Glossary

Alphabetical list of key terms. Links point to relevant docs where useful.

.maestro: Project directory holding intent.yaml, gates.yaml, evidence ledger, and other compiler config. Learn more.
Agent mode: Interactive mode where the agent can read, search, edit, and run commands to complete tasks. Learn more.
AGENTS.md: Markdown file in project root or .midcore/ with persistent instructions the agent reads each session. Learn more.
API_INTERNAL_URL: Server-side env var: base URL of the FastAPI service used by Next.js rewrites (e.g. http://api:8000 in Docker). Learn more.
Ask mode: Read-only mode for exploration and Q&A; no file edits or code changes. Learn more.
AUTH_INTERNAL_URL: Server-side env var: base URL of the auth (OAuth) service for Next.js /auth/* rewrites. Learn more.
Automation memory: Persistent key-value store scoped to an automation for facts across runs. Learn more.
Automation trigger: Event that starts an automation run: cron, GitHub, Slack, webhook, or manual trigger. Learn more.
Autonomy API (HTTP): HTTP routes under /api/v1/autonomy on the API host, often called from the web app via /api/backend/autonomy/… or /api/autonomy/… rewrites. Learn more.
BFF (backend-for-frontend): Pattern where the browser talks only to the web origin; Next.js forwards selected paths to internal APIs so base URLs stay server-side. Learn more.
Brain pipeline: Sense → think → act → learn loop for Puble: perception fusion, planning, control, and post-run learning with audit trails. Learn more.
BugBot pass: One parallel review pass in the BugBot pipeline; multiple passes feed majority voting. Learn more.
Capability graph: Versioned graph of capabilities and their contract edges; built from intent by the Outcome Compiler. Learn more.
Capability passport: A document that proves a capability ships with contracts, tests, and evidence. Learn more.
Checkpoint: Persisted snapshot of agent state for long sessions so work can resume after interruption. Learn more.
Cloud handoff: Serialized conversation + repo snapshot transferred between local and cloud agent sessions. Learn more.
Contract: A formal specification (e.g. API or event shape) that is agreed before implementation. Learn more.
Debug mode: Mode for multi-hypothesis debugging with instrumentation and targeted fixes. Learn more.
Delivery tier: Level (prototype, pilot, production) that determines how many gates and how much hardening are required. Learn more.
Denial budget: Circuit breaker for repeated command/tool denials that escalates to human review when thresholds hit. Learn more.
Digital twin: Simulated or replayed model of a physical system used to rehearse scenarios before live execution. Learn more.
DRL (deep reinforcement learning): Training policies from simulation or logged trajectories with reward signals and exportable artifacts. Learn more.
E-stop: Emergency stop that halts actuation and records an audit event for robotics and automation. Learn more.
Evidence: Machine-verifiable artifact (e.g. gate result, timestamp, scope) recorded in the evidence ledger. Learn more.
Evidence ledger: Append-only store of gate results and audit trail used to verify completion. Learn more.
Fleet management: Registering, monitoring, and scheduling work across many robots or workers with shared policy and evidence. Learn more.
Gate: A deterministic check (e.g. tests, lint, contract validation) that must pass for a release. Learn more.
HAL (hardware abstraction layer): Stable interface between autonomy logic and vendor-specific drivers so policies and tests stay portable. Learn more.
Hardening: Rounds of checks (identity, audit, secrets, build, data, integration, launch) required for production. Learn more.
Ignore file: File (e.g. .gitignore, .midcoreignore) that defines paths the agent should not read or edit. Learn more.
Instant grep: Trigram-backed index for sub-50ms search over large repositories without full rescans. Learn more.
Intent: Structured description of what you want to achieve; drives the Outcome Compiler and agent. Learn more.
Majority voting: Aggregating findings across independent review passes so only consensus bugs surface. Learn more.
MCP OAuth: OAuth 2.1 PKCE flow for MCP servers so tool access is authenticated without static secrets in prompts. Learn more.
MIMO: Multi-input multi-output experiment design mapping multiple controlled factors to measured outcomes. Learn more.
OpenAPI: Machine-readable description of HTTP operations (paths, schemas); served at /openapi.json on the API service in this stack. Learn more.
Outcome Compiler: System that turns intent into a capability graph, contracts, and evidence-backed releases. Learn more.
Outcome composer training: Optional, evidence-backed training loop for outcome composition policies, using sandboxed rollouts and evaluation gates. Learn more.
Pilot: Delivery tier with standard gates and most hardening required. Learn more.
Plan mode: Mode for design and trade-off discussion without executing edits. Learn more.
Planner–worker–judge: Multi-agent pattern: planner assigns work, workers execute, judge evaluates continuation vs replan. Learn more.
Production: Delivery tier with all gates mandatory and full hardening. Learn more.
Proof-carrying: Releases and capabilities that ship with machine-verifiable evidence (gates, ledger), not narrative. Learn more.
Prototype: Delivery tier with relaxed gates and optional hardening escrow. Learn more.
Puble: Operational robotics stack in Midcore: fleet, brain, safety, telemetry, simulation, and training. Learn more.
Research wizard: Guided UI to assemble hypotheses, datasets, and evaluation hooks before running a study. Learn more.
Scenario: Named configuration of inputs, environment, and success criteria for robotics or research runs. Learn more.
Scope: The set of files, areas, or capabilities that a task or release is limited to. Learn more.
Scope lock: Freezing scope and contracts before implementation so work stays bounded. Learn more.
Scout track: Research product track for discovery-oriented studies and literature-aligned workflows. Learn more.
Self-hosted worker: Agent runtime in your infrastructure that executes tools close to data with fleet control-plane health. Learn more.
Server-side compaction: Provider-side context compaction API that preserves critical facts while reducing token load. Learn more.
SIL (safety integrity level): Framework-aligned rigor for safety-critical paths; interlocks separate estimation from actuation. Learn more.
Simulation track: Research track for Monte Carlo, digital twins, and synthetic experiments with reproducible seeds. