1.12.0

• fix: chat-completions contract handling — reasoning is sent back to the server complete and unmodified, in the field it arrived in (reasoning, reasoning_content, or OpenRouter reasoning_details), and never inside the message content; reasoning_content streams (DeepSeek, llama.cpp, LM Studio) are now read
• fix: every tool call gets its tool message, including when a tool throws; missing tool-call ids are generated, chunks without an index are joined, empty or unparsable arguments go back as {}
• fix: finish_reason length continues a cut-off reply (up to three times), content_filter is reported as an error, and refusals are shown
• fix: SSE parsing accepts CRLF and bare-CR line endings, data: without a space, and a last event without a trailing blank line
• fix: HTTP status is kept for retry decisions, Retry-After is honoured, 408 is retried
• fix: a tool call the model writes inside its reasoning (DeepSeek call syntax) is recovered and run instead of counting as an empty reply; a reply with no action no longer crashes the run
• fix: compaction carries MI_API_PARAMS and drops chat_template_kwargs when a server rejects it
• feat: MI_SESSION keeps a one-shot (-p or piped) conversation in a file (mode 0600) and continues it on the next call under the current system message
• feat: goal loop — one worker conversation for the whole budget; a review passes only when its last line is exactly VERDICT: PASS; reviews get a pristine copy of the working directory for literal change-set checks; request cap and retry under a deadline; MI_GOAL_PLAN=0 skips the planner
• feat: MI_DENY refuses tool calls whose arguments match a regex
• add: benchmarking on the Harbor framework (bench/, docs/benchmarking.md, docs/benchmark-history.md), including an optional network allowlist for evaluation runs
• docs: docs/api-contracts.md — the audit of mi's model API handling against the official contracts, with live evidence from local model servers

Known limitations

• The contract fixes were verified live only against local servers (Ollama and llama.cpp). Behaviour that needs a paid API is covered by tests against the documented wire shapes only: OpenRouter's reasoning_details round trip, DeepSeek's rejection of a tool-calling request without reasoning_content, and OpenAI's rejection of unknown request arguments.
• Not exercised against a real server: CRLF streams, tool calls without ids or index, refusals, Retry-After.
• reasoning_details that arrive without an index across several deltas can be ambiguous to reassemble.
• Reasoning is now sent back in full, which makes requests larger on servers that count it as prompt tokens.
• When a conversation already exceeds the context window, the compaction request exceeds it too and fails.
• mi speaks the Chat Completions format only; there is no Responses API or Anthropic Messages support.