ada-agent 0.6.1 → 0.7.0

package/README.md

@@ -245,6 +245,13 @@ See **[docs/architecture.md](docs/architecture.md)** for the design (adapters, r
 flow, file layout), **[docs/orchestration.md](docs/orchestration.md)** for the agent strategies, and
 **[docs/integrations.md](docs/integrations.md)** for the HTTP API / SDK / ACP.
 
+## Enterprise
+
+`ada-server` doubles as an org **control plane**: per-user seat keys, an org policy (server-enforced
+model allowlist + tool rules pushed to every client), per-user usage metering, and an audit log —
+activated only when you create seats, file-backed, self-hosted in your own network. See
+**[docs/enterprise.md](docs/enterprise.md)** for the 2-minute bootstrap.
+
 ## Benchmarks
 
 ada can run **SWE-bench Verified** — it generates patches for real GitHub issues (one isolated repo

package/docs/enterprise.md

@@ -0,0 +1,83 @@
+# Enterprise: seats, policy, metering, audit
+
+`ada-server` doubles as an org **control plane**: per-user seat keys, an org policy the backend
+enforces (and clients apply locally), per-user usage metering, and an audit log. It's all
+file-backed under `~/.ada/server/` (override with `ADA_DATA_DIR`) — fine to ~50 seats; a database
+is the upgrade path, not the starting point.
+
+**Enterprise mode activates only when a seat exists or `ADA_ADMIN_KEY` is set.** With neither, the
+backend behaves exactly as before (dev-open, or `ADA_CLIENT_KEYS`/login).
+
+## Bootstrap (2 minutes)
+
+```bash
+# 1. start the backend with a bootstrap admin key (any long random string)
+export ADA_ADMIN_KEY=$(openssl rand -hex 24)
+export CLOUDFLARE_ACCOUNT_ID=... CLOUDFLARE_API_TOKEN=...     # your provider keys
+ada-server                     # banner shows: [ENTERPRISE (0 seats + admin key)]
+
+# 2. create a seat per developer — the key is shown ONCE
+curl -s -X POST -H "Authorization: Bearer $ADA_ADMIN_KEY" \
+     -d '{"name":"alice"}' http://localhost:8787/v1/users
+# → { "key": "ada_sk_…", "name": "alice", "role": "dev", "note": "shown once — store it now" }
+```
+
+Each developer sets their seat key and points at the backend:
+
+```bash
+export ADA_BACKEND_URL=https://ada.yourcompany.com/v1
+export ADA_CLIENT_KEY=ada_sk_…
+ada
+```
+
+## Seats
+
+```bash
+GET    /v1/users               # list: name, role, keyPrefix, created, disabled   (admin)
+POST   /v1/users               # {"name":"bob","role":"dev"|"admin"} → full key, once   (admin)
+DELETE /v1/users/<keyPrefix>   # disable a seat (≥12 chars of its key; kept for the audit trail)   (admin)
+```
+
+Full keys are never listed after creation — only a 14-char prefix. `ADA_ADMIN_KEY` is the
+break-glass admin; create an admin *seat* for day-to-day and keep the env key in a vault.
+
+## Org policy
+
+```bash
+curl -X PUT -H "Authorization: Bearer $ADA_ADMIN_KEY" http://localhost:8787/v1/policy -d '{
+  "models": ["@cf/*", "claude-*"],
+  "permissions": [
+    { "tool": "web_*",  "action": "deny" },
+    { "tool": "bash",   "pattern": "*curl*", "action": "ask" }
+  ]
+}'
+```
+
+- **`models`** — allowlist (`*` wildcards). Enforced **server-side** (403 + audit entry), so a
+  modified client can't route around it. Empty/absent = all models.
+- **`permissions`** — tool rules **pushed to clients** (fetched from `GET /v1/policy` at startup —
+  interactive, `-p` headless, `serve`, and `acp` alike). Merged restrictive-wins with local config:
+  an org `deny` beats any local `allow`, an org `ask` upgrades a local `allow`, and an org `allow`
+  can never *loosen* a local deny or the default gating. **Honest caveat:** tool rules run in the
+  *client*, so they govern well-behaved clients — and only **model-allowlist** denials are audited
+  server-side; tool-rule outcomes are not visible to `/v1/audit`. The **hard, server-enforced**
+  guarantees are: authentication, the model allowlist, provider pinning (when an allowlist is set,
+  the client's `provider` hint is ignored so a request can't be re-routed off-policy), and metering.
+
+## Usage & audit
+
+```bash
+GET /v1/usage?days=30    # totals + per-user + per-model {requests, promptTokens, completionTokens}   (admin)
+GET /v1/audit?limit=200  # seat_created / seat_disabled / policy_updated / policy_denied_model …      (admin)
+```
+
+Metering is captured server-side by teeing every chat response (streamed or not) and recording the
+upstream's reported token usage per user — clients can't underreport. Join with
+`ada catalog` prices for cost.
+
+## Deployment notes
+
+- Run behind TLS (caddy/nginx) — seat keys travel as bearer tokens.
+- `ADA_DATA_DIR` on a persistent volume; back it up (it's 4 small JSON/JSONL files).
+- One deployment = one org. Multi-org/SaaS is deliberately out of scope for v1.
+- Compliance paperwork (SOC 2, DPA) is process, not code — start it when a buyer asks.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ada-agent",
-  "version": "0.6.1",
+  "version": "0.7.0",
   "description": "A from-zero terminal coding agent with a Cursor-style routing backend, ~285 skills, MCP connectors, and ask/plan/auto modes",
   "type": "module",
   "license": "MIT",

package/src/client/cli.ts

@@ -3,7 +3,8 @@
 import { createInterface } from "node:readline/promises";
 import { spawnSync } from "node:child_process";
 import { basename, dirname, join, resolve } from "node:path";
-import { readFileSync } from "node:fs";
+import { mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { homedir } from "node:os";
 import { fileURLToPath } from "node:url";
 import { stdin, stdout } from "node:process";
 import OpenAI from "openai";
@@ -13,7 +14,7 @@ import { expandPrompt, loadPrompts } from "./prompts.ts";
 import { Session, list, type SessionMeta } from "./session.ts";
 import { deleteCredential, getCredential, listCredentials } from "../server/credentials.ts";
 import { deviceLogin, oauthConfig } from "../server/oauth.ts";
-import { addTrust, isTrusted, loadSettings, setActiveAgentPermissions, type Settings } from "./settings.ts";
+import { addTrust, isTrusted, loadSettings, setActiveAgentPermissions, setOrgPermissions, type PermRule, type Settings } from "./settings.ts";
 import { getCommands, loadExtensions } from "./extensions.ts";
 import { registerTool, setAsker } from "./tools.ts";
 import { addRemoteSkill, loadSkills, registerSkillTool } from "./skills.ts";
@@ -368,6 +369,38 @@ async function loginFlow(provider: string): Promise<boolean> {
   }
 }
 
