@parall/claude-agent 1.58.0 → 1.58.2

package/dist/dispatch.d.ts

@@ -10,6 +10,8 @@ type ClaudeCodeAdapterOptions = Pick<ClaudeAgentConfig, 'additionalDirs' | 'allo
     orgId: string;
     /** Catalog context size for the selected model. */
     contextWindow?: number;
+    /** Catalog max output tokens for the selected model (input-budget subtrahend). */
+    maxTokens?: number;
     contextFilePathForSession?: (sessionKey: string) => string;
     /** @deprecated Use contextFilePathForSession. */
     stepIdFilePathForSession?: (sessionKey: string) => string;
@@ -41,6 +43,7 @@ export declare class ClaudeCodeAdapter implements DispatchAdapter {
     private _model;
     private _effortLevel;
     private _contextWindow;
+    private _maxTokens;
     constructor(opts: ClaudeCodeAdapterOptions);
     get currentModel(): string | undefined;
     get currentEffort(): string | undefined;
@@ -48,6 +51,7 @@ export declare class ClaudeCodeAdapter implements DispatchAdapter {
         model?: string | null;
         effort?: string | null;
         contextWindow?: number | null;
+        maxTokens?: number | null;
     }): void;
     requestProcessRestart(): void;
     enqueueDuringDispatch(sessionKey: string, body: string, inputLifecycle?: DispatchInputLifecycle): boolean;

