AgentPort 3.1.26: Haiku tool-loading and token safeguards

AgentPort 3.1.26

Fixes oversized Haiku tool prompts. When Codex supplies its Code Mode execution tool, AgentPort keeps app and MCP definitions behind that client-owned mechanism instead of repeating every schema in the provider request. Core coding tools, collaboration, tools referenced by the conversation and explicit tool choices remain available. The router does not execute tools itself.

Older clients that omit the execution tool no longer trigger injection of the static app catalogue. If their remaining tool definitions exceed 256 KB, AgentPort refuses the request locally with a clear message. It does not silently drop tools or send the oversized request to Haiku. Restart Codex to refresh its model capabilities if this warning appears.

This safeguard applies to the Haiku Code Mode route. Native GPT models, their context windows, provider token accounting, and other model routes are unchanged. Existing allowance consumption cannot be undone.

Offline regression checks cover 1,400 MCP tool definitions being kept behind Codex execution, preservation of required tools, native routing, and an oversized request returning HTTP 400 without reaching the mock provider. No paid Haiku request was sent to test this release. Actual quota savings still require confirmation from a subsequent user request.

Windows x64 release. The Apple Silicon download remains on 3.1.24 pending a working Mac build environment.