+/** Fetch org policy from an enterprise backend and apply its tool rules locally (restrictive-wins
+ *  merge in settings.permissionFor; the backend enforces the model allowlist regardless). Caches the
+ *  last-good policy under ~/.ada so a transient fetch failure falls back to known rules instead of
+ *  silently dropping them. No-op against a non-enterprise backend. */
+async function applyOrgPolicy(): Promise<void> {
+  const cacheFile = join(homedir(), ".ada", "org-policy.json");
+  const enterprise = clientKey().startsWith("ada_sk_"); // a seat key ⇒ this is an enterprise backend
+  try {
+    const r = await fetch(`${BACKEND}/policy`, { headers: { authorization: `Bearer ${clientKey()}` }, signal: AbortSignal.timeout(3000) });
+    if (!r.ok) throw new Error(`HTTP ${r.status}`);
+    const policy = (await r.json()) as { permissions?: PermRule[] };
+    setOrgPermissions(policy.permissions ?? null);
+    try {
+      mkdirSync(join(homedir(), ".ada"), { recursive: true });
+      writeFileSync(cacheFile, JSON.stringify(policy));
+    } catch {
+      /* cache best-effort */
+    }
+    if (policy.permissions?.length) console.error(`\x1b[2m↳ org policy applied (${policy.permissions.length} rule${policy.permissions.length === 1 ? "" : "s"})\x1b[0m`);
+  } catch (e) {
+    if (!enterprise) return; // non-enterprise backend — local rules only, silently
+    // Enterprise backend unreachable — fall back to the last policy we saw, and say so loudly.
+    try {
+      const cached = JSON.parse(readFileSync(cacheFile, "utf8")) as { permissions?: PermRule[] };
+      setOrgPermissions(cached.permissions ?? null);
+      console.error(`\x1b[33m[warn] could not fetch org policy (${e instanceof Error ? e.message : e}) — using cached rules.\x1b[0m`);
+    } catch {
+      console.error(`\x1b[33m[warn] could not fetch org policy (${e instanceof Error ? e.message : e}) and no cache — org tool rules NOT applied this session.\x1b[0m`);
+    }
+  }
+}
+
 /** Startup login check: probe the backend; if it says 401, offer to sign in and rebuild the client. */
 async function ensureAuth(rl: RL, client: OpenAI): Promise<OpenAI> {
   let status: number;
@@ -658,6 +691,7 @@ async function main(): Promise<void> {
     registerSkillTool(loadSkills(trusted));
     await loadMcpServers(trusted);
     const client = makeClient();
+    await applyOrgPolicy(); // enterprise org rules apply to acp sessions too
     let model = process.env.ADA_MODEL || settings.model || "";
     if (!model) {
       try {
@@ -741,6 +775,7 @@ async function main(): Promise<void> {
     registerSkillTool(loadSkills(trusted));
     await loadMcpServers(trusted);
     const client = makeClient();
+    await applyOrgPolicy(); // enterprise org rules apply to serve sessions too
     let model = (process.argv[3] && !process.argv[3].startsWith("--") ? process.argv[3] : "") || process.env.ADA_MODEL || settings.model || "";
     if (!model) {
       try {
@@ -1082,6 +1117,7 @@ async function main(): Promise<void> {
   if (flags.print !== undefined) {
     const trusted = isTrusted(process.cwd());
     const settings = loadSettings(trusted);
+    await applyOrgPolicy(); // org tool rules bind headless runs too (CI is the classic bypass path)
     let pm = flags.model ?? process.env.ADA_MODEL ?? settings.model ?? scoped[0] ?? "";
     if (!pm) {
       try {
@@ -1125,6 +1161,7 @@ async function main(): Promise<void> {
   const mcp = await loadMcpServers(includeProject);
 
   client = await ensureAuth(rl, client); // always check login at startup; prompt if the backend says 401
+  await applyOrgPolicy(); // enterprise backends push org tool rules; no-op otherwise
 
   let session: Session;
   let history: Msg[] = [];

package/src/client/settings.ts

@@ -66,9 +66,15 @@ export function setActiveAgentPermissions(rules: PermRule[] | null): void {
   activeAgentPerms = rules;
 }
 
-/** Evaluate the configured permission rules for a tool call. null = no matching rule (use defaults). */
-export function permissionFor(toolName: string, summary: string): PermAction | null {
-  const rules = activeAgentPerms ?? loadSettings(isTrusted(process.cwd())).permissions ?? [];
+// Org policy pushed by an enterprise backend (fetched from /v1/policy at startup). Merged
+// restrictive-wins: an org "deny" beats any local "allow"; an org "ask" upgrades a local "allow".
+// A local "deny" always stands — the org can tighten a user's setup, never loosen it.
+let orgPerms: PermRule[] | null = null;
+export function setOrgPermissions(rules: PermRule[] | null): void {
+  orgPerms = rules?.length ? rules : null;
+}
+
+function evalRules(rules: PermRule[], toolName: string, summary: string): PermAction | null {
   let result: PermAction | null = null;
   for (const r of rules) {
     const toolOk = !r.tool || r.tool === toolName || globMatch(r.tool, toolName);
@@ -78,6 +84,18 @@ export function permissionFor(toolName: string, summary: string): PermAction | n
   return result;
 }
 
+const STRICTNESS: Record<PermAction, number> = { allow: 0, ask: 1, deny: 2 };
+
+/** Evaluate the configured permission rules for a tool call. null = no matching rule (use defaults). */
+export function permissionFor(toolName: string, summary: string): PermAction | null {
+  const local = evalRules(activeAgentPerms ?? loadSettings(isTrusted(process.cwd())).permissions ?? [], toolName, summary);
+  if (!orgPerms) return local;
+  const org = evalRules(orgPerms, toolName, summary);
+  if (org === null) return local;
+  if (local === null) return org === "allow" ? null : org; // org can't LOOSEN the default gating, only tighten
+  return STRICTNESS[org] > STRICTNESS[local] ? org : local;
+}
+
 export function addTrust(dir: string): void {
   const g = readJson(GLOBAL);
   const dirs = new Set(g.trustedDirs ?? []);

package/src/selfcheck.ts

@@ -293,6 +293,79 @@ async function main(): Promise<void> {
     assert.equal(route("anything-else"), "openrouter", "unmatched → openrouter");
   }
 
