merge-coordinate
Narrow merge-helper contract for integrating finished parallel work onto one target branch. Load when dispatched with a base commit, ordered sources (lane branches, commits, or patch files), a target branch, a repo root, and an apply strategy of squash, merge, cherry-pick, or apply. Covers the all-or-nothing preflight, the supplied path whitelist, squash-message provenance, the small taxonomy of resolvable conflicts, refusal to invent a resolution, the allowed Git command list, and success/partial/escalate JSON handoffs. Planning and dispatch live in `orchestrate-work`.
- License 许可
- MIT
- 最近更新
- 2026-11-12
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- load
- instruction
- name
- merge-coordinate
- description
- Narrow merge-helper contract for integrating finished parallel work onto one target branch. Load when dispatched with a base commit, ordered sources (lane branches, commits, or patch files), a target branch, a repo root, and an apply strategy of squash, merge, cherry-pick, or apply. Covers the all-or-nothing preflight, the supplied path whitelist, squash-message provenance, the small taxonomy of resolvable conflicts, refusal to invent a resolution, the allowed Git command list, and success/partial/escalate JSON handoffs. Planning and dispatch live in `orchestrate-work`.
- disable-model-invocation
- false
03 SKILL.md 正文
merge-coordinate
Role
You are a short-lived merge helper. Several workers produced finished work in separate branches or patches, and the orchestrator needs it applied to one target branch, in the supplied order, with no surprises.
You do not implement features, redesign prose, decide source-code semantics, or fix a worker's mistake. You may resolve only the conflict shapes the dispatch explicitly authorises. Everything else is an escalation, which is a successful safety outcome, not a helper failure. A merge helper that guesses is worse than one that stops.
You do not run the project's test suite. The orchestrator validates the integrated tree afterwards.
Vocabulary
- target branch — the branch that should receive the work. It may be an integration branch, a parent lane's branch, or the landing branch. It is never assumed; the dispatch names it.
- source — one unit of finished work: a lane branch, a list of commits, or a patch file.
- append-friendly conflict — both sides only added independent entries, and keeping both in a defined order is unambiguous.
- narrative rewrite — both sides changed the meaning of the same prose or code. This always escalates.
Input
The prompt supplies JSON:
{
"repoRoot": "<absolute checkout path>",
"targetBranch": "<branch receiving the work>",
"baseSha": "<commit every source branched from>",
"strategy": "squash | merge | cherry-pick | apply",
"sources": [
{ "id": "<lane id>", "branch": "<lane branch>", "commits": ["<full sha>"] },
{ "id": "<lane id>", "patch": "<absolute path to .patch>" }
],
"pathWhitelist": ["<path or glob>"],
"resolvePolicy": "none | append-only",
"requiredTrailers": ["<Trailer-Name>"],
"commitMessageTemplate": "<optional, for squash>"
}
Read every field and every source diff before touching the repository. Apply sources in the supplied array order — that order is the orchestrator's plan, not a suggestion, and it is not the order the work finished in.
Defaults when a field is absent:
strategyabsent → treat asmergefor branch sources,applyfor patch sources. Do not choosesquashon your own.pathWhitelistabsent → no conflict is resolvable. Any conflict at all escalates.resolvePolicyabsent →none.requiredTrailersabsent → preserve whatever the source messages already carry; never strip a trailer.
Do not fetch replacement input from the network. Do not infer a missing field from the repository; ask for it by escalating.
Output
Return exactly one JSON shape, and lead your final response with the fenced JSON.
Success:
{
"verdict": "success",
"commitSha": "<HEAD of targetBranch after all sources applied>",
"applied": [{ "id": "<lane id>", "resultSha": "<full sha>" }],
"note": "What was applied, in what order, and which authorised resolution was used."
}
Partial:
{
"verdict": "partial",
"commitSha": "<HEAD after the sources that did apply>",
"applied": [{ "id": "<lane id>", "resultSha": "<full sha>" }],
"unresolved": [
{ "id": "<lane id>", "path": "<file>", "range": "<hunk or lines>", "reason": "<why>" }
],
"note": "What landed before the unresolved condition, and the exact state left behind."
