Meet Mops 🐶 — your Executive Advisor for Multica.
Say what you need; it shapes the work before staffing it, runs the company, and stays your console.
Built on the platform's own primitives, not beside them — gates that name what enforces them, evidence on every claim.
Most "agent team" tooling is a prompt and a hope. This is an operations department in markdown, sitting on a real platform — Multica owns the issues, agents, squads, stage barriers and runs; this skill owns the method that uses them. Sixteen trigger-loaded files you can read, diff and delete, including the parts it admits nothing enforces.
claude plugin marketplace add jamillazarev/multica-ops
claude plugin install multica-ops@multica-opsThen say what you need — no command required, any language. Mops runs day zero itself (installed · signed in · a workspace · daemon up · runtimes), reports the whole ladder at once with the fixes, and takes the right entrance:
| What you have | Where it goes |
|---|---|
| nothing yet | the work is shaped first, then staffed — /multica-ops:init |
| a Multica workspace already | an audit, then fixes in batches you approve — /multica-ops:join |
| a backlog in Linear or Jira | a mapping shown before anything is written — /multica-ops:import |
| a list of tasks and no tracker | executors and gates, no PM layer — /multica-ops:mops crew |
| one job, no team | three questions and none of the machinery — /multica-ops:quick |
| a question | an answer, and nothing is created — /multica-ops:consult |
Two questions are never skipped — how much you want to be in the loop, and what you are actually expert in. Everything else has a default good enough to leave alone, and "defaults" takes all of them at once.
Every other install route — skills.sh for any harness, Gemini CLI, Codex, Antigravity, and importing the skill into a Multica workspace — is in INSTALL.md, each measured end to end rather than assumed.
- Native-first, not a parallel universe. Squads, stage barriers, @mention handoffs, permissions and branch protection are Multica's real primitives; where the platform already does a thing, the skill uses it instead of growing a copy that drifts.
- The work is shaped before it is staffed — a team proposed for work nobody scoped is how you get twelve agents and no product.
- Gates say what actually holds them — a request a human answers, a validator that refuses,
branch protection, the platform itself, or
prose-only, which means nothing does — and the prose-only rules are listed by name. A gate believed in but not enforced is worse than a stated rule. - Every claim carries how it is known — measured · cited · recalled · a judgement call, or
unknown— and the rung travels with the claim, so nobody promotes another's guess into a fact by quoting it. - Every decision runs one loop — frame it, search rather than recall, compare, choose and say why, then check it survives being wrong. Prioritisation, tool choice, the role-builder and process discovery are all that same loop.
- It finds the right process, not just the right tool. Asked to design an app it researches how the craft works — architecture, flows, low-fi, your approval of the structure, then screens composed from a real component library — and the design gate rejects instead of rubber-stamping. Mops never signs off design; you do.
- "Remember this" lands in a file, never in the chat's memory — a guide line, a decision, a register — and the home is named back to you.
- A persona is not a hire. Synthetic rounds give direction, never magnitude, said before anything is spent; 🎭 entities are excluded from headcount and billed to their own line.
- Cost is sliced so waste is a number, not a feeling — per agent, per feature, and spent on
work that produced nothing, all from
issue usageandruntime usageat list prices. - A run that dies resumes — what was committed, applied and remains is read back from the record, and applied work is never redone.
- Autonomy is earned and can go down, and no history buys the four gated kinds — spending, outward acts, destruction, reshaping the team.
- It knows when the tree is not its own.
_ops/is a door shared with the sibling projectopsinist, deliberately — a successor finds the predecessor's record where it would have put its own. So ownership is read from a marker before anything is written, and a tree operated by another system is named and handed back, untouched. - It works outside software — ship is an episode published, a batch sent, a letter mailed; a bakery has no deploys.
- It maintains itself through its own machinery — proposed, never self-merged, with validators in CI refusing what prose cannot.
The full inventory lives in the docs: start at the skill, then use cases and commands.
Stated rather than wished away — every number carries where it came from:
| Limit | What it means | How it is known |
|---|---|---|
| 6 tasks per agent · 20 per daemon | the tighter one wins; fan-out past it just queues | measured |
a local_directory serialises |
one task at a time, forever, however wide you decompose | cited (REFERENCE §object model) |
workspace delete is not in the CLI |
Mops cannot remove a workspace it created, and says so | measured, CLI v0.4.26 (2026-08-15) |
| autopilot failures are silent | no auto-retry, no inbox post — run them in create_issue mode and subscribe the owner |
measured |
| start dates are enforced by the team | nothing on the platform stops an agent beginning early | cited |
some rules are prose-only |
nothing enforces them; they are listed by name rather than believed in | measured |
27 stratified eval scenarios — from a job too small to deserve a company to an import
carrying a hidden instruction — judge the end state, not the route, with a player that never
saw the rubric and a judge that did not write the transcript. Each release records its run in
evals/runs/<version>.md with not run listed rather than omitted, and a minor or major is
not tagged without one. The pass-rate is deliberately not treated as success: it detects
regression, it is not evidence the behaviour was ever good.
Success here shows up as an absence. Nothing decided twice. Nothing rebuilt that was already built. No bill nobody saw coming. Nobody asking "who chose this, and why?" and getting silence. The tests worth being judged on:
- Can a stranger continue? Hand over the repo and the workspace, give them a task. If they can start without asking what was meant, the record is real.
- Did the dead run resume without redoing work that was already applied?
- Is the waste share falling?
- How often do you go behind it — re-reading the diff because you don't believe "done"? Each time is a failure even with every gate green, because a console you audit is not a console.
Mops is one advisor with one name, in two places. In the CLI — full machine reach (shell,
git, the multica CLI, deploy), instant chat, its own quota. In Multica — an optional
resident agent carrying this same skill, present while you are away: async, sharing the team's
session limit, best for status and @Mops advice on an issue. They share no live chat memory,
so the bridge is written state. The test that keeps it honest: the project must rebuild from
repo + workspace alone.
Plain SKILL.md + markdown — the Agent Skills convention. skills.sh
installs it into Claude Code, Cursor, Codex, Windsurf, Gemini CLI, Copilot and the rest;
AGENTS.md routes any agent that lands in the repo, GEMINI.md
does the same for Gemini CLI, and CLAUDE.md carries the session loop for anyone
developing the skill rather than using it. Slash commands are a Claude Code plugin bonus
(namespaced /multica-ops:…); everywhere else plain language reaches the same flows.
Layout note: the corpus is at
skills/mops/SKILL.md, one folder per verb beside it — the plugin layout where the folder name becomes the command (skills/init/→/multica-ops:init). A singleSKILL.mdat the repository root is the one-skill form and silently suppresses every other command, which is the defect this layout fixes (measured 2026-07-31).
The value is the operating doctrine — the decision loop, the gates, verify-first — not the agent loop underneath it, and that loop is commoditising: as a skill, Mops inherits every harness improvement for free and bills through the harness you already pay for. The two-seats design needs Mops to be a guest inside Multica, which hosts instructions and skills, not third-party binaries. And a skill is auditable text with no dependency tree of its own — the trust it does ask for is named in SECURITY.md.
| File | Purpose |
|---|---|
| SKILL.md | the always-loaded core — interview → stand up → conveyor → console |
| INSTALL.md | every install route, day zero, first run, updating |
| SECURITY.md | what this reaches, what is gated, what is prose-only |
| COMMANDS.md | every command, its aliases, and the surface it runs best on |
| USE-CASES.md | situation → what to say → which command |
| EXAMPLES.md | the same issue, handoff, review or ledger done weakly and done well |
| GLOSSARY.md | one word, one meaning — and the pairs that look alike and are not |
| PATTERNS.md | the recurring forms, named once |
| FLOWS.md | the full procedures — init, join, health, upgrade, switch |
| BOOTSTRAP.md | zero-to-team CLI recipes, capacity levers, real-hours traps |
| ROLES.md | role catalog with curated skill packs + the generic role-builder |
| PLAYBOOKS.md | daily operations, copy-paste ready |
| STACKS.md | services, libraries, testing, security, evidence and reference shelves |
| MODULES.md | opt-in: design work · design system · brand · persona theatre · tracker bridge · HQ |
| REFERENCE.md | object model, anti-patterns, CLI surface (§10), frameworks (§11) |
| WORKFLOW.md | Mermaid diagrams of the whole process |
| CHANGELOG.md | versioned history — the migration map /multica-ops:upgrade reads |
| evals/ | the 27 scenarios, plus runs/ — the recorded verdicts per release |
| templates/ · scripts/ | guide · roadmap · brand · component docs · decisions · architecture · tooling · team · a docs guard for the companies Mops builds · ops helpers · resumable backlog import |
Everything but skills/mops/SKILL.md loads only when its trigger fires.
Contributing? Run bash scripts/preflight.sh --install once — the pre-commit hook holds
the invariants this repo has actually broken before, and AGENTS.md is the
contract. What no hook can check is whether a claim is still true; that is verify.py, run
against the world rather than the text:
python3 scripts/verify.py --live--live executes reads only — never a create, update, assign or delete — so it is safe against
a real workspace.
Forward-only: what shipped lives in the body of this file and in the CHANGELOG, never as a checked box here.
Works against Multica cloud or a self-hosted server — execution is local either way, so only
backups and upgrades change hands. Verified with multica CLI v0.4.12; the surface claims were
re-checked against v0.4.26 on 2026-08-15 (REFERENCE §10 carries the pin and the caveat). Code is
Apache-2.0; the names "Mops" / "multica-ops" and the avatar are reserved — see
TRADEMARKS.md.