Skip to content

Ark Managed Agent Issue-Fix Qualification v0

This protocol qualifies LoopX issue-fix work when one LoopX goal prompt is submitted to an Ark Managed Agent Goal host. It composes existing issue-fix and host contracts; it does not introduce another runtime or another prompt family.

The qualification is deliberately staged. A repaired file is useful evidence, but it is not proof that the Goal host reached a durable terminal state.

Evidence stages

Stage Required evidence Pass condition
1. Intake and route Public-safe issue metadata and an issue_fix_feasibility_v0 packet Exactly one route is selected. A fix_pr route has a named repro, bounded scope, and named validation.
2. Worker repair Repro-before, minimal patch, focused validation-after, and changed-file summary The repro fails before the patch, validation passes after it, and the artifact is review-ready without claiming an external write.
3. Durable LoopX closure Validated todo writeback plus explicit successor or no_followup The selected todo is durably closed and a new quota should-run no longer asks the worker to repeat the repair.
4. Goal host closure Host events for Goal evaluation and terminal session state Evaluation ends satisfied, the Goal becomes achieved, and the session returns idle without a provider, evaluator, or transport error.
5. Review handoff Review packet and explicit authority state The packet is ready for review. PR creation, review request, merge, and publish remain false until separately authorized.

Stages 1–3 qualify the issue-fix worker path. Stages 1–4 qualify the one-shot Goal host path end to end. Stage 5 qualifies the handoff boundary; it does not grant publication authority.

Verdict rules

Use these verdicts instead of one overloaded success bit:

Observed result Worker verdict Goal-host verdict Meaning
Repro, patch, validation, and durable todo closure all pass; Goal evaluation ends satisfied pass pass End-to-end host case passed.
Worker evidence passes; Goal evaluator or provider fails before satisfied pass fail The repair is real, but the host case is not complete. Diagnose the evaluator/provider boundary.
Code changes, but focused validation or durable writeback is missing fail not reached Do not treat a plausible diff as a completed issue fix.
Goal reports satisfied without a validated repair artifact fail invalid The evaluator result is insufficient and the case must be rejected.
Review packet is ready but external-write authority is absent pass pass or not applicable Stop at draft/review handoff; do not publish.

issue_fix_validated_fix_artifact_v0 intentionally owns worker evidence only. It does not contain a Goal terminal-state field. Goal satisfaction must come from the host event stream or an equivalent host readback, never from the presence of a patch.

Representative matrix

An L2 qualification run should cover more than one happy-path issue:

  1. a bounded issue that selects fix_pr, reproduces locally, applies a minimal patch, and passes focused validation;
  2. an issue with useful diagnosis but insufficient repair evidence that routes to comment_only and remains externally gated;
  3. an issue with neither a safe repair nor useful comment payload that routes to triage_only with no fabricated follow-up;
  4. a one-shot Goal-host run using the same generated task_body as the cloud host, including durable todo closure and terminal Goal readback; and
  5. a review-ready handoff that proves no comment, PR, merge, or publish action occurred without authority.

The deterministic repository fixture covers the repair mechanics. At least one authenticated host run is still required for stage 4; a mock provider or a successful patch alone cannot close that requirement.

Current live finding

A fresh Goal-host parity replay installed the CLI and workflow skills from one clean revision, then submitted one generated task body. The real issue-fix worker reproduced the retry-delay defect, applied the one-line repair, passed all four focused tests, wrote the review handoff, and closed its LoopX todo with explicit no_followup.

The next quota read returned terminal_no_followup with complete user and agent todo sources and no acceptance gaps. A state-aware Goal evaluator then observed that durable terminal state, emitted a satisfied evaluation, and the session returned idle. The run completed 14 authenticated provider exchanges. The correct classification is therefore repair=pass, durable_closure=pass, goal_host=pass; publication remains a separate gated handoff.

Private prompts, credentials, provider payloads, local paths, and raw traces are intentionally excluded from this public protocol.

Durable checks

Run the focused contract:

python -m pytest -q tests/test_ark_managed_agent_issue_fix_matrix.py

The test composes the current feasibility, deterministic repair, Goal-host, and review-handoff contracts. It also prevents the validated repair artifact from silently acquiring or implying Goal terminal-state authority.