package/dist/dispatch.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"dispatch.d.ts","sourceRoot":"","sources":["../src/dispatch.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EACV,eAAe,EACf,eAAe,EACf,sBAAsB,EACtB,YAAY,EACZ,QAAQ,EAER,YAAY,EACb,MAAM,oBAAoB,CAAC;AAK5B,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAOrD,OAAO,KAAK,EAAuB,oBAAoB,EAAE,MAAM,sBAAsB,CAAC;AACtF,OAAO,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAG/C,OAAO,EAAE,aAAa,EAAE,CAAC;AAEzB,KAAK,wBAAwB,GAAG,IAAI,CAClC,iBAAiB,EACf,gBAAgB,GAChB,aAAa,GACb,cAAc,GACd,oBAAoB,GACpB,WAAW,GACX,YAAY,GACZ,iBAAiB,GACjB,OAAO,GACP,gBAAgB,GAChB,cAAc,CACjB,GAAG;IACF,cAAc,EAAE,oBAAoB,CAAC;IACrC,MAAM,EAAE,MAAM,CAAC;IACf,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,mDAAmD;IACnD,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,yBAAyB,CAAC,EAAE,CAAC,UAAU,EAAE,MAAM,KAAK,MAAM,CAAC;IAC3D,iDAAiD;IACjD,wBAAwB,CAAC,EAAE,CAAC,UAAU,EAAE,MAAM,KAAK,MAAM,CAAC;IAC1D;;;;;;;OAOG;IACH,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB;;;;;;OAMG;IACH,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC3B,CAAC;AAqCF,qBAAa,iBAAkB,YAAW,eAAe;IAU3C,OAAO,CAAC,QAAQ,CAAC,IAAI;IATjC,QAAQ,CAAC,kBAAkB,EAAG,UAAU,CAAU;IAClD,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAmC;IAC7D,OAAO,CAAC,eAAe,CAAC,CAAgB;IACxC,OAAO,CAAC,qBAAqB,CAAC,CAAsB;IACpD,OAAO,CAAC,YAAY,CAAS;IAC7B,OAAO,CAAC,MAAM,CAAqB;IACnC,OAAO,CAAC,YAAY,CAAqB;IACzC,OAAO,CAAC,cAAc,CAAqB;gBAEd,IAAI,EAAE,wBAAwB;IAK3D,IAAI,YAAY,IAAI,MAAM,GAAG,SAAS,CAErC;IACD,IAAI,aAAa,IAAI,MAAM,GAAG,SAAS,CAEtC;IACD,YAAY,CAAC,MAAM,EAAE;QACnB,KAAK,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QACtB,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QACvB,aAAa,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;KAC/B,GAAG,IAAI;IAmBR,qBAAqB,IAAI,IAAI;IAM7B,qBAAqB,CACnB,UAAU,EAAE,MAAM,EAClB,IAAI,EAAE,MAAM,EACZ,cAAc,CAAC,EAAE,sBAAsB,GACtC,OAAO;IA2BV,aAAa,CAAC,UAAU,EAAE,MAAM,GAAG,IAAI;IAcvC,oBAAoB,CAAC,UAAU,EAAE,MAAM,GAAG,OAAO;IAM1C,QAAQ,CAAC,EACd,KAAK,EACL,YAAY,EACZ,UAAU,EACV,OAAO,EACP,cAAc,EACd,YAAY,GACb,EAAE,YAAY,GAAG,aAAa,CAAC,YAAY,CAAC;IAiE7C,cAAc,CAAC,WAAW,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS;IAIvD,qBAAqB,CAAC,UAAU,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS;IAc7D,WAAW,CAAC,EAAE,UAAU,EAAE,EAAE,QAAQ;IAIpC,WAAW,CAAC,EAAE,IAAI,EAAE,EAAE,eAAe;IAQrC,cAAc,IAAI,IAAI;IAOhB,QAAQ,IAAI,OAAO,CAAC,IAAI,CAAC;YAUhB,OAAO;IAuCtB,OAAO,CAAC,uBAAuB;IAa/B;;;;;OAKG;YACW,sBAAsB;IAwCpC;;;;;OAKG;YACY,eAAe;IA4L9B,OAAO,CAAC,aAAa;IAmCrB,OAAO,CAAC,YAAY;IA6DpB,OAAO,CAAC,WAAW;IAYnB,OAAO,CAAC,eAAe;IAiBvB,OAAO,CAAC,oBAAoB;IAU5B,OAAO,CAAC,gBAAgB;IAaxB,OAAO,CAAC,SAAS;IAwCjB,OAAO,CAAC,uBAAuB;CAmBhC"}
+{"version":3,"file":"dispatch.d.ts","sourceRoot":"","sources":["../src/dispatch.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EACV,eAAe,EACf,eAAe,EACf,sBAAsB,EACtB,YAAY,EACZ,QAAQ,EAER,YAAY,EACb,MAAM,oBAAoB,CAAC;AAK5B,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAOrD,OAAO,KAAK,EAAuB,oBAAoB,EAAE,MAAM,sBAAsB,CAAC;AACtF,OAAO,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAG/C,OAAO,EAAE,aAAa,EAAE,CAAC;AAEzB,KAAK,wBAAwB,GAAG,IAAI,CAClC,iBAAiB,EACf,gBAAgB,GAChB,aAAa,GACb,cAAc,GACd,oBAAoB,GACpB,WAAW,GACX,YAAY,GACZ,iBAAiB,GACjB,OAAO,GACP,gBAAgB,GAChB,cAAc,CACjB,GAAG;IACF,cAAc,EAAE,oBAAoB,CAAC;IACrC,MAAM,EAAE,MAAM,CAAC;IACf,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,mDAAmD;IACnD,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,kFAAkF;IAClF,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,yBAAyB,CAAC,EAAE,CAAC,UAAU,EAAE,MAAM,KAAK,MAAM,CAAC;IAC3D,iDAAiD;IACjD,wBAAwB,CAAC,EAAE,CAAC,UAAU,EAAE,MAAM,KAAK,MAAM,CAAC;IAC1D;;;;;;;OAOG;IACH,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB;;;;;;OAMG;IACH,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC3B,CAAC;AAqCF,qBAAa,iBAAkB,YAAW,eAAe;IAW3C,OAAO,CAAC,QAAQ,CAAC,IAAI;IAVjC,QAAQ,CAAC,kBAAkB,EAAG,UAAU,CAAU;IAClD,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAmC;IAC7D,OAAO,CAAC,eAAe,CAAC,CAAgB;IACxC,OAAO,CAAC,qBAAqB,CAAC,CAAsB;IACpD,OAAO,CAAC,YAAY,CAAS;IAC7B,OAAO,CAAC,MAAM,CAAqB;IACnC,OAAO,CAAC,YAAY,CAAqB;IACzC,OAAO,CAAC,cAAc,CAAqB;IAC3C,OAAO,CAAC,UAAU,CAAqB;gBAEV,IAAI,EAAE,wBAAwB;IAM3D,IAAI,YAAY,IAAI,MAAM,GAAG,SAAS,CAErC;IACD,IAAI,aAAa,IAAI,MAAM,GAAG,SAAS,CAEtC;IACD,YAAY,CAAC,MAAM,EAAE;QACnB,KAAK,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QACtB,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QACvB,aAAa,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QAC9B,SAAS,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;KAC3B,GAAG,IAAI;IAqBR,qBAAqB,IAAI,IAAI;IAM7B,qBAAqB,CACnB,UAAU,EAAE,MAAM,EAClB,IAAI,EAAE,MAAM,EACZ,cAAc,CAAC,EAAE,sBAAsB,GACtC,OAAO;IA2BV,aAAa,CAAC,UAAU,EAAE,MAAM,GAAG,IAAI;IAcvC,oBAAoB,CAAC,UAAU,EAAE,MAAM,GAAG,OAAO;IAM1C,QAAQ,CAAC,EACd,KAAK,EACL,YAAY,EACZ,UAAU,EACV,OAAO,EACP,cAAc,EACd,YAAY,GACb,EAAE,YAAY,GAAG,aAAa,CAAC,YAAY,CAAC;IAiE7C,cAAc,CAAC,WAAW,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS;IAIvD,qBAAqB,CAAC,UAAU,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS;IAc7D,WAAW,CAAC,EAAE,UAAU,EAAE,EAAE,QAAQ;IAIpC,WAAW,CAAC,EAAE,IAAI,EAAE,EAAE,eAAe;IAQrC,cAAc,IAAI,IAAI;IAOhB,QAAQ,IAAI,OAAO,CAAC,IAAI,CAAC;YAUhB,OAAO;IAuCtB,OAAO,CAAC,uBAAuB;IAa/B;;;;;OAKG;YACW,sBAAsB;IAwCpC;;;;;OAKG;YACY,eAAe;IA4L9B,OAAO,CAAC,aAAa;IAmCrB,OAAO,CAAC,YAAY;IA8DpB,OAAO,CAAC,WAAW;IAYnB,OAAO,CAAC,eAAe;IAiBvB,OAAO,CAAC,oBAAoB;IAU5B,OAAO,CAAC,gBAAgB;IAaxB,OAAO,CAAC,SAAS;IAwCjB,OAAO,CAAC,uBAAuB;CAmBhC"}

