@phnx-labs/agents-cli 1.22.1 → 1.22.4

package/dist/lib/secrets/read-backoff.js

@@ -0,0 +1,64 @@
+/**
+ * Negative memo for failed/cancelled macOS keychain reads (the "back-off").
+ *
+ * A cancelled Touch ID sheet is the user saying "not now" — but a polling
+ * caller (the Factory extension host's `agents view` loop, a watch script)
+ * retries the same read a few seconds later and pops the sheet again, forever.
+ * The headless guard (index.ts `assertRawKeychainReadAllowed`) covers the
+ * no-TTY case; this memo covers a context that CAN prompt but just had its
+ * prompt cancelled or fail: the next read of the same item within the TTL
+ * throws the back-off error instead of re-prompting. Any successful read (or
+ * write) of the item clears the memo.
+ *
+ * Stored as regenerable state under `~/.agents/.cache/keychain-read-backoff/`,
+ * one file per item (filename is a hash of the item name; the file carries no
+ * secret material — a name and a deadline only). All operations are
+ * best-effort: a lost memo costs at most one extra prompt, never a read.
+ */
+import { createHash } from 'node:crypto';
+import * as fs from 'fs';
+import * as os from 'os';
+import * as path from 'path';
+/** How long a failed/cancelled read suppresses retries of the same item. */
+export const KEYCHAIN_READ_BACKOFF_TTL_MS = 5 * 60 * 1000;
+let dirOverride = null;
+/** Test seam: point the memo at a temp dir so tests never touch the real cache. */
+export function setKeychainReadBackoffDirForTest(dir) {
+    dirOverride = dir;
+}
+function backoffDir() {
+    return dirOverride ?? path.join(os.homedir(), '.agents', '.cache', 'keychain-read-backoff');
+}
+function backoffFile(key) {
+    const digest = createHash('sha256').update(key).digest('hex').slice(0, 24);
+    return path.join(backoffDir(), `${digest}.json`);
+}
+/** True while `key` is inside the back-off window opened by a failed/cancelled read. */
+export function isKeychainReadBackedOff(key, now = Date.now()) {
+    try {
+        const parsed = JSON.parse(fs.readFileSync(backoffFile(key), 'utf8'));
+        return typeof parsed.until === 'number' && parsed.until > now;
+    }
+    catch {
+        return false; // absent or malformed memo → no back-off
+    }
+}
+/** Open (or refresh) the back-off window for `key` after a failed/cancelled read. */
+export function noteKeychainReadFailure(key, now = Date.now()) {
+    try {
+        fs.mkdirSync(backoffDir(), { recursive: true, mode: 0o700 });
+        fs.writeFileSync(backoffFile(key), JSON.stringify({ item: key, until: now + KEYCHAIN_READ_BACKOFF_TTL_MS }), { mode: 0o600 });
+    }
+    catch {
+        /* best-effort — the cache dir is regenerable; a lost memo costs one prompt */
+    }
+}
+/** Clear the memo: a successful read or write of the item resets the back-off. */
+export function clearKeychainReadBackoff(key) {
+    try {
+        fs.rmSync(backoffFile(key), { force: true });
+    }
+    catch {
+        /* best-effort */
+    }
+}

