@reclaimprotocol/agent 0.1.23 → 0.1.25

package/src/mcp/tools/agent/session.ts

@@ -2,17 +2,21 @@ import type { ConfigStore } from '../../../config/store.ts'
 import { defineTool, type RegisteredTool } from '../../server.ts'
 import { type AttachState, disposeAttached } from './attach.ts'
 import { type CaptureStore, resetCaptures } from './capture.ts'
+import { type LiveShareRef, stopLiveShare } from './live-view.ts'
 
 export function sessionTools(
 	attachRef: { current?: AttachState, config: ConfigStore },
 	captures: CaptureStore,
+	liveShare: LiveShareRef,
 ): RegisteredTool[] {
 	return [
 		defineTool<Record<string, never>>(
 			{
 				name: 'get_session',
 				description:
-					'Return current attach state, active captures, and persisted config.',
+					'Report the current attach state, the live captureIds, and the '
+					+ 'persisted browser config. Call it to check whether a browser is '
+					+ 'still attached before reattaching, or to recover a captureId.',
 				inputSchema: { type: 'object', properties: {} },
 			},
 			async() => ({
@@ -25,6 +29,13 @@ export function sessionTools(
 				orgId: attachRef.current?.orgId,
 				captures: [...captures.sessions.keys()],
 				activeTab: captures.activeTab,
+				// The public share, when one is open — so a caller that lost the
+				// URL can recover it instead of opening a second tunnel.
+				...(liveShare.current
+					? {
+						sharedView: { url: liveShare.current.url },
+					}
+					: {}),
 			}),
 		),
 		defineTool<Record<string, never>>(
@@ -33,11 +44,12 @@ export function sessionTools(
 				description:
 					'Stop captures, disconnect the current browser, and immediately '
 					+ 'delete a Builder-hosted Popcorn session. Custom CDP sessions '
-					+ 'are disconnected but never deleted by Reclaim.',
+					+ 'are disconnected but never deleted by Reclaim. Also closes a '
+					+ 'public live view opened by share_browser_view.',
 				inputSchema: { type: 'object', properties: {} },
 			},
 			async() => {
-				await resetAll(attachRef, captures)
+				await resetAll(attachRef, captures, liveShare)
 				return { disposed: true }
 			},
 		),
@@ -46,21 +58,36 @@ export function sessionTools(
 				name: 'reset_session',
 				description:
 					'Stop all captures and clear capture state — drops every '
-					+ 'session\'s drafts and their in-memory secrets.',
+					+ 'session\'s drafts and their in-memory secrets, and closes a '
+					+ 'public live view opened by share_browser_view.',
 				inputSchema: { type: 'object', properties: {} },
 			},
 			async() => {
-				await resetAll(attachRef, captures)
+				await resetAll(attachRef, captures, liveShare)
 				return { reset: true }
 			},
 		),
 	]
 }
 
-async function resetAll(
+/**
+ * Release everything a session holds: the public tunnel, the captures, and the
+ * browser itself.
+ *
+ * Exported because a session can end WITHOUT anyone calling `dispose_browser` —
+ * the developer closes their agent, the client kills the server, the transport
+ * drops. Nothing runs on its own then, so `start.ts` runs this. Without it a
+ * container keeps running and, worse, a public URL to a live browser outlives
+ * the session that created it.
+ */
+export async function resetAll(
 	attachRef: { current?: AttachState, config: ConfigStore },
 	captures: CaptureStore,
+	liveShare: LiveShareRef,
 ) {
+	// Before the captures: a public URL must never outlive the browser it points
+	// at.
+	await stopLiveShare(liveShare)
 	await resetCaptures(captures)
 	await disposeAttached(attachRef)
 }

