| name | op |
| description | Route an idea or request through the Agentic Operator — classify route × tier and dispatch to rf, arc, skillmeat, meatywiki, intenttree, or multi-hop workflows. Use when the user brings an idea to route, requests escalation to a new tier, approves/rejects a pending gate, or asks for run status/history. Triggers: "op new", "route this", "which subsystem", "escalate", "approve", "reject", "what's the status", "spin up". Do NOT use for: direct subsystem command invocation (use rf, intenttree, council-review, meatywiki, skillmeat-cli skills instead). |
| version | 0.1 |
| app_version | 2026-06-14 |
| updated | "2026-06-14T00:00:00.000Z" |
| spec | ./SPEC.md |
op Skill
The Agentic Operator entry point and control-plane router. One skill to bring an idea; the operator classifies it and dispatches to the right subsystem at the right effort tier.
1. Overview
The operator is a two-dial classifier and run-record keeper. It receives an unclassified idea, decides the route (which subsystem) and tier (how much machinery), shows a plan at a gate, and dispatches to agents-as-tools. It owns the durable run record so you can resume work across sessions, and it enforces three gates (plan, writeback, task-graph mutation) only on the expensive/irreversible hops.
2. The Two-Dial Model (route × tier)
| Dial | Name | Values | Meaning |
|---|
| A | ROUTE | research / council / knowledge / artifact / tasks / multi / clarify | Which subsystem to invoke |
| B | TIER | 0–4 | How much machinery to spin up |
Route enum (each maps 1:1 to a sibling skill):
research → rf (find evidence, verify claims, state-of-the-art)council → arc / council-review (red-team, critique, scorecard)knowledge → meatywiki (recall, file a note, what do we know)artifact → skillmeat (deploy, find, bundle, register)tasks → intenttree (plan, break down, track, dispatch)multi → operator decomposes (spans 2+ subsystems)clarify → ask the user one disambiguating question (never guess on expensive work)
3. The Tier Ladder (T0–T4, each a superset of the one below)
- T0 — Capture / Idea. "Just hold this." → IntentTree off-tree Ideas capture (+ optional MeatyWiki note). ~free, trivially reversible.
- T1 — Spike / Question. A bounded question or lookup. → One small
rf pass or a meatywiki query. Low cost, reversible.
- T2 — Workflow / Council. "Make this repeatable" or "judge this." → Registered SkillMeat workflow artifact and/or an ARC council scorecard. Medium cost.
- T3 — Initiative / Feature. Chartered work inside an existing project. → IntentTree workspace + planning docs (charter→PRD→plan) + optional research/council seed. Med-high cost.
- T4 — New Project. A genuinely new thing. → Full scaffold: repo skeleton + project
CLAUDE.md + SkillMeat bundle + IntentTree workspace + MeatyWiki namespace + seed research/council pass. High cost.
The tier is also the budget: each tier carries a default token ceiling and fan-out width (T0=0 agents, T1=1–3, T2=3–5, T3/T4=full wave).
4. Decision Tree
INTENT COMMAND
─────────────────────────────────────────────────────────────────────
"op research X" / "op council Y" (explicit verb) → forced route, default tier
"capture this" / short noun phrase, no verb → op capture "<idea>" (T0)
Question, "look into", "is X true" → op "<idea>" (classifier → T1)
"Make a reusable…" / "review this" → op "<idea>" (classifier → T2)
"Build the X feature" (existing project) → op new "<idea>" --tier 3
"New app/tool/project", "start…" → op new "<idea>" --tier 4
Pending gate / "approve <run_id>" → op approve <run_id> [--yes]
Cancel paused run → op reject <run_id> [--reason "..."]
"What's the status" / one-line summary → op status <run_id>
Need full run record JSON → op show <run_id>
List all runs → op list
Subsystem availability check → op doctor
─────────────────────────────────────────────────────────────────────
5. Command Map (the real op CLI surface)
| Command | Purpose | Key flags | Notes |
|---|
op "<idea>" | Bare-idea form: classify (route × tier) + run to first gate | --tier 0..4 --yes --json --home | The default entrypoint. No subcommand needed. |
op run "<text>" | Same as bare form, explicit | (as above) | Identical handler. |
op research "<text>" | Force research route | --tier --yes --json --home | Routes to rf. |
op council "<text>" | Force council route | --tier --yes --json --home | Routes to arc/council-review. |
op knowledge "<text>" | Force knowledge route | --tier --yes --json --home | Routes to meatywiki. |
op artifact "<text>" | Force artifact route | --tier --yes --json --home | Routes to skillmeat. |
op tasks "<text>" | Force tasks route | --tier --yes --json --home | Routes to intenttree. |
op capture "<idea>" | Force a T0 Ideas capture | --yes --json --home | Always tier=0, always tasks route. |
op new "<idea>" [--tier N] | Tiered "new thing" scaffolder | --tier 0..4 --json --home | Proportional output (T0=note → T4=repo). |
op status <run_id> | One-line synthesis (NO model call) | --json --home | Read-only; ICA-friendly. |
op show <run_id> | Full run record (JSON) | --home | Read-only. |
op list | List run ids | --json --home | Read-only. |
op approve <run_id> | Approve the pending gate and continue | --yes --json --home | Resumes paused run. |
op reject <run_id> | Cancel a paused run | --reason "..." --json --home | Records reason in run record. |
op doctor | Probe subsystem availability + env | --json --home | Lists adapters + routes; never calls a model. |
Global flags (every command)
--json — machine-readable output (Principle §2.4: read paths never call a model).--home <dir> — run-record home (default $OP_HOME or ./.operator/runs).--yes / -y — auto-approve gates (non-interactive).--tier <0..4> — override the classifier's tier choice.
