@inetafrica/open-claudia 3.5.20 → 3.5.24

package/test-fixtures/fake-agent-cli.js

@@ -13,7 +13,19 @@ const { spawn, spawnSync } = require("child_process");
 
 const argv = process.argv.slice(2);
 const kind = process.env.FAKE_AGENT_KIND;
-const scenario = process.env.FAKE_AGENT_SCENARIO || "success";
+// "provider-error": the run fails with FAKE_AGENT_ERROR_TEXT (a dead sign-in,
+// Claude Code's refresh-lock race). With FAKE_AGENT_ERROR_ONCE_FILE set, only
+// the first run fails (the file marks it) and later runs work.
+function effectiveScenario() {
+  const named = process.env.FAKE_AGENT_SCENARIO || "success";
+  const once = process.env.FAKE_AGENT_ERROR_ONCE_FILE;
+  if (named === "provider-error" && once) {
+    if (fs.existsSync(once)) return "success";
+    fs.writeFileSync(once, "failed once\n");
+  }
+  return named;
+}
+const scenario = effectiveScenario();
 let text = process.env.FAKE_AGENT_TEXT || "fixture says café 👋";
 let invocationCapturePath = null;
 let invocationCapture = null;
@@ -290,6 +302,19 @@ function claudeEvents(sessionId) {
     ];
   }
 
+  if (scenario === "provider-error") {
+    return [
+      init,
+      {
+        type: "result",
+        subtype: "error_during_execution",
+        is_error: true,
+        session_id: sessionId,
+        result: process.env.FAKE_AGENT_ERROR_TEXT || "provider error fixture",
+      },
+    ];
+  }
+
   if (scenario === "usage-limit") {
     // Claude Code 2.1.280 at a plan limit: the rejected rate_limit_event, the
     // synthetic notice (error "rate_limit") and an error result.
@@ -390,6 +415,9 @@ function codexEvents(sessionId) {
       { type: "turn.failed", error: { message: "Authentication failed" } },
     ];
   }
+  if (scenario === "provider-error") {
+    return [started, { type: "turn.failed", error: { message: process.env.FAKE_AGENT_ERROR_TEXT || "provider error fixture" } }];
+  }
 
   const events = [started, { type: "turn.started" }];
   if (process.env.FAKE_AGENT_THINKING) {
@@ -526,6 +554,19 @@ async function holdUntilReleased(stdin) {
   while (!fs.existsSync(file) && Date.now() < deadline) await sleep(20);
 }
 
+// FAKE_AGENT_PROMPT_PROBES: {"name": "<regex source>", ...}; the run log says
+// which of them matched the prompt (never the prompt itself).
+function promptProbes(text) {
+  let spec = null;
+  try { spec = JSON.parse(process.env.FAKE_AGENT_PROMPT_PROBES || "null"); } catch (_) { return null; }
+  if (!spec || typeof spec !== "object") return null;
+  const out = {};
+  for (const [name, source] of Object.entries(spec)) {
+    try { out[name] = new RegExp(source).test(String(text || "")); } catch (_) { out[name] = false; }
+  }
+  return out;
+}
+
 function logRun(event, stdin) {
   const logPath = process.env.FAKE_AGENT_RUN_LOG;
   if (!logPath) return;
@@ -538,6 +579,9 @@ function logRun(event, stdin) {
     threadId: process.env.OC_THREAD_ID || null,
     runId: process.env.OC_RUN_ID || null,
     tag: tag ? tag[1] : null,
+    // Every [[tag:…]] in the prompt, in order (a turn can carry several).
+    tags: [...String(stdin || "").matchAll(/\[\[tag:([\w-]{1,40})\]\]/g)].map((m) => m[1]),
+    probes: promptProbes(stdin),
     fork: argv.includes("--fork-session"),
     resume: argv.indexOf("--resume") >= 0 ? argv[argv.indexOf("--resume") + 1] || null : null,
     sessionEnv: process.env.OC_SESSION_ID || null,
@@ -610,7 +654,7 @@ async function main() {
     setInterval(() => {}, 1000);
     return;
   }
-  if (scenario === "auth-failure" || scenario === "refusal" || scenario === "usage-limit") {
+  if (scenario === "auth-failure" || scenario === "refusal" || scenario === "usage-limit" || scenario === "provider-error") {
     process.exitCode = 1;
     return;
   }

