@parall/claude-agent 1.36.1 → 1.37.0

package/dist/index.js

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
 import * as os from 'node:os';
-import { ParallAgentGateway, createPlatformConfigManager, createLogger, createOtelLogger, childLogger, isPlatformManagedProfile, isPlatformModelOverride, resolveRuntimeModel, parseShutdownDeadlineMs, parseForkDeadlineMs, parseDispatchDeadlineMs, parseProviderConfig, clearAllProviderCreds, llmSource, initAgentTelemetry, } from '@parall/agent-core';
+import { ParallAgentGateway, createPlatformConfigManager, createLogger, createOtelLogger, childLogger, deriveModelIsPin, resolveRuntimeModel, parseShutdownDeadlineMs, parseForkDeadlineMs, parseDispatchDeadlineMs, parseProviderConfig, clearAllProviderCreds, llmSource, initAgentTelemetry, } from '@parall/agent-core';
 import { ApiError, ParallClient, ParallWs } from '@parall/sdk';
 import { buildClaudeRuntimeKey, contextFilePathForSession, resolveClaudeAgentConfig, resolveWsUrl, sessionStateFilePathForRuntime, stepIdFilePathForSession, } from './config.js';
 import { ClaudeCodeAdapter } from './dispatch.js';
@@ -81,26 +81,19 @@ async function main() {
             log: agentLog,
         });
         const platformDefaults = await configMgr.fetch();
-        let platformManaged = isPlatformManagedProfile(me.agent_profile);
-        // Model precedence: operator override (platform-managed) > env > server floor.
-        // The server delivers a model floor for Parall-proxy routes even when
-        // model_management=self, so the CLI never falls back to a built-in default
-        // the catalog might reject — but that floor must lose to an explicit env
-        // override (env > floor). A profile is an operator-override only when it is
-        // platform-managed AND not explicitly self-managed: this treats legacy hosted
-        // agents (model_management=null, resolved to platform server-side) as
-        // overrides, while excluding local self agents whose delivered model is a
-        // floor. effort stays an operator override (platform).
-        let platformModelOverride = isPlatformModelOverride(me.agent_profile);
+        // Model precedence: operator PIN (override) > env > server FLOOR. The server
+        // now says which it is via model_is_pin (deriveModelIsPin presence-gates the
+        // dual-read: old servers omit it → fall back to model_management). A PIN beats
+        // an explicit env override; a FLOOR loses to env. effort is delivered by the
+        // server only when pinned, so we no longer gate it on a local flag.
+        const platformModelOverride = deriveModelIsPin(platformDefaults, me.agent_profile);
         const resolvedModel = resolveRuntimeModel(platformModelOverride, platformDefaults.model, config.model);
-        const resolvedEffort = platformManaged
-            ? (platformDefaults.thinkingEffort ??
-                (process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || undefined))
-            : process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || undefined;
+        const resolvedEffort = platformDefaults.thinkingEffort ??
+            (process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || undefined);
         if (platformDefaults.model)
-            agentLog.info(`platform config: model=${platformDefaults.model} → resolved=${resolvedModel ?? 'cli-default'} (${platformModelOverride ? 'platform-managed' : 'floor/self'})`);
+            agentLog.info(`platform config: model=${platformDefaults.model} → resolved=${resolvedModel ?? 'cli-default'} (${platformModelOverride ? 'pin/override' : 'floor'})`);
         if (platformDefaults.thinkingEffort)
-            agentLog.info(`platform config: thinking_effort=${platformDefaults.thinkingEffort} (${platformModelOverride ? 'applied' : 'ignored, model_management=self'})`);
+            agentLog.info(`platform config: thinking_effort=${platformDefaults.thinkingEffort}`);
         const adapter = new ClaudeCodeAdapter({
             claudeBin: config.claudeBin,
             claudeHome: config.claudeHome,
@@ -125,6 +118,32 @@ async function main() {
         if (resolvedEffort) {
             adapter.updateConfig({ effort: resolvedEffort });
         }
+        // Shared by onConfigUpdate + onSessionReady. /agents/me is only consumed as
+        // deriveModelIsPin's legacy fallback when the server omits model_is_pin
+        // (old server) — skip that round trip otherwise.
+        const refreshPlatformConfig = async () => {
+            const updated = await configMgr.fetch();
+            let refreshedProfile;
+            if (updated.modelIsPin === undefined) {
+                try {
+                    refreshedProfile = (await getAgentMeWithLegacyFallback(client, config.orgId))
+                        .agent_profile;
+                }
+                catch (err) {
+                    // configMgr.fetch() degrades to cache on failure; mirror that for the
+                    // legacy fallback so a transient /agents/me error doesn't reject this
+                    // config-refresh callback. No profile → deriveModelIsPin treats the
+                    // delivered model as a floor (safe default).
+                    agentLog.warn(`platform config refresh: legacy /agents/me fallback failed: ${String(err)}`);
+                }
+            }
+            const isPin = deriveModelIsPin(updated, refreshedProfile);
+            const localEffort = process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || null;
+            adapter.updateConfig({
+                model: resolveRuntimeModel(isPin, updated.model, config.model) ?? null,
+                effort: updated.thinkingEffort ?? localEffort,
+            });
+        };
         const gateway = new ParallAgentGateway({
             accountId: agentUserId,
             client,
@@ -151,28 +170,8 @@ async function main() {
             dispatchDeadlineMs: parseDispatchDeadlineMs(process.env.PRLL_DISPATCH_DEADLINE_MS),
             contextFilePathForSession: (sessionKey) => contextFilePathForSession(config.stateDir, sessionKey),
             stepIdFilePathForSession: (sessionKey) => stepIdFilePathForSession(config.stateDir, sessionKey),
-            onConfigUpdate: async () => {
-                const refreshed = await getAgentMeWithLegacyFallback(client, config.orgId);
-                platformManaged = isPlatformManagedProfile(refreshed.agent_profile);
-                platformModelOverride = isPlatformModelOverride(refreshed.agent_profile);
-                const updated = await configMgr.fetch();
-                const localEffort = process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || null;
-                adapter.updateConfig({
-                    model: resolveRuntimeModel(platformModelOverride, updated.model, config.model) ?? null,
-                    effort: platformManaged ? (updated.thinkingEffort ?? localEffort) : localEffort,
-                });
-            },
-            onSessionReady: async () => {
-                const refreshed = await getAgentMeWithLegacyFallback(client, config.orgId);
-                platformManaged = isPlatformManagedProfile(refreshed.agent_profile);
-                platformModelOverride = isPlatformModelOverride(refreshed.agent_profile);
-                const updated = await configMgr.fetch();
-                const localEffort = process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || null;
-                adapter.updateConfig({
-                    model: resolveRuntimeModel(platformModelOverride, updated.model, config.model) ?? null,
-                    effort: platformManaged ? (updated.thinkingEffort ?? localEffort) : localEffort,
-                });
-            },
+            onConfigUpdate: refreshPlatformConfig,
+            onSessionReady: refreshPlatformConfig,
             // Long-lived `claude` subprocesses outlive a single dispatch, so the
             // gateway's shuttingDown flag alone does not tear them down. Piggyback on
             // onBeforeDisconnect to close stdin and SIGTERM any survivors after

