1.11.0

• feat: tool loading supports named exports — modules that export { name, description, parameters, handler } as named exports (not just default) now load correctly
• add: OpenRouter DeepSeek-V4-Flash Terminal-Bench 2.0 stratified 10-task estimator preset

v1.10.1

Broken user tools degrade gracefully instead of crashing the agent.

When a tool in ~/.mi/tools/ fails to import (missing package, syntax error, etc.), it becomes a stub with [broken] in its description. Calling the stub returns the error message, so the model can diagnose and self-fix — install the missing dep, edit the file, then retry on the next loop iteration.

v1.10.0

Load user tools from ~/.mi/tools/ in addition to the bundled tools directory.

• loadTools() scans both MI_DIR/tools/ and MI_HOME/tools/; user tools override bundled tools by name
• Agent creates new tools in ~/.mi/tools/ instead of the install directory
• Deduplicates tool schemas so overrides don't produce duplicates

v1.9.0

What's new

• Multi-file -f — pass multiple -f flags to attach several files; text goes to system prompt, images become image_url content parts
• Image support — -f photo.png reads as base64 and sends as vision-compatible image_url (png/jpg/jpeg/gif/webp)
• Reactive compaction — on context overflow, summarizes the conversation via API and splices history to keep working (dumps full history to /tmp for recovery)
• Fetch retry — automatic retry with exponential backoff on 429/5xx/network errors
• REPL /help — lists available commands and env vars
• Error .code preservation — enables reliable detection of context_length_exceeded from structured API errors

Fixes

• -f no longer eats subsequent flags as filenames
• -f with nonexistent file gives a clean error instead of stack trace
• MI_API_PARAMS with invalid JSON validated at startup
• /help now lists all documented env vars

Stats

• 30 meaningful LOC (unchanged)
• 60 tests (was 50)

v1.8.0

Breaking

• Goal tool: check parameter is now a judge prompt, not a bash command. A mi subagent evaluates the work (can read files, run commands) and responds ACK/NACK. For bash checks, include the command in the prompt: check="run npm test — ACK if pass, NACK otherwise".

Added

• -g / -c CLI flags: Goal loop as a first-class execution mode. mi -g "goal" -c "judge criteria" iterates worker/judge subagent pairs until the judge ACKs.

Changed

• Persona: terse, unimpressed, one-sentence cap for non-task input. Drier responses to philosophical and off-topic questions.

v1.7.10

Add MI_API_PARAMS env var — a JSON object merged into the chat completion payload, allowing arbitrary API parameter overrides (e.g. MI_API_PARAMS='{"temperature":0,"max_tokens":4096}').

v1.7.9

• REPL: /new and /clear added as aliases for /reset
• Sandbox image ships with procps, util-linux, bind-tools (dig), iputils (ping), tree, rsync, vim
• README: features reordered by impact; sandbox toolset documented

v1.7.8

Sandbox utility upgrade

Adds missing GNU/Linux utilities to the sandbox image:

• procps (full ps), util-linux, bind-tools (dig/nslookup), iputils (ping)
• tree, rsync, vim, wget
• tzdata, ca-certificates, openssl
• Symlinks awk → gawk for consistency

Image size: 222 MB → 267 MB (+45 MB).

v1.7.7

• Forward OPENAI_API_KEY to sandbox container via -e
• Prevent recursive sandbox when MI_SANDBOX is set in config: container gets MI_SANDBOXED=1 to skip the sandbox code path

v1.7.6

• Add -v / --version flag
• Restructured --help output (modes, flags, env sections)
• Fix sandbox mode: absolute entrypoint path prevents "cannot find module /work/index.mjs" when -w /work overrides WORKDIR
• Updated self skill with config file docs, env vars table, current tool list