package/test-provider-boot-matrix.js

@@ -297,7 +297,9 @@ if (process.argv[2] === "--probe") {
         DISAPPEAR_PROVIDER: scenario.disappear || "",
       },
       encoding: "utf8",
-      timeout: 30000,
+      // A boot takes about 1 s; loaded CI runners pushed one past 30 s (4 and 5 Oct,
+      // 9 Oct twice), so allow 2 min: this checks what boots, not how fast.
+      timeout: 120000,
     });
 
     try {
@@ -374,7 +376,9 @@ if (process.argv[2] === "--probe") {
         WEB_FIXTURE_WORKSPACE: workspace,
       },
       encoding: "utf8",
-      timeout: 30000,
+      // A boot takes about 1 s; loaded CI runners pushed one past 30 s (4 and 5 Oct,
+      // 9 Oct twice), so allow 2 min: this checks what boots, not how fast.
+      timeout: 120000,
     });
     assert.ifError(webChild.error);
     assert.strictEqual(webChild.status, 0, webChild.stderr || webChild.stdout);

package/test-provider-capabilities.js

@@ -162,7 +162,7 @@ async function commandProbe() {
   assert.ok(modelControls.includes("mb:codex:gpt-5.6-sol"));
   assert.ok(modelControls.includes("mb:codex:gpt-5.6-luna"));
   assert.ok(!modelControls.includes("mb:codex:o4-mini"));
-  assert.ok(modelControls.includes("mb:claude:claude-sonnet-4-6"));
+  assert.ok(modelControls.includes("mb:claude:claude-sonnet-5-5"));
 }
 
 async function callbackProbe() {

package/test-provider-claude.js

@@ -48,7 +48,7 @@ assert.strictEqual(provider.defaultModel(), "claude-default-fixture");
 assert.ok(provider.modelChoices().includes("claude-fable-5-1"));
 assert.ok(provider.modelChoices().includes("claude-opus-5-5"), "opus 5.5 is offered in the model picker");
 assert.ok(provider.modelChoices().includes("claude-fable-5"));
-assert.ok(provider.modelChoices().includes("claude-sonnet-4-6"));
+assert.ok(provider.modelChoices().includes("claude-sonnet-5-5"));
 // 3.4.21: the superseded opus 4.x entries are gone from the picker. They are
 // BUTTONS only - /model takes a typed name verbatim (core/handlers.js reads
 // tail without consulting this list), so the rollback path survives removal.
@@ -57,7 +57,9 @@ for (const retired of ["claude-opus-4-8", "claude-opus-4-7", "claude-opus-4-6"])
 }
 // These two are NOT picker garnish - they are what tierModel returns for
 // utility work, so they must survive any picker trim.
-assert.ok(provider.modelChoices().includes("claude-haiku-4-5-20251001"), "haiku stays: it is the low tier");
+assert.ok(provider.modelChoices().includes("claude-haiku-5-5"), "haiku stays: it is the low tier");
+assert.ok(!provider.modelChoices().includes("claude-haiku-4-5-20251001"), "Haiku 4.5 left the picker for Haiku 5.5 (typed /model still works)");
+assert.ok(!provider.modelChoices().includes("claude-sonnet-4-6"), "Sonnet 4.6 left the picker for Sonnet 5.5 (typed /model still works)");
 
