Writ

worktrees-hives Skill

Installable agent skill for the worktrees-hives Rust enforcement core.

AGENTS.md is the authoritative repository contribution and autonomy contract. This portable skill supplies platform-neutral procedures and must not broaden or relax that policy.

When to use

Use this skill when:

• Discovering work from GitHub or Linear issues
• Spawning worker subagents for code changes
• Running Safe Issue → Verified Commit then Safe Verified Commit → PR
• Handing a pull request to the installed companion babysit-pr skill when interactive monitoring is needed
• Executing a human-requested one-shot merge after the automated workflows end
• Reporting results back to the operator

Authoritative safety policy

Before any mutation, read and apply the corresponding AGENTS.md sections:

• core prohibitions and deny-list
• human-authorized one-shot merge protocol
• force-with-lease allow-list
• attribution semantics
• team-maintainer operating model

This skill never grants an exception to those rules. Worker, orchestrator, scheduled, discovery, issue-to-PR, and companion-skill monitoring flows never merge, including by locally merging another PR or stacked/peer branch. If the authoritative policy is unavailable, contradictory, or cannot be enforced by the Rust boundary -- wh itself, or an enforcing wrapper that routes the mutation through wh-core's allowlist and branch verification -- stop the mutating flow and report the blocker.

Branch/worktree pre-edit checklist

Before making any code change, the agent MUST verify:

1. Worktree isolation: pwd is inside the assigned worktree path ({worktree_root}/{owner}/{repo}/{job_id}).
2. Branch correctness: git branch --show-current matches the assigned feature branch.
3. Clean state: git status shows no uncommitted changes from other work. A dirty or stale primary checkout is preserved and is not a reason to abort isolated work.
4. Remote alignment: For a newly created, unpublished assigned branch, fetch the intended remote base and prove that the branch equals that exact remote-base commit before edits; it may lack an upstream only for this creation proof. For a published assigned branch, fetch and verify its expected upstream and the expected local/remote relationship instead of comparing the branch with the base. Stop on an unexpected upstream, unexpected remote commit, behind state, or divergence.
5. No cross-boundary edits: No file outside the worktree is modified (no ../ paths, no absolute paths outside the worktree root).

Repair a clean bootstrap source or a newly created unpublished branch's verified-base alignment before editing. Abort and report an unsafe identity or path mismatch, a genuine ownership collision, or any other non-recoverable failure.

Validation and final publication sequence

Run focused, task-relevant gates while working. After the first tested implementation and before final publication, obtain one independent review matched to the change's risk. Additional review is required only for a named high-risk boundary or a reproduced finding that warrants follow-up.

Before first publication of an unpublished assigned branch, fetch the verified remote base and complete final alignment or rebase. For a published branch, fetch and reconcile it with its expected upstream; do not require equality with the base. Then run exactly one complete native gate suite on the exact HEAD that would be pushed. An issue may add focused checks; it must not replace or reduce that final suite. Any later tree change invalidates that run and requires restoring the applicable alignment and rerunning the suite before push.

Final status guidance

When handing off a pull request, report:

• PR status: Open / Ready for review / Blocked
• Residual issues: List of unresolved CI failures, review comments, or blockers
• Agent attribution: Every PR comment and commit message includes agent identification

A worker or companion-skill monitoring agent MUST NOT claim it merged the PR. A primary interactive agent may claim a merge only after it performed and verified the authorized one-shot operation. If another actor merged the PR, report that without taking credit.

Platform-neutral worker prompt template

When spawning a worker subagent, include these safety instructions in the prompt:

SAFETY RULES (non-negotiable):
- NEVER merge a PR or invoke any merge API/CLI
- NEVER locally merge another PR or stacked/peer branch
- NEVER use bare `git push --force` or `git push -f`
- `git push --force-with-lease` is allowed only for rebasing your own branch
- NEVER edit files outside your assigned worktree
- One writable worker per assigned worktree and branch
- Before editing, verify: worktree path, branch name, clean assigned state, and remote alignment; exact remote-base equality applies only to a newly created unpublished branch, while a published branch must match its expected upstream relationship
- Repair a clean bootstrap source or unpublished verified-base alignment; abort on unsafe identity or path mismatch
- After the first tested implementation and before publication, obtain one independent risk-matched review; add review only for a named high-risk boundary or an actual finding
- After pushing, reply with SHA and agent attribution

Worker prompts remain strictly non-merging. Do not forward the primary agent's merge authorization to a worker or subagent.

Enforcement routing

This skill is portable procedure, not a security boundary. Route orchestrated mutations through wh, with Rust enforcing the runtime boundary as defined in AGENTS.md. The separate interactive host merge path is available only to the primary agent after it completes the linked one-shot authorization protocol; never forward that authority to a worker.