v0.11.0

v0.11.0 - Invocation-native runtime

Status: breaking release guidance

This release changes MCO from a findings/decision pipeline into a raw-answer invocation runtime. Legacy findings-driven design documents remain available for context but are superseded by the invocation runtime v1 contract; this note is the migration guide for callers adopting the new public surface.

New behavior

• Use repeatable --agent [alias=]provider:model for explicit Agent identity and multiple models from one Provider.
• Use --providers as the invocation-native shorthand for one configured/default model per Provider.
• Consume raw answer text plus operational invocation records. Task status is complete, partial, or failed, with exit codes 0, 1, and 2.
• Use text mode, --json, or --stream jsonl; stdout contains answers or the selected machine protocol, while diagnostics go to stderr.
• Use --chain, --debate, and --synthesize for LLM-driven file-backed stages.
• Use --result-mode artifact|stdout|both for temporary or persistent Markdown artifacts.
• A synthesis manifest retains run and debate records, including failures or missing entries, but never includes synthesis itself. Root result.md groups actual stages, places synthesis first, and keeps every original answer and failure record.

Breaking removals

The following surfaces are removed:

• mco findings and the findings JSON schema;
• semantic normalization, deduplication, confidence, consensus, passive lifecycle, and findings-driven memory runtime;
• Markdown-PR and SARIF renderers;
• content-based INCONCLUSIVE status;
• --format, --strict-contract, --memory, --space, --diff, --staged, --unstaged, and --diff-base.

Removed flags return migration guidance rather than being silently ignored.

Migration examples

Old structured review callers should send the desired scope and output shape as a raw prompt:

mco review --target-paths runtime --providers claude,codex \
  --prompt "Review runtime/ for correctness and security risks; report your reasoning in natural language." \
  --result-mode both --task-id runtime-review

Old multi-model callers should use explicit invocations:

mco run --prompt "Compare these implementations." \
  --agent fast=pi:model-a \
  --agent careful=pi:model-b \
  --stream jsonl

Old persistent-memory callers should save and pass Markdown artifacts through chain/debate/synthesis stages:

mco run --prompt "Analyze the repository." \
  --providers claude,codex --chain --execution-mode read_only \
  --result-mode artifact --task-id repository-analysis

Write safety

MCO does not create or manage worktrees. If a user explicitly chooses parallel writers, the calling Agent must partition non-overlapping --target-paths and warn about overlapping edits.