 // The SHIPPED default, asserted on the real module rather than the fixture above,
 // so changing the constant cannot pass unnoticed. Read from source because the
@@ -72,8 +74,8 @@ assert.match(
 );
 // The default governs main turns only: utility tiers stay cheap. These are what
 // stop a default bump from silently repricing recall, pack review and the enforcer.
-assert.strictEqual(claudeProvider.tierModel("low"), "claude-haiku-4-5-20251001", "low-tier utility work stays on haiku");
-assert.strictEqual(claudeProvider.tierModel("medium"), "claude-sonnet-4-6", "medium-tier utility work stays on sonnet");
+assert.strictEqual(claudeProvider.tierModel("low"), "claude-haiku-5-5", "low-tier utility work stays on haiku (Haiku 5.5 since 9 Oct 2026)");
+assert.strictEqual(claudeProvider.tierModel("medium"), "claude-sonnet-5-5", "medium-tier utility work stays on sonnet (Sonnet 5.5 since 9 Oct 2026)");
 // Guard on the RESOLVED config, not process.env: loadEnv folds ~/.open-claudia/.env
 // over process.env, so an operator pin lives in config.config and never appears in
 // process.env. The 3.4.20 version of this check tested the wrong one and passed
@@ -205,7 +207,7 @@ assert.deepStrictEqual(utility.args.slice(utility.args.indexOf("--allowedTools")
 assert.deepStrictEqual(utility.args.slice(utility.args.indexOf("--permission-mode"), utility.args.indexOf("--permission-mode") + 2), ["--permission-mode", "plan"]);
 assert.strictEqual(utility.args.filter((arg) => arg === "/fixture/memory").length, 1, "read-only utility directories are deduplicated");
 assert.deepStrictEqual(utility.args.slice(utility.args.indexOf("/fixture/memory") - 1, utility.args.indexOf("/fixture/memory") + 1), ["--add-dir", "/fixture/memory"]);
-assert.strictEqual(utility.args[utility.args.indexOf("--model") + 1], "claude-haiku-4-5-20251001");
+assert.strictEqual(utility.args[utility.args.indexOf("--model") + 1], "claude-haiku-5-5");
 assert.strictEqual(utility.args[utility.args.indexOf("--json-schema") + 1], JSON.stringify({ type: "object", required: ["ok"] }));
 assert.strictEqual(utility.stdin, "UTILITY PROMPT", "utility prompt uses protected stdin transport");
 assert.ok(!utility.args.includes("UTILITY PROMPT"), "utility prompt is absent from argv");

package/test-provider-env-isolation.js

@@ -112,6 +112,30 @@ assert.strictEqual(hash(claude.CLAUDE_CONFIG_DIR), hash(source.CLAUDE_CONFIG_DIR
 assert.strictEqual(hash(codex.OPENAI_API_KEY), hash(source.OPENAI_API_KEY));
 assert.strictEqual(hash(codex.CODEX_HOME), hash(source.CODEX_HOME));
 
+// agent-browser settings reach the agent unchanged, so its browser commands and Open Claudia's
+// Browser app share one set of launch settings (a mismatch relaunches the browser on a pod).
+const browserSource = {
+  ...source,
+  AGENT_BROWSER_CDP: "9222",
+  AGENT_BROWSER_PIN_TAB: "1",
+  AGENT_BROWSER_DOWNLOAD_PATH: "/browser/downloads",
+  AGENT_BROWSER_EXECUTABLE_PATH: "/usr/bin/chromium",
+  AGENT_BROWSER_ARGS: "--no-first-run",
+  AGENT_BROWSERX: "not-ours",
+  agent_browser_cdp: "lowercase-not-ours",
+};
+for (const builder of [baseAgentEnv, claudeAgentEnv, codexAgentEnv]) {
+  const env = builder(browserSource);
+  for (const key of ["AGENT_BROWSER_CDP", "AGENT_BROWSER_PIN_TAB", "AGENT_BROWSER_DOWNLOAD_PATH", "AGENT_BROWSER_EXECUTABLE_PATH", "AGENT_BROWSER_ARGS"]) {
+    assert.strictEqual(env[key], browserSource[key], `${key} reaches the agent`);
+  }
+  assert.strictEqual(env.AGENT_BROWSERX, undefined, "only the AGENT_BROWSER_ prefix");
+  assert.strictEqual(env.agent_browser_cdp, undefined);
+  assert.strictEqual(env.TELEGRAM_BOT_TOKEN, undefined, "control-plane tokens stay stripped");
+  assert.strictEqual(env.OPEN_CLAUDIA_KEYRING_CANARY, undefined);
+}
+assert.deepStrictEqual(Object.keys(baseAgentEnv({ PATH: "/bin" })), ["PATH"], "nothing extra without browser settings");
+
 assert.match(NATIVE_ISOLATION_NOTICE, /environment isolation/i);
 assert.match(NATIVE_ISOLATION_NOTICE, /same-UID/i);
 assert.match(NATIVE_ISOLATION_NOTICE, /not filesystem isolation/i);

package/test-provider-events.js

@@ -5,7 +5,7 @@
 const assert = require("assert");
 const fs = require("fs");
 const path = require("path");
-const { NORMALIZED_EVENT_TYPES, selectTerminalText } = require("./core/providers/events");
+const { NORMALIZED_EVENT_TYPES, providerErrorFlags, providerFailureCode, selectTerminalText } = require("./core/providers/events");
 const { createClaudeParserState, normalizeClaudeEvent } = require("./core/providers/claude-events");
 const { createCodexParserState, normalizeCodexEvent } = require("./core/providers/codex-events");
 
