1.13.0

Merge pull request #3 from av/fix/model-api-contracts

Fulfil the chat-completions API contracts (reasoning, tool calls, streaming, errors); Harbor benchmark work and goal-loop changes from bench/tb2-reference.

1.11.0

feat: tool loading supports named exports (not just default export)
add: openrouter deepseek-v4-flash tb2-10 benchmark preset

1.10.1

1.10.0

feat: multi-file -f, image support, compaction, retry, /help

- Multi-file -f: accept multiple -f flags, skip flag-like values, clean
  error on missing files
- Image -f: detect .png/.jpg/.jpeg/.gif/.webp, send as base64 image_url
  content in user message with correct MIME mapping
- Reactive compaction: on context-overflow errors (isCtx), summarize
  conversation via non-streaming API call and splice history to
  [system, user-recap]. Dumps full history to /tmp for recovery.
  Single retry guard per run() call.
- Fetch retry: exponential backoff (500ms base) on 429/5xx/network
  errors, 3 retries max
- REPL /help: lists available commands and env vars
- Error .code preservation: thrown errors carry API error codes so
  isCtx can match structured context_length_exceeded responses
- MI_API_PARAMS: validate JSON at startup instead of raw SyntaxError
- 10 new tests covering all new features and edge cases

Persona: terse/unimpressed traits, structural one-sentence cap for no…

…n-task input

Bump to v1.8.0 — goal tool check param is now a judge prompt (breaking)

Add MI_API_PARAMS env var for chat completion payload overrides

Release v1.7.9

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Sandbox: add procps, util-linux, bind-tools, iputils, tree, rsync, vim

Closes ergonomic gaps in the sandbox image — ps/dig/tree/ping/etc.
were missing or busybox-only. +45MB on the final image (222→267MB).