Purpose: To provide a clear plan and verifiable results when Codex handles complex long tasks, making it easy to track progress and to resume from the recorded state after an interruption instead of starting over.
Main capabilities: Split requests into minimal parts that are bounded and have dependencies; execute ready parts in a safe order; dedicated agents are optional, and when unavailable the Coordinator completes them inline following the same process; record local state with revisions; before retrying, reconcile work that may still be running and preserve evidence of what has been completed; at wrap-up, for non-exempt graphs, rerun the positive and negative proof commands and require all to succeed; record the five dimensions of complexity (breadth, change surface, coupling, novelty, verification) and ambiguity factors; exceeding the default threshold requires splitting or refinement.
Usage: Install via the prompt from the repository's bundled installer; the installation process clones the repository to a temporary directory, records the commit, copies the self-contained skill/ directory to the current user, deletes the temporary checkout, and reports the installation path and source commit, replacing only when the target identifies itself as Coordinator; requires Python 3.11+, with no third-party runtime dependencies. The documentation recommends starting each coordination task with the $coordinator prefix and stating the completion criteria. The persisted schema-v8 workflow documentation is located at ~/.agent-coordinator/workflows, and on Windows at %USERPROFILE%.agent-coordinator\workflows; v6 and v7 are automatically upgraded, and their existing nodes are permanently exempt from proof.
Applicable scenarios: Tasks such as implementation, diagnosis, recovery, and evidence-only comparison; for example, adding saved searches in an application without breaking the checkout flow, requiring three fixes ordered by impact, each citing project evidence, and without modifying files.
Limitations: Installation does not edit Codex settings, nor does it register a global custom agent configuration; it does not add Coordinator-owned persistent files to the target project (during initialization it only temporarily creates and deletes a privately named file to probe filesystem case behavior; Windows requires no probe); it does not invoke or inspect a version control system; routing only arranges candidates published by the current runtime, and when there is no directory or selection fails, it inherits the parent model and effort; each executable part needs acceptance criteria, one role, and zero or more normalized repository-relative write_scopes; an empty scope means evidence-only work and requires change_surface=0; independent active work must not overlap scopes. The demo deletes and rebuilds demo/home/, so outputs must be saved in advance; running the demo requires setting OPENAI_API_KEY in demo/.env.
