E2E test: implementation recipe on a seeded sandbox issue #5233

Maintainer setup

• Seed an open issue in TalonT-Org/autoskillit-e2e-sandbox describing a small, self-contained feature with an obvious test.

Everything else comes from #5230 (MINIMAX_API_KEY, E2E_SANDBOX_TOKEN). No Anthropic credential.

Depends on: #5230; #5232 (smoke) should pass first.

What it proves

The bundled implementation recipe completes on MiniMax from issue to open PR: make-plan, dry-walkthrough, implement, test, merge into the run's branch, audit, PR.

Cost and limits

• Steps run one at a time, so the peak is the orchestrator, one step's session and that step's children: 6 during implement (implement-worktree-no-merge names four exploration responsibilities) and up to 8 during resolve-review in the local review loop (three to six groups on a typical PR). make-plan's exploration fan-out depends on the task.
• audit-impl starts one deviation evaluator per plan deviation, so its fan-out depends on the plan; the ceiling below bounds it.
• Architecture lenses stay off (arch_lenses defaults to false); they would run several sessions in parallel.
• The harness's CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS=6 must not go lower, because skills require their declared children to run concurrently.
• For the timeout: a MiniMax implement session took a median 7.0 minutes in the local session index (n = 175).

Scope

• Catalog entry: selected only for larger PRs touching recipes, skills, the server or execution code; never more than one pipeline-class test per PR.
• Run with open_pr=true, auto_merge=false; the harness closes the PR and deletes the branch afterwards.
• Assertions: envelope success; a PR open on the sandbox referencing the seeded issue; the diff touches the sandbox's source and tests; sandbox tests green on the PR branch.

Each defect found gets its own bug issue; the catalog entry is marked expected-failure with the link until fixed. The first candidate to watch for: the fleet run orchestrator ignores providers.default_provider, which the settings-file route is meant to make irrelevant.

Acceptance criteria

• Completes in CI on MiniMax with no Anthropic credential.
• Peak concurrent sessions measured from the session logs and recorded in the catalog entry.

Live commissioning status: 2026-10-09

The first implementation attempt was run 37854525992, using PR #5255 commit 017b3f1 and reusable sandbox issue 1. E2E_IMPLEMENTATION_ISSUE_URL is configured.

Docker setup and the sandbox baseline tests passed. The bundled implementation recipe explicitly supplies output_dir: . to make-plan, whose declared scope is {{AUTOSKILLIT_TEMP}}/make-plan/. The run failed before implementation with: run_skill output_dir is outside the skill's declared write scope.

A resource-limited, network-isolated Docker probe of the exact published source reproduced the mismatch: the root path is rejected, and .autoskillit/temp/make-plan/ is admitted. The probe used about 55 MB peak Python RSS and ran no model or test suite. This is a deterministic planner-dispatch contract mismatch, not an observed memory failure. The separate reporter crash remains tracked in #5249.

No implementation PR or sandbox PR CI success was produced. The seed stays OPEN with a fail label and no claim label. A matching branch remains at the same commit as main, while ownership.json records no owned branch; teardown is not fully verified.

Commissioning remains incomplete. Recommended gate: repair planner admission independently, verify it with the Docker reproducer, then obtain two genuine successful MiniMax implementation runs and complete ownership/teardown evidence before recording full-pipeline peak_sessions_evidence. Expected-failure routing, if explicitly adopted for a provisional harness merge, must link to the defect and must not be described as successful commissioning.

Planner blocker: #5256

The deterministic planner output-path mismatch is now filed separately as bug #5256, labeled bug and recipe:remediation. It includes the exact tested revision, source contract mismatch, live failure evidence, complete self-contained Docker probe/build/run commands, resource limits, observed results, and artifact retrieval commands.

The reporter defect remains #5249. PR #5255 remains open while commissioning is incomplete; the planning-only Docker reproduction is not a passing full implementation run. Next: fix planner admission independently, verify in isolated Docker, and rerun MiniMax twice with complete PR CI and teardown evidence.