Skip to content

Feature inventory

mini-ork ships 119 capabilities across seven pillars; five more are specced and not yet mainline, for 124 catalogued entries. The full, code-anchored catalogue — every entry with a file:line anchor, a shipped/specced marker and a lens-consensus score — is docs/reference/FEATURE-INVENTORY.md.

This page is the map: the capabilities most likely to matter to an operator, grouped by pillar, with a pointer to the page that explains each one.

At a glance

Pillar Count What it covers
Orchestration core 13 The run lifecycle, recipes, planning, scheduling
Heterogeneous model dispatch 18 Providers, lanes, routing, fallbacks, cost guardrails
Runtime reliability 14 Durable DAG, leases, checkpoints, resume, rollback
Verification & gates 11 Gate registry, verifiers, oracle, evidence-cited vetoes
Self-improvement & learning 17 GRPO writeback, bandit routing, reflection, apply loop
Observability 22 Web app, live stream, run detail, why-did-this-fail, OTel
Operator & dev ergonomics 24 Install, init, doctor, validate, garden, worktrees, review

Orchestration core

Capability What it does Docs
run lifecycle Classify → profile → plan → execute → verify → reflect in one dispatch run
Subcommand registry Third-party subcommands register without editing the dispatcher Extending
Keyword task classifier Regex and alias scoring over task classes, no model call Task classes
Planner with repair Retries recoverable planner verdicts and preserves the raw response Run lifecycle
Profile Q&A gate Blocks planning when confidence is below the floor, auto-answers when it can Your first run
Recovery DAG Parses workflow nodes and edges; operator edges are excluded from data flow Recovery
Multi-epic scheduler Bounded parallel pool with dependency cascade Scheduler
Recipe resolution Maps a task class to a recipe, with _ ↔ - fallback Recipes
Run-profile generator Extracts scope, commands and lanes from the kickoff into run_profile.json Core concepts
Conductor Picks topology and lane hints from learning tables Learning loop
Workflow compiler Compiles nodes, edges and bindings with explicit output visibility Artifact graph
Recipe set Shipped recipes, each with all three recipe files Recipes
Idea-tree persistence Stores the conductor's recursive breakdown Architecture

Heterogeneous model dispatch

Capability What it does Docs
Provider registry Five provider kinds, bring your own endpoints Providers and lanes
dispatch_model Preflight, lane health, tool grants, transport Providers and lanes
Routing policy registry Six policies, selected by MO_ROUTING_POLICY Providers and lanes
Daily cost circuit breaker Sums 24 h spend against MO_DAILY_BUDGET_USD Cost
Cost-pause sentinel Writes a pause marker every MO_PAUSE_EVERY_USD, operator-resumable Cost
Dispatch primitive Prompt on stdin, structured result, never raises Extending
Fallback chains First non-empty success wins across a lane chain Providers and lanes
Codex native transport Sidecar transport with token and cost accounting Providers and lanes
Role → lane mapping Each node role resolves to a configured lane agents.yaml
Lane fuse Three strikes remove a lane from rotation Scheduler
Throttle guard Per-provider cool-down after rate limits Scheduler
Owner-only secret store Credentials at mode 0600, never in argv Vault
Tool-grant hermetic dispatch A node sees only its declared tools and MCP servers Sandboxing
CWD guard Refuses dispatch from a framework tree Security

Runtime reliability

Capability What it does Docs
Durable-DAG resume Skips succeeded nodes, re-dispatches the rest Recovery
Liveness circuit breaker Stops work that cannot make progress Scheduler
Single-writer lease + fencing One writer per run, stale writers fenced out Run lifecycle
Tool receipts Idempotency envelope around side effects Run lifecycle
Deadline budget Wall-clock cap for a run Cost
Checkpoint writer Per-step ledger plus a validity check Recovery
Failure-class state machine Separates infra, code and policy failures Failure triage
Publisher commit gate Strict-child path before a change lands Run lifecycle
Rollback Workflow or agent rollback to the previous stable point Recovery
Repo integrity guard Detects a home/engine mismatch before writing Troubleshooting
RunContext Typed run environment with scoped overrides Environment
File-surface lease --owns claims prevent concurrent agents racing a file Worktrees

