praxis-agent 0.62.0 → 0.62.1

package/README.md

@@ -98,6 +98,13 @@ cd /path/to/project
 praxis
 ```
 
+Anthropic models use a 200,000-token context window by default, including
+unknown model IDs. Add the exact terminal `[1m]` suffix (for example,
+`claude-sonnet-4-20250514[1m]`) to request a 1,000,000-token context window;
+Praxis keeps that selected model public, removes the suffix on the wire, and
+adds the `context-1m-2025-08-07` Anthropic beta once. An explicit
+`PRAXIS_CONTEXT_WINDOW_TOKENS` value overrides either inferred window.
+
 Common non-interactive operations:
 
 ```sh
@@ -177,7 +184,7 @@ troubleshooting. Run `praxis --help` for the authoritative command surface.
   patches and readable binary/conflict/transient-path notes, semantic plan/question decision panels with complete
   screen-reader actions, semantic screen projection across selectable surfaces,
   deterministic resize-aware URL/form elicitation rendering, and measured
-  context budgets; print mode,
+  context budgets with base64-payload-independent image estimates; print mode,
   structured JSON/JSONL, context compaction, tool loops, and bounded execution.
 - **Built-in tools** — read, write, edit, `ApplyPatch` for bounded ordered exact
   multi-file replacements, configured plugin LSP navigation with fresh bounded
@@ -294,7 +301,9 @@ troubleshooting. Run `praxis --help` for the authoritative command surface.
   route, whether primary or fallback, stays sticky only through that logical
   Turn's tool continuations; incompatible routes fail closed, and the next
   independent Turn starts from primary. Recovery may persist only an optional
-  selected model, never provider route or wire state.
+  selected model, never provider route or wire state. Anthropic uses a
+  200,000-token default or exact terminal `[1m]` model syntax for 1,000,000
+  tokens; `PRAXIS_CONTEXT_WINDOW_TOKENS` overrides the advertised window.
 - **Transactional self-update** — `praxis update` verifies the package before
   installing it, rejects concurrent updates, and can roll back after an
   interruption or crash.

package/dist/core/context-budget.js

@@ -24,6 +24,12 @@ export function estimateTextTokens(value) {
     }
     return Math.ceil(ascii / 4) + nonAscii;
 }
+const IMAGE_VISUAL_TOKEN_ESTIMATE = 1_600;
+// This provider-neutral conservative fallback never interprets billed usage;
+// observed provider usage remains authoritative at a ContextBudget watermark.
+function estimateImageTokens(mediaType) {
+    return 8 + estimateTextTokens(mediaType) + IMAGE_VISUAL_TOKEN_ESTIMATE;
+}
 function estimateMessageTokens(message) {
     let tokens = 4 + estimateTextTokens(message.role);
     if (message.role === 'tool') {
@@ -31,16 +37,14 @@ function estimateMessageTokens(message) {
             estimateTextTokens(message.toolCallId) +
                 estimateTextTokens(message.content);
         for (const image of message.images ?? []) {
-            tokens +=
-                8 + estimateTextTokens(image.mediaType) + estimateTextTokens(image.data);
+            tokens += estimateImageTokens(image.mediaType);
         }
         return tokens + (message.isError ? 1 : 0);
     }
     tokens += estimateTextTokens(message.content);
     if (message.role === 'user') {
         for (const image of message.images ?? []) {
-            tokens +=
-                8 + estimateTextTokens(image.mediaType) + estimateTextTokens(image.data);
+            tokens += estimateImageTokens(image.mediaType);
         }
         for (const document of message.documents ?? []) {
             tokens += 8 + estimateTextTokens(document.mediaType) + 2000;

package/dist/providers/anthropic-compatible.d.ts

@@ -4,6 +4,10 @@ export interface AnthropicCompatibleProviderOptions {
     baseUrl: string;
     apiKey: string;
     model: string;
+    promptCacheResolver?: (target: {
+        baseUrl: string;
+        model: string;
+    }) => AnthropicPromptCachePolicy;
     maxOutputTokens?: number;
     anthropicVersion?: string;
     webSearch?: boolean;
@@ -32,6 +36,8 @@ export declare class AnthropicCompatibleProvider implements ModelProvider {
     private readonly maxToolMetadataBytes;
     private readonly maxErrorBodyBytes;
     private readonly thinking;
+    private readonly wireModel;
+    private readonly providerBetas;
     private readonly promptCaching;
     private readonly streaming;
     constructor(options: AnthropicCompatibleProviderOptions);