package/src/mcp/tools/authenticate/authenticate.ts

@@ -18,17 +18,12 @@ export function authenticateTool(
 			name,
 			description:
 				(name === 'authenticate_builder'
-					? 'This tool only exists in OLD-DEVTOOLS mode — Builder mode has '
-					+ 'no separate version of it; its own reclaim_authenticate '
-					+ 'already covers everything below. In old-devtools mode, this '
-					+ 'authenticates with Builder for exactly two Builder-exclusive '
-					+ 'features: allocating a Popcorn remote browser (attach_browser '
-					+ 'mode="builder") and listing Builder organizations '
-					+ '(list_builder_organizations — old-devtools has no generated '
-					+ 'org tool of its own). It does NOT authorize '
-					+ 'create_provider_version_from_capture or get_me_providers, '
-					+ 'which need their OWN, separate login via the old-devtools '
-					+ 'reclaim_authenticate tool. Check both with get_devtools_mode. '
+					? 'OLD-DEVTOOLS MODE ONLY, and narrow: it authorizes exactly two '
+					+ 'Builder-exclusive features — allocating a remote browser '
+					+ '(attach_browser mode="builder") and list_builder_organizations. '
+					+ 'It does NOT authorize create_provider_version_from_capture or '
+					+ 'get_me_providers; those need the separate old-devtools login, '
+					+ 'reclaim_authenticate. Check both with get_devtools_mode. '
 					: '')
 				+ 'Authorize the CLI / MCP to make authenticated API calls. Opens '
 				+ 'the '
@@ -37,8 +32,9 @@ export function authenticateTool(
 				'the user to sign in and approve. On success the access token is ' +
 				'cached in ~/.reclaim/config.json and activated for this session.\n\n' +
 				'Call with no arguments. This bearer token authenticates ' +
-				'dashboard/org-management calls (e.g. list_orgs, issuing an org ' +
-				'secret). Programmatic verification calls use a separate org secret ' +
+				'dashboard/org-management calls (for example, list_orgs, or ' +
+				'issuing an org secret). Programmatic verification calls use a ' +
+				'separate org secret ' +
 				'(rorg_…, set as RECLAIM_ORG_SECRET).',
 			inputSchema: { type: 'object', properties: {} },
 		},
@@ -85,6 +81,18 @@ export async function runOAuthLoopback(
 	const codeChallenge = await oauth.calculatePKCECodeChallenge(codeVerifier)
 	const state = oauth.randomState()
 
+	// openid-client/oauth4webapi refuse any non-HTTPS issuer by default
+	// ("only requests to HTTPS are allowed"). Better Auth's own dev default
+	// (`canonicalOrigin()`, packages/app/src/consts.ts) is plain
+	// `http://localhost:4001` — the server side has no HTTPS requirement at
+	// all — so a local `RECLAIM_API_URL=http://localhost:...` would otherwise
+	// always fail here. Loosen it ONLY for loopback hosts, matching what the
+	// server already trusts (`authTrustedOrigins()` defaults to
+	// localhost:4000/4001 outside production); a non-loopback http:// origin
+	// still gets the strict check.
+	const isLoopbackAuthBase =
+		['localhost', '127.0.0.1', '[::1]'].includes(new URL(authBase).hostname)
+
 	// Bind the callback port first: the client is registered for this exact
 	// loopback redirect.
 	const callback = await listenForCallback()
@@ -119,6 +127,13 @@ export async function runOAuthLoopback(
 				// `authorizationCodeGrant` fails instead (confirmed by hitting
 				// exactly that when this only checked method === 'POST').
 				[oauth.customFetch]: _customFetchForOauth,
+				// Applies to the `Configuration` this call returns, so the later
+				// `buildAuthorizationUrl`/`authorizationCodeGrant` calls below (which
+				// reuse that same `config`) stop enforcing HTTPS too — no separate
+				// opt-in needed for them.
+				...(isLoopbackAuthBase
+					? { execute: [oauth.allowInsecureRequests] }
+					: {}),
 			},
 		)
 		openBrowserFn(

