Pipeline Repair Agent

Task

Build a LangGraph agent that investigates a failed DAG, repairs its source code, runs checks, and prepares a merge request.

Requirements

1. Return a new state with status, steps, pending, observations, answer, seen; never mutate input. max_steps is an integer from 1 to 20, not a boolean. Invalid values raise ValueError containing max_steps. done and stopped states are unchanged.
2. Only ready accepts a decision event; otherwise raise ValueError containing pending. A tool action contains {type: "tool", id, name, arguments}. Allow inspect_dag, read_logs, read_file, apply_patch, run_checks, and open_merge_request. Require a nonblank unused string ID and dictionary arguments; otherwise raise ValueError containing action. Store only id, name, arguments in pending, append id to seen, and increment steps. Reads and checks set tool_pending. apply_patch and open_merge_request set awaiting_approval.
3. Before scheduling open_merge_request, require a previous run_checks observation whose result has passed: true and revision equal to action.arguments.revision. Otherwise raise ValueError containing checks. The supplied tool also rechecks the current workspace revision before publishing; a past passing test is not enough after a new edit.
4. An approval event contains call_id and approved (a boolean). Only the exact pending call in awaiting_approval can be approved. Reject stale IDs, wrong states, or nonboolean approval values with ValueError containing approval. Approval cannot replace tool arguments. True sets tool_pending. False returns ready, clears pending, and records {call_id, name, result: {error: "approval_denied"}}. Approval consumes no model step.
5. A tool_result event contains call_id and a dictionary result. Only the matching pending call in tool_pending can finish; otherwise raise ValueError containing call_id. Append {call_id, name, result}, clear pending, and return ready. Results already in observations are unchanged replays. Failed checks remain observations so the agent can repair and test again.
6. A final action is {type: "final", text}; require nonblank text or raise ValueError containing action. Store {text}, add one step, set done. Before accepting any decision, stop and clear pending if steps >= max_steps. Unknown event types raise ValueError containing event. This function schedules work; it must not read files, run commands, call a model, or publish an MR.