);
                 res.end();
             }
-            metric(status, message);
+            metric(status, { error: message });
         }
     });
     // OpenAI Responses API — Codex speaks ONLY this after codex#7782 removed wire_api="chat". Codex
@@ -73,7 +77,7 @@ export function mountOpenAI(app, router, onMetric) {
         const canon = responsesRequestToCanonical(req.body);
         canon.model = router.resolveModel(canon.model);
         const provider = router.pick(canon.model);
-        const metric = (status, error) => onMetric({ endpoint: "/openai/responses", model: canon.model, status, latencyMs: Date.now() - start, error });
+        const metric = (status, opts = {}) => onMetric({ endpoint: "/openai/responses", model: canon.model, status, latencyMs: Date.now() - start, tokensIn: opts.tokensIn, tokensOut: opts.tokensOut, error: opts.error });
         try {
             if (canon.stream) {
                 res.setHeader("content-type", "text/event-stream");
@@ -119,11 +123,12 @@ export function mountOpenAI(app, router, onMetric) {
                 for (const f of sse.finish(usage, finish, argsByIdx))
                     res.write(f);
                 res.end();
-                metric(200, runawayReason ? `runaway stream cut (${runawayReason}) — model degenerated, ended early` : undefined);
+                metric(200, { tokensIn: usage?.promptTokens, tokensOut: usage?.completionTokens, error: runawayReason ? `runaway stream cut (${runawayReason}) — model degenerated, ended early` : undefined });
             }
             else {
-                res.json(canonicalToResponsesResponse(await provider.complete(canon)));
-                metric(200);
+                const resp = await provider.complete(canon);
+                res.json(canonicalToResponsesResponse(resp));
+                metric(200, { tokensIn: resp.usage?.promptTokens, tokensOut: resp.usage?.completionTokens });
             }
         }
         catch (err) {
@@ -138,7 +143,7 @@ export function mountOpenAI(app, router, onMetric) {
                 res.write(`data: ${JSON.stringify({ type: "error", message })}\n\n`);
                 res.end();
             }
-            metric(status, message);
+            metric(status, { error: message });
         }
     });
 }

package/dist/worker/router.js

@@ -1,5 +1,6 @@
 import { bestModelMatch } from "../core/fuzzy.js";
 import { FALLBACK_MODELS } from "../providers/copilot/models.js";
+import { stripOneM } from "../core/model-canonical.js";
 // M1: single provider. Model name is remapped to the provider's actual id.
 export class Router {
     providers;
@@ -16,12 +17,12 @@ export class Router {
     listModels() { return this.available.length ? this.available : FALLBACK_MODELS; }
     resolveModel(requested) {
         // Claude Code appends [1m] to signal its 1M context window; Copilot doesn't know that id, so
-        // strip it back to the real model before mapping/forwarding.
-        requested = requested.endsWith("[1m]") ? requested.slice(0, -4) : requested;
+        // strip it back to the canonical model before mapping/forwarding.
+        requested = stripOneM(requested);
         const mapped = this.modelMap[requested];
         if (mapped)
             return mapped;
-        // Fuzzy-match a near-miss id (e.g. claude-opus-4-8-20251101 -> claude-opus-4.8) to a real model.
+        // Fuzzy-match a near-miss id (e.g. canonical claude-opus-4-8 -> Copilot claude-opus-4.8) to a real model.
         if (this.available.length) {
             const match = bestModelMatch(requested, this.available);
             if (match)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "copilot-reverse",
-  "version": "0.6.0",
+  "version": "0.8.0",
   "description": "Interactive terminal app that exposes your GitHub Copilot subscription as local OpenAI- and Anthropic-compatible endpoints, with a self-healing daemon and a built-in assistant.",
   "type": "module",
   "license": "MIT",