package/dist/providers/anthropic-compatible.js

@@ -3,6 +3,7 @@ import { transportFailureKind } from './provider-errors.js';
 import { reportProviderTransportActivity } from './provider-transport-activity.js';
 import { markNonStreamingFallbackEligible } from './non-streaming-fallback-provider.js';
 import { createAnthropicPromptCachePolicyResolver, } from './anthropic-prompt-cache.js';
+import { resolveAnthropicModelSpec } from './anthropic-model-spec.js';
 function isRecord(value) {
     return typeof value === 'object' && value !== null && !Array.isArray(value);
 }
@@ -643,21 +644,24 @@ export class AnthropicCompatibleProvider {
     maxToolMetadataBytes;
     maxErrorBodyBytes;
     thinking;
+    wireModel;
+    providerBetas;
     promptCaching;
     streaming;
     constructor(options) {
         this.options = options;
-        if (options.contextWindowTokens !== undefined) {
-            positiveInteger(options.contextWindowTokens, 'Context window tokens');
-        }
+        const modelSpec = resolveAnthropicModelSpec(options.model, options.contextWindowTokens);
+        positiveInteger(modelSpec.contextWindowTokens, 'Context window tokens');
         this.endpoint = `${options.baseUrl.replace(/\/+$/, '')}/messages`;
-        this.model = options.model;
+        this.model = modelSpec.model;
+        this.wireModel = modelSpec.wireModel;
+        this.providerBetas = Object.freeze([...modelSpec.betas]);
         this.streaming = options.streaming ?? true;
         this.fetchImplementation = options.fetchImplementation ?? fetch;
         this.maxOutputTokens = positiveInteger(options.maxOutputTokens ??
-            (options.model.includes('claude-opus-4-6')
+            (this.wireModel.includes('claude-opus-4-6')
                 ? 64_000
-                : options.model.startsWith('claude-')
+                : this.wireModel.startsWith('claude-')
                     ? 32_000
                     : 8192), 'Max output tokens');
         this.capabilities = {
@@ -671,19 +675,20 @@ export class AnthropicCompatibleProvider {
                 modes: ['enabled', 'adaptive', 'disabled'],
                 maxTokens: true,
             },
-            ...(options.contextWindowTokens === undefined
-                ? {}
-                : { contextWindowTokens: options.contextWindowTokens }),
+            contextWindowTokens: modelSpec.contextWindowTokens,
             maxOutputTokens: this.maxOutputTokens,
             terminalReasons: true,
         };
         this.thinking = validateThinking(options.thinking);
-        const promptCaching = options.promptCaching === false
-            ? undefined
-            : (options.promptCaching ??
+        const promptCaching = options.promptCaching !== undefined
+            ? options.promptCaching
+            : (options.promptCacheResolver?.({
+                baseUrl: options.baseUrl,
+                model: this.wireModel,
+            }) ??
                 createAnthropicPromptCachePolicyResolver({}, 'native')({
                     baseUrl: options.baseUrl,
-                    model: options.model,
+                    model: this.wireModel,
                 }));
         this.promptCaching = promptCaching
             ? cacheControl(promptCaching.ttl)
@@ -714,6 +719,7 @@ export class AnthropicCompatibleProvider {
                     budget_tokens: thinking.maxTokens ?? maxTokens - 1,
                 };
         const betas = [
+            ...this.providerBetas,
             ...(request.betas ?? []),
             ...(thinking && thinking.mode !== 'disabled'
                 ? ['interleaved-thinking-2025-05-14']
@@ -730,7 +736,7 @@ export class AnthropicCompatibleProvider {
                 ...(betas.length ? { 'anthropic-beta': betas.join(',') } : {}),
             },
             body: JSON.stringify({
-                model: this.options.model,
+                model: this.wireModel,
                 max_tokens: maxTokens,
                 messages: serialized.messages,
                 stream: this.streaming,

package/dist/providers/anthropic-model-spec.d.ts

@@ -0,0 +1,9 @@
+export declare const ANTHROPIC_LONG_CONTEXT_BETA = "context-1m-2025-08-07";
+export interface ResolvedAnthropicModelSpec {
+    readonly model: string;
+    readonly wireModel: string;
+    readonly contextWindowTokens: number;
+    readonly betas: readonly string[];
+}
+export declare function resolveAnthropicModelSpec(model: string, explicitContextWindowTokens?: number): ResolvedAnthropicModelSpec;
+//# sourceMappingURL=anthropic-model-spec.d.ts.map

package/dist/providers/anthropic-model-spec.js

@@ -0,0 +1,15 @@
+export const ANTHROPIC_LONG_CONTEXT_BETA = 'context-1m-2025-08-07';
+export function resolveAnthropicModelSpec(model, explicitContextWindowTokens) {
+    const longContext = model.endsWith('[1m]');
+    const wireModel = longContext ? model.slice(0, -'[1m]'.length) : model;
+    if (longContext && wireModel.trim().length === 0) {
+        throw new Error('Anthropic [1m] model spec must include a base model name');
+    }
+    return Object.freeze({
+        model,
+        wireModel,
+        contextWindowTokens: explicitContextWindowTokens ?? (longContext ? 1_000_000 : 200_000),
+        betas: Object.freeze(longContext ? [ANTHROPIC_LONG_CONTEXT_BETA] : []),
+    });
+}
+//# sourceMappingURL=anthropic-model-spec.js.map

package/dist/providers/provider-registry.js

@@ -175,10 +175,7 @@ class NativeProviderRegistry {
                 ...(this.options.anthropicPromptCacheResolver === undefined
                     ? {}
                     : {
-                        promptCaching: this.options.anthropicPromptCacheResolver({
-                            baseUrl: target.baseUrl,
-                            model: target.modelId,
-                        }),
+                        promptCacheResolver: this.options.anthropicPromptCacheResolver,
                     }),
                 ...(this.options.providerEnvironment?.maxOutputTokens === undefined
                     ? {}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "praxis-agent",
-  "version": "0.62.0",
+  "version": "0.62.1",
   "description": "Local-first, single-user general agent for the command line.",
   "license": "MIT",
   "author": "wuqisen",