slop-fix-batch git:20261012.c2ad695A

Immutable. This exact content never changes and is served at /api/v1/blob/42f0dafbfdee0d76.

---
name: slop-fix-batch
description: "Fix a fixed batch of open repository issues selected by human-supplied conditions using parallel subagents and scoped stack executors, group related fixes, publish each verified group immediately in dependency-based PR stacks with scoped lower-layer repairs during construction, and directly manage ongoing repairs and conditional whole-stack merging after verified publication. Start only when the human explicitly invokes $slop-fix-batch in Codex or /slop-fix-batch in Claude Code."
disable-model-invocation: true
---

# Slop Fix Batch

Start a batch only on an explicit human `$slop-fix-batch` (Codex) or `/slop-fix-batch` (Claude Code) invocation with issue-selection conditions. Loading, editing, testing or planning this skill does not start a batch. Collect the complete canonical fixed list once; never add later discoveries. `$bulk` and `$slop-fix-issue` publish separately and are not batch workers. Never create implementation or repair sidebar chats.

The invocation authorizes scoped investigation/implementation, evidence-backed closure of already-resolved issues, verified stacked publication, and conditional management/merging of this batch's ready stacks. The objective is verified merging, not PR creation or merge-request acceptance. Before each new assignment at a safe boundary, process due ready-stack gates, merges and result verification. External CI/review waits permit independent construction.

Keep management in this coordinator chat through construction, repair and verified completion. Do not create automations, heartbeat jobs or background polling processes. Use interruptible waits of at most 30 seconds only when no independent work is eligible; honor refresh deadlines and shared cooldowns. Recover the same ledger and reconcile owners/effects after interruption.

## Read only the current procedure

Read a reference once, then reuse it unless its contents change. Locate headings with `rg` and read the relevant sections instead of loading every reference or dependency. All roles use [validation](references/validation.md), [runtime tools](references/runtimes.md)' common rules plus the recorded runtime section, and worker-prompts' [settings/context](references/worker-prompts.md#worker-settings-and-context), [assignment verification](references/worker-prompts.md#render-and-verify-assignments), and [shared API policy](references/worker-prompts.md#shared-api-policy-for-every-handoff).

- **Coordinator:** use [construction](references/construction.md) for selection/grouping/base decisions and closure, [scheduling](references/scheduling.md) at allocation boundaries, and [delegation](references/delegation.md) for reservations/grants/recovery. Read the [helper contracts](references/helpers.md) sections for commands in use. Read maintenance's direct loop/checkpoints/accounting when the first verified prefix is published. For early repair signals, use [published building-prefix repair](references/batch-maintenance.md#repair-a-published-building-prefix); do not execute an assigned actor's tail.
- **Investigation/preparation:** read only the assigned [investigation](references/worker-prompts.md#investigation-assignment) or [preparation](references/worker-prompts.md#dependent-group-preparation) handoff. Inspect assigned issues/code/history and report evidence, acceptance, overlaps and dependencies. No source or GitHub writes.
- **Implementation:** read the [implementation handoff](references/worker-prompts.md#implementation-assignment). Implement/check only its confirmed group at the exact assigned prerequisite SHA; no PR publication, issue writes or merging.
- **Stack executor:** one checked group in one building stack. Read [stack execution](references/delegation.md#stack-execution-job), its [handoff](references/worker-prompts.md#stack-execution-assignment), and relevant ledger/publication helper contracts. Invoke `$create-pr` and `$write-ste` at publication. Own integration, scoped ledger events, registration, gh-stack prefix submission/verification and complete-stack ready transition; never merge.
- **Maintenance:** one ready stack. Read [bounded maintenance](references/batch-maintenance.md#each-bounded-batch-pass), ownership/recovery and the applicable merge-gate/result sections. Read repair dispatch/alignment sections only when those signals arise. Load `$merge-pr` when merging, `$repair-pr` when dispatching repairs, and `$manage-stacked-prs` only for proven alignment work. Repair children need coordinator-reserved slots; without one, return the verified signal and slot.
- **Inventory:** one captured read-only pass. Read maintenance's [inventory scope](references/batch-maintenance.md#inventory-scope-and-candidate-queue), [checkpoints](references/batch-maintenance.md#construction-checkpoints), and [checkpoint helper contract](references/helpers.md#checkpoints-and-inventory). Record complete versus partial coverage and return the slot; never mutate GitHub or replace fresh merge gates.

In Plan Mode, use [planning investigation and handoff](references/planning.md) before collecting/assigning issues. It permits read-only investigation only and requires complete fixed-list coverage before the final plan.

## Shared execution constraints

Record `context.runtime` as `codex` or `claude-code`; an absent field in an old ledger means Codex. Resume that runtime without switching workers. Keep this skill in the coordinator conversation without `context: fork`. Resolve installed dependency/reference paths when needed; attachment requirements follow runtime policy and GitHub verification always applies.

New execution ledgers activate scoped delegation at an idle pre-dispatch boundary; existing ledgers remain coordinator-only until explicit safe-boundary activation. Use worker-settings.mjs for every new worker, including repairs. Workers start without parent history; pass a compact verified handoff and full evidence file paths. Preserve existing actors' creation settings. No default total concurrency cap is added.

Every active/reserved/nested/uncertain actor consumes human/runtime capacity, including maintenance and inventory. Use scheduling's lane priorities, publication ceiling, progress exception and ownership accounting. Preserve explicit human allocations. Bounded actors return only after known effects are recorded; unknown effects retain slots/ownership. Verified standalone auto-merge waits follow delegation's retained-ownership rule. Retain building-stack protection between jobs.

Preserve protected ancestry: no prerequisite copying across stacks or incidental construction rebases. A verified repair may change a published lower layer and replay its affected upper layers under transferred whole-stack ownership. Pause same-stack construction at a safe boundary and resume only after verified repair results restore build ownership; completion of every group is required for merging, not repairing. Worker checks cover their exact SHA; unchanged cherry-picks reuse that coverage, integration changes need relevant checks, and no cumulative/final-tip validation runs. Disclose skipped real-world validation. Direct issue closure requires all acceptance evidence.

Originally multi-PR stacks prefer native whole-stack squash merging after fresh complete gates for every remaining layer; sequential fallback requires verified unsupported capability. Only originally standalone PRs may enable auto-merge. Complete review reads pass only with no effective change request or unresolved non-outdated finding. Use maintenance's full gates and recovery procedure.

All roles share one batch GitHub queue/cache/cooldown state and complete paginated observations. Cache TTL, candidates and scheduler hints never authorize merging. Invalidate representations around every write attempt; reconcile uncertain writes before retrying. Preserve Git/stack ownership and serialized catalog mutation. Record direct gh/extension coverage gaps and missing-header limits. Follow GitHub's [REST best practices](https://docs.github.com/en/rest/using-the-rest-api/best-practices-for-using-the-rest-api) and [rate-limit rules](https://docs.github.com/en/graphql/overview/rate-limits-and-query-limits-for-the-graphql-api). Use helpers for deterministic collection, ledger transitions, construction, publication verification, checkpoints and handoffs; helpers do not decide root causes or merge authority.

Keep complete state/results in batch files. CLI summaries and worker messages carry new facts, changed identities and blockers; do not echo full ledgers, unchanged remote responses or prior reports. See [output and evidence contracts](references/helpers.md).