+  // --- enterprise control plane: seats, policy, metering, audit (temp data dir, no HTTP) ---
+  {
+    const dir = join(tmpdir(), `ada-ent-${Date.now()}`);
+    const ent = await import("./server/enterprise.ts");
+    process.env.ADA_DATA_DIR = dir;
+    try {
+      assert.equal(ent.enterpriseMode(dir), false, "no seats + no admin key → enterprise mode off");
+      const key = ent.createSeat("alice", "admin", dir);
+      assert.ok(key.startsWith("ada_sk_") && key.length > 40, "seat keys are long and prefixed");
+      assert.equal(ent.enterpriseMode(dir), true, "a seat activates enterprise mode");
+      assert.deepEqual(ent.identifySeat(key, dir), { user: "alice", role: "admin" }, "seat key resolves to its identity");
+      assert.equal(ent.identifySeat("ada_sk_wrong", dir), null, "unknown key → null");
+      // The auth-bypass the review caught: Object.prototype keys must NOT authenticate.
+      for (const evil of ["toString", "constructor", "__proto__", "valueOf", "hasOwnProperty"]) {
+        assert.equal(ent.identifySeat(evil, dir), null, `prototype key "${evil}" must not authenticate`);
+      }
+      assert.equal(ent.listSeats(dir)[0]!.keyPrefix.length, 14, "listing exposes only a key prefix");
+      assert.equal(ent.disableSeat(key.slice(0, 8), dir), null, "too-short prefix refused");
+      assert.equal(ent.disableSeat(key.slice(0, 14), dir), "alice", "disable by unique prefix");
+      assert.equal(ent.identifySeat(key, dir), null, "disabled seat no longer authenticates");
+
+      assert.ok(ent.modelAllowed("claude-opus-4-8", {}), "empty policy allows everything");
+      const pol = { models: ["@cf/*", "claude-*"] };
+      assert.ok(ent.modelAllowed("@cf/moonshotai/kimi-k2.7-code", pol), "wildcard allowlist matches");
+      assert.ok(!ent.modelAllowed("gpt-5", pol), "non-listed model denied");
+
+      ent.appendUsage({ ts: Date.now(), user: "alice", model: "m1", provider: "p", promptTokens: 100, completionTokens: 20 }, dir);
+      ent.appendUsage({ ts: Date.now(), user: "alice", model: "m1", provider: "p", promptTokens: 50, completionTokens: 10 }, dir);
+      ent.appendUsage({ ts: Date.now() - 90 * 86_400_000, user: "old", model: "m1", provider: "p", promptTokens: 999, completionTokens: 999 }, dir);
+      const sum = ent.usageSummary(30, dir);
+      assert.equal(sum.byUser.alice!.requests, 2, "usage aggregates per user");
+      assert.equal(sum.totals.promptTokens, 150, "old rows fall outside the window");
+
+      assert.ok(ent.auditTail(10, dir).some((e) => e.event === "seat_created"), "audit log records seat creation");
+
+      const sse = 'data: {"choices":[]}\n\ndata: {"choices":[],"usage":{"prompt_tokens":11,"completion_tokens":7,"completion_tokens_details":{"reasoning_tokens":2}}}\n\ndata: [DONE]\n\n';
+      assert.deepEqual(ent.extractLastUsage(sse), { promptTokens: 11, completionTokens: 7 }, "usage extracted from SSE tail (nested details ok)");
+      assert.equal(ent.extractLastUsage("no usage here"), null, "no usage → null");
+      // A trailing "usage": null must not hide the real one earlier in the stream.
+      assert.deepEqual(ent.extractLastUsage('{"usage":{"prompt_tokens":5,"completion_tokens":3}}\n{"usage":null}'), { promptTokens: 5, completionTokens: 3 }, "trailing usage:null skipped, real one found");
+
+      // policy validation rejects malformed shapes, accepts good ones
+      assert.ok("error" in ent.validatePolicy({ models: [1, 2] }), "non-string models rejected");
+      assert.ok("error" in ent.validatePolicy({ permissions: [{ tool: "x" }] }), "permission without action rejected");
+      assert.ok("policy" in ent.validatePolicy({ models: ["@cf/*"], permissions: [{ tool: "bash", action: "deny" }] }), "valid policy accepted");
+
+      // corrupt users.json → CorruptStore (fail-closed), NOT an empty map that unlocks the backend
+      writeFileSync(join(dir, "users.json"), "{ this is not json");
+      assert.throws(() => ent.loadSeats(dir), (e: unknown) => e instanceof ent.CorruptStore, "corrupt users.json throws CorruptStore");
+      assert.equal(ent.enterpriseMode(dir), true, "corrupt store → still enterprise (locked), never open");
+    } finally {
+      delete process.env.ADA_DATA_DIR;
+      rmSync(dir, { recursive: true, force: true });
+    }
+  }
+
+  // --- org policy merge: restrictive wins, org can tighten but never loosen ---
+  {
+    const { permissionFor, setActiveAgentPermissions, setOrgPermissions } = await import("./client/settings.ts");
+    setActiveAgentPermissions([{ tool: "bash", action: "allow" }]);
+    setOrgPermissions([{ tool: "bash", action: "deny" }]);
+    assert.equal(permissionFor("bash", "x"), "deny", "org deny beats local allow");
+    setOrgPermissions([{ tool: "bash", action: "ask" }]);
+    assert.equal(permissionFor("bash", "x"), "ask", "org ask upgrades local allow");
+    setActiveAgentPermissions([{ tool: "bash", action: "deny" }]);
+    setOrgPermissions([{ tool: "bash", action: "allow" }]);
+    assert.equal(permissionFor("bash", "x"), "deny", "org allow cannot loosen a local deny");
+    setActiveAgentPermissions([]);
+    assert.equal(permissionFor("bash", "x"), null, "org allow cannot loosen the default gating");
+    setOrgPermissions(null);
+    setActiveAgentPermissions(null);
+  }
+
   // --- @codebase semantic search: pure parts (no network / no embedding model needed) ---
   {
     const { chunkText, cosine, walkFiles } = await import("./client/embed-index.ts");