package/dist/lib/secrets/session-store.js

@@ -89,7 +89,8 @@ function sessionBlobItem(name, harness) {
 /** Read the session index by its fixed name. `{bundles:{}}` when absent/unreadable. */
 export function readIndex() {
     try {
-        const raw = getKeychainToken(SESSION_INDEX_ITEM);
+        // Written no-ACL (writeIndex below) — attest so a headless resolve stays silent.
+        const raw = getKeychainToken(SESSION_INDEX_ITEM, { silentNoAcl: true });
         const parsed = JSON.parse(raw);
         if (parsed && typeof parsed === 'object' && parsed.bundles)
             return parsed;
@@ -145,7 +146,8 @@ export function loadSession(name, now = Date.now(), harness = GLOBAL_HARNESS) {
     if (!shouldPersist())
         return null;
     try {
-        const raw = getKeychainToken(sessionBlobItem(name, harness));
+        // Written no-ACL (saveSession) — attest so a headless resolve stays silent.
+        const raw = getKeychainToken(sessionBlobItem(name, harness), { silentNoAcl: true });
         const entry = JSON.parse(raw);
         if (!entry || typeof entry !== 'object' || !entry.bundle || !entry.env)
             return null;
@@ -227,7 +229,7 @@ export function rehydrateSessions(now = Date.now()) {
             if (key.includes(':'))
                 continue;
             try {
-                const raw = getKeychainToken(`${SESSION_ITEM_PREFIX}${key}`);
+                const raw = getKeychainToken(`${SESSION_ITEM_PREFIX}${key}`, { silentNoAcl: true });
                 const legacy = JSON.parse(raw);
                 setKeychainToken(sessionBlobItem(key, GLOBAL_HARNESS), JSON.stringify({ ...legacy, harness: GLOBAL_HARNESS }), { noAcl: true });
                 deleteKeychainToken(`${SESSION_ITEM_PREFIX}${key}`);
@@ -265,7 +267,7 @@ export function rehydrateSessions(now = Date.now()) {
                 continue;
             }
             try {
-                const raw = getKeychainToken(sessionBlobItem(bundleName, 'cli'));
+                const raw = getKeychainToken(sessionBlobItem(bundleName, 'cli'), { silentNoAcl: true });
                 const legacy = JSON.parse(raw);
                 setKeychainToken(sessionBlobItem(bundleName, GLOBAL_HARNESS), JSON.stringify({ ...legacy, harness: GLOBAL_HARNESS }), { noAcl: true });
                 deleteKeychainToken(sessionBlobItem(bundleName, 'cli'));

package/dist/lib/secrets/vault.js

@@ -209,7 +209,9 @@ export function clearVaultKey() {
 export function getVaultSession() {
     let raw;
     try {
-        raw = getKeychainToken(VAULT_SESSION_ITEM);
+        // Written no-ACL (cacheVaultKey) — the vault session probe runs inside bundleBackend
+        // for EVERY bundle op, headless included, so it must stay silent.
+        raw = getKeychainToken(VAULT_SESSION_ITEM, { silentNoAcl: true });
     }
     catch {
         return { loggedIn: false };

package/dist/lib/share/config.d.ts

@@ -32,12 +32,20 @@ export declare function readWriteTokenFromBundle(): string;
 /** Read the raw write token from injected env first, then the local bundle.
  * Throws with an actionable message if absent (run setup/join first). */
 export declare function readWriteToken(): string;
-/** Best-effort runtime env for spawned agents. Never throws: a missing/locked
- * bundle should not block unrelated agent runs, but an already-unlocked bundle
- * or injected token lets ephemeral agents publish with no local setup. */
-export declare function shareRuntimeEnv(opts?: {
-    agentOnly?: boolean;
-}): Record<string, string> | undefined;
+/** Best-effort runtime env for spawned agents. Never throws AND never prompts.
+ *
+ * Auto-injecting the share write token on every `agents run` is a background
+ * convenience, NOT a user-initiated secret access — so it MUST NOT raise a Touch
+ * ID sheet (SEC-13: an agent launch never pops biometry on its own). This was the
+ * per-run prompt storm: `share` is a keychain bundle that is rarely broker-held,
+ * so an interactive read here spawned the helper and popped Touch ID on EVERY
+ * launch. The read is now always `agentOnly` — it resolves the token only from the
+ * injected env or an already-held / no-ACL bundle, and silently returns undefined
+ * otherwise (the caller runs without auto-share; the agent can still publish via
+ * its own `agents share`). To get zero-friction auto-share with no prompt: unlock
+ * once (`agents secrets unlock share`) or make it no-ACL (`agents secrets policy
+ * share never`). */
+export declare function shareRuntimeEnv(): Record<string, string> | undefined;
 /** Cloudflare API credentials for provisioning, read from `cloudflare` (or a
  * user-named bundle). Fuzzy-matches key names so it works across bundle layouts. */
 export declare function readCloudflareCreds(bundle?: string, override?: {

