pr-batch

Files (109)

• SKILL.md
• openai.yaml · agents/
• agent-claimed-mirror-contract-test.rb · bin/
• agent-coord-bounded · bin/
• agent-coord-bounded-test.rb · bin/
• assignee-exclusion-contract-test.rb · bin/
• autonomous-merge-calibrate · bin/
• autonomous-merge-calibrate-test.rb · bin/
• autonomous-merge-closeout · bin/
• autonomous-merge-closeout-test.rb · bin/
• autonomous-merge-contract-test.rb · bin/
• autonomous-merge-eligibility · bin/
• autonomous-merge-eligibility-test.rb · bin/
• autonomous-merge-evidence-test.rb · bin/
• batch-usage-receipt · bin/
• batch-usage-receipt-test.rb · bin/
• coordination-declaration · bin/
• coordination-declaration-contract-test.rb · bin/
• coordination-telemetry-contract-test.rb · bin/
• current-integration-evidence-test.rb · bin/
• dependency-conflict-policy-contract-test.rb · bin/
• diff-identity · bin/
• diff-identity-test.rb · bin/
• dispatcher-capability-preflight · bin/
• dispatcher-capability-preflight-test.rb · bin/
• file-expansion-contract-test.rb · bin/
• github-actor-trust-test.rb · bin/
• github-comment-envelope · bin/
• github-comment-envelope-test.rb · bin/
• goal-completion-contract-test.rb · bin/
• goal-state-change-monitor · bin/
• goal-state-change-monitor-test.rb · bin/
• hosted-qa-gate-contract-test.rb · bin/
• hosted-qa-readiness · bin/
• hosted-qa-readiness-test.rb · bin/
• human-attention · bin/
• human-attention-contract-test.rb · bin/
• human-attention-test.rb · bin/
• integration-closeout-contract-test.rb · bin/
• merge-assurance · bin/
• merge-assurance-test.rb · bin/
• model-routing-contract-test.rb · bin/
• pr-body-human-first-contract-test.rb · bin/
• pr-ci-readiness · bin/
• pr-ci-readiness-test.rb · bin/
• pr-merge-submit · bin/
• pr-merge-submit-test.rb · bin/
• pr-security-preflight · bin/
• pr-security-preflight-test.rb · bin/
• production-release-contract-test.rb · bin/
• review-wave-contract-test.rb · bin/
• security-floor-contract-test.rb · bin/
• single_target_entrypoint_test.rb · bin/
• skill-stage-source-test.rb · bin/
• stage-dependency-gate · bin/
• stage-dependency-gate-test.rb · bin/
• stale-assignment-sweep · bin/
• stale-assignment-sweep-test.rb · bin/
• target-membership-guard · bin/
• target-membership-guard-test.rb · bin/
• …and 49 more

SKILL.md (first lines)

# PR Batch

For new Codex planning, resolve the advisory `astra-pilot-v1` profile from
[central routing data](../plan-pr-batch/references/model-routing-profiles.json) with the plan skill's
`bin/model-routing-profile --role <role>`. Its named preferences supersede the
GPT-5.6 recommendations below for the listed roles; those recommendations and
planning tables remain the established comparison baseline. Keep explicit user
routes, verified host support, portable fallback, and independent evidence rules.
This is an unmeasured pilot, not a measured promotion.
If the current host is Cursor, resolve the advisory `cursor-grok-v0` profile
the same way; every role stays portable/generic until measured.
If a partial or pinned installation lacks the resolver or data, continue with
established or portable advisory routes; use the complete pack to access the pilot.

Run one or more PR work lanes through one canonical process. A single target is
a batch of one, not a separate workflow.

Resolve writing style before authoring human-facing prose. Run
`agent-workflow-writing-style --repo-root <trusted-repository-root> --format json`
using the resolution, provenance, warning, trusted-base, and evidence-preserving
contract in `workflows/pr-processing.md` → **Writing Style Resolution**. Apply
the guide to PR descriptions and updates, issue or PR comments, review-facing
explanations, and final handoffs. Never let style remove repository template
sections, required evidence, or machine-readable receipts.

Use `docs/coordination-backend.md` as the canonical vocabulary for private
backend, public fallback, no-backend mode, and `UNKNOWN` coordination state.