package/src/mcp/tools/mode/mode.ts

@@ -24,7 +24,7 @@ function hasValidBuilderSession(config: ConfigStore): boolean {
  * `~/.reclaim/config.json` and applied the NEXT time the server starts —
  * the registered tool set, server instructions, and backend client are all
  * fixed at connect time, so the switch requires a restart/reconnect. Follows
- * the `attach_chrome` persist-through-config pattern.
+ * the `attach_browser` persist-through-config pattern.
  *
  * Registered in BOTH modes (so the user can switch either direction) and
  * outside `buildAgentTools` (so it survives `RECLAIM_AGENT_DISABLED=1`). An
@@ -74,27 +74,15 @@ export function buildModeTools(
 			{
 				name: 'get_devtools_mode',
 				description:
-					'Report which backend this MCP server is currently targeting: '
-					+ '"old" (devtools) or "builder" (the builder '
-					+ 'backend). Also reports WHY that mode is in effect (source: '
-					+ 'a USE_OLD_DEVTOOLS env override, the persisted config, or '
-					+ 'the default) and which mode is persisted for the next '
-					+ 'start. Use set_devtools_mode to change it. Also reports '
-					+ 'AUTH status for BOTH auth systems — builderAuthenticated (the '
-					+ 'Builder OAuth login). In BUILDER mode this is the only login '
-					+ 'that exists and covers everything, including org listing '
-					+ '(list_orgs) and remote-browser allocation. In OLD-DEVTOOLS '
-					+ 'mode it is a SEPARATE, additional login (authenticate_builder) '
-					+ 'needed only for two Builder-exclusive features: remote-browser '
-					+ '(Popcorn) allocation and org listing (list_builder_organizations '
-					+ '— old-devtools has no generated org tool of its own). '
-					+ 'oldDevtoolsAuthenticated (the OLD-mode reclaim_authenticate) is '
-					+ 'required separately for create_provider_version_from_capture '
-					+ 'and get_me_providers, and only exists/matters in old-devtools '
-					+ 'mode. Both are presence checks, not '
-					+ 'validity checks — a stale old-devtools token still reports '
-					+ 'true here; a real publish/list call will surface staleness '
-					+ 'with a clear error pointing back at reclaim_authenticate.',
+					'Report which backend this server targets ("old" devtools or '
+					+ '"builder"), WHY (a USE_OLD_DEVTOOLS env override, the persisted '
+					+ 'config, or the default), and which mode is persisted for the '
+					+ 'next start. Change it with set_devtools_mode. Also reports auth '
+					+ 'status for both login systems: builderAuthenticated and '
+					+ 'oldDevtoolsAuthenticated. Both are PRESENCE checks, not validity '
+					+ 'checks — a stale token still reports true; a real call surfaces '
+					+ 'the staleness with an error naming the tool to re-run. Which '
+					+ 'login authorizes what: how_it_works({ topic: "auth" }).',
 				inputSchema: { type: 'object', properties: {} },
 			},
 			async() => {

package/src/old/client.ts

@@ -50,7 +50,7 @@ export interface RegisterProviderBody {
 	/** Interception mechanism (`NONE` = standard replay flow, the default;
 	 *  `HAWKEYE`/`MSWJS`/`XHOOK` = JS-context interception; `CDP` = browser-
 	 *  level). Column default is `HAWKEYE` when omitted entirely — unlike
-	 *  `stepsToFollow` etc. above, whether the register endpoint actually
+	 *  `stepsToFollow`, and so on. above, whether the register endpoint actually
 	 *  PERSISTS this field is UNVERIFIED (no source access to confirm); sent
 	 *  here on the chance it does, but the publish tool always follows up
 	 *  with a config (add-version) call regardless, which is confirmed to
@@ -370,7 +370,7 @@ export class ReclaimOldClient {
 		// An identity ALREADY SET (as opposed to `#requireIdentity`'s "never
 		// authenticated" case, which never reaches the backend) plus either a
 		// 401/403 OR a message that itself names token verification means the
-		// backend rejected a token we had — i.e. it's stale/invalid, not merely
+		// backend rejected a token we had — that is, it's stale/invalid, not merely
 		// absent. The message check matters because this backend does NOT
 		// reliably use 401/403 for it — confirmed live: an actually-expired
 		// Firebase token surfaced as "Error while verifying token" with a

package/src/old/instructions.ts

@@ -2,14 +2,8 @@
  * MCP server instructions shown to the LLM when running in OLD-devtools mode
  * (`USE_OLD_DEVTOOLS` unset or not `false`). Describes the REDUCED tool
  * surface — there are no org/credential-link/review flows and no generated
- * API tools here. The prose lives in `skill/old-mode.md` (published via the
- * `files` array); the path resolves identically from `src/old` (dev) and
- * `lib/old` (published).
+ * API tools here. Both modes' instructions load together in
+ * `mcp/instructions` (alongside the topic-split authoring guide); re-exported
+ * here so old-mode callers keep importing from their own module.
  */
-import { readFileSync } from 'node:fs'
-import { fileURLToPath } from 'node:url'
-
-export const OLD_MODE_INSTRUCTIONS = readFileSync(
-	fileURLToPath(new URL('../../skill/old-mode.md', import.meta.url)),
-	'utf8',
-).trim()
+export { OLD_MODE_INSTRUCTIONS } from '../mcp/instructions.ts'