package/src/server/enterprise.ts

@@ -0,0 +1,346 @@
+// Enterprise control plane: seats (per-user client keys), org policy, usage metering, audit log.
+// One deployment = one org (it's self-hosted; multi-org is the SaaS upgrade path, not v1).
+//
+// Enterprise mode ACTIVATES when a seat exists or ADA_ADMIN_KEY is set — with neither, the backend
+// behaves exactly as before (dev-open / ADA_CLIENT_KEYS / login). Bootstrap:
+//
+//   ADA_ADMIN_KEY=<random> ada-server
+//   curl -X POST -H "Authorization: Bearer $ADA_ADMIN_KEY" localhost:8787/v1/users -d '{"name":"alice"}'
+//
+// Security posture (hardened after an adversarial review):
+//   - lookups are own-property + format-guarded (no prototype-key auth bypass);
+//   - writes are atomic (tmp + rename); a corrupt/unreadable store fails CLOSED (never dev-open);
+//   - key comparisons are timing-safe.
+// ponytail: file-backed under ~/.ada/server — fine to ~50 seats. Postgres + rotating usage logs are
+// the upgrade path when an org outgrows files (usageSummary/auditTail read whole files: OK to
+// low-millions of rows, then rotate).
+
+import { randomBytes, timingSafeEqual } from "node:crypto";
+import { appendFileSync, existsSync, mkdirSync, readFileSync, renameSync, writeFileSync } from "node:fs";
+import { homedir } from "node:os";
+import { join, resolve } from "node:path";
+
+// Resolved once: an empty ADA_DATA_DIR means "unset", and a relative path can't scatter the auth
+// store across working directories (which would itself be a fail-open).
+const DATA_DIR = resolve(process.env.ADA_DATA_DIR || join(homedir(), ".ada", "server"));
+function dataDir(): string {
+  return DATA_DIR;
+}
+
+/** Thrown when a store file exists but can't be read/parsed — callers must fail CLOSED, never open. */
+export class CorruptStore extends Error {}
+
+export interface Seat {
+  name: string;
+  role: "admin" | "dev";
+  created: string;
+  disabled?: boolean;
+}
+export interface PolicyRule {
+  tool?: string;
+  pattern?: string;
+  action: "allow" | "ask" | "deny";
+}
+export interface Policy {
+  models?: string[];
+  permissions?: PolicyRule[];
+}
+export interface UsageRow {
+  ts: number;
+  user: string;
+  model: string;
+  provider: string;
+  promptTokens: number;
+  completionTokens: number;
+}
+export interface Identity {
+  user: string;
+  role: "admin" | "dev";
+}
+
+const usersFile = (dir: string): string => join(dir, "users.json");
+const policyFile = (dir: string): string => join(dir, "policy.json");
+const usageFile = (dir: string): string => join(dir, "usage.jsonl");
+const auditFile = (dir: string): string => join(dir, "audit.jsonl");
+
+function atomicWrite(file: string, data: string): void {
+  mkdirSync(join(file, ".."), { recursive: true });
+  const tmp = `${file}.tmp.${process.pid}`;
+  writeFileSync(tmp, data);
+  renameSync(tmp, file); // atomic on the same filesystem — a crash can't leave a torn file
+}
+
+function errno(e: unknown): string | undefined {
+  return (e as NodeJS.ErrnoException).code;
+}
+
+/** Seats keyed by full key. Missing file → empty (prototype-free) map. Any OTHER read/parse error
+ *  → CorruptStore (callers fail closed — a torn users.json must not silently disable auth). */
+export function loadSeats(dir = dataDir()): Record<string, Seat> {
+  const map: Record<string, Seat> = Object.create(null); // no Object.prototype — belt-and-suspenders with the own-property check
+  let text: string;
+  try {
+    text = readFileSync(usersFile(dir), "utf8");
+  } catch (e) {
+    if (errno(e) === "ENOENT") return map;
+    throw new CorruptStore(`users.json unreadable: ${e instanceof Error ? e.message : e}`);
+  }
+  try {
+    const parsed = (JSON.parse(text) as { users?: Record<string, Seat> }).users ?? {};
+    for (const [k, v] of Object.entries(parsed)) map[k] = v;
+    return map;
+  } catch (e) {
+    throw new CorruptStore(`users.json corrupt: ${e instanceof Error ? e.message : e}`);
+  }
+}
+
+function saveSeats(seats: Record<string, Seat>, dir = dataDir()): void {
+  atomicWrite(usersFile(dir), JSON.stringify({ users: seats }, null, 2));
+}
+
+/** Enterprise mode = admin key set, or seats exist. A corrupt seat store counts as enterprise
+ *  (locked), never as "no seats" — fail closed. */
+export function enterpriseMode(dir = dataDir()): boolean {
+  if (process.env.ADA_ADMIN_KEY) return true;
+  try {
+    return Object.keys(loadSeats(dir)).length > 0;
+  } catch {
+    return true;
+  }
+}
+
+function timingEqual(a: string, b: string): boolean {
+  const ab = Buffer.from(a);
+  const bb = Buffer.from(b);
+  return ab.length === bb.length && timingSafeEqual(ab, bb);
+}
+
+/** Resolve a bearer token to a seat identity (or the bootstrap admin). Null = not a seat. Throws
+ *  CorruptStore if the seat store can't be read (caller returns 503, never dev-open). */
+export function identifySeat(token: string, dir = dataDir()): Identity | null {
+  const admin = process.env.ADA_ADMIN_KEY;
+  if (admin && timingEqual(token, admin)) return { user: "admin", role: "admin" };
+  if (!token.startsWith("ada_sk_")) return null; // format guard — "toString"/"__proto__"/… never reach the map
+  const seats = loadSeats(dir); // may throw CorruptStore
+  if (!Object.prototype.hasOwnProperty.call(seats, token)) return null; // own-property only
+  const seat = seats[token]!;
+  return seat.disabled ? null : { user: seat.name, role: seat.role };
+}
+
+/** Create a seat; returns its full key (shown once — only a prefix is ever listed again). */
+export function createSeat(name: string, role: "admin" | "dev" = "dev", dir = dataDir()): string {
+  const key = `ada_sk_${randomBytes(24).toString("hex")}`;
+  const seats = loadSeats(dir);
+  seats[key] = { name, role, created: new Date().toISOString() };
+  saveSeats(seats, dir);
+  appendAudit({ ts: Date.now(), user: "-", event: "seat_created", detail: `${name} (${role})` }, dir);
+  return key;
+}
+
+/** Disable (not delete — the audit trail keeps the history) the seat whose key starts with prefix. */
+export function disableSeat(prefix: string, dir = dataDir()): string | null {
+  if (prefix.length < 12) return null; // too short to be safely unique
+  const seats = loadSeats(dir);
+  const keys = Object.keys(seats).filter((k) => k.startsWith(prefix));
+  if (keys.length !== 1) return null;
+  seats[keys[0]!]!.disabled = true;
+  saveSeats(seats, dir);
+  appendAudit({ ts: Date.now(), user: "-", event: "seat_disabled", detail: seats[keys[0]!]!.name }, dir);
+  return seats[keys[0]!]!.name;
+}
+
+/** Key prefixes + metadata for listing — full keys are never returned after creation. Display-only,
+ *  so a corrupt store yields [] rather than crashing the banner. */
+export function listSeats(dir = dataDir()): Array<Seat & { keyPrefix: string }> {
+  try {
+    return Object.entries(loadSeats(dir)).map(([k, s]) => ({ ...s, keyPrefix: k.slice(0, 14) }));
+  } catch {
