fix(ci): bypass delegation-guard during merge conflict resolution

[aram-devdocs]

Summary

• Allow .claude/hooks/delegation-guard.sh to permit root Rust source edits while the worktree has unmerged paths.
• Add a regression test that creates a temporary conflicted Git repository and verifies root Rust edits are allowed only in that merge-conflict state.

Why

Merge conflict marker arbitration is mechanical cleanup, not new Rust implementation work, so forcing it through 01-implementer conflicts with that subagent's TDD-oriented contract.

Validation

• bash -n .claude/hooks/delegation-guard.sh .claude/hooks/tests/test_delegation_guard_subagent.sh
• bash .claude/hooks/tests/test_delegation_guard_subagent.sh
• git diff --check
• cargo fmt --check passed after installing the pinned Rust 1.95.0 toolchain under /workspace.
• cargo clippy --workspace --all-targets --all-features -- -D warnings could not complete in this container because no C linker (cc, gcc, or clang) is installed. A bundled rust-lld retry also failed due missing libc development linker names/startup environment, so cargo test was not run here for the same environment reason.

[aram-devdocs]

Updated delegation-guard.sh so root Rust source edits are allowed while the worktree has unmerged paths, and added a regression test that creates a temporary merge conflict to verify that behavior. Local shell validation passed (bash -n, delegation guard regression test, git diff --check) and cargo fmt --check passed using the pinned Rust 1.95.0 toolchain installed under /workspace; cargo clippy and cargo test need a human or CI rerun because this container lacks a usable C linker/libc development environment (cc/gcc/clang unavailable, rust-lld could not link libc-backed build scripts).

[github-actions]

I have a full picture. Here is the review.

---

PR #133 — Delegation Guard: Merge-Conflict Bypass

Buckets touched: ci (hooks), gitignore, docs (crates/plumb-mcp/AGENTS.md)
Rust code changed: none — all five blocker categories (unsafe, unwrap, println!, SystemTime, HashMap-in-output) are N/A.

---

Punch list

.claude/hooks/delegation-guard.sh:49–56 — conflict-bypass logic

Correct and narrowly scoped:

• --diff-filter=U lists only files with unresolved merge conflicts; it drops staged/modified files, so the exception disappears the moment Claude resolves and git adds the file. This is verified by test case 3.
• grep -qxF "$rel_path" is exact-line matching — a partial path like lib.rs cannot match crates/plumb-core/src/lib.rs. No false-positive risk.
• || true prevents set -euo pipefail from aborting when the hook runs outside a git repo.
• Placement is correct: the check runs after subagent and infra-path early-exits, so only root-context Rust files under crates/*/src/ reach this branch.
• Security surface is minimal: engineering a fake merge conflict to exploit this bypass requires controlling the git state of the worktree, which is a much stronger position than what this hook defends against.

.claude/hooks/tests/test_delegation_guard_subagent.sh:121–174 (committed version) — three new test cases

• Case 1 (allow conflicted file): correct.
• Case 2 (block unrelated file during conflict): correct — crates/plumb-cli/src/main.rs is not in conflicted_paths, so falls through to the hook_block role check.
• Case 3 (block after git add): correct — once staged, --diff-filter=U stops listing the file, so the bypass no longer fires.

The test repo setup (git merge feature producing a real conflict) is the right approach — it exercises the actual git plumbing rather than mocking it.

crates/plumb-mcp/AGENTS.md

Removing the 09-mcp-tool-author one-liner is fine; the agent is still registered at .claude/agents/09-mcp-tool-author.md and referenced in the subagent-driven-development and gh-issue skills. No dead-reference problem.

The phrase "for the handoff path" is slightly opaque but not wrong. Non-blocking.

.gitignore — .claude-pr/ entry

Clean, no issues.

Working-tree state (not a PR defect)

The review context shows staged/unstaged reverts of the bypass code in the working tree. These are not part of the PR commits and appear to be artifacts of a prior review run. The PR commits at 39ed3d7 are correct.

HOOK_CWD / rel_path interaction

HOOK_CWD is set from $CLAUDE_PROJECT_DIR (or git rev-parse --show-toplevel) in _lib.sh. The test sets CLAUDE_PROJECT_DIR="$merge_repo" so git -C "$HOOK_CWD" runs against the scratch repo with the live conflict. Works as intended. Pre-existing edge case (subdirectory CLAUDE_PROJECT_DIR) is not introduced by this PR.

---

Verdict: APPROVE

[aram-devdocs]

PR #133 is ready to merge: the PR title now uses the allowed ci scope, crates/plumb-mcp/AGENTS.md is back under the 60-line preflight cap, .claude-pr/ is ignored, and Claude review punch-list items were addressed by narrowing the merge-conflict bypass to only the unresolved Rust file plus adding negative regression coverage. GitHub CI is green after rerunning one transient setup-layer Determinism failure, and the latest Claude review verdict is APPROVE.