package/src/old/login-cdp.ts

@@ -97,8 +97,8 @@ export interface CdpLoginResult {
  * finish signing in, read the resulting Firebase ID token from page storage,
  * then clean up. If we had to launch a dedicated Chrome for this, we close the
  * whole browser afterwards; if we attached to a Chrome that was already up
- * (e.g. the capture flow's), we only close the tab we opened so we don't
- * disrupt it.
+ * (for example, the capture flow's), we only close the tab we opened so we
+ * don't disrupt it.
  */
 export async function captureFirebaseTokenViaCdp(opts: {
 	loginUrl?: string
@@ -107,7 +107,13 @@ export async function captureFirebaseTokenViaCdp(opts: {
 }): Promise<CdpLoginResult> {
 	const loginUrl =
 		opts.loginUrl || oldLoginUrl() || DEFAULT_LOGIN_URL
-	const timeoutMs = opts.timeoutMs ?? 180_000
+	// 6 minutes, not 3. On a FRESH Chrome profile the dashboard is a cold load
+	// of an ~8.6 MB uncompressed bundle before the login form is even
+	// interactive — tens of seconds on a slow link — and only then does the
+	// developer start typing credentials and a 2FA code. The old 3-minute
+	// budget was routinely spent before they could finish, and the retry that
+	// followed looked like a 6-minute hang with no output.
+	const timeoutMs = opts.timeoutMs ?? 360_000
 	const pollMs = opts.pollMs ?? 1000
 
 	const cfg = new ConfigStore().read()?.browser_agent
@@ -123,7 +129,7 @@ export async function captureFirebaseTokenViaCdp(opts: {
 			throw new Error(
 				`attach mode: nothing answering CDP on port ${port}. Launch Chrome ` +
 				`with --remote-debugging-port=${port} AND a non-default ` +
-				'--user-data-dir, or switch to dedicated mode (attach_chrome).',
+				'--user-data-dir, or switch to dedicated mode (attach_browser).',
 			)
 		}
 
@@ -139,6 +145,10 @@ export async function captureFirebaseTokenViaCdp(opts: {
 	try {
 		const session = await handles.attachToTab(targetId)
 		const deadline = Date.now() + timeoutMs
+		// Tracks how far the page got, so a timeout can say WHICH thing stalled:
+		// a bundle that never finished downloading is a different problem from a
+		// developer who walked away mid-sign-in.
+		let sawInteractivePage = false
 		while(Date.now() < deadline) {
 			const result = await session.send('Runtime.evaluate', {
 				expression: EXTRACT_TOKEN_JS,
@@ -150,12 +160,24 @@ export async function captureFirebaseTokenViaCdp(opts: {
 				return { token: value, closedBrowser: launchedByUs }
 			}
 
+			if(!sawInteractivePage) {
+				sawInteractivePage = await pageIsInteractive(session)
+			}
+
 			await new Promise((resolve) => setTimeout(resolve, pollMs))
 		}
 
+		const waited = Math.round(timeoutMs / 1000)
 		throw new Error(
-			'Timed out waiting for login. Finish signing in at the opened ' +
-			'page, or re-run reclaim_authenticate with a longer timeoutMs.',
+			`Timed out after ${waited}s waiting for login. `
+			+ (sawInteractivePage
+				? 'The page finished loading, so sign-in was never completed — '
+					+ 'ask the developer to finish signing in at the opened window, '
+					+ 'then re-run reclaim_authenticate.'
+				: 'The page never became interactive, so the dashboard was still '
+					+ 'loading the whole time. On a fresh Chrome profile its bundle '
+					+ 'is a large cold download. Retry (the profile is now warm), or '
+					+ 'pass a bigger timeoutMs.'),
 		)
 	} finally {
 		// Best-effort cleanup — never mask the real result/error.
@@ -173,6 +195,28 @@ export async function captureFirebaseTokenViaCdp(opts: {
 	}
 }
 
+/**
+ * Whether the login page has reached a usable state. `readyState` alone is not
+ * enough for a single-bundle SPA: the document is 'complete' while the app is
+ * still parsing, and nothing is clickable yet. Requiring a rendered root as
+ * well is what distinguishes "still loading" from "waiting for the human".
+ */
+async function pageIsInteractive(
+	session: { send: (m: string, p?: object) => Promise<unknown> },
+): Promise<boolean> {
+	try {
+		const result = await session.send('Runtime.evaluate', {
+			expression:
+				'document.readyState === "complete" '
+				+ '&& !!document.body && document.body.innerText.trim().length > 0',
+			returnByValue: true,
+		})
+		return runtimeResultValue(result) === true
+	} catch{
+		return false
+	}
+}
+
 function runtimeResultValue(value: unknown): unknown {
 	if(typeof value !== 'object' || !value || !('result' in value)) {
 		return undefined

