All notable changes to minga_claude_bridge are listed here. The format follows Keep a Changelog, and the project uses Semantic Versioning.

0.1.0 - Unreleased

The first release. It runs the Claude models of your Claude subscription inside Minga's agent through the claude CLI you already have installed and signed in. Minga keeps running the agent: its tools, approvals, and transcript.

Requires Minga with model routes and plan usage (jsmestad/minga#3734 and #3756), and Claude Code 2.1.330 or newer.

Added

  • Protocol codec (#6): decodes and encodes Claude Code's stream-json frames: system init, stream events, assistant messages, results, rate-limit events, and control requests and responses. Unknown frame types decode to a value instead of failing.
  • Tool routing (#7): serves Minga's tools to Claude Code as the in-process MCP server minga, pairs each tool call with its tool_use id, flattens schemas Claude Code would drop, and compares history to decide whether a conversation can continue its Claude Code session.
  • Session process (#8): one process per conversation drives a pure state machine. It suspends a turn on tool calls, resumes with Minga's tool results, aborts within a bounded time, and stops Claude Code before the model's first output when it reports any tool other than Minga's.
  • Real claude process (#9): starts the resolved claude binary without a shell, in an empty private working directory, with built-in tools, hooks, plugins, skills, slash commands, and other MCP servers turned off. Strips provider keys and parent-session variables from its environment, applies the routes.claude_bridge.env credential entry, and never leaves a claude process behind.
  • Minga model route (#10): the extension registers the claude_bridge route. /model lists Claude's models with readiness ("Ready ยท Claude Pro") and actionable messages for a missing CLI, a signed-out CLI, an old CLI, and a reached plan limit. Follow-up turns continue the same Claude Code session; edits, branches, and model changes start a fresh one with the earlier history.
  • Context windows and plan usage (#12): measures each model's real context window and reads plan usage from Claude Code in the background, reports five-hour and seven-day plan windows to Minga, and streams thinking with adaptive thinking.

Fixed

  • A get_usage answer without windows now reports an empty list, so Minga clears stale plan windows.
  • A rate-limit event's utilization now updates its plan window.
  • The usage refresh after a turn now runs under the bridge's task supervisor and stops with the extension.
  • A plan label learned from usage now only updates the readiness of the credential it was read with.