package/dist/dispatch.js

@@ -36,10 +36,12 @@ export class ClaudeCodeAdapter {
     _model;
     _effortLevel;
     _contextWindow;
+    _maxTokens;
     constructor(opts) {
         this.opts = opts;
         this._model = opts.model;
         this._contextWindow = opts.contextWindow;
+        this._maxTokens = opts.maxTokens;
     }
     get currentModel() {
         return this._model;
@@ -51,13 +53,16 @@ export class ClaudeCodeAdapter {
         const modelChanged = config.model !== undefined && config.model !== this._model;
         const effortChanged = config.effort !== undefined && config.effort !== this._effortLevel;
         const contextWindowChanged = config.contextWindow !== undefined && config.contextWindow !== this._contextWindow;
+        const maxTokensChanged = config.maxTokens !== undefined && config.maxTokens !== this._maxTokens;
         if (modelChanged)
             this._model = config.model ?? undefined;
         if (effortChanged)
             this._effortLevel = config.effort ?? undefined;
         if (contextWindowChanged)
             this._contextWindow = config.contextWindow ?? undefined;
-        if (modelChanged || effortChanged || contextWindowChanged) {
+        if (maxTokensChanged)
+            this._maxTokens = config.maxTokens ?? undefined;
+        if (modelChanged || effortChanged || contextWindowChanged || maxTokensChanged) {
             this.requestProcessRestart();
         }
     }
@@ -513,11 +518,12 @@ export class ClaudeCodeAdapter {
         const env = buildSpawnEnv(process.env, this.opts.claudeHome, this.buildPlaceholderContext(sessionKey), {
             allowApiKey: this.opts.allowApiKey,
             contextWindow: this._contextWindow,
+            maxTokens: this._maxTokens,
             effortLevel: this._effortLevel,
             wikiMountRoot: this.opts.workspaceDir,
             capabilityBinDir: this.opts.capabilityBinDir,
         });
-        log?.info(`spawn long-lived ${this.opts.claudeBin} (session ${sessionKey}, model=${this._model || 'default'}, context=${this._contextWindow || 'cli-default'}, effort=${this._effortLevel || 'default'}, mode=${this.opts.permissionMode})`);
+        log?.info(`spawn long-lived ${this.opts.claudeBin} (session ${sessionKey}, model=${this._model || 'default'}, context=${this._contextWindow || 'cli-default'}, maxOut=${this._maxTokens || 'cli-default'}, effort=${this._effortLevel || 'default'}, mode=${this.opts.permissionMode})`);
         const proc = spawn(IS_WIN32 ? quoteWin32Arg(this.opts.claudeBin) : this.opts.claudeBin, IS_WIN32 ? args.map(quoteWin32Arg) : args, {
             cwd: this.opts.workspaceDir,
             env,

