fix: tool-call schemas, predefined provider base_url, i18n#8166
Merged
Conversation
The base-url field was already hidden from the UI for predefined providers, leaving the per-provider settings entry as vestigial data with two sources of truth. Anthropic shipped with a /v1 mismatch between top-level base_url and the settings value as a result. Drop the base-url setting entry from all predefined providers; the top-level base_url constant is now authoritative. Custom providers added via "+" still carry an editable base-url. Migration v14 overwrites stale persisted base_urls with the canonical constant and strips leftover base-url settings.
llama.cpp's json-schema-to-grammar emits PCRE shorthands (\d, \w, \s) when it sees `format: date-time|date|time` or any `pattern` using those escapes. GBNF rejects them with "error parsing grammar: unknown escape at \d\d", grammar compilation fails, llama-server silently falls back to unconstrained generation, and tool-capable models (e.g. Qwen3) emit their native <tool_call> XML instead of OpenAI JSON. The downstream OAI tool-call parser then reports a misleading "Failed to parse input at pos N". Strip those from tool input schemas in both the TS client-side normalizer and its Rust proxy counterpart. The model still has the type and description; only the unenforceable validation hints are dropped, and only when they would have broken GBNF. Caught with caldav-mcp's create-event/update-event tools, whose recurrenceRule.until / start / end ship both `format: date-time` and a redundant \d-laden pattern.
Each locale's common.json already carried `mcp-servers` with the right translation; the parallel `connectors` key was only used by the settings menu link and was still "Connectors" in en/common.json while every other locale had already been rewritten to "MCP Servers". Point SettingsMenu (and its test) at the canonical key and delete the redundant lines.
Contributor
Minh141120
approved these changes
May 20, 2026
Minh141120
approved these changes
May 20, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Describe Your Changes
Predefined providers no longer carry a user-editable
base-urlsetting. The settings card was already hidden from the UI for predefined providers, so the per-providerbase-urlentry was vestigial data with two sources of truth — and that mismatch had already shipped a bug (Anthropic'sbase_urlwas…/v1but the settings entry was…without/v1). Top-levelbase_urlis now authoritative for all predefined providers; custom providers added via "+" keep their editablebase-url. Migration v14 overwrites stale persistedbase_urls with the canonical constant and strips any leftoverbase-urlsetting.Strip GBNF-incompatible
formatandpatternfrom tool input schemas. llama.cpp'sjson-schema-to-grammaremits PCRE shorthands (\d,\w,\s) when it seesformat: date-time|date|timeor anypatternusing those escapes. GBNF rejects them ("error parsing grammar: unknown escape at \d\d"), grammar compilation fails, llama-server silently falls back to unconstrained generation, and tool-capable models (Qwen3 in particular) emit their native<tool_call>XML instead of OpenAI JSON. The downstream parser then reports a misleading "Failed to parse input at pos N". Fix is applied symmetrically in the TS client-side normalizer and the Rust proxy. Caught withcaldav-mcp'screate-event/update-eventtools, whoserecurrenceRule.until/start/endship both aformatand a redundant\d-ladenpattern.Drop the duplicate
connectorsi18n key. Every locale'scommon.jsonalready carriedmcp-serverswith the right translation; the parallelconnectorskey was only used by the settings menu and was still rendering "Connectors" in English while every other locale had already been rewritten to "MCP Servers". PointSettingsMenu(and its test) at the canonicalmcp-serverskey and delete the redundant lines.Fixes Issues
Self Checklist