Configure or diagnose the staged-only multi-model harness, including selectable implementation workers, review advisors, high-assurance packet policy, provider probes, and optional Claude model-matrix output style. Use when the user says "set up the agent harness", "configure multi-model reviewers", "choose worker models", "check harness setup", "skonfiguruj harness", or "sprawdź modele".
Permissions
Files
Configure or diagnose the staged-only multi-model harness, including selectable implementation workers, review advisors, high-assurance packet policy, provider probes, and optional Claude model-matrix output style. Use when the user says "set up the agent harness", "configure multi-model reviewers", "choose worker models", "check harness setup", "skonfiguruj harness", or "sprawdź modele".
Configure the optional agentHarness section without changing the existing PR
pipeline skills. Preserve the rest of .ai/agentic.config.json, store no
secrets, probe every selected adapter, and leave repository-owned setup changes
staged but uncommitted.
--check (optional) — validate and probe the current configuration without writing.Version history
--preset <cross-model-jury|custom> (optional) — start from the bundled jury or an empty custom registry; default cross-model-jury.--defaults (optional) — enable the detected members of the bundled cross-model jury without asking questions.--no-output-style (optional) — do not install the optional Claude output style.--no-hooks (optional) — do not install the prevention hooks into .claude/hooks/ and .claude/settings.json.ALWAYS check first: Apply .ai/skills/om-setup-agent-harness/SKILL.md when present; safety rules still win.
references/agentic-setup.md before any action: load approved configuration, apply local extensions and shared rules, and preserve the untrusted-content boundary.Read references/provider-catalog.md and start from the bundled jury in
references/configuration-template.json: Codex, DeepSeek, Kimi, GLM, and MiMo.
Detect Claude as the host. Use the harness probe operation to check executable
versions and credential availability without printing or copying credentials.
The built-in DeepSeek and OpenCode Zen adapters may check their official local
auth-store entries internally; the setup agent never reads or displays the key.
When no bundled provider is detected, say so plainly and lead with the
bring-your-own-reviewer path from the provider catalog: any OpenAI-compatible
endpoint or any locally installed CLI can be bound as a reviewer through the
generic adapters — the bundled jury is a default, not a requirement. The only
fixed element is Claude as the host. A user with no external provider at all
still has a working standard profile (fresh Claude review only); offer that
explicitly instead of failing, but only after explaining what the multi-model
profiles would have added, so choosing it is an informed decision rather than
a silent default.
Unless --defaults is set, ask which detected models should act as workers,
reviewers, or both; which are required; the desired concurrency, timeout, and
retry policy (attempts, backoff, timeout escalation); and whether to enable
the high-assurance profile. Offer every bundled model independently, then offer “add custom
model” using the generic contracts in the provider catalog. Require command
adapters for workers. Keep HTTP and built-in subscription/API presets review-only.
Accept a worker only when its provider or OS sandbox demonstrably disables
network, remote/tracker writes, and Git ref writes; record that mechanism in
workerSecurity. Use only runtime-recognized audited adapters; version 1
supports codex-workspace-write-sandbox. A prompt-only promise or arbitrary
adapter declaration is insufficient.
Build an agentHarness object from references/configuration-template.json and
the user's choices. Remove unselected bundled models from the active profiles;
do not delete custom models. Keep the multi* profiles on all-required with
requiredReviewers recalculated to exactly the selected reviewer set; use
advisory mode only when no reviewer is selected at all. If high-assurance is enabled, require a worker plus enough reviewers
outside its model family for the configured risk and verification rules; do not
silently weaken those rules. Detect the current Codex model from non-secret CLI
configuration, but keep the shipped compatible default when that model fails a
smoke review. Preserve every existing top-level config value. Resolve the
installed om-harness skill (its directory sits beside this skill's own install
directory; when it is absent, stop and ask the user to install the collection
with om-harness included), then run node <om-harness>/scripts/harness.mjs configure against the working-tree config. Show the exact diff before staging it.
Keep portable team bindings in the repository object; put personal executable
paths or endpoints in the user-local overlay named by
OM_AGENT_HARNESS_CONFIG, and do not stage that overlay.
Keep every selected advisor in the profile reviewer list. Under the shipped
all-required policy every selected reviewer must complete (the runtime
retries failures automatically); selecting a reviewer is a commitment to hear
from it, never a best-effort hint.
Write only satisfiable profiles. Keep standard always; include optimized,
multi-optimized, and high-assurance only when an audited worker is bound,
and multi/multi-optimized only when at least one reviewer is selected. A
wrapper that later requests a profile this configuration does not define
reroutes back here instead of running with an empty jury.
Run validate-config, then probe for every selected profile. Run one harmless
structured review per selected advisor using the complete installed
om-code-review rubric so a valid binary but expired subscription is reported
correctly. This is an adapter smoke test, not an operational bound council; the
wrappers supply the required fresh Claude artifact and exact-subject packet.
A
required unavailable model or failed review policy is blocking; an optional
unavailable model is reported as skipped with its concrete smoke-test reason.
For high-assurance, require validate-config to accept its packet policy and
probe its full worker/reviewer readiness. Do not run a real worker against
repository files during setup.
Unless disabled, copy assets/claude-review-output-style.md to
.claude/output-styles/om-harness-review.md. Never overwrite a locally edited
copy silently; show a diff and ask whether to replace, merge, or keep it.
Offer to install the structural stop-before-PR and test-freeze guards (default
yes). Copy guard.mjs, block-push-and-pr.sh, and freeze-tests.sh from the installed
om-harness skill's hooks/ directory to .claude/hooks/om-harness/, keep
them executable, and merge the PreToolUse entries from the sibling
hooks-settings-snippet.json into .claude/settings.json without disturbing
existing hooks. These are the preventive layer of the stage-only contract: the
Bash guard is inactive until capture starts a run in the current worktree and
denies git push and PR creation until successful stage or explicit
end-run ends it. Other linked worktrees and ordinary publication workflows
remain usable. Keep both hook executable paths quoted so installation works
in directories containing spaces. The Edit/Write/MultiEdit/NotebookEdit
guard denies test-file edits while a .om-freeze-tests sentinel is present.
Never overwrite locally edited hook files or settings silently; show a diff and
ask.
Stage only .ai/agentic.config.json, the output-style file when created, and
the hook files plus .claude/settings.json when the hooks were installed.
Never commit, push, or publish. Print the readiness table defined in
references/provider-catalog.md, the configured profiles, skipped adapters,
and the staged paths. With --check, write and stage nothing.
Shared communication, marker, label, and safety rules: references/rules.md. They always apply.
Keep Claude as host, conductor, and final reconciler.
Require the operational wrappers to prove a new Claude review context; never configure the conductor's current context as its replacement.
Keep delivery.mode fixed to stage-only.
Keep secrets out of repository and user-local JSON; store only environment-variable names.
Never replace existing pipeline config values that the user did not ask to change.
Never install a worker binding whose command has not been confirmed to edit only the current worktree.
Never commit, push, open a pull request, or mutate tracker state.
Hook support requires Bash and Node on macOS, Linux, or WSL. Native PowerShell hook wiring is unsupported. Document that limitation during setup and block a staged run when the host cannot enforce publication permissions.
In these kits
More from @openmercato
Works with
Claude, Codex, Cursor & more