package/dist/index.js

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
 import * as os from 'node:os';
-import { capabilityBinDir, childLogger, clearAllProviderCreds, configureHttpKeepAlive, createLogger, createOtelLogger, createPlatformConfigManager, deriveModelIsPin, identityFromMe, initAgentTelemetry, llmSource, materializeChannelCapabilities, ParallAgentGateway, parseDispatchDeadlineMs, parseForkDeadlineMs, parseProviderConfig, parseShutdownDeadlineMs, resolveRuntimeContextWindow, resolveRuntimeModel, } from '@parall/agent-core';
+import { capabilityBinDir, childLogger, clearAllProviderCreds, configureHttpKeepAlive, createLogger, createOtelLogger, createPlatformConfigManager, deriveModelIsPin, identityFromMe, initAgentTelemetry, llmSource, materializeChannelCapabilities, ParallAgentGateway, parseDispatchDeadlineMs, parseForkDeadlineMs, parseProviderConfig, parseShutdownDeadlineMs, resolveRuntimeContextWindow, resolveRuntimeMaxTokens, resolveRuntimeModel, } from '@parall/agent-core';
 import { ApiError, ParallClient, ParallWs } from '@parall/sdk';
 import { buildClaudeRuntimeKey, contextFilePathForSession, dispatchContextDirPath, resolveClaudeAgentConfig, resolveWsUrl, sessionStateFilePathForRuntime, stepIdFilePathForSession, } from './config.js';
 import { ClaudeCodeAdapter } from './dispatch.js';
@@ -113,6 +113,7 @@ async function main() {
         const platformModelOverride = deriveModelIsPin(platformDefaults, me.agent_profile);
         const resolvedModel = resolveRuntimeModel(platformModelOverride, platformDefaults.model, config.model);
         const resolvedContextWindow = resolveRuntimeContextWindow(resolvedModel, platformDefaults, activeLLMSource);
+        const resolvedMaxTokens = resolveRuntimeMaxTokens(resolvedModel, platformDefaults, activeLLMSource);
         const resolvedEffort = platformDefaults.thinkingEffort ??
             (process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || undefined);
         if (platformDefaults.model)
@@ -121,12 +122,15 @@ async function main() {
             agentLog.info(`platform config: thinking_effort=${platformDefaults.thinkingEffort}`);
         if (resolvedContextWindow)
             agentLog.info(`platform config: context_window=${resolvedContextWindow}`);
+        if (resolvedMaxTokens)
+            agentLog.info(`platform config: max_tokens=${resolvedMaxTokens}`);
         const adapter = new ClaudeCodeAdapter({
             claudeBin: config.claudeBin,
             claudeHome: config.claudeHome,
             workspaceDir: config.workspaceDir,
             model: resolvedModel,
             contextWindow: resolvedContextWindow,
+            maxTokens: resolvedMaxTokens,
             permissionMode: config.permissionMode,
             allowedTools: config.allowedTools,
             disallowedTools: config.disallowedTools,
@@ -179,6 +183,7 @@ async function main() {
             adapter.updateConfig({
                 model,
                 contextWindow: resolveRuntimeContextWindow(model, updated, activeLLMSource) ?? null,
+                maxTokens: resolveRuntimeMaxTokens(model, updated, activeLLMSource) ?? null,
                 effort: updated.thinkingEffort ?? localEffort,
             });
             // Channel capabilities heat-update with the config: re-materialize the