+    return [];
+  }
+}
+
+let lastGoodPolicy: Policy | null = null;
+/** Missing file → {} (no policy = allow all, legitimate). A corrupt EXISTING file → last-known-good
+ *  if we have one, else CorruptStore (fail closed — a security control must not degrade to allow-all). */
+export function loadPolicy(dir = dataDir()): Policy {
+  let text: string;
+  try {
+    text = readFileSync(policyFile(dir), "utf8");
+  } catch (e) {
+    if (errno(e) === "ENOENT") return {};
+    if (lastGoodPolicy) return lastGoodPolicy;
+    throw new CorruptStore(`policy.json unreadable: ${e instanceof Error ? e.message : e}`);
+  }
+  try {
+    lastGoodPolicy = JSON.parse(text) as Policy;
+    return lastGoodPolicy;
+  } catch (e) {
+    if (lastGoodPolicy) return lastGoodPolicy;
+    throw new CorruptStore(`policy.json corrupt: ${e instanceof Error ? e.message : e}`);
+  }
+}
+
+export function savePolicy(p: Policy, dir = dataDir()): void {
+  atomicWrite(policyFile(dir), JSON.stringify(p, null, 2));
+  lastGoodPolicy = p;
+  appendAudit({ ts: Date.now(), user: "-", event: "policy_updated", detail: JSON.stringify(p).slice(0, 300) }, dir);
+}
+
+/** Validate a policy shape from the wire. Returns the typed policy or an error message. */
+export function validatePolicy(raw: unknown): { policy: Policy } | { error: string } {
+  if (!raw || typeof raw !== "object" || Array.isArray(raw)) return { error: "policy must be a JSON object" };
+  const r = raw as Record<string, unknown>;
+  const out: Policy = {};
+  if (r.models !== undefined) {
+    if (!Array.isArray(r.models) || r.models.some((m) => typeof m !== "string" || !m.trim())) return { error: "models must be an array of non-empty strings" };
+    out.models = r.models as string[];
+  }
+  if (r.permissions !== undefined) {
+    if (!Array.isArray(r.permissions)) return { error: "permissions must be an array" };
+    for (const p of r.permissions) {
+      const rule = p as Record<string, unknown>;
+      if (!rule || typeof rule !== "object" || !["allow", "ask", "deny"].includes(rule.action as string)) return { error: "each permission needs action: allow|ask|deny" };
+      if (rule.tool !== undefined && typeof rule.tool !== "string") return { error: "permission.tool must be a string" };
+      if (rule.pattern !== undefined && typeof rule.pattern !== "string") return { error: "permission.pattern must be a string" };
+    }
+    out.permissions = r.permissions as PolicyRule[];
+  }
+  return { policy: out };
+}
+
+function globMatch(pattern: string, s: string): boolean {
+  const re = new RegExp(`^${pattern.split("*").map((p) => p.replace(/[.*+?^${}()|[\]\\]/g, "\\$&")).join(".*")}$`, "i");
+  return re.test(s);
+}
+
+/** Is this model allowed by org policy? No/empty allowlist = everything allowed. */
+export function modelAllowed(model: string, policy: Policy): boolean {
+  if (!Array.isArray(policy.models) || !policy.models.length) return true;
+  return policy.models.some((p) => globMatch(p, model));
+}
+
+export function appendUsage(row: UsageRow, dir = dataDir()): void {
+  try {
+    mkdirSync(dir, { recursive: true });
+    appendFileSync(usageFile(dir), `${JSON.stringify(row)}\n`);
+  } catch {
+    /* metering is best-effort; never fail a request over it */
+  }
+}
+
+export interface AuditRow {
+  ts: number;
+  user: string;
+  event: string;
+  detail: string;
+}
+
+export function appendAudit(row: AuditRow, dir = dataDir()): void {
+  try {
+    mkdirSync(dir, { recursive: true });
+    appendFileSync(auditFile(dir), `${JSON.stringify(row)}\n`);
+  } catch {
+    /* best-effort */
+  }
+}
+
+export function auditTail(limit = 200, dir = dataDir()): AuditRow[] {
+  let lines: string[];
+  try {
+    lines = readFileSync(auditFile(dir), "utf8").split("\n").filter(Boolean);
+  } catch {
+    return [];
+  }
+  const out: AuditRow[] = [];
+  for (const l of lines.slice(-limit)) {
+    try {
+      out.push(JSON.parse(l) as AuditRow); // skip a torn last line instead of losing the whole view
+    } catch {
+      /* skip corrupt line */
+    }
+  }
+  return out;
+}
+
+interface Bucket {
+  requests: number;
+  promptTokens: number;
+  completionTokens: number;
+}
+export interface UsageSummary {
+  since: number;
+  totals: Bucket;
+  byUser: Record<string, Bucket>;
+  byModel: Record<string, Bucket>;
+}
+
+export function usageSummary(days = 30, dir = dataDir()): UsageSummary {
+  const since = Date.now() - days * 86_400_000;
+  const zero = (): Bucket => ({ requests: 0, promptTokens: 0, completionTokens: 0 });
+  const out: UsageSummary = { since, totals: zero(), byUser: Object.create(null), byModel: Object.create(null) };
+  let lines: string[] = [];
+  try {
+    lines = readFileSync(usageFile(dir), "utf8").split("\n").filter(Boolean);
+  } catch {
+    return out;
+  }
+  for (const l of lines) {
+    let r: UsageRow;
+    try {
+      r = JSON.parse(l) as UsageRow;
+    } catch {
+      continue;
+    }
+    if (r.ts < since) continue;
+    for (const b of [out.totals, (out.byUser[r.user] ??= zero()), (out.byModel[r.model] ??= zero())]) {
+      b.requests++;
+      b.promptTokens += r.promptTokens || 0;
+      b.completionTokens += r.completionTokens || 0;
+    }
+  }
+  return out;
+}
+
+function matchBraces(text: string, start: number): string | null {
+  let depth = 0;
+  let inStr = false;
+  let esc = false;
+  for (let i = start; i < text.length; i++) {
+    const c = text[i];
+    if (inStr) {
+      if (esc) esc = false;
+      else if (c === "\\") esc = true;
+      else if (c === '"') inStr = false;
+    } else if (c === '"') inStr = true;
+    else if (c === "{") depth++;
+    else if (c === "}" && --depth === 0) return text.slice(start, i + 1);
+  }
+  return null;
+}
+
+/** Pull the LAST real `"usage": { … }` object out of streamed/response text. Skips a trailing
+ *  `"usage": null` and keeps scanning backwards, so a null in a late frame doesn't hide a real one. */
+export function extractLastUsage(text: string): { promptTokens: number; completionTokens: number } | null {
+  let at = text.lastIndexOf('"usage"');
+  while (at >= 0) {
+    const brace = text.indexOf("{", at + 7);
+    const colon = text.indexOf(":", at + 7);
+    if (brace >= 0 && colon >= 0 && text.slice(colon + 1, brace).trim() === "") {
+      const obj = matchBraces(text, brace);
+      if (obj) {
+        try {
+          const u = JSON.parse(obj) as { prompt_tokens?: number; completion_tokens?: number };
+          if (u.prompt_tokens != null || u.completion_tokens != null) return { promptTokens: u.prompt_tokens ?? 0, completionTokens: u.completion_tokens ?? 0 };
+        } catch {
+          /* malformed — keep looking backwards */
+        }
+      }
+    }
+    at = text.lastIndexOf('"usage"', at - 1);
+  }
+  return null;
+}
+
+export function storeExists(dir = dataDir()): boolean {
+  return existsSync(usersFile(dir)) || existsSync(policyFile(dir));
+}