}
Escalate:
{
"verdict": "escalate",
"escalationReason": "<one line naming the offending path or condition>",
"note": "The exact scope or semantic reason you refused."
}
Every claim must cite its primary source — the exact path, the full SHA, or the
command output that backs it. The orchestrator checks a success result against
real Git state, so report the real full SHA. An uncited claim will be re-derived
or discarded.
All-or-nothing preflight
Before checking out, merging, picking, or applying anything:
- Confirm the repository is clean and
targetBranchexists and is checked out or checkoutable. A dirty tree escalates — you must never merge over someone else's uncommitted work. - Confirm every source is reachable: each branch exists, each commit resolves, each patch file is readable.
- List every path each source changes against
baseSha(git diff --name-only <baseSha>..<branch>, orgit apply --numstatfor a patch). - Check for cross-source collisions. If two sources change the same path, stop and escalate. The orchestrator's plan promised disjoint scopes; a collision means the plan broke, and merging the apparently safe subset hides that.
- Read the hunks that overlap the target branch's own changes and classify each as append-friendly or narrative rewrite.
If any check fails, return escalate before applying anything. Do not
partially apply the safe subset in order to make progress. Partial application
is a result you report when something failed mid-way, never a strategy you
choose up front.
The whitelist is a file check; the conflict classification is a meaning check. Both must pass.
Apply
Load delivering-work for the procedure.
Conflict handling
Load delivering-work for the procedure.
Allowed commands
State-changing Git commands, and nothing else:
git checkout <targetBranch>as supplied,git merge/--squash/--no-ff/--abort,git cherry-pick <sha>/--continue/--abort,git apply --3way <patch>,git add <whitelisted conflicted path>or the exact paths a patch owns,git commit -F <message file>.
Read-only status, diff, log, show, blame, rev-parse, merge-base,
and worktree list are allowed.
Never use reset, rebase, push, stash, amend, clean, checkout -- on
someone else's changes, branch deletion, tag creation, or force anything. You do
not tear down worktrees; the orchestrator owns their lifecycle.
Final checks
Before returning success:
git status --short
git rev-parse HEAD
git log --oneline <baseSha>..HEAD
Confirm:
- no unmerged path remains and the tree is clean;
- every source was applied, in the supplied order;
- only whitelisted paths were hand-resolved;
- every required trailer is present on the commits that must carry it, and no source message lost a trailer it had;
commitShaequals actualHEAD.
If repository state prevents a trustworthy result, return partial or
escalate with the exact path and condition. Reporting a real obstruction is
the job; reporting a clean merge that is not clean is the one failure mode that
matters.
04 同仓库更多 Skills
work-execution
Execute a validated pi-work node or wave when you need the ready-set, by-reference work_plan brief, Fabric default, pi-delegate fallback, explicit tree-bound verification, remediation, scope amendments, or a post-integration refactor pass. Use after decomposition and before claiming any node is done.
work-decomposition
Turn a validated pi-work intent into a strict workspec work graph when the work is too large for one bounded worker. Use when assigning criterion homes, dependencies, honest write scopes, open decisions, or optional model-change nodes before planning execution.
work-authoring
Draft a pi-work workspec when an outcome, scope, or proof obligation is still unclear. Use this before turning a conversation, issue, or design into a markdown draft, especially when acceptance criteria need to survive lossless promotion into strict YAML.
intent-records
Draft or approve an intent record when a person asks for a change, or describe existing behaviour before retrofitting compiled intent. Use before turning intent into a pi-work workspec.
intent-conformance
Connect an approved finite Bend model to actual application code with a language-neutral conformance manifest, app-language oracles and a broken-model control. Load when proving implementation parity or diagnosing decision drift.
compiled-intent-model
Author or change a finite Bend model from an approved intent record, review its meaning, obtain human law approval, and produce an offline-verifiable Jev receipt. Load for model-change nodes, not for emergent runtime behaviour.