zack3d/big-agi
1
0
Fork 0
mirror of https://github.com/enricoros/big-AGI.git synced 2026-05-10 21:50:14 -07:00
Code Issues Packages Projects Releases Wiki Activity

AIX: Anthropic: do not turn on code execution just for dynamic filtering. #1087

Browse Source
This commit is contained in:
Enrico Ros
2026-04-28 18:24:00 -07:00
parent ac69c62020
commit e5de61d682
2 changed files with 12 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/modules/aix/server/dispatch/chatGenerate/adapters/anthropic.messageCreate.ts
+10 -3
View File
@@ -35,7 +35,7 @@ export function aixAnthropicHostedFeatures(model: AixAPI_Model, chatGenerate: Ai
const _hasAixToolRestrictivePolicy = chatGenerate.toolsPolicy?.type === 'any' || chatGenerate.toolsPolicy?.type === 'function_call';
// Dynamic web tools (20260209) require code execution for programmatic tool calling
const hasDynamicWebTools = model.vndAntWebDynamic === true && (model.vndAntWebSearch === 'auto' || model.vndAntWebFetch === 'auto');
// const hasDynamicWebTools = model.vndAntWebDynamic === true && (model.vndAntWebSearch === 'auto' || model.vndAntWebFetch === 'auto');
// Programmatic Tool Calling - tools with allowed_callers or input_examples
const programmaticToolCalling = chatGenerate.tools?.some(tool =>
@@ -45,10 +45,15 @@ export function aixAnthropicHostedFeatures(model: AixAPI_Model, chatGenerate: Ai
),
) ?? false;
// [Anthropic, issue #1087] Dynamic web tools (20260209) have INTERNAL code execution. We do not
// explicitly add the code_execution tool nor the beta header for them: Anthropic enables what is
// needed implicitly behind the scenes. Adding our own creates two execution environments and
// confuses the model (e.g. bash searches for inlined attachments, web tools called from scripts
// without `allowed_callers` permission, runaway tool loops).
return {
disableAllHostedTools: !!(_hasAixCustomTools && _hasAixToolRestrictivePolicy),
enable1MContext: model.vndAnt1MContext === true,
enableCodeExecution: !!model.vndAntSkills || !!model.vndAntContainerId || hasDynamicWebTools || programmaticToolCalling,
enableCodeExecution: !!model.vndAntSkills || !!model.vndAntContainerId || programmaticToolCalling,
enableFastMode: model.vndAntInfSpeed === 'fast',
enableSkills: !!model.vndAntSkills,
enableStrictOutputs: !!model.strictJsonOutput || !!model.strictToolInvocations,
@@ -284,7 +289,9 @@ export function aixToAnthropicMessageCreate(model: AixAPI_Model, _chatGenerate:
name: 'tool_search_tool_bm25',
});
// Code Execution tool - required for dynamic filtering, Skills, etc.
// Code Execution tool - for Skills, container reuse, and Programmatic Tool Calling.
// Note: NOT added for dynamic web tools (_20260209) - they execute code internally and adding
// a standalone environment confuses the model (issue #1087).
if (enableCodeExecution)
hostedTools.push({ type: 'code_execution_20260120', name: 'code_execution' });
src/modules/llms/server/anthropic/anthropic.access.ts
+2 -1
View File
@@ -177,7 +177,8 @@ export function anthropicBetaFeatures(options?: AnthropicHostedFeatures): string
if (options?.enable1MContext)
bf.add('context-1m-2025-08-07');
// Code execution (for dynamic web tools PFC, or Skills) + files API for container downloads
// Code execution (for Skills, container reuse, Programmatic Tool Calling) + files API for container downloads.
// NOT enabled for dynamic web tools (_20260209): those have internal code execution managed by Anthropic.
// Note: SDK defines code-execution-2025-05-22; we use 2025-08-25 (newer iteration, not yet in SDK types).
// Code execution may be GA now (most SDK examples skip the beta namespace), but keeping for safety.
if (options?.enableCodeExecution) {