package/dist/spawn-env.d.ts

@@ -2,6 +2,7 @@ import type { DispatchOpts } from '@parall/agent-core';
 export declare function buildSpawnEnv(parentEnv: NodeJS.ProcessEnv, claudeHome: string, context: DispatchOpts['context'], opts: {
     allowApiKey: boolean;
     contextWindow?: number;
+    maxTokens?: number;
     effortLevel?: string;
     wikiMountRoot?: string;
     capabilityBinDir?: string;

package/dist/spawn-env.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"spawn-env.d.ts","sourceRoot":"","sources":["../src/spawn-env.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAEvD,wBAAgB,aAAa,CAC3B,SAAS,EAAE,MAAM,CAAC,UAAU,EAC5B,UAAU,EAAE,MAAM,EAClB,OAAO,EAAE,YAAY,CAAC,SAAS,CAAC,EAChC,IAAI,EAAE;IACJ,WAAW,EAAE,OAAO,CAAC;IACrB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC3B,GACA,MAAM,CAAC,UAAU,CA8DnB"}
+{"version":3,"file":"spawn-env.d.ts","sourceRoot":"","sources":["../src/spawn-env.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAEvD,wBAAgB,aAAa,CAC3B,SAAS,EAAE,MAAM,CAAC,UAAU,EAC5B,UAAU,EAAE,MAAM,EAClB,OAAO,EAAE,YAAY,CAAC,SAAS,CAAC,EAChC,IAAI,EAAE;IACJ,WAAW,EAAE,OAAO,CAAC;IACrB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,gBAAgB,CAAC,EAAE,MAAM,CAAC;CAC3B,GACA,MAAM,CAAC,UAAU,CAiFnB"}

package/dist/spawn-env.js

@@ -35,10 +35,27 @@ export function buildSpawnEnv(parentEnv, claudeHome, context, opts) {
     if (typeof opts.contextWindow === 'number' &&
         Number.isSafeInteger(opts.contextWindow) &&
         opts.contextWindow > 0) {
-        // Claude Code otherwise assumes 200k for an unrecognized model ID. Its
-        // own auto-compaction policy reserves the output buffer before this
-        // declared boundary; Parall supplies the model fact, not a second ratio.
-        result.CLAUDE_CODE_MAX_CONTEXT_TOKENS = String(opts.contextWindow);
+        // Claude Code otherwise assumes 200k for an unrecognized model ID.
+        //
+        // Declared as the model's INPUT budget (context_window - max_tokens; both
+        // raw catalog facts), not the full window. The upstream validates
+        // input + max_tokens <= context_window, while the CLI requests up to its
+        // own output default (32k) yet reserves at most 20k below this declared
+        // boundary before auto-compacting (2.1.251) — so a session declared the
+        // full window can grow into a band the provider already rejects but the
+        // CLI still considers legal: it retries forever, auto-compact never
+        // fires, and once past the boundary the compaction call itself is
+        // over-limit (the ParallSRE 08-31 lockup). Subtracting the catalog
+        // max_tokens keeps the compaction trigger strictly inside what the
+        // provider accepts.
+        let inputBudget = opts.contextWindow;
+        if (typeof opts.maxTokens === 'number' &&
+            Number.isSafeInteger(opts.maxTokens) &&
+            opts.maxTokens > 0 &&
+            opts.maxTokens < opts.contextWindow) {
+            inputBudget = opts.contextWindow - opts.maxTokens;
+        }
+        result.CLAUDE_CODE_MAX_CONTEXT_TOKENS = String(inputBudget);
     }
     // Channel-capability shims resolve ahead of any globally-installed CLI of
     // the same name (the shim owns auth; the real binary is found further down