package/src/server/index.ts

@@ -5,6 +5,7 @@
 import { createServer } from "node:http";
 import type { IncomingMessage, ServerResponse } from "node:http";
 import { PORT, PROVIDERS, clientKeys, configuredProviders, isConfigured } from "./config.ts";
+import { CorruptStore, type Identity, appendAudit, appendUsage, auditTail, createSeat, disableSeat, enterpriseMode, extractLastUsage, identifySeat, listSeats, loadPolicy, modelAllowed, savePolicy, usageSummary, validatePolicy } from "./enterprise.ts";
 import { allowedUsers, isAllowed, verifyIdentity } from "./identity.ts";
 import { adapterFor } from "./providers/registry.ts";
 import { route } from "./router.ts";
@@ -19,20 +20,32 @@ function readBody(req: IncomingMessage): Promise<string> {
 }
 
 function locked(): boolean {
-  return clientKeys() !== null || allowedUsers() !== null || !!process.env.ADA_REQUIRE_LOGIN;
+  return enterpriseMode() || clientKeys() !== null || allowedUsers() !== null || !!process.env.ADA_REQUIRE_LOGIN;
 }
 
-/** A request is allowed if it carries a known static client key, OR a valid GitHub/Google
- *  login token (allowlisted). With nothing configured, the backend is open (dev mode). */
-async function authorized(req: IncomingMessage): Promise<boolean> {
-  if (!locked()) return true; // dev mode: no auth configured
+/** Resolve a request to WHO is making it. Order: seat key / ADA_ADMIN_KEY (enterprise), legacy
+ *  static client key, GitHub/Google login. With no auth configured, the backend is open (dev mode).
+ *  Returns "corrupt" if the seat store can't be read — the caller MUST 503, never fall through to
+ *  dev-open. Null = unauthorized. */
+async function identify(req: IncomingMessage): Promise<Identity | "corrupt" | null> {
   const h = req.headers["authorization"];
   const token = typeof h === "string" && h.startsWith("Bearer ") ? h.slice(7) : "";
-  if (!token) return false;
-  const keys = clientKeys();
-  if (keys?.includes(token)) return true; // static client key
-  const id = await verifyIdentity(token); // GitHub / Google login
-  return !!id && isAllowed(id.user);
+  if (token) {
+    let seat: Identity | null;
+    try {
+      seat = identifySeat(token);
+    } catch (e) {
+      if (e instanceof CorruptStore) return "corrupt";
+      throw e;
+    }
+    if (seat) return seat;
+    // Legacy ADA_CLIENT_KEYS are NOT honored once seats/admin-key exist — enterprise supersedes them,
+    // so a disabled seat can't be resurrected via a still-configured shared key.
+    if (!enterpriseMode() && clientKeys()?.includes(token)) return { user: "team", role: "dev" };
+    const id = await verifyIdentity(token); // GitHub / Google login
+    if (id && isAllowed(id.user)) return { user: id.user, role: "dev" };
+  }
+  return locked() ? null : { user: "dev", role: "dev" }; // dev mode: open
 }
 
 function json(res: ServerResponse, status: number, obj: unknown): void {
@@ -49,7 +62,7 @@ async function handleModels(res: ServerResponse): Promise<void> {
   json(res, 200, { object: "list", data });
 }
 
-async function handleChat(req: IncomingMessage, res: ServerResponse): Promise<void> {
+async function handleChat(req: IncomingMessage, res: ServerResponse, who: Identity): Promise<void> {
   const raw = await readBody(req);
   let body: Record<string, unknown>;
   try {
@@ -61,32 +74,90 @@ async function handleChat(req: IncomingMessage, res: ServerResponse): Promise<vo
   const model = String(body.model ?? "");
   if (!model) return json(res, 400, { error: { message: "missing 'model'" } });
 
-  const provider = route(model, typeof body.provider === "string" ? body.provider : undefined);
+  // Org policy: model allowlist (enterprise). Enforced server-side so a modified client can't skip it.
+  let policy: import("./enterprise.ts").Policy;
+  try {
+    policy = loadPolicy();
+  } catch (e) {
+    if (e instanceof CorruptStore) return json(res, 503, { error: { message: "org policy unreadable — refusing requests (fail-closed)" } });
+    throw e;
+  }
+  if (!modelAllowed(model, policy)) {
+    appendAudit({ ts: Date.now(), user: who.user, event: "policy_denied_model", detail: model });
+    return json(res, 403, { error: { message: `model '${model}' is not allowed by org policy (allowed: ${policy.models!.join(", ")})` } });
+  }
+
+  // When an allowlist is active, IGNORE the client's `provider` hint — else a seat holder could
+  // send an allowlisted model id with a different provider and leak the body to it before the
+  // upstream rejects the id. Route by the model id only.
+  const explicit = policy.models?.length ? undefined : typeof body.provider === "string" ? body.provider : undefined;
+  const provider = route(model, explicit);
   if (!isConfigured(provider)) {
     return json(res, 400, {
       error: { message: `provider '${provider}' not configured — set ${PROVIDERS[provider].keyEnv} on the backend` },
     });
   }
 
