agent-merge-conflict-arbiter 精选
Neutral arbiter for merge conflicts between two agents.
安装方式
cp -r optional-skills/autonomous-ai-agents/agent-merge-conflict-arbiter ~/.claude/skills/agent-merge-conflict-arbiter把该 skill 目录放进客户端 skills 目录。不同客户端的目录可能不同,Claude Code 默认是 ~/.claude/skills/
https://raw.githubusercontent.com/NousResearch/hermes-agent/0b89435e7e4f7fab88a34eb69f13bf339eb2fd00/optional-skills/autonomous-ai-agents/agent-merge-conflict-arbiter/SKILL.md下载后另存为 ~/.claude/skills/agent-merge-conflict-arbiter/SKILL.md。链接锁定采集时的 commit,内容与本站展示一致。
整仓一次性获取:git clone --depth 1 https://github.com/NousResearch/hermes-agent.git
依赖与权限
- 无需 API Key
- 无需联网
- 依赖外部命令:
gitmake - 以上为按正文与 frontmatter 自动推断的信号,仅供参考;安装前建议先阅读原文。
SKILL.md 正文
Agent Merge-Conflict Arbiter
Resolve a git merge conflict between two AGENTS' branches as an impartial third
party. Agents resolving conflicts against a peer's work reliably either
overwrite the peer or abandon their own change — they lack the peer's context
and are biased toward their own side. This skill is the fix: a neutral
reconciler that receives both diffs plus both sides' stated intents and
produces a merged result, like a merge-queue arbiter.
When to Use
- Two agent branches/worktrees collide during a parallel campaign (kanban
engineering pipeline, parallel-PR wave, multi-worktree refactor).
git mergeorgit rebasehalts on conflicts between two agents' work and
neither original agent should self-adjudicate.
- Do NOT use for conflicts within a single agent's own work, or for trivial
lockfile/generated-file conflicts (regenerate those instead).
Prerequisites
- A repo checkout containing the halted merge, or the two branch names plus
permission to run the merge yourself.
- Both sides' intent sources: kanban completion summaries (
terminalrunning
hermes kanban show <task-id>), PR bodies, or at minimum each branch's
commit messages.
- The project's build/test command, if one exists.
How to Run
Standalone — a human (or agent) invokes this skill inside the conflicted
repo: load the skill, then follow the Procedure top to bottom.
Spawned neutral agent — the preferred shape in multi-agent campaigns:
delegate_task: spawn a subagent whose task message contains the repo path,
both branch names, and both sides' intent summaries verbatim, plus an
instruction to follow this skill.
- Kanban-native: create a reconciliation card assigned to a third profile
(not either worker's profile) with BOTH conflicted cards linked as parents —
`kanban_create(title="reconcile branch-a x branch-b", assignee="reconciler",
parents=["t_a", "t_b"])`. The parent links carry both sides' completion
summaries into the reconciler's context automatically; the card body should
name the repo path and the two branches.
Quick Reference
| Hunk class | Definition | Resolution |
|---|---|---|
| disjoint-intent | The two changes serve different goals and can coexist | Combine both |
| same-question-different-answer | Both sides answered one design question differently | Pick ONE per stated intents; surface the decision |
| superseded | One side's premise no longer holds after the other's change | Keep the surviving side; note why |
Impartiality contract: never favor the side that spawned you; touch ONLY
conflicted regions (no drive-by edits); every design-question pick must appear
explicitly in the hand-back summary.
Procedure
1. Gather both sides
- Run via
terminal:git status(confirm the conflicted state and list
conflicted files), git merge-base <A> <B>, then for each side
git log --oneline <base>..<side> and git diff <base>..<side> -- <file>
for every conflicted file. In a halted merge, HEAD is one side and
MERGE_HEAD is the other.
- Collect each side's intent:
hermes kanban show <task-id>for completion
summaries/metadata, or the PR body, or the commit messages from the log
above. Write down one sentence of intent per side before touching any file.
- Done when: you can state both intents in your own words and have both diffs
for every conflicted file.
2. Classify every conflicted hunk
- Open each conflicted file with
read_fileand locate each
<<<<<<</=======/>>>>>>> block.
- Assign each hunk exactly one class from the Quick Reference table, judging
by the stated intents — not by which change looks nicer.
- If a single hunk contains multiple independent decisions (e.g., new logic
that combines cleanly PLUS a styling/rounding choice both sides answered
differently), decompose it into sub-decisions and classify each one.
- A single file often mixes classes: one hunk may be a design collision while
a neighboring hunk is disjoint. Classify per hunk, not per file.
- Done when: every hunk has a written class and a one-line rationale.
3. Resolve under the impartiality contract
- Edit each hunk with
patch(orwrite_filefor whole-file rewrites): - disjoint-intent → merge both changes so each intent is fully served.
- same-question-different-answer → pick the answer that best serves the
STATED intents (e.g., an intent of "strict validation" beats "quick
default" if the task required correctness). Never split the difference
into a hybrid neither side asked for.
- superseded → keep the surviving side; delete the dead premise.
- Never favor the side that spawned you. If intents genuinely tie, escalate
(block the kanban card / report back) rather than guess.
- Change nothing outside conflict markers — no formatting, renames, or
opportunistic fixes.
git addeach resolved file viaterminal.- Done when:
search_filesfinds no<<<<<<<markers in the repo and every
resolved file is staged.
4. Verify
- Run the project's build/tests via
terminal; at minimum import/execute the
touched modules. Both intents must be observable in the merged behavior
(e.g., side A's new semantics AND side B's disjoint addition both present).
- Complete the merge:
git commit(the default merge message plus a body
listing hunk decisions is fine).
- Done when: verification passes and the merge commit exists.
5. Hand back
- Produce a completion summary naming EVERY hunk decision:
file:lines — class — which side(s) kept — rationale. For every
same-question-different-answer hunk, state the design question and the
answer you picked so a human can veto it — never bury a design call.
- Kanban:
kanban_complete(summary=...). Standalone: print the summary. - Done when: the summary is delivered and lists all hunks.
Pitfalls
- Self-favoring: if you were spawned by one of the conflicting agents,
you are structurally biased — state this and weigh the other side's intent
deliberately. Prefer the third-profile shape so this never arises.
- Splitting the difference on a design collision produces a hybrid nobody
designed; pick one answer and surface it.
- Per-file classification: files usually mix hunk classes; classifying a
whole file as one class silently drops a disjoint change.
- Drive-by edits make the merge unreviewable and steal decisions from the
original agents.
- Missing intents: commit messages alone can be thin; prefer kanban
completion summaries or PR bodies. If neither side's intent is recoverable,
escalate instead of guessing.
- Repeat offenders: repeated conflicts on the SAME file across rounds are
a hotspot signal, not routine reconciliation work — flag it (e.g. a
hotspot: <path> — <reason> kanban comment) so the orchestrator decomposes
that file, rather than serially reconciling every new collision on it.
Verification
git statusshows a clean tree on the target branch with a merge commit.- No conflict markers remain (
search_filespattern<<<<<<<). - Build/tests pass; both sides' intents are demonstrably present or the
dropped one is explicitly named in the summary.
- The hand-back summary enumerates every hunk with class and rationale.
质量与来源
- 所属仓库
- NousResearch/hermes-agent
- 仓库星标
- 25 万
- 最近更新
- 今天
- 采集时间
- 今天
- commit
- 0b89435e7e4f
- 内容指纹
- sha256:f8b1a0cf2392a1d
- 文件路径
- optional-skills/autonomous-ai-agents/agent-merge-conflict-arbiter/SKILL.md