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@parall/claude-agent",
-  "version": "1.36.1",
+  "version": "1.37.0",
   "description": "Claude Code bridge runtime for self-hosted Parall agents",
   "license": "MIT",
   "repository": {
@@ -25,9 +25,9 @@
     "src"
   ],
   "dependencies": {
-    "@parall/agent-core": "1.36.1",
-    "@parall/cli": "1.36.1",
-    "@parall/sdk": "1.36.1"
+    "@parall/agent-core": "1.37.0",
+    "@parall/sdk": "1.37.0",
+    "@parall/cli": "1.37.0"
   },
   "devDependencies": {
     "@types/node": "^22.0.0",

package/src/index.ts

@@ -7,8 +7,7 @@ import {
   createLogger,
   createOtelLogger,
   childLogger,
-  isPlatformManagedProfile,
-  isPlatformModelOverride,
+  deriveModelIsPin,
   resolveRuntimeModel,
   parseShutdownDeadlineMs,
   parseForkDeadlineMs,
@@ -112,34 +111,26 @@ async function main() {
       log: agentLog,
     });
     const platformDefaults = await configMgr.fetch();
-    let platformManaged = isPlatformManagedProfile(me.agent_profile);
-    // Model precedence: operator override (platform-managed) > env > server floor.
-    // The server delivers a model floor for Parall-proxy routes even when
-    // model_management=self, so the CLI never falls back to a built-in default
-    // the catalog might reject — but that floor must lose to an explicit env
-    // override (env > floor). A profile is an operator-override only when it is
-    // platform-managed AND not explicitly self-managed: this treats legacy hosted
-    // agents (model_management=null, resolved to platform server-side) as
-    // overrides, while excluding local self agents whose delivered model is a
-    // floor. effort stays an operator override (platform).
-    let platformModelOverride = isPlatformModelOverride(me.agent_profile);
+    // Model precedence: operator PIN (override) > env > server FLOOR. The server
+    // now says which it is via model_is_pin (deriveModelIsPin presence-gates the
+    // dual-read: old servers omit it → fall back to model_management). A PIN beats
+    // an explicit env override; a FLOOR loses to env. effort is delivered by the
+    // server only when pinned, so we no longer gate it on a local flag.
+    const platformModelOverride = deriveModelIsPin(platformDefaults, me.agent_profile);
     const resolvedModel = resolveRuntimeModel(
       platformModelOverride,
       platformDefaults.model,
       config.model,
     );
-    const resolvedEffort = platformManaged
-      ? (platformDefaults.thinkingEffort ??
-        (process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || undefined))
-      : process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || undefined;
+    const resolvedEffort =
+      platformDefaults.thinkingEffort ??
+      (process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || undefined);
     if (platformDefaults.model)
       agentLog.info(
-        `platform config: model=${platformDefaults.model} → resolved=${resolvedModel ?? 'cli-default'} (${platformModelOverride ? 'platform-managed' : 'floor/self'})`,
+        `platform config: model=${platformDefaults.model} → resolved=${resolvedModel ?? 'cli-default'} (${platformModelOverride ? 'pin/override' : 'floor'})`,
       );
     if (platformDefaults.thinkingEffort)
-      agentLog.info(
-        `platform config: thinking_effort=${platformDefaults.thinkingEffort} (${platformModelOverride ? 'applied' : 'ignored, model_management=self'})`,
-      );
+      agentLog.info(`platform config: thinking_effort=${platformDefaults.thinkingEffort}`);
 
