harness

name

harness

description

Run a change through the full Murmur harness — isolated worktree, writer, deterministic checks, independent adversarial + risk reviews, hash-bound PASS attestation, guarded commit. Use PROACTIVELY when a change is risky, multi-step, or you want the earned safety net (lock/crypto/egress/protocol, or anything you want independently verified). Skip it for docs, chores, and low-risk edits — those commit normally.

/harness — opt-in rigor

Murmur's harness is opt-in. Normal commits run freely (only secret-scan and trunk-push protection are always on). Invoke the harness deliberately, via this skill, when a change deserves independent verification.

When to reach for it

• Anything touching the lock model / crypto / secrets / storage / MCP / egress / the sharing protocol (the risk_classification paths in .agents/harness/config.json).
• A multi-step feature or refactor where you want a fresh adversarial reviewer to try to break the change before it lands.
• Any time you want the hash-bound Definition-of-Done receipt on the commit.

For a docs fix, a chore, a version bump, or a small low-risk edit: do not use this — just commit normally.

How it works

The switch is physical: the harness runs in an isolated sibling worktree that carries a task manifest. While that task is active, finish-guard enforces the full attestation and the resource lane is required for heavy commands. Your main murmur checkout never has a task, so work there is unconstrained.

Run it

# 1. Create the contract + isolated worktree (task_id is positional;
#    --prompt and --owned are required, --owned may repeat for multiple paths)
scripts/agent-harness init <task-id> --kind <bug|feature|refactor|docs|harness> \
  --prompt "<what>" --owned <path> [--owned <path> ...] \
  [--risk <lock|egress|protocol|runtime|ui|performance|release>] \
  [--agent <codex|claude>] [--reviewer <codex|claude>]

# 2. Drive writer → checks → independent reviews → PASS attestation
scripts/agent-harness run <task-id>

# 3. Commit from the attested index (QueaT identity, no AI trailers) and open a PR
scripts/agent-harness commit <task-id> -m "<type>(<scope>): <subject>"
gh pr create -R murmur-io/murmur --base murmur
gh pr merge --merge

# 4. Close the task
scripts/agent-harness close <task-id>

Choosing the writer/reviewer pair

• --agent sets the writer vendor (default: config.json default_writer).
• --reviewer sets the reviewer vendor (default: config.json default_reviewer).
• All four pairs are allowed: claude→codex, codex→claude, claude→claude, codex→codex. The shipped default is claude→claude — the harness runs entirely on Claude, no Codex dependency.
• Whatever the pair, the reviewer is ALWAYS a fresh, independent session with no writer context — so same-vendor is still a real adversarial review, not self-grading, and the implementer never owns the verdict.
• Trade-off: same-vendor loses cross-model-family diversity (Codex and Claude catch different failure modes) — a fresh same-family session recovers the writer's self-attribution blind spot but not its family-level blind spots. So lock/egress/protocol tasks auto-escalate a same-vendor reviewer to the opposite vendor (pass --allow-same-vendor-high-risk to keep same-vendor there, printed with a loud warning). Other paths stay same-vendor by default.

Verify the harness itself

bash .claude/hooks/selftest.sh
scripts/agent-harness selftest --ci
scripts/agent-config-audit --ci