@@ -118,6 +118,7 @@ assert.deepStrictEqual(claudeAuth, [{
   message: "Authentication failed",
   authError: true,
   usageLimit: false,
+  transient: false,
   refusal: false,
 }]);
 
@@ -241,6 +242,7 @@ assert.deepStrictEqual(limitFailure, [{
   message: limitText,
   authError: false,
   usageLimit: true,
+  transient: false,
   refusal: false,
   limitResetsAt: 1791205800000,
   limitType: "five_hour",
@@ -306,6 +308,7 @@ assert.deepStrictEqual(codexAuth, [{
   message: "401 Unauthorized",
   authError: true,
   usageLimit: false,
+  transient: false,
   refusal: false,
 }]);
 
@@ -346,4 +349,43 @@ assert.deepStrictEqual(normalizeClaudeEvent(samples.claude.session, isolatedB),
   );
 }
 
+// What a failure means, read one way for chat turns and background calls
+// (9 Oct 2026: the 07:00 dream failed on 18 of 24 fleet bots; Claude Code
+// 2.1.280's own wording below).
+{
+  const deadRefresh = "Failed to authenticate: OAuth session expired and could not be refreshed";
+  assert.deepStrictEqual(
+    [providerErrorFlags(deadRefresh).authError, providerErrorFlags(deadRefresh).transient],
+    [true, false],
+    "a sign-in that died is a sign-in error (it used to read as a plain provider error)",
+  );
+  for (const text of [
+    "OAuth refresh token is no longer valid; run /login to re-authenticate",
+    "API Error: 401 OAuth access token is invalid",
+    "OAuth access token is invalid",
+    "Not logged in · Please run /login",
+    "Your authentication token has expired. Please try signing in again.",
+  ]) {
+    assert.strictEqual(providerErrorFlags(text).authError, true, text);
+    assert.strictEqual(providerFailureCode(providerErrorFlags(text), "UTILITY_PROVIDER_ERROR"), "PROVIDER_UNAUTHENTICATED", text);
+  }
+  for (const text of [
+    "Failed to refresh OAuth token: another Claude Code process is refreshing it or exited mid-refresh. This is usually transient; retry in a minute, and if it persists close other Claude Code processes or sign in again",
+    "another Claude Code process is refreshing it (or exited mid-refresh)",
+  ]) {
+    const flags = providerErrorFlags(text);
+    assert.strictEqual(flags.transient, true, text);
+    assert.strictEqual(flags.authError, false, "the refresh race is not a dead sign-in");
+    assert.strictEqual(providerFailureCode(flags), "PROVIDER_TRANSIENT", text);
+  }
+  const weekly = providerErrorFlags("You've hit your weekly limit · resets Oct 9, 5am (UTC)");
+  assert.strictEqual(providerFailureCode(weekly, "UTILITY_PROVIDER_ERROR"), "PROVIDER_USAGE_LIMIT");
+  assert.strictEqual(providerFailureCode({ refusal: true, authError: true }), "PROVIDER_REFUSAL", "a refusal leads");
+  assert.strictEqual(providerFailureCode({}, "UTILITY_PROVIDER_ERROR"), "UTILITY_PROVIDER_ERROR");
+  assert.strictEqual(providerFailureCode({}), "PROVIDER_EVENT_ERROR");
+  // Codex's turn.failed carries the same flags.
+  const codexExpired = normalizeCodexEvent({ type: "turn.failed", error: { message: "Your authentication token has expired. Please try signing in again." } }, createCodexParserState());
+  assert.strictEqual(codexExpired[0].authError, true);
+}
+
 console.log("provider events OK");