package/src/old/to-register.ts

@@ -39,7 +39,8 @@ export interface OldProviderOpts {
 	/** Page-injection script — the devtools name for the builder's
 	 *  `webSettings.jsUserScripts`. Top-level on register, nested under
 	 *  `providerConfig` on config (add-version). The backend rejects a few
-	 *  deprecated injection APIs (e.g. `window.payloadData`) with a 400. */
+	 *  deprecated injection APIs (for example, `window.payloadData`) with a
+	 *  400. */
 	customInjection?: string
 	/** Accepted at register too (unlike the four fields below, which the
 	 *  backend only accepts on config/add-version). */
@@ -355,7 +356,7 @@ export interface SemverTriplet {
 /**
  * Pull `{major,minor,patch}` triplets out of a `/versions` response, tolerant
  * of the exact envelope (array, `{ providerVersions: [...] }`, `{ items }`,
- * etc.). Each entry's `version` is the triplet.
+ *, and so on). Each entry's `version` is the triplet.
  */
 export function extractVersions(resp: unknown): SemverTriplet[] {
 	const list = pickArray(resp)

package/src/old/tools/authenticate.ts

@@ -46,6 +46,11 @@ export function authenticateTool(client: ReclaimOldClient): RegisteredTool {
 				'and closes the tab. A previously stored identity that is still ' +
 				'valid is reused first (pass `useBrowser: true` to force a fresh ' +
 				'login). \n\n' +
+				'EXPECT THIS TO TAKE A WHILE ON A FIRST RUN: a fresh Chrome ' +
+				'profile downloads the dashboard cold, so tell the developer it is ' +
+				'loading and that the window is waiting for them — do not re-run ' +
+				'the tool while it is still waiting. It polls for up to 6 minutes ' +
+				'and reports whether the page ever became interactive.\n\n' +
 				'Manual overrides:\n' +
 				'- `token`: paste a Firebase ID token (skips the browser).\n' +
 				'- `ethAddress`: an `eth:0x…` / `0x…` uid for an already-provisioned ' +
@@ -73,7 +78,10 @@ export function authenticateTool(client: ReclaimOldClient): RegisteredTool {
 					timeoutMs: {
 						type: 'number',
 						description:
-							'How long to wait for sign-in, in ms (default 180000).',
+							'How long to wait for sign-in, in ms (default 360000, i.e. 6 '
+							+ 'minutes). A fresh Chrome profile loads the dashboard cold, '
+							+ 'which can take tens of seconds before the form is usable — '
+							+ 'that time comes out of this budget.',
 					},
 					useBrowser: {
 						type: 'boolean',

package/src/old/tools/get-provider-info.ts

@@ -3,8 +3,8 @@ import { defineTool, type RegisteredTool } from '../../mcp/server.ts'
 import type { ReclaimSdkClient } from '../sdk-client.ts'
 
 /** Best-effort call: resolve to the response, or to an error string instead
- *  of throwing — one endpoint 404ing (e.g. a bad `versionNumber`) shouldn't
- *  hide a result the other endpoint DID have. */
+ *  of throwing — one endpoint 404ing (for example, a bad `versionNumber`)
+ *  shouldn't hide a result the other endpoint DID have. */
 async function attempt(
 	label: string,
 	call: () => Promise<unknown>,
@@ -51,8 +51,9 @@ export function getProviderInfoTool(client: ReclaimSdkClient): RegisteredTool {
 				'computed `isScriptRequestingClaim` flag). Read `provider` here ' +
 				'BEFORE create_provider_version_from_capture so you can pass the ' +
 				'whole existing config back — that tool always creates a new ' +
-				'version and carries nothing forward. One endpoint failing (e.g. ' +
-				'an invalid `versionNumber`) is reported under `errors` without ' +
+				'version and carries nothing forward. One endpoint failing (for ' +
+				'example, an invalid `versionNumber`) is reported under `errors` ' +
+				'without ' +
 				'hiding a result the other endpoint did return.',
 			inputSchema: {
 				type: 'object',
@@ -64,7 +65,7 @@ export function getProviderInfoTool(client: ReclaimSdkClient): RegisteredTool {
 					versionNumber: {
 						type: 'string',
 						description:
-							'Specific version, e.g. "1.0.2". Omit for the ' +
+							'Specific version, for example, "1.0.2". Omit for the ' +
 							"provider's current latest version.",
 					},
 				},

package/src/old/tools/issue-app-credentials.ts

@@ -24,8 +24,9 @@ export function issueAppCredentialsTool(
 				'starts UNLINKED and sandboxed (100 sessions) — call ' +
 				'link_app_to_account to attach it to your dashboard account and ' +
 				'check_app_status to check its state. Store `appSecret` like a ' +
-				'password (e.g. in a gitignored .env as RECLAIM_APP_SECRET) — it ' +
-				'is returned once and cannot be recovered from the server.',
+				'password (for example, in a gitignored .env as ' +
+				'RECLAIM_APP_SECRET) — it is returned once and cannot be ' +
+				'recovered from the server.',
 			inputSchema: { type: 'object', properties: {} },
 		},
 		async() => {