odu-mcp

The agent face of odu — an MCP stdio server that re-exposes a live CI run as agent tools (run, node_rerun, node_cancel, lane_cancel, wait_for_settle, cancel, lease, release) and subscribable resources (surface://streams/nodes, surface://collections/logs/{id}), so Claude Code / Codex / opencode / Gemini CLI drive CI with structured calls instead of scraping terminal output.

lease / release are the agent-held venue layer: hold a free box across discrete tool calls without re-queuing between runs. lease returns immediately (held or waiting); re-call or inventory to observe the line. run reuses held hosts and does not release them on exit.

wait_for_settle defaults to fail-fast: it returns the instant the first node goes red (fail_fast_tripped: true, settled: false), so the agent drills into the failure without blocking on the slow lanes — its failed[] is only what's red so far, and only passed: true (a fully settled run) is a trustworthy green. A verdict about an observed run is stamped with that run's identity — sha7 always, seq whenever the coordinator reserved an ordinal — so it's clear which run it describes (seq is null only when none was reserved: a wait that saw no frame, or the rare case the coordinator couldn't reserve one); and unposted[] carries full owed GitHub status rows ({context, lastError, attempts}) not yet confirmed (reporting debt never blocks settle — the test verdict stays the truth). Called with no run live it fails loud (an error mirroring odu status, not an empty settled: false), and an optional expected_sha (prefix-matched against the run's sha7) refuses loud when the live run's commit doesn't match. A run that is still PROVISIONING — run returns as soon as the coordinator serves its socket, which is now before it claims a machine — is a live run: wait_for_settle blocks on it, the nodes resource shows _ci-setup@<platform> running, and its log resource carries the runner closure's copying path … progress. A claim that never succeeds arrives as a red _ci-setup@<platform> verdict rather than as a run-time error. If the coordinator's socket closes before it publishes a terminal frame, the verdict comes from the run's finalized record on disk — never green for a run torn down mid-flight. The nodes resource carries the same unposted. MCP run tees coordinator stdout/stderr to .ci/<sha7>/runs/<seq>.log.

A settled node's surface://collections/logs/{id} is the last 64KB of its output and now reliably ENDS at the recipe's final line: a run that settles on its own holds its lanes open until every node has finished streaming, because a node's status and its output travel on different streams and the status one arrives first — so the recipe summary the agent came for used to be the part that never made it. Read .ci/<sha7>/<platform>/<node>.log for the whole thing. A lane that stops streaming with output still owed stamps [odu] log truncated: … into the log, so a drill-in never reads a cut log as a complete one.

cancel stops the live run and waits until it's torn down; node_cancel stops one node (ci::fmt@plat) and lane_cancel drops a whole platform while the rest of the run settles (status cancelled, not errored/failed). run's supersede cancels a run already live here before starting (the "stop this, run the fixed commit" move), linger keeps the coordinator serving past settle so a node can be rerun afterwards, and no_wait fails immediately when every host in a venue pool is busy (default: wait in line). Together they let the agent loop call off or replace a run instead of stranding it or hitting "a run is already in progress".

bin/serve is self-contained — it resolves odu via nix run and serves over stdio in the consumer's repo (dialing .ci/odu.sock). Set ODU_FLAKE to override the odu flake-ref (default github:juspay/odu); a repo that re-exports odu can point it at its own pinned output with ODU_FLAKE=.#odu.

Full docs in the repo README.

This skill primitive exists for APM's deployment convention — it lands bin/serve at .agents/skills/odu-mcp/bin/serve in the consumer's working tree (APM's skills-convergence path), which keeps the launcher available even before apm install runs on a fresh clone. The package is mechanically a "skill" in APM's primitive vocabulary; semantically it's a tool launcher.