+  // Metering must not be client-suppressible: force the upstream to emit a usage object on streams
+  // (OpenAI-compat only sends it when include_usage is set). Harmless for providers that ignore it.
+  if (body.stream) body.stream_options = { ...((body.stream_options as Record<string, unknown>) ?? {}), include_usage: true };
+
+  // Usage metering: tee the response (streamed or not) and record the last usage object the
+  // upstream reported. Wrapping res keeps this in ONE place for every adapter.
+  {
+    let tail = "";
+    const scan = (c: unknown): void => {
+      if (typeof c === "string" || Buffer.isBuffer(c)) tail = (tail + c.toString()).slice(-16_384);
+    };
+    const write = res.write.bind(res);
+    const end = res.end.bind(res);
+    res.write = ((c: never, ...a: never[]) => {
+      scan(c);
+      return write(c, ...a);
+    }) as typeof res.write;
+    res.end = ((c?: never, ...a: never[]) => {
+      scan(c);
+      const u = extractLastUsage(tail);
+      if (u) appendUsage({ ts: Date.now(), user: who.user, model, provider, promptTokens: u.promptTokens, completionTokens: u.completionTokens });
+      return end(c, ...a);
+    }) as typeof res.end;
+  }
+
   delete body.provider; // our routing hint; never forward it upstream
   await adapterFor(provider).chat({ provider, model, body, res });
 }
 
 /** Embeddings for @codebase semantic search — forwarded to the ollama provider's
- *  OpenAI-compatible endpoint (embedding models only live there for now). */
-async function handleEmbeddings(req: IncomingMessage, res: ServerResponse): Promise<void> {
+ *  OpenAI-compatible endpoint (embedding models only live there for now). Subject to the same org
+ *  model allowlist as chat, and metered/attributed. */
+async function handleEmbeddings(req: IncomingMessage, res: ServerResponse, who: Identity): Promise<void> {
   const raw = await readBody(req);
+  let body: Record<string, unknown>;
   try {
-    JSON.parse(raw);
+    body = JSON.parse(raw);
   } catch {
     return json(res, 400, { error: { message: "invalid JSON body" } });
   }
+  const model = String(body.model ?? "");
+  let policy: import("./enterprise.ts").Policy;
+  try {
+    policy = loadPolicy();
+  } catch (e) {
+    if (e instanceof CorruptStore) return json(res, 503, { error: { message: "org policy unreadable — refusing requests" } });
+    throw e;
+  }
+  if (model && !modelAllowed(model, policy)) {
+    appendAudit({ ts: Date.now(), user: who.user, event: "policy_denied_model", detail: `embeddings:${model}` });
+    return json(res, 403, { error: { message: `embedding model '${model}' is not allowed by org policy` } });
+  }
   const upstream = await fetch(`${PROVIDERS.ollama.baseURL}/embeddings`, {
     method: "POST",
     headers: { "content-type": "application/json" },
     body: raw,
   });
   const text = await upstream.text();
+  const u = extractLastUsage(text); // embedding responses report prompt_tokens
+  if (u) appendUsage({ ts: Date.now(), user: who.user, model, provider: "ollama", promptTokens: u.promptTokens, completionTokens: 0 });
   res.writeHead(upstream.status, { "content-type": "application/json" });
   res.end(text);
 }
@@ -98,22 +169,85 @@ const server = createServer(async (req, res) => {
       res.writeHead(200, { "content-type": "text/plain" });
       return res.end("ada backend ok");
     }
+    const who = await identify(req);
+    if (who === "corrupt") return json(res, 503, { error: { message: "auth store unreadable — refusing all requests (fail-closed). Fix ~/.ada/server/users.json." } });
+    if (!who) return json(res, 401, { error: { message: "unauthorized — invalid client key, seat key, or login" } });
+
     if (req.method === "GET" && url.pathname === "/v1/whoami") {
-      if (!(await authorized(req))) return json(res, 401, { error: { message: "not logged in" } });
-      return json(res, 200, { ok: true });
+      return json(res, 200, { ok: true, user: who.user, role: who.role });
     }
     if (req.method === "GET" && url.pathname === "/v1/models") {
-      if (!(await authorized(req))) return json(res, 401, { error: { message: "unauthorized — invalid client key or login" } });
       return await handleModels(res);
     }
     if (req.method === "POST" && url.pathname === "/v1/chat/completions") {
-      if (!(await authorized(req))) return json(res, 401, { error: { message: "unauthorized — invalid client key or login" } });
-      return await handleChat(req, res);
+      return await handleChat(req, res, who);
     }
     if (req.method === "POST" && url.pathname === "/v1/embeddings") {
-      if (!(await authorized(req))) return json(res, 401, { error: { message: "unauthorized — invalid client key or login" } });
-      return await handleEmbeddings(req, res);
+      return await handleEmbeddings(req, res, who);
     }
+
+    // ---- enterprise control plane ----
+    if (url.pathname === "/v1/policy") {
+      if (req.method === "GET") {
+        // any seat — clients fetch this and apply the tool rules locally
+        let policy: unknown;
+        try {
+          policy = loadPolicy();
+        } catch (e) {
+          if (e instanceof CorruptStore) return json(res, 503, { error: { message: "org policy unreadable" } });
+          throw e;
+        }
+        appendAudit({ ts: Date.now(), user: who.user, event: "policy_fetched", detail: "" }); // spot seats that never fetch
+        return json(res, 200, policy);
+      }
+      if (req.method === "PUT") {
+        if (who.role !== "admin") return json(res, 403, { error: { message: "admin only" } });
+        let parsed: unknown;
+        try {
+          parsed = JSON.parse(await readBody(req));
+        } catch {
+          return json(res, 400, { error: { message: "invalid JSON body" } });
+        }
+        const v = validatePolicy(parsed);
+        if ("error" in v) return json(res, 400, { error: { message: v.error } });
+        savePolicy(v.policy);
+        return json(res, 200, { ok: true });
+      }
+    }
+    if (url.pathname === "/v1/users") {
+      if (who.role !== "admin") return json(res, 403, { error: { message: "admin only" } });
+      if (req.method === "GET") return json(res, 200, { users: listSeats() });
+      if (req.method === "POST") {
+        let name = "";
+        let role: "admin" | "dev" = "dev";
+        try {
+          const b = JSON.parse(await readBody(req)) as { name?: string; role?: string };
+          name = String(b.name ?? "").trim();
+          if (b.role === "admin") role = "admin";
+        } catch {
+          /* falls through to the name check */
+        }
+        if (!name) return json(res, 400, { error: { message: "missing 'name'" } });
+        return json(res, 200, { key: createSeat(name, role), name, role, note: "shown once — store it now" });
+      }
+    }
+    {
+      const m = req.method === "DELETE" && url.pathname.match(/^\/v1\/users\/([\w]+)$/);
+      if (m) {
+        if (who.role !== "admin") return json(res, 403, { error: { message: "admin only" } });
+        const name = disableSeat(m[1]!);
+        return json(res, name ? 200 : 404, name ? { ok: true, disabled: name } : { error: { message: "unknown or ambiguous key prefix (send ≥12 chars)" } });
+      }
+    }
+    if (req.method === "GET" && url.pathname === "/v1/usage") {
+      if (who.role !== "admin") return json(res, 403, { error: { message: "admin only" } });
+      return json(res, 200, usageSummary(Math.min(Number(url.searchParams.get("days")) || 30, 365)));
+    }
+    if (req.method === "GET" && url.pathname === "/v1/audit") {
+      if (who.role !== "admin") return json(res, 403, { error: { message: "admin only" } });
+      return json(res, 200, { events: auditTail(Math.min(Number(url.searchParams.get("limit")) || 200, 2000)) });
+    }
+
     return json(res, 404, { error: { message: "not found" } });
   } catch (err) {
     if (!res.headersSent) json(res, 500, { error: { message: err instanceof Error ? err.message : String(err) } });
@@ -127,9 +261,13 @@ const server = createServer(async (req, res) => {
 });
 