package/dist/lib/share/config.js

@@ -56,6 +56,14 @@ export function storeWriteToken(token) {
         bundle = {
             name: SHARE_BUNDLE,
             description: 'agents share — write token for the R2 share endpoint',
+            // A NEW share bundle defaults to the `never` tier (no biometry ACL). The R2
+            // write token is low-sensitivity automation infra that is auto-read on EVERY
+            // `agents run` (shareRuntimeEnv) — a biometry ACL there is what produced the
+            // per-run Touch ID storm. `never` stores it no-ACL so auto-share is silent
+            // and needs no unlock. An EXISTING bundle keeps its tier (we never silently
+            // downgrade one the user already made); change it explicitly with
+            // `agents secrets policy share <tier>`.
+            policy: 'never',
             vars: {},
         };
     }
@@ -82,10 +90,20 @@ export function readWriteTokenFromBundle() {
 export function readWriteToken() {
     return readWriteTokenEnv() ?? readWriteTokenFromBundle();
 }
-/** Best-effort runtime env for spawned agents. Never throws: a missing/locked
- * bundle should not block unrelated agent runs, but an already-unlocked bundle
- * or injected token lets ephemeral agents publish with no local setup. */
-export function shareRuntimeEnv(opts = {}) {
+/** Best-effort runtime env for spawned agents. Never throws AND never prompts.
+ *
+ * Auto-injecting the share write token on every `agents run` is a background
+ * convenience, NOT a user-initiated secret access — so it MUST NOT raise a Touch
+ * ID sheet (SEC-13: an agent launch never pops biometry on its own). This was the
+ * per-run prompt storm: `share` is a keychain bundle that is rarely broker-held,
+ * so an interactive read here spawned the helper and popped Touch ID on EVERY
+ * launch. The read is now always `agentOnly` — it resolves the token only from the
+ * injected env or an already-held / no-ACL bundle, and silently returns undefined
+ * otherwise (the caller runs without auto-share; the agent can still publish via
+ * its own `agents share`). To get zero-friction auto-share with no prompt: unlock
+ * once (`agents secrets unlock share`) or make it no-ACL (`agents secrets policy
+ * share never`). */
+export function shareRuntimeEnv() {
     if (!readShareConfig())
         return undefined;
     const fromEnv = readWriteTokenEnv();
@@ -97,7 +115,7 @@ export function shareRuntimeEnv(opts = {}) {
         const { env } = readAndResolveBundleEnv(SHARE_BUNDLE, {
             caller: 'share',
             keys: [SHARE_TOKEN_KEY],
-            agentOnly: opts.agentOnly,
+            agentOnly: true, // never raise a Touch ID sheet on an agent launch (SEC-13)
         });
         const token = env[SHARE_TOKEN_KEY];
         return token ? { [SHARE_TOKEN_ENV_KEY]: token } : undefined;

package/dist/lib/types.d.ts

@@ -764,6 +764,16 @@ export interface Meta {
      */
     isolatedAgents?: Partial<Record<AgentId, string>>;
     run?: RunConfig;
+    /**
+     * Daemon watchdog config. `rotate` (default `on`) lets the watchdog rotate a
+     * rate-limited session IN PLACE onto a healthy account/harness via
+     * `agents run auto` — see lib/watchdog/rotate.ts. Set `off` to keep the
+     * nudge-only behavior (the Factory `agents.watchdog.autoRotate: false`
+     * migration writes `off` here).
+     */
+    watchdog?: {
+        rotate?: 'on' | 'off';
+    };
     /**
      * `agents run --lease` config. `secretsBundle` names the keychain secrets bundle
      * whose provider token (e.g. `HCLOUD_TOKEN`) crabbox uses to reach the cloud API.

package/dist/lib/usage.js

@@ -1246,7 +1246,9 @@ function claudeOauthCacheActive() {
  * or the token itself has expired — in which case the stale entry is dropped. */
 function readCachedClaudeOauth(service) {
     try {
-        const entry = JSON.parse(getKeychainToken(claudeOauthCacheItem(service)));
+        // The cache item is written no-ACL (writeCachedClaudeOauth) — its whole purpose is
+        // serving the token prompt-free, so attest that to the raw-read storm guard.
+        const entry = JSON.parse(getKeychainToken(claudeOauthCacheItem(service), { silentNoAcl: true }));
         if (!entry || typeof entry.accessToken !== 'string' || !entry.accessToken)
             return null;
         const now = Date.now();

package/dist/lib/watchdog/rotate.d.ts

@@ -0,0 +1,218 @@
+/**
+ * Watchdog rotate — in-place rotation of a rate-limited session onto a healthy
+ * account/harness, inside the SAME terminal tab (one-watchdog; follow-up to
+ * RUSH-2132 / PR #1875).
+ *
+ * When a stalled session's transcript tail shows a hard limit ("You've hit your
+ * weekly limit · resets …"), the daemon watchdog rotates it instead of nudging:
+ *
+ *   1. DETECT  — classifyTailForRotate() matches the tail against the limit
+ *      patterns (ported from apps/factory/src/core/autoRotate.ts) and parses the
+ *      `resets <time>` clause when present.
+ *   2. GATE    — defaultRotateGate() runs the SAME first-party selection
+ *      `agents run auto` would (collectHarnessCandidates + pickHarnessWeighted,
+ *      ../rotate.ts). Zero healthy → ONE `rotate` skip event per cooldown window
+ *      and the terminal is left untouched. No `agents view` subprocess anywhere.
+ *   3. RELAUNCH — the per-harness exit sequence (ported from apps/factory
+ *      prewarm.ts PREWARM_CONFIGS) is injected, then
+ *      `agents run auto --interactive --session-id <uuid>`.
+ *   4. REPLAY   — when the new session's TUI is live (bounded wait, default
+ *      60s), the resume replay is injected. On timeout the session is flagged
+ *      and the machine stops — never blind-type into a dead shell.
+ *
+ * The machine spans ticks (the exit sequence kills the old session, so it drops
+ * out of the active-session list before the new TUI is live): state persists at
+ * <watchdog-state>/rotate/<sessionId>.json as
+ * exiting → launching → awaiting-tui → replaying → done | failed.
+ *
+ * Config: `watchdog.rotate: on|off` in agents.yaml (default on), read per tick.
+ */
+import type { InjectTarget } from '../terminal/inject.js';
+import type { ActiveSession } from '../session/active.js';
+/**
+ * Agent-reported hard-limit texts, matched against a session transcript tail.
+ * Ported verbatim from apps/factory/src/core/autoRotate.ts RATE_LIMIT_PATTERNS —
+ * kept specific on purpose: a transcript carries prose, so a loose "rate limit"
+ * match would rotate terminals whose agent merely DISCUSSED limits. The first
+ * two patterns cover the weekly/session variants, including claude's
+ * "You've hit your weekly limit · resets <time>" form.
+ */
+export declare const ROTATE_LIMIT_PATTERNS: RegExp[];
+export type RotateTailVerdict = {
+    kind: 'none';
+} | {
+    kind: 'rate_limited';
+    resetsAtMs?: number;
+};
+/**
+ * Classify a transcript tail for the rotate decision: does it show a hard
+ * account limit (rotate this session) or not (leave it to the nudge path)?
+ * Unlike the retired extension path there is NO `no healthy` tail parsing here —
+ * the health gate is a first-party function call (defaultRotateGate), not a
+ * cross-package string contract.
+ */
+export declare function classifyTailForRotate(tailLines: string[], nowMs: number): RotateTailVerdict;
+/**
+ * Parse the `resets <time>` clause of a limit line into an epoch-ms horizon.
+ * Ported from apps/factory/src/core/autoRotate.ts parseResetTimeMs (behavior
+ * verbatim): the ISO form (milliseconds + Z) is matched EXPLICITLY and first —
+ * a generic capture stops at the milliseconds dot and drops the Z, which makes
+ * Date.parse read LOCAL time (the suppression would end hours off). Time-of-day
+ * forms like `7am` / `7:30pm` with an optional `(Area/City)` IANA zone cover
+ * claude's own limit text. Returns undefined when no usable reset is present or
+ * the parsed time is already past (caller falls back to its default cooldown).
+ */
+export declare function parseRotateResetMs(text: string, nowMs: number): number | undefined;
+/**
+ * Clean-exit key sequences per harness, ported verbatim from apps/factory
+ * prewarm.ts PREWARM_CONFIGS. Injected as RAW BYTES with no trailing Enter — a
+ * literal \x03 written to the pty IS Ctrl+C (SIGINT), \x1b IS Esc. claude's Ink
+ * TUI needs the Esc first to leave any open mode before the interrupt pair.
+ */
+export declare const ROTATE_EXIT_SEQUENCES: Record<string, string[]>;
+/** Unknown harnesses get the common denominator: Ctrl+C twice. */
+export declare const DEFAULT_ROTATE_EXIT_SEQUENCE: string[];
+export declare function exitSequenceFor(agent: string): string[];
+/**
+ * The rotate relaunch, typed into the same tab: full auto — the CLI resolves
+ * host (affinity) → harness (cross-harness headroom) → account (balanced) and
+ * exits nonzero when every layer is exhausted. Ported from apps/factory
+ * autoRotate.ts buildAutoRotateLaunchCommand. A terminal on a REMOTE device
+ * rotates ON that device (`--host`); a local terminal omits it. `--session-id`
+ * is honored only when the CLI picks claude (existing claude-only semantics)
+ * and ignored otherwise — passing it unconditionally keeps the terminal's
+ * AGENT_SESSION_ID aligned with the session Claude actually creates.
+ */
+export declare function buildRotateLaunchCommand(opts: {
+    host?: string;
+    sessionId: string;
+}): string;
+/**
+ * The harness-agnostic replay injected once the new TUI is live: load the OLD
+ * session's transcript, assess, continue. This is the same instruction shape
+ * the CLI's own `continue` flow uses.
+ */
+export declare function buildRotateReplayText(oldSessionId: string): string;
+export type RotatePhase = 'exiting' | 'launching' | 'awaiting-tui' | 'replaying' | 'done' | 'failed';
+/** Persisted at <watchdog-state>/rotate/<sessionId>.json — keyed by the OLD session id. */
+export interface RotateState {
+    /** The OLD (rate-limited) session id — the file key and the replay target. */
+    sessionId: string;
+    /** The id passed to `--session-id` on the relaunch. */
+    newSessionId: string;
+    /** The harness that was rate-limited (drives the exit-sequence table). */
+    agent: string;
+    phase: RotatePhase;
+    /** The resolved inject target — serializable, so the sweep can replay without re-resolving. */
+    target: InjectTarget;
+    /** Remote device the terminal lives on, when provenance says ssh. */
+    host?: string;
+    /**
+     * The old session's cwd — correlates the readiness fallback: a fresh active
+     * session only counts as the relaunched TUI when it runs in the SAME project.
+     */
+    cwd?: string;
+    /**
+     * The machine the old session runs on (provenance host = os.hostname()) —
+     * the second half of the readiness-fallback correlation, so a fresh session
+     * on ANOTHER box never satisfies it.
+     */
+    machineHost?: string;
+    startedAtMs: number;
+    updatedAtMs: number;
+    /** awaiting-tui deadline: startedAtMs + readiness budget. */
+    deadlineMs: number;
+    error?: string;
+    /**
+     * Set on the transition to `failed`: the tick will not re-begin a rotate for
+     * this session until then (default +15m). Without it a session whose old TUI
+     * ignored the exit sequence re-enters begin → deadline → failed every tick.
+     */
+    suppressUntilMs?: number;
+}
+/** Bounded wait for the relaunched TUI to come live (readiness). */
+export declare const DEFAULT_ROTATE_READINESS_MS = 60000;
+/** Zero-healthy skip cooldown when neither the gate nor the tail carries a reset. */
+export declare const DEFAULT_ROTATE_SKIP_COOLDOWN_MS: number;
+/** Retry cooldown after a FAILED rotate — honored at begin via the state file. */
+export declare const DEFAULT_ROTATE_FAILED_COOLDOWN_MS: number;
+export declare function rotateStatePath(dir: string, sessionId: string): string;
+export declare function readRotateState(dir: string, sessionId: string): RotateState | null;
+export declare function writeRotateState(dir: string, state: RotateState): void;
+export declare function clearRotateState(dir: string, sessionId: string): void;
+/** A phase the machine still has work to do in (done/failed are terminal). */
+export declare function isInflightPhase(phase: RotatePhase): boolean;
+/** Every persisted rotate state (any phase) — for `watchdog status`. */
+export declare function listRotateStates(dir: string): RotateState[];
+/** In-flight rotates only — the set a tick's sweep must advance. */
+export declare function listInflightRotates(dir: string): RotateState[];
+/**
+ * One `rotate` skip event per cooldown window, tracked as
+ * <watchdog-state>/rotate-skips.json: { [sessionId]: suppressUntilMs }. A skip
+ * inside the window logs nothing and touches nothing.
+ */
+export declare function readRotateSkipLedger(dir: string): Record<string, number>;
+/** True when a skip for this session is OUTSIDE its suppression window (log it). */
+export declare function shouldLogRotateSkip(dir: string, sessionId: string, nowMs: number): boolean;
+/** Suppress further skip events for this session until suppressUntilMs. */
+export declare function recordRotateSkip(dir: string, sessionId: string, suppressUntilMs: number): void;
+/**
+ * `watchdog.rotate` in agents.yaml (default ON — it is safe now: the health gate
+ * is first-party and the readiness wait is bounded). Read fresh per tick so a
+ * flip mid-run is honored on the next pass.
+ */
+export declare function isWatchdogRotateEnabled(): boolean;
+/**
+ * Persist `watchdog.rotate: on|off`. Called by the `agents watchdog rotate
+ * on|off` subcommand (commands/watchdog.ts) — the rotate-only switch the
+ * Factory migration uses so a user who opted out of autoRotate keeps nudging
+ * (rather than `agents watchdog disable`, which kills the whole watchdog).
+ */
+export declare function setWatchdogRotateEnabled(on: boolean): void;
+export interface RotateGateResult {
+    /** True when at least one harness has a healthy account to rotate INTO. */
+    healthy: boolean;
+    /** Earliest future window reset across all candidates, when any snapshot carries one. */
+    resetsAtMs?: number;
+    /** Human detail for the skip event (the zero-healthy error text). */
+    detail: string;
+}
+/**
+ * The first-party health gate: run the SAME selection `agents run auto` would —
+ * collectHarnessCandidates over every installed harness, pickHarnessWeighted.
+ * Zero healthy → the caller suppresses rotation until earliestResetAcross (or
+ * the parsed tail reset, or the default cooldown) and leaves the terminal alone.
+ * No `agents view` subprocess, no Keychain probe: collection is cache-only
+ * (collectRunCandidates reads daemon-written snapshots, readOnly).
+ */
+export declare function defaultRotateGate(): Promise<RotateGateResult>;
+/**
+ * Default TUI-liveness probe for the relaunched session: the new session's
+ * transcript resolves under any known harness layout. `--session-id` is honored
+ * on a claude pick; for other harnesses the runner's readiness check ALSO
+ * accepts a fresh active session (started after the rotate began), so a codex
+ * pick with an unknown id is still detected.
+ */
+export declare function defaultRotateTranscriptLive(newSessionId: string): boolean;
+/**
+ * The readiness FALLBACK correlation. A fresh active session counts as the
+ * relaunched TUI only when ALL of these hold:
+ *   - it is not the old session and started at/after the rotate began;
+ *   - it runs in the SAME cwd (trailing-slash normalized); and
+ *   - it runs on the SAME machine (provenance host = os.hostname(), the same
+ *     field provenance.ts populates).
+ * An unrelated fresh session — another project, another host, a remote
+ * teammate — must NEVER satisfy readiness: on a busy fleet box an
+ * uncorrelated "any new session" match fires on the first sweep regardless of
+ * whether the relaunch came up, and when `agents run auto` failed loud after
+ * the gate that types the replay into a bare shell. When the state lacks cwd
+ * or host the fallback cannot correlate and only the transcript probe counts.
+ */
+export declare function isCorrelatedRelaunch(state: RotateState, s: ActiveSession): boolean;
+/**
+ * The default TUI-liveness probe. The new-session-id transcript is PRIMARY (a
+ * claude pick honors `--session-id`); the correlated fresh-session fallback
+ * (isCorrelatedRelaunch) covers non-claude picks whose id we can't know a
+ * priori.
+ */
+export declare function defaultTuiLiveFor(state: RotateState, sessions: ActiveSession[]): boolean;