package/test-provider-language.js

@@ -595,7 +595,22 @@ assert.match(pkg.description, /provider-agnostic coding-agent harness/i);
 // also ensure you clear the images after they all done"): the Browser app in
 // Kazee Chat (bot apps v1): Take over, Teach me with its pictures deleted when
 // the work is done, and the bot paused while a person drives. A canary bot first.
-assert.strictEqual(pkg.version, "3.5.20", "release version must remain unchanged without explicit approval");
+// 3.5.23 approved by Sumeet 2026-10-09 15:56 EAT ("yeh release and upgrade all
+// the bots, include rahil for the updates going forward"), after 14:18 ("there's
+// something about the cancel that you seem to have broke … one for a qued message
+// a small card with delete queued … get all the messages that are queued at once"),
+// 14:51 ("have it shifted to the new haiku 5.5") and 15:36 ("sonnet should also be
+// just 5.5"): usage limits as one wait-for-reset rule, /stop as a cancel, the
+// queue's ✕ and run-together, Haiku 5.5 and Sonnet 5.5 tiers. Carries 3.5.20.
+// Tagged first as 3.5.21 (cbfd547), whose CI build failed on test-stop-queue's
+// registration before anything published; 3.5.22 (bd95677) failed on two
+// load-sensitive timing checks, now given more time; the release ships as 3.5.23.
+// 3.5.24 approved under Sumeet's 2026-10-09 12:02 EAT Browser go-ahead ("fine go ahead
+// and push"), after 21:27 EAT ("I can't change the url on the mobile app"): the Browser
+// app's address bar, agent-browser's settings for the agent, and container mode for Agent
+// Space's browser container. A canary build for the Browser canary bot; the fleet waits
+// for his OK.
+assert.strictEqual(pkg.version, "3.5.24", "release version must remain unchanged without explicit approval");
 for (const keyword of ["claude", "codex", "coding-agent", "provider-agnostic"]) {
   assert.ok(pkg.keywords.includes(keyword), `package keyword missing: ${keyword}`);
 }

package/web.js

@@ -379,9 +379,12 @@ const USER_ACCESS_ROUTES = new Set(["/api/auth/approve", "/api/auth/deny", "/api
 //   POST /api/apps/check                                    may this person open the app (C1)
 //   POST /api/browser/control  GET /api/browser/control     take / renew / release, status (C2)
 //   POST /api/browser/input                                 the driver's input events (C3)
+//   POST /api/browser/navigate                              the driver's address bar: open <url>
+//                                                           (public addresses only), back,
+//                                                           forward, reload
 //   POST /api/browser/teach                                 Teach me: start / stop / cancel (C4)
 //   GET  /api/browser/teach[/<id>[/pictures/<n>]]?userId=   that person's recordings only
-const BROWSER_APP_ROUTE = /^\/api\/(?:apps\/check|browser\/(?:control|input|teach))(?:[/?]|$)/;
+const BROWSER_APP_ROUTE = /^\/api\/(?:apps\/check|browser\/(?:control|input|navigate|teach))(?:[/?]|$)/;
 const BROWSER_APP_MAX_BODY = 64 * 1024;
 const BROWSER_APP_USER_ID = /^[A-Za-z0-9._:-]{1,100}$/;
 
@@ -404,6 +407,7 @@ async function handleBrowserApp(req, res, body) {
   if (route === "/api/browser/control" && req.method === "POST") return answer(await control.control(parsed));
   if (route === "/api/browser/control" && req.method === "GET") return answer(control.status());
   if (route === "/api/browser/input" && req.method === "POST") return answer(await control.input(parsed));
+  if (route === "/api/browser/navigate" && req.method === "POST") return answer(await control.navigate(parsed));
   if (route === "/api/browser/teach" && req.method === "POST") return answer(await control.teachAction(parsed));
   if (req.method === "GET" && (route === "/api/browser/teach" || route.startsWith("/api/browser/teach/"))) {
     const teach = require("./core/browser-teach");
@@ -700,7 +704,8 @@ async function handleAPI(req, res, body, dependencies = {}) {
     return res.end(JSON.stringify({ content: loadLogs() }));
   }
 
-  // Browser app in Kazee Chat (bot apps v1): who may open it, Take over, input and Teach me.
+  // Browser app in Kazee Chat (bot apps v1): who may open it, Take over, input, the address bar
+  // and Teach me.
   // Machine-to-machine for Agent Space, so the control token only: a dashboard session is
   // refused, like /api/otp/mint.
   if (BROWSER_APP_ROUTE.test(url)) {