 server.listen(PORT, () => {
-  const auth = locked()
-    ? `auth ON (client keys + GitHub/Google login${allowedUsers() ? `, allowlist: ${allowedUsers()!.length}` : ""})`
-    : "AUTH DISABLED (dev) — set ADA_CLIENT_KEYS or ADA_ALLOWED_USERS to lock down";
+  if (enterpriseMode() && clientKeys()) console.warn("\x1b[33m[warn] ADA_CLIENT_KEYS is set but ignored in enterprise mode (seats supersede it) — unset it to avoid confusion.\x1b[0m");
+  const seats = listSeats().filter((s) => !s.disabled).length;
+  const auth = enterpriseMode()
+    ? `ENTERPRISE (${seats} seat${seats === 1 ? "" : "s"}${process.env.ADA_ADMIN_KEY ? " + admin key" : ""})`
+    : locked()
+      ? `auth ON (client keys + GitHub/Google login${allowedUsers() ? `, allowlist: ${allowedUsers()!.length}` : ""})`
+      : "AUTH DISABLED (dev) — set ADA_CLIENT_KEYS or ADA_ADMIN_KEY to lock down";
   const provs = configuredProviders();
   console.log(`ada backend → http://localhost:${PORT}  [${auth}]`);
   console.log(`providers: ${provs.length ? provs.join(", ") : "(none configured — set provider API keys)"}`);

package/src/server/providers/anthropic.ts

@@ -114,6 +114,8 @@ export const anthropicAdapter: Adapter = {
 
     let stop = "stop";
     let toolIndex = -1;
+    let inTokens = 0; // Anthropic reports input on message_start, cumulative output on message_delta
+    let outTokens = 0;
 
     try {
       const client = await getClient();
@@ -148,7 +150,9 @@ export const anthropicAdapter: Adapter = {
       );
 
       for await (const event of stream) {
-        if (event.type === "content_block_start") {
+        if (event.type === "message_start") {
+          inTokens = (event.message as { usage?: { input_tokens?: number } }).usage?.input_tokens ?? 0;
+        } else if (event.type === "content_block_start") {
           const cb = event.content_block as { type: string; id?: string; name?: string };
           if (cb.type === "tool_use") {
             toolIndex++;
@@ -161,10 +165,15 @@ export const anthropicAdapter: Adapter = {
         } else if (event.type === "message_delta") {
           const reason = (event.delta as { stop_reason?: string | null }).stop_reason;
           if (reason) stop = mapStop(reason);
+          const ot = (event as { usage?: { output_tokens?: number } }).usage?.output_tokens;
+          if (typeof ot === "number") outTokens = ot; // cumulative — take the latest
         }
       }
 
       chunk({}, stop);
+      // Emit an OpenAI-shaped usage chunk so the backend's metering (and the client's own token
+      // counters) work for Claude too — Anthropic doesn't send one in this wire format.
+      writeChunk(res, { id, object: "chat.completion.chunk", created, model, choices: [], usage: { prompt_tokens: inTokens, completion_tokens: outTokens, total_tokens: inTokens + outTokens } });
       endStream(res);
     } catch (err) {
       chunk({ content: `\n[backend: anthropic error: ${err instanceof Error ? err.message : String(err)}]` }, "stop");

package/src/server/providers/openai-compat.ts

@@ -42,14 +42,25 @@ export const openAICompatAdapter: Adapter = {
     // endpoint wants the bare id. (Cloudflare's "@cf/…" ids aren't "cloudflare/…", so they pass through.)
     const prefix = `${provider}/`;
     const outBody = typeof body.model === "string" && body.model.startsWith(prefix) ? { ...body, model: body.model.slice(prefix.length) } : body;
+    // If the client goes away, abort the upstream too — else the full completion is generated,
+    // billed, and (for enterprise) metered against a request nobody is reading.
+    const ac = new AbortController();
+    res.on("close", () => {
+      if (!res.writableEnded) ac.abort();
+    });
     let upstream: Awaited<ReturnType<typeof fetch>>;
     try {
       upstream = await fetch(`${def.baseURL}/chat/completions`, {
         method: "POST",
         headers: { "content-type": "application/json", ...(await authHeaders(provider)) },
         body: JSON.stringify(outBody),
+        signal: ac.signal,
       });
     } catch (e) {
+      if (ac.signal.aborted) {
+        res.end();
+        return;
+      }
       res.writeHead(502, { "content-type": "application/json" });
       res.end(
         JSON.stringify({
@@ -71,10 +82,18 @@ export const openAICompatAdapter: Adapter = {
     if (body.stream) {
       res.writeHead(200, SSE_HEADERS);
       const reader = upstream.body.getReader();
-      for (;;) {
-        const { done, value } = await reader.read();
-        if (done) break;
-        if (value) res.write(Buffer.from(value));
+      try {
+        for (;;) {
+          const { done, value } = await reader.read();
+          if (done) break;
+          if (res.destroyed) {
+            await reader.cancel(); // client gone → stop pulling tokens from upstream
+            break;
+          }
+          if (value) res.write(Buffer.from(value));
+        }
+      } catch {
+        /* aborted mid-stream (client closed) — nothing more to do */
       }
       res.end();
     } else {