6. The Three Gates (only the hops that matter)
- Plan gate — before any fan-out: show route + tier + hop plan + estimated token/effort budget. Auto-skipped for T0/T1 below threshold.
- Writeback gate — before writing into MeatyWiki or deploying via SkillMeat (irreversible knowledge/artifact mutations).
- Task-graph mutation gate — reuses IntentTree's existing approvable
AgentRun gate; nothing new to build.
Gates persist to the run record and are resumable across sessions — "approve tomorrow" works. Use op approve <run_id> to continue, op reject <run_id> --reason "..." to cancel.
7. When to escalate to op new (vs. op vs. op capture)
| User signal | Use | Why |
|---|
| "Just remember this", short noun phrase | op capture "<idea>" | T0, zero ceremony, no gates, IntentTree Ideas capture only. |
| Question / "look into" / ambiguous | op "<idea>" | Bare form runs the cascade classifier; lets the operator pick route × tier. |
| "Build feature X in " | op new "<idea>" --tier 3 | T3 scaffolder: workspace + planning docs + optional seed. |
| "New app", "start a project", "launch X" | op new "<idea>" --tier 4 | T4 scaffolder: full repo + bundle + namespace + seed pass. |
| User explicitly names a subsystem | op <route> "<text>" | Skip classifier; force the route. Faster, deterministic. |
| "What happened with run X?" | op status X / op show X | Read-only, no model. |
| "Approve / reject the pending gate" | op approve X / op reject X | Resumes the paused run from the persisted gate. |
8. Deferred / Do Not Say (do NOT invoke)
| Feature | Status | What the skill says |
|---|
| Atomic dispatch+await on tasks route | Not shipped | Operator dispatches but does not yet await intenttree completion atomically; the adapter is the harness-exec bridge for now. |
| Fan-out swarm primitive | Not shipped | No built-in N-way dispatch; the op composes fan-out from sequential hops + a parent aggregation node. The multi route is staged for Phase 3. |
| Real research model adapters | Not installed | rf doctor reports adapters 0/6 available. Real discovery goes through the rf-run-execute.js Claude workflow; the deterministic spine works offline. |
| Command Center UI / "spin up swarm" button | Later | The visual front-end is a Phase-5 deliverable over the same operator core. Do not promise a UI. |
| Tier override that conflicts with classifier | Reject + clarify | If --tier is far from the classifier's confidence band, the operator asks one clarifying question rather than blindly clamping. |
9. References Pointer Table
| File | Load when | Max lines |
|---|
./references/command-reference.md | Need full flag/exit-code/example detail | 600 |
./references/routing-cascade.md | Cascade classifier disagrees; need heuristics | 400 |
./references/tier-selection.md | Helping the user pick a tier | 350 |
/Users/miethe/dev/homelab/development/agentic_meta_dev/docs/agentic-operator/00-AGENTIC-OPERATOR-DESIGN.md | Design doc; canonical reference for §4 (route × tier) and §5 (tier ladder) | full |
/Users/miethe/dev/homelab/development/agentic_meta_dev/src/operator_core/cli.py | Verb behavior is unclear; read the real handlers | full |
10. Contract Pointer
See SPEC.md for the run-record schema, gate sequencing, adapter contracts, and the cascade classifier's confidence thresholds.
Key References
- /Users/miethe/dev/homelab/development/agentic_meta_dev/docs/agentic-operator/00-AGENTIC-OPERATOR-DESIGN.md
- /Users/miethe/dev/homelab/development/agentic_meta_dev/src/operator_core/cli.py
- /Users/miethe/dev/homelab/development/agentic_meta_dev/docs/agentic-operator/contracts/skill-structure.md