Verification & gates

Capability What it does Docs
Gate registry Eight gate types, extensible by registration Verification stack
Promotion gate Only measurable improvements are promoted Self-improvement
Native oracle gates In-process, no network round-trip Verification
Verifier dispatcher Routes a verifier to the right runner Verification stack
Per-recipe verifier scripts Deterministic checks shipped with the recipe Artifact contract
Artifact-contract validator Checks expected artifact kind and outputs Artifact contract
Grounded rejection A rejection must cite the evidence that caused it Verification
Rubric pre-screen Grades a run before deeper checks Verification stack
Citation verifier Mechanically checks cited sources exist Results
Krippendorff-α panel gates Detects judge-panel disagreement Verification stack
Coord gate Advisory or strict denial on concurrent scope concord

Self-improvement & learning

Capability What it does Docs
GRPO writeback Group-relative advantage written back per outcome Learning loop
Anti-Goodhart reward contract Reward cannot be satisfied by self-report alone Learning loop
Contextual-bandit router Cost-free UCB routing over lane outcomes Learning loop
Advantage store Domain, region and slice tables Learning loop
Reflection pipeline Extracts textual gradients and patterns after a run Learning loop
Apply loop Turns a learned change into a scored candidate apply
Self-improve outer loop Worktree and branch per iteration, gated self-improve
Semantic long-term memory Reconciles memories rather than appending blindly Learning loop
Context assembler Builds the per-node prompt surface Learning loop
Role evolver Retires, splits or renames lanes Self-improvement
Process-reward heuristic Scores intermediate steps when a judge is unavailable Learning loop
Group evolver Proposes workflow candidates improve
Per-task no-regression gate A candidate must not regress any task in the set Self-improvement
Topology telemetry Reports panel shape and inductive distance Observability

Observability

Capability What it does Docs
FastAPI observability app Fourteen routers over the state database Web app
SPA mount with API fallback The UI works even when the bundle is absent Web app
SSE live stream Events as they happen Observability
Run control Stop, kill, pause, resume, steer Web app
Token-based auth Writes are default-deny Security
Detached run launch Start a run without holding the terminal run
Task-run list + summary cache Fast board reads Inspecting state
Run detail Correlated queries across a run Web app
DAG status overlay Live node state on the graph Web app
Why-this-failed Aggregates the evidence for a failure Failure triage
Active-run fleet Heartbeats for running work Observability
Learning endpoints Fifteen read surfaces over learning state Observability
Recipe fingerprint Lane and verifier topology per recipe Observability
Trajectory charts Run-level trajectory views Observability
Recovery projection Shows what a resume would do Recovery
Project registry Switch between project homes Web app
Dispatch endpoint Reports outcomes with confidence intervals Observability
OTel + Langfuse export Span buffer flushed over OTLP Observability
Path-escape rejection Artifact endpoints refuse paths outside the run Security

Operator & dev ergonomics

Capability What it does Docs
install Cross-platform launcher Installation
init Scaffolds a project home maintenance
doctor Tools, engine version, credentials, vault maintenance
validate Pre-run static checks with fix hints maintenance
garden Drift detection: collisions, orphans, stale runs maintenance
update Applies migrations and reports drift maintenance
providers Workflow-aware credential status and configure providers
eval Benchmarks a workflow candidate Lifecycle commands
recipe-eval Author-time recipe lint maintenance
spawn Worktree per slug with --owns Worktree workflow
coord / concord File-surface leases and mailbox concord
review Pre-push critic panel Tests and CI
metrics / usage-report Lifetime and usage statistics Observability
Config layering Project over home over repository defaults Environment

Specced, not mainline

Five catalogue entries are roadmap items rather than shipped mainline behaviour: metamorphic and jury stages in eval-in-run-flow, live GEPA reflection at scale, a full React bundle for the web app, the pre-push review panel as a default merge gate, and adaptive conductor auto-engagement. Treat them as design intent, not as available features.