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@parall/claude-agent",
-  "version": "1.58.0",
+  "version": "1.58.2",
   "description": "Claude Code bridge runtime for self-hosted Parall agents",
   "license": "MIT",
   "repository": {
@@ -25,9 +25,9 @@
     "src"
   ],
   "dependencies": {
-    "@parall/agent-core": "1.58.0",
-    "@parall/cli": "1.58.0",
-    "@parall/sdk": "1.58.0"
+    "@parall/agent-core": "1.58.2",
+    "@parall/cli": "1.58.2",
+    "@parall/sdk": "1.58.2"
   },
   "devDependencies": {
     "@types/node": "^22.0.0",

package/src/dispatch.ts

@@ -47,6 +47,8 @@ type ClaudeCodeAdapterOptions = Pick<
   orgId: string;
   /** Catalog context size for the selected model. */
   contextWindow?: number;
+  /** Catalog max output tokens for the selected model (input-budget subtrahend). */
+  maxTokens?: number;
   contextFilePathForSession?: (sessionKey: string) => string;
   /** @deprecated Use contextFilePathForSession. */
   stepIdFilePathForSession?: (sessionKey: string) => string;
@@ -113,10 +115,12 @@ export class ClaudeCodeAdapter implements DispatchAdapter {
   private _model: string | undefined;
   private _effortLevel: string | undefined;
   private _contextWindow: number | undefined;
+  private _maxTokens: number | undefined;
 
   constructor(private readonly opts: ClaudeCodeAdapterOptions) {
     this._model = opts.model;
     this._contextWindow = opts.contextWindow;
+    this._maxTokens = opts.maxTokens;
   }
 
   get currentModel(): string | undefined {
@@ -129,15 +133,18 @@ export class ClaudeCodeAdapter implements DispatchAdapter {
     model?: string | null;
     effort?: string | null;
     contextWindow?: number | null;
+    maxTokens?: number | null;
   }): void {
     const modelChanged = config.model !== undefined && config.model !== this._model;
     const effortChanged = config.effort !== undefined && config.effort !== this._effortLevel;
     const contextWindowChanged =
       config.contextWindow !== undefined && config.contextWindow !== this._contextWindow;
+    const maxTokensChanged = config.maxTokens !== undefined && config.maxTokens !== this._maxTokens;
     if (modelChanged) this._model = config.model ?? undefined;
     if (effortChanged) this._effortLevel = config.effort ?? undefined;
     if (contextWindowChanged) this._contextWindow = config.contextWindow ?? undefined;
-    if (modelChanged || effortChanged || contextWindowChanged) {
+    if (maxTokensChanged) this._maxTokens = config.maxTokens ?? undefined;
+    if (modelChanged || effortChanged || contextWindowChanged || maxTokensChanged) {
       this.requestProcessRestart();
     }
   }
@@ -664,6 +671,7 @@ export class ClaudeCodeAdapter implements DispatchAdapter {
       {
         allowApiKey: this.opts.allowApiKey,
         contextWindow: this._contextWindow,
+        maxTokens: this._maxTokens,
         effortLevel: this._effortLevel,
         wikiMountRoot: this.opts.workspaceDir,
         capabilityBinDir: this.opts.capabilityBinDir,
@@ -671,7 +679,7 @@ export class ClaudeCodeAdapter implements DispatchAdapter {
     );
 
     log?.info(
-      `spawn long-lived ${this.opts.claudeBin} (session ${sessionKey}, model=${this._model || 'default'}, context=${this._contextWindow || 'cli-default'}, effort=${this._effortLevel || 'default'}, mode=${this.opts.permissionMode})`,
+      `spawn long-lived ${this.opts.claudeBin} (session ${sessionKey}, model=${this._model || 'default'}, context=${this._contextWindow || 'cli-default'}, maxOut=${this._maxTokens || 'cli-default'}, effort=${this._effortLevel || 'default'}, mode=${this.opts.permissionMode})`,
     );
 
     const proc = spawn(

package/src/index.ts

@@ -20,6 +20,7 @@ import {
   parseProviderConfig,
   parseShutdownDeadlineMs,
   resolveRuntimeContextWindow,
+  resolveRuntimeMaxTokens,
   resolveRuntimeModel,
 } from '@parall/agent-core';
 import { ApiError, ParallClient, ParallWs } from '@parall/sdk';
@@ -159,6 +160,11 @@ async function main() {
       platformDefaults,
       activeLLMSource,
     );
+    const resolvedMaxTokens = resolveRuntimeMaxTokens(
+      resolvedModel,
+      platformDefaults,
+      activeLLMSource,
+    );
     const resolvedEffort =
       platformDefaults.thinkingEffort ??
       (process.env.CLAUDE_CODE_EFFORT_LEVEL?.trim() || undefined);
@@ -170,6 +176,7 @@ async function main() {
       agentLog.info(`platform config: thinking_effort=${platformDefaults.thinkingEffort}`);
     if (resolvedContextWindow)
       agentLog.info(`platform config: context_window=${resolvedContextWindow}`);
+    if (resolvedMaxTokens) agentLog.info(`platform config: max_tokens=${resolvedMaxTokens}`);
 
     const adapter = new ClaudeCodeAdapter({
       claudeBin: config.claudeBin,
@@ -177,6 +184,7 @@ async function main() {
       workspaceDir: config.workspaceDir,
       model: resolvedModel,
       contextWindow: resolvedContextWindow,
+      maxTokens: resolvedMaxTokens,
       permissionMode: config.permissionMode,
       allowedTools: config.allowedTools,
       disallowedTools: config.disallowedTools,
@@ -233,6 +241,7 @@ async function main() {
       adapter.updateConfig({
         model,
         contextWindow: resolveRuntimeContextWindow(model, updated, activeLLMSource) ?? null,
+        maxTokens: resolveRuntimeMaxTokens(model, updated, activeLLMSource) ?? null,
         effort: updated.thinkingEffort ?? localEffort,
       });
       // Channel capabilities heat-update with the config: re-materialize the

package/src/spawn-env.ts

@@ -8,6 +8,7 @@ export function buildSpawnEnv(
   opts: {
     allowApiKey: boolean;
     contextWindow?: number;
+    maxTokens?: number;
     effortLevel?: string;
     wikiMountRoot?: string;
     capabilityBinDir?: string;
@@ -50,10 +51,29 @@ export function buildSpawnEnv(
     Number.isSafeInteger(opts.contextWindow) &&
     opts.contextWindow > 0
   ) {
-    // Claude Code otherwise assumes 200k for an unrecognized model ID. Its
-    // own auto-compaction policy reserves the output buffer before this
-    // declared boundary; Parall supplies the model fact, not a second ratio.
-    result.CLAUDE_CODE_MAX_CONTEXT_TOKENS = String(opts.contextWindow);
+    // Claude Code otherwise assumes 200k for an unrecognized model ID.
+    //
+    // Declared as the model's INPUT budget (context_window - max_tokens; both
+    // raw catalog facts), not the full window. The upstream validates
+    // input + max_tokens <= context_window, while the CLI requests up to its
+    // own output default (32k) yet reserves at most 20k below this declared
+    // boundary before auto-compacting (2.1.251) — so a session declared the
+    // full window can grow into a band the provider already rejects but the
+    // CLI still considers legal: it retries forever, auto-compact never
+    // fires, and once past the boundary the compaction call itself is
+    // over-limit (the ParallSRE 08-31 lockup). Subtracting the catalog
+    // max_tokens keeps the compaction trigger strictly inside what the
+    // provider accepts.
+    let inputBudget = opts.contextWindow;
+    if (
+      typeof opts.maxTokens === 'number' &&
+      Number.isSafeInteger(opts.maxTokens) &&
+      opts.maxTokens > 0 &&
+      opts.maxTokens < opts.contextWindow
+    ) {
+      inputBudget = opts.contextWindow - opts.maxTokens;
+    }
+    result.CLAUDE_CODE_MAX_CONTEXT_TOKENS = String(inputBudget);
   }
   // Channel-capability shims resolve ahead of any globally-installed CLI of
   // the same name (the shim owns auth; the real binary is found further down