swe-marathon-five-arm
Migrate historical SWE-Marathon agent configurations to the shared Codex runtime.
Use when LoopX must manage a multi-PR or multi-MR delivery program across one or more repositories: inventory current change requests, reconcile new/merged/closed or retargeted work, preserve requirement and dependency priorities, maintain a roadmap document, or monitor material
$ npx -y skills add loopx-project/loopx --skill loopx-pr-program --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/loopx-pr-programContext preview
The summary Claude sees to decide when to auto-load this skill.
Use when LoopX must manage a multi-PR or multi-MR delivery program across one or more repositories: inventory current change requests, reconcile new/merged/closed or retargeted work, preserve requirement and dependency priorities, maintain a roadmap document, or monitor material
name: loopx-pr-program description: "Use when LoopX must manage a multi-PR or multi-MR delivery program across one or more repositories: inventory current change requests, reconcile new/merged/closed or retargeted work, preserve requirement and dependency priorities, maintain a roadmap document, or monitor material lifecycle/check/review changes over time. Use provider-neutral snapshots and one grouped continuous monitor; do not use for deep per-PR code review, approval, commenting, or merge actions."
Manage a delivery program as durable LoopX state instead of rebuilding a queue from chat memory. Keep source acquisition provider-local, normalize observations into one public contract, and write back only material transitions.
Use this workflow when the user asks to manage, prioritize, reconcile, document, or monitor several pull requests or merge requests. Route a deep review of each selected PR to `loopx-pr-review`. Route approval, comments, reruns, branch retargeting, closing, or merging to the normal provider-specific workflow and require the corresponding authority.
Treat document maintenance and monitor creation as writes. A request to update the roadmap or keep monitoring the program authorizes those scoped writes; an ordinary status question remains read-only.
Start or join the project goal before material work. Represent the program with one advancement todo for the current reconciliation and one `continuous_monitor` todo for recurring observation. Do not create one monitor todo per change request.
Use a stable monitor identity:
task_class=continuous_monitor action_kind=pr_program_reconcile target_key=pr-program-<stable-program-id> cadence=<user cadence or 30m>
Preserve `claimed_by`, `last_checked_at`, `next_due_at`, `result_hash`, `consecutive_no_change`, and `material_change`. Quiet monitor polls keep liveness but do not count as delivery progress, rewrite roadmap documents, or spend progress quota.
Use any authorized source-control read interface available in the current environment. Prefer one batch query for inventory and targeted reads for the items that changed. Never encode a private transport command, executable name, credential, internal hostname, or document token in this skill, repository examples, committed fixtures, or public evidence.
Normalize observations using [`references/snapshot-contract.md`](references/snapshot-contract.md). Mark `result_completeness.complete=true` only after proving the requested repository, author, state, and time-window inventory is exhaustive. An incomplete current snapshot must not make absent rows look closed or removed. Persist the structured scope fingerprint with the baseline. Do not advance the durable baseline or grouped-monitor `result_hash` from an incomplete snapshot or when the previous and current scope fingerprints differ; otherwise a partial page or narrowed query can create false remove/re-add transitions on the next poll.
Store raw and normalized snapshots under an ignored owner-local directory such as `.local/loopx/pr-program/<program-id>/`. Verify the path with `git check-ignore` before writing. If no ignored path is available, use a temporary directory and keep only a redacted evidence summary in LoopX state.
Run the bundled delta helper before manually comparing rows:
python skills/loopx-pr-program/scripts/diff_snapshot.py \ --previous <previous.json> \ --current <current.json> \ --output <delta.json>
Omit `--previous` for the first baseline. The helper treats lifecycle, draft, target branch, head revision, checks, review, work item, requirement, theme, priority, and dependency changes as material. Timestamp-only movement is observation noise. Read the actual description, latest review context, checks, and changed-file evidence for every added or materially changed row before updating the program judgment.
Do not infer motivation or priority from title, number, author, age, or green CI alone. Product requirements set priority. Correctness dependencies and real merge gates determine order within a priority. Record a requirement gap explicitly when no change request implements part of the requested behavior; do not call the requirement complete because a neighboring parameter or feature landed.
When several selected change requests belong to one repository, compose this program view with LoopX's `integration-branch-reconcile` capability. Keep the ownership split explicit:
proves what exact code is composed.
Do not populate the plan from every open or P0 change request automatically. Select sources from explicit program scope, dependency order, and current review intent. Refresh the local refs through the authorized host workflow, then verify that each selected ref resolves to the observed `head_sha` before configuring or syncing the integration branch. A mismatch is source evidence drift, not permission to merge a stale local ref.
Use the same grouped program monitor to read both the normalized change-request delta and `loopx integration-branch status --format json`. Treat base/source movement, an unexpected integration head, and merge conflict as material program evidence. A monitor poll remains read-only: it may report that the candidate needs reconciliation, but it must not run `sync --execute` by itself. That command is a separate, explicit local write and never grants authority to push, retarget, approve, or merge a remote change request.
Keep one canonical program projection with three sections:
1. completed work, grouped by product theme; 2. pending work, grouped by product theme, with archived or superseded development s
A control plane with a durable state kernel for long-horizon agents and teams. Keep work moving and improving across sessions, with less human attention.
Repo: loopx-project/loopx
Migrate historical SWE-Marathon agent configurations to the shared Codex runtime.
在 Terminal-Bench 4.0 上做 codex harness 五臂对照(裸 codex / 原生 /goal / LoopX 三模式)。复用 SWE-Marathon…
Inspect authorized LoopX Goals, Todos and deliveries to explain progress, identify owner…
Use when acting as the operator or post-run analyst of a LoopX-managed benchmark experiment…
Qualify the exact final diff for a LoopX-managed goal. Use when goal policy enables…
Use when a connected LoopX project is asked to read, remember, record, index, register, or…