     const adapter = new ClaudeCodeAdapter({
       claudeBin: config.claudeBin,
@@ -168,6 +159,36 @@ async function main() {
       adapter.updateConfig({ effort: resolvedEffort });
     }
 
+    // Shared by onConfigUpdate + onSessionReady. /agents/me is only consumed as
+    // deriveModelIsPin's legacy fallback when the server omits model_is_pin
+    // (old server) — skip that round trip otherwise.
+    const refreshPlatformConfig = async () => {
+      const updated = await configMgr.fetch();
+      let refreshedProfile:
+        | Awaited<ReturnType<typeof getAgentMeWithLegacyFallback>>['agent_profile']
+        | undefined;
+      if (updated.modelIsPin === undefined) {
+        try {
+          refreshedProfile = (await getAgentMeWithLegacyFallback(client, config.orgId))
+            .agent_profile;
+        } catch (err) {
+          // configMgr.fetch() degrades to cache on failure; mirror that for the
+          // legacy fallback so a transient /agents/me error doesn't reject this
+          // config-refresh callback. No profile → deriveModelIsPin treats the
+          // delivered model as a floor (safe default).
+          agentLog.warn(
+            `platform config refresh: legacy /agents/me fallback failed: ${String(err)}`,
+          );
+        }
+      }
+      const isPin = deriveModelIsPin(updated, refreshedProfile);
+      const localEffort = process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || null;
+      adapter.updateConfig({
+        model: resolveRuntimeModel(isPin, updated.model, config.model) ?? null,
+        effort: updated.thinkingEffort ?? localEffort,
+      });
+    };
+
     const gateway = new ParallAgentGateway({
       accountId: agentUserId,
       client,
@@ -196,28 +217,8 @@ async function main() {
         contextFilePathForSession(config.stateDir, sessionKey),
       stepIdFilePathForSession: (sessionKey) =>
         stepIdFilePathForSession(config.stateDir, sessionKey),
-      onConfigUpdate: async () => {
-        const refreshed = await getAgentMeWithLegacyFallback(client, config.orgId);
-        platformManaged = isPlatformManagedProfile(refreshed.agent_profile);
-        platformModelOverride = isPlatformModelOverride(refreshed.agent_profile);
-        const updated = await configMgr.fetch();
-        const localEffort = process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || null;
-        adapter.updateConfig({
-          model: resolveRuntimeModel(platformModelOverride, updated.model, config.model) ?? null,
-          effort: platformManaged ? (updated.thinkingEffort ?? localEffort) : localEffort,
-        });
-      },
-      onSessionReady: async () => {
-        const refreshed = await getAgentMeWithLegacyFallback(client, config.orgId);
-        platformManaged = isPlatformManagedProfile(refreshed.agent_profile);
-        platformModelOverride = isPlatformModelOverride(refreshed.agent_profile);
-        const updated = await configMgr.fetch();
-        const localEffort = process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || null;
-        adapter.updateConfig({
-          model: resolveRuntimeModel(platformModelOverride, updated.model, config.model) ?? null,
-          effort: platformManaged ? (updated.thinkingEffort ?? localEffort) : localEffort,
-        });
-      },
+      onConfigUpdate: refreshPlatformConfig,
+      onSessionReady: refreshPlatformConfig,
       // Long-lived `claude` subprocesses outlive a single dispatch, so the
       // gateway's shuttingDown flag alone does not tear them down. Piggyback on
       // onBeforeDisconnect to close stdin and SIGTERM any survivors after