If a skill picker only exposes installed/global skills, treat this skill as an
entry point. After fetching, prefer repo-local `.agents/skills/...` and
`.agents/workflows/...` files when they exist; otherwise use the installed
shared files adjacent to this skill, especially `../../workflows/pr-processing.md`.

The completed-batch closeout validation contract requires `pr-batch` and
`post-merge-audit` from the same Agent Workflows pack revision. Its contract
test intentionally loads the production receipt parser from the sibling
`post-merge-audit` skill; an isolated pinned copy must include that companion
or stop with a precise missing-companion blocker.
Apply [Ordinary PR Closeout](../../workflows/pr-batch-integration-closeout.md#ordinary-pr-closeout)
first. Only for a required audit, execute receipt and archive replay through the canonical
[Completed-Batch Audit Receipt And Archive Replay](../../workflows/pr-batch-integration-closeout.md#completed-batch-audit-receipt-and-archive-replay)
component; do not mirror that policy here.

Use the trusted-base `hosted-qa-readiness` helper and the canonical hosted QA
contract in `workflows/pr-batch-integration-closeout.md`; do not reproduce or reinterpret that
contract here.

Memorable invocation:

```text
$pr-batch
Run this task as one PR lane
Run an agent batch
Run a Codex batch
Run a Claude batch
```

<!-- host-branch: cursor-only start -->
Cursor: `/pr-batch` or pin as a Custom Mode. Sequential Task workers plus
`git worktree add` are the supported path; stop if a file-editing worker cannot
isolate. Overnight multi-lane fleets stay blocked until isolation matching
`isolation: 'worktree'` is proven. Never write into `skills-cursor`.
<!-- host-branch: cursor-only end -->

## User-Facing Coordination Contract

The current task is the sole user-facing coordinator. Subagents, lane workers,
reviewers, and QA agents are internal workers owned by the current task, never
separate chats whose mechanics the user must coordinate. External tasks may
send evidence or requests without gaining ownership, and automations only wake
the current task. Apply authority decisions and separate-scope routing through
the shared
[user-facing coordination contract](../../docs/user-facing-coordination.md).

For a heartbeat or monitor, a no-change wake produces no user-visible
notification. Notify only for an HST-v1 actionable material state change: a
decision or action is required, a target is ready for walkthrough or approval,
a blocker exhausted its bounded retries and needs intervention, or
closeout/archive completed; delete the heartbeat when its gate clears or
becomes durably terminal. The automation never owns the task or next action.

## Coordinator Output Contract

<!-- Keep this summary in sync with `.agents/workflows/pr-processing.md` -> `### Coordinator Output Contract`. -->

`OC-v1` bounds coordinator narration volume. Use the canonical
[Coordinator Output Contract](../../workflows/pr-processing.md#coordinator-output-contract)
instead of restating its rules here. In short: bound coordinator narration to
the five typed checkpoints `dispatch`, `pr-open`, `decision-required`,
`merge-decision`, and `final-handoff`; keep recaps delta-only, findings
single-surface, and corrections proportional; and report the shadow-only
`coordinator-narration-volume v1` marker in FYI / decisions made at closeout.
`decision-required` is limited to blockers that need user action; a
coordinator-clearable blocked target is a `merge-decision`.
Four message kinds stay allowed outside those checkpoints and count in the
marker's scalar `always_allowed` bucket: a direct answer, an explicitly
requested status report, a turn another contract requires the coordinator to
show, and a required safety stop. The marker also reports
`always_allowed_detail` per kind and classifies overlaps in this
most-specific-first order: `safety-stop`, `required-turn`, `requested-status`,
then `direct-answer`.

`OC-v1` is presentation only. It relaxes no evidence, verification, or
`UNKNOWN`-honesty rule, drops no required exact string, deletes no durable copy
another contract requires, and collapses no closing structure except for the
compact terminal structure allowed for single-repo batches at or below
`compact_terminal_structure_max_lanes`; the required receipt and final
`Conversation status:` line still stay intact.

## Single-Target Mode

Use this mode for one GitHub issue, existing pull request, or durably overridden
direct-prompt task after the Canonical Launch Target Gate passes. It keeps
the same security, coordination, validation, review, QA, readiness, handoff, and
closeout gates as a multi-target batch; only batch packing and collision analysis
collapse to one lane.

### Prompt Intake

From shakacode/agent-workflows, MIT.

More from shakacode/agent-workflows