Releases·Tool Release
[P] Tool schema drift in FastMCP / AI Agents: How to update Pydantic schemas without breaking active agent calls
A developer building FastMCP tools for LLM agents released Pyverge, a small library that handles tool schema versioning by migrating incoming payloads between versions before they reach tool handlers. It addresses validation crashes that occur when agents with cached prompts keep sending old-format payloads after a tool schema changes. The library is open-sourced on GitHub with a runnable FastMCP showcase.
1 engagement1 sourceTue, Sep 29, 2026, 04:08 PM
AI Overview
- Developer releases Pyverge (github.com/JoHa-HQ/pyverge), which migrates incoming payloads like v1 to v2 via versioned Pydantic models and JSON patch transitions before tool handlers run
- Pyverge includes a runnable FastMCP showcase, keeping versioning plumbing out of tool functions
- The author reports cached agent contexts keep sending v1 payloads like location: str after schemas require lat and lng, causing validation crashes
Related
News·Incident
Tool schema drift: when the function changes but the registration doesn't, your agent fails silently
A common agentic failure occurs when a tool's implementation changes (e.g., adding a required parameter) but its registration schema is not updated. The model then calls the tool with the old arguments, causing silent failures. This pattern highlights the need for automated schema synchronization.
A common agentic failure occurs when a tool's implementation changes (e.g., adding a required parameter) but its registration schema is not updated. The model then calls the tool with the old arguments, causing silent failures. This pattern highlights the need for automated schema synchronization.
1 engagement1 sourceredditWed, Jul 22, 2026, 02:41 AM
Releases·Tool Release
pydantic-ai ships v2.35.0–v2.35.3 with capability API deprecation and provider fixes
pydantic-ai released v2.35.0 on 2026-08-25, deprecating RunContext.capability_loaded and available_capability_ids in favor of capability_active and active_capability_ids, reducing default Temporal metric export frequency, and fixing tool-description handling. Three follow-up patch releases (v2.35.1 through v2.35.3) landed within days, fixing Bedrock guardrail traces, Temporal cancellation behavior, Heroku GLM model routing, and deferred-tool result resolution. Developers using capability APIs or durable execution backends should review the deprecations and pin updates.
pydantic-ai released v2.35.0 on 2026-08-25, deprecating RunContext.capability_loaded and available_capability_ids in favor of capability_active and active_capability_ids, reducing default Temporal metric export frequency, and fixing tool-description handling. Three follow-up patch releases (v2.35.1 through v2.35.3) landed within days, fixing Bedrock guardrail traces, Temporal cancellation behavior, Heroku GLM model routing, and deferred-tool result resolution. Developers using capability APIs or durable execution backends should review the deprecations and pin updates.
Updated Sat, Aug 29, 2026, 01:25 AM — v2.35.3 released 2026-08-27, adding Heroku GLM routing, cookie scoping, and deferred-tool fixes.
0400 engagement4 sourcesgithubTue, Aug 25, 2026, 12:00 AM
News·Incident
Pydantic patches moderate security flaw in pydantic-ai v1.107.1
Pydantic released v1.107.1 of pydantic-ai on 2026-07-10, fixing a moderate-severity vulnerability (GHSA-jpr8-2v3g-wgf9, CWE-863) in the AG-UI serving path. The bug could cause unresolved tool calls to be incorrectly retained when client messages were dropped during sanitization.
Pydantic released v1.107.1 of pydantic-ai on 2026-07-10, fixing a moderate-severity vulnerability (GHSA-jpr8-2v3g-wgf9, CWE-863) in the AG-UI serving path. The bug could cause unresolved tool calls to be incorrectly retained when client messages were dropped during sanitization.
1 engagement1 sourcegithubFri, Jul 10, 2026, 12:00 AM
Releases·Tool Release
LangChain 1.3.15 adds trace_policy to AgentMiddleware
LangChain released version 1.3.15 with new features and fixes, including exposing trace_policy on AgentMiddleware and preserving history on SummarizationMiddleware summary failure. The release also includes type error fixes and dependency bumps.
LangChain released version 1.3.15 with new features and fixes, including exposing trace_policy on AgentMiddleware and preserving history on SummarizationMiddleware summary failure. The release also includes type error fixes and dependency bumps.
0 engagement1 sourcegithubTue, Aug 11, 2026, 07:11 PM
Social·Community
Community argues tool schemas are part of the model under evaluation
A Reddit post argues that agent evaluation reports must treat tool definitions as part of the model under test, not incidental plumbing. Renaming tools, changing parameter descriptions, or altering return shapes changes the model's reasoning interface, so results from old harnesses don't transfer.
A Reddit post argues that agent evaluation reports must treat tool definitions as part of the model under test, not incidental plumbing. Renaming tools, changing parameter descriptions, or altering return shapes changes the model's reasoning interface, so results from old harnesses don't transfer.
19 engagement1 sourceredditTue, Aug 18, 2026, 01:40 PM