@songsid/agend 2.1.4-beta.13 → 2.1.4-beta.15

package/dist/fleet-manager.d.ts

@@ -36,6 +36,25 @@ export declare function selectLruEvictions(warm: string[], cap: number, opts: {
     isIdle: (name: string) => boolean;
     lastInboundAt: (name: string) => number;
 }): string[];
+/**
+ * Answer shape for `list_models`. `scope` reports where the LIST came from —
+ * "instance" only when it was read through that instance's own backend config,
+ * "global" for the account/CLI catalog — so a caller can tell an authoritative
+ * per-instance list from a best-effort account-wide one.
+ */
+export interface ModelCatalog {
+    backend: string;
+    scope: "instance" | "global";
+    /** Set whenever an instance was asked about, even if the list is global. */
+    instance?: string;
+    current_model: string | null;
+    models: import("./backend/types.js").ModelOption[];
+    /** "cache" = startup probe cache, "live" = probed now, "fallback" = none available. */
+    source: "cache" | "live" | "fallback";
+    probed_at?: string;
+    /** Caveat the caller should read before trusting the list. */
+    note?: string;
+}
 export interface DeliveryOptions {
     /** Explicitly identify agent-to-agent delivery when metadata is unavailable. */
     isCrossInstance?: boolean;
@@ -270,6 +289,13 @@ export declare class FleetManager implements FleetContext, LifecycleContext, Arc
      */
     private enforceWarmCap;
     private waitForInstanceIdle;
+    /**
+     * Ask the daemon to capture the pane before answering, then wait for any state
+     * report produced after this request. The ordinary query is intentionally
+     * cache-only (idle gates call it frequently); this opt-in refresh is reserved
+     * for lifecycle decisions where a stale state would strand UI.
+     */
+    private refreshInstanceExecutionState;
     private deliverWithIdleGate;
     /**
      * Hand a payload to an instance's IPC, waiting out a *transient* disconnect.
@@ -693,10 +719,14 @@ export declare class FleetManager implements FleetContext, LifecycleContext, Arc
     private startProgressTicker;
     /**
      * After a reply: give the instance REPLY_RETIRE_GRACE_MS to resume working; if
-     * it has not, retire its button. Re-arming replaces the previous timer, so a
-     * burst of replies ends with exactly one pending check.
+     * it has not, retire its button. The daemon is asked for a fresh pane capture
+     * before deciding. Reading only the transition cache stranded the first
+     * post-restart bubble when its startup "working" report never got a matching
+     * idle edge. Re-arming replaces the previous timer, so a burst of replies ends
+     * with exactly one pending check.
      */
     private armReplyGrace;
+    private finishReplyGrace;
     /** Retire (delete) every cancel button belonging to an instance. */
     private retireInstanceButtons;
     /** Begin retiring one button (delete + bounded retry on failure). Idempotent:
@@ -848,6 +878,34 @@ export declare class FleetManager implements FleetContext, LifecycleContext, Arc
     private probeCliEnvs;
     /** Best-effort model list for `/model`: cached CLI env first, else live probe. Never throws. */
     private getModelOptions;
+    /**
+     * Model catalog behind the `list_models` tool.
+     *
+     * The two scopes are not cosmetic. "global" is the account/CLI catalog served
+     * from the startup probe cache; "instance" is resolved through that instance's
+     * OWN backend config, and for a Codex instance on a custom provider that is a
+     * different catalog entirely — `listModels()` reads models_cache.json out of
+     * the instance's private CODEX_HOME. Answering such an instance with the
+     * account list would name models its CLI rejects, which is exactly the
+     * mistake this tool exists to prevent.
+     *
+     * `scope` always describes where the returned LIST came from, not what was
+     * asked for: an instance query that falls back to the account catalog reports
+     * scope "global" and says so in `note`, rather than implying instance-level
+     * accuracy it does not have.
+     *
+     * Never throws — a model listing is an aid, and failing it must not fail a turn.
+     */
+    listModelCatalog(opts?: {
+        backend?: string;
+        instanceName?: string;
+    }): Promise<ModelCatalog>;
+    /** The custom provider an instance overrides its backend with, if any. */
+    private customProviderFor;
+    /** Ask a backend for its catalog using ONE instance's real config. Never throws. */
+    private instanceScopedModels;
+    /** Account-wide catalog: probe cache first, live probe on miss. */
+    private globalModelCatalog;
     /** `/model` slash handler (admin only). No arg → DC menu; `/model <name>` → apply directly. */
     /** Label an effort choice, marking the one currently configured. */
     private effortChoiceLabel;

package/dist/fleet-manager.js

@@ -107,6 +107,8 @@ const CANCEL_BTN_IDLE_RETIRE_GRACE_MS = 2_000;
  * button alone (the idle edge retires it when the run really ends).
  */
 const REPLY_RETIRE_GRACE_MS = 2 * 60_000;
+/** Bound for the daemon to capture the pane and answer a post-reply state query. */
+const REPLY_STATE_REFRESH_TIMEOUT_MS = 2_000;
 /**
  * The daemon only broadcasts execution state on TRANSITIONS, so a long
  * single-state run sends nothing for hours. The idle backstop therefore pokes a
@@ -899,6 +901,48 @@ export class FleetManager {
             query();
         });
     }
+    /**
+     * Ask the daemon to capture the pane before answering, then wait for any state
+     * report produced after this request. The ordinary query is intentionally
+     * cache-only (idle gates call it frequently); this opt-in refresh is reserved
+     * for lifecycle decisions where a stale state would strand UI.
+     */
+    refreshInstanceExecutionState(instanceName, timeoutMs) {
+        const ipc = this.instanceIpcClients.get(instanceName);
+        if (!ipc?.connected)
+            return Promise.resolve(false);
+        const previous = this.instanceStateCache.get(instanceName);
+        return new Promise(resolve => {
+            let settled = false;
+            const finish = (refreshed) => {
+                if (settled)
+                    return;
+                settled = true;
+                clearTimeout(timeout);
+                const waiters = this.instanceIdleWaiters.get(instanceName);
+                waiters?.delete(check);
+                if (waiters?.size === 0)
+                    this.instanceIdleWaiters.delete(instanceName);
+                resolve(refreshed);
+            };
+            const check = () => {
+                if (this.instanceStateCache.get(instanceName) !== previous)
+                    finish(true);
+            };
+            const waiters = this.instanceIdleWaiters.get(instanceName) ?? new Set();
+            waiters.add(check);
+            this.instanceIdleWaiters.set(instanceName, waiters);
+            const timeout = setTimeout(() => finish(false), timeoutMs);
+            timeout.unref?.();
+            const sent = ipc.send({
+                type: "query_instance_state",
+                requestId: `reply-grace-${Date.now()}`,
+                refresh: true,
+            });
+            if (!sent)
+                finish(false);
+        });
+    }
     async deliverWithIdleGate(instanceName, payload, timeoutMs) {
         let idleObservedAfter = this.lastDeliveryAt.get(instanceName) ?? 0;
         if (this.lifecycle.isPaused(instanceName)) {
@@ -5424,8 +5468,11 @@ export class FleetManager {
     }
     /**
      * After a reply: give the instance REPLY_RETIRE_GRACE_MS to resume working; if
-     * it has not, retire its button. Re-arming replaces the previous timer, so a
-     * burst of replies ends with exactly one pending check.
+     * it has not, retire its button. The daemon is asked for a fresh pane capture
+     * before deciding. Reading only the transition cache stranded the first
+     * post-restart bubble when its startup "working" report never got a matching
+     * idle edge. Re-arming replaces the previous timer, so a burst of replies ends
+     * with exactly one pending check.
      */
     armReplyGrace(instanceName) {
         for (const entry of this.cancelButtons.values()) {
@@ -5437,14 +5484,26 @@ export class FleetManager {
                 entry.replyGraceTimer = undefined;
                 if (!this.cancelButtons.has(entry.messageId))
                     return;
-                if (!this.getInstanceIdle(instanceName))
-                    return; // resumed — a long run keeps its button
-                this.logger.info({ instanceName, messageId: entry.messageId }, "Cancel button retired — no work resumed after reply");
-                this.retireButton(entry);
+                void this.finishReplyGrace(instanceName, entry);
             }, REPLY_RETIRE_GRACE_MS);
             entry.replyGraceTimer.unref?.();
         }
     }
+    async finishReplyGrace(instanceName, entry) {
+        const refreshed = await this.refreshInstanceExecutionState(instanceName, REPLY_STATE_REFRESH_TIMEOUT_MS);
+        if (!this.cancelButtons.has(entry.messageId) || entry.retiring)
+            return;
+        if (!refreshed) {
+            // Fail safe: without an authoritative answer, retain a potentially live
+            // Cancel button. The 5-minute/30-minute/24-hour safety nets still apply.
+            this.logger.debug({ instanceName, messageId: entry.messageId }, "Cancel reply-grace state refresh timed out");
+            return;
+        }
+        if (!this.getInstanceIdle(instanceName))
+            return; // genuine long run keeps its button
+        this.logger.info({ instanceName, messageId: entry.messageId }, "Cancel button retired — no work resumed after reply");
+        this.retireButton(entry);
+    }
     /** Retire (delete) every cancel button belonging to an instance. */
     retireInstanceButtons(instanceName) {
         // Snapshot first — retireButton may delete entries from the map on success.
@@ -6457,6 +6516,17 @@ Plus the operational skills (fleet-health, instance-lifecycle, scheduling, sessi
                 return null;
             const probed = await be.probeCLIEnv({ workingDirectory: "", instanceDir: join(getAgendHome(), "cli-env"), instanceName: `probe-${backend}`, mcpServers: {} });
             const env = { backend, probedAt: Date.now(), ...probed };
+            // An empty result must never overwrite a catalog we already have. Some
+            // probes hit the network (`agy models` fetches, 5s cap), so a slow moment
+            // returns [] — and writing that would blank the list for the whole 24h
+            // TTL, long after the CLI recovered. Observed live: a good 11-model
+            // antigravity cache replaced by an empty one. Keep the known models and
+            // let the fresher currentModel/version through.
+            if (!env.models?.length) {
+                const previous = this.readCliEnv(backend);
+                if (previous?.models?.length)
+                    env.models = previous.models;
+            }
             const path = this.cliEnvPath(backend);
             mkdirSync(dirname(path), { recursive: true });
             writeFileSync(path, JSON.stringify(env, null, 2));
@@ -6495,6 +6565,108 @@ Plus the operational skills (fleet-health, instance-lifecycle, scheduling, sessi
         const env = await this.probeBackend(backendName);
         return env?.models ?? [];
     }
+    /**
+     * Model catalog behind the `list_models` tool.
+     *
+     * The two scopes are not cosmetic. "global" is the account/CLI catalog served
+     * from the startup probe cache; "instance" is resolved through that instance's
+     * OWN backend config, and for a Codex instance on a custom provider that is a
+     * different catalog entirely — `listModels()` reads models_cache.json out of
+     * the instance's private CODEX_HOME. Answering such an instance with the
+     * account list would name models its CLI rejects, which is exactly the
+     * mistake this tool exists to prevent.
+     *
+     * `scope` always describes where the returned LIST came from, not what was
+     * asked for: an instance query that falls back to the account catalog reports
+     * scope "global" and says so in `note`, rather than implying instance-level
+     * accuracy it does not have.
+     *
+     * Never throws — a model listing is an aid, and failing it must not fail a turn.
+     */
+    async listModelCatalog(opts = {}) {
+        const { instanceName } = opts;
+        if (instanceName) {
+            const backend = this.backendNameForInstance(instanceName);
+            const resolved = this.resolveInstanceModel(instanceName);
+            const currentModel = resolved.source === "unresolved" ? null : resolved.model;
+            const provider = this.customProviderFor(instanceName, backend);
+            const scoped = await this.instanceScopedModels(instanceName, backend);
+            if (scoped.length) {
+                return {
+                    backend, scope: "instance", instance: instanceName,
+                    current_model: currentModel, models: scoped, source: "live",
+                    ...(provider ? { note: `Catalog read through this instance's ${backend} provider "${provider}" — it may differ from the account catalog.` } : {}),
+                };
+            }
+            // No instance-local catalog (never launched, or the backend has no
+            // per-instance list). The account catalog is the best available answer,
+            // but it is labelled honestly rather than dressed up as instance scope.
+            const global = await this.globalModelCatalog(backend);
+            return {
+                ...global, instance: instanceName, current_model: currentModel,
+                note: provider
+                    ? `No instance-local catalog yet; showing the account catalog, which may NOT match this instance's ${backend} provider "${provider}".`
+                    : "No instance-local catalog yet; showing the account catalog.",
+            };
+        }
+        return this.globalModelCatalog(opts.backend ?? this.fleetConfig?.defaults?.backend ?? "claude-code");
+    }
+    /** The custom provider an instance overrides its backend with, if any. */
+    customProviderFor(instanceName, backend) {
+        const opts = this.fleetConfig?.instances?.[instanceName]?.backend_options?.[backend]
+            ?? this.fleetConfig?.defaults?.backend_options?.[backend];
+        const provider = opts?.provider;
+        return typeof provider === "string" && provider.trim() ? provider.trim() : null;
+    }
+    /** Ask a backend for its catalog using ONE instance's real config. Never throws. */
+    async instanceScopedModels(instanceName, backend) {
+        try {
+            const inst = this.fleetConfig?.instances?.[instanceName];
+            const instanceDir = this.getInstanceDir(instanceName);
+            const be = createBackend(backend, instanceDir);
+            if (!be.listModels)
+                return [];
+            return await be.listModels({
+                workingDirectory: inst?.working_directory ?? "",
+                instanceDir,
+                instanceName,
+                mcpServers: {},
+                model: inst?.model,
+                backendOptions: inst?.backend_options?.[backend] ?? this.fleetConfig?.defaults?.backend_options?.[backend],
+            }) ?? [];
+        }
+        catch {
+            // listModels is documented never to throw, but a backend constructor can
+            // (missing binary). A catalog is an aid; degrade to the account list.
+            return [];
+        }
+    }
+    /** Account-wide catalog: probe cache first, live probe on miss. */
+    async globalModelCatalog(backend) {
+        const cached = this.readCliEnv(backend);
+        if (cached?.models?.length) {
+            return {
+                backend, scope: "global", current_model: cached.currentModel ?? null,
+                models: cached.models, source: "cache",
+                probed_at: new Date(cached.probedAt).toISOString(),
+            };
+        }
+        const env = await this.probeBackend(backend);
+        if (env?.models?.length) {
+            return {
+                backend, scope: "global", current_model: env.currentModel ?? null,
+                models: env.models, source: "live",
+                probed_at: new Date(env.probedAt).toISOString(),
+            };
+        }
+        // Reported rather than thrown: "we could not enumerate" is a useful answer,
+        // and the caller can still set a model by name (AgEnD passes it through).
+        return {
+            backend, scope: "global", current_model: env?.currentModel ?? null,
+            models: [], source: "fallback",
+            note: `Could not enumerate models for ${backend} (CLI missing, not logged in, or it offers no list). Model names are passed through to the CLI, so a known-good name still works.`,
+        };
+    }
     /** `/model` slash handler (admin only). No arg → DC menu; `/model <name>` → apply directly. */
     /** Label an effort choice, marking the one currently configured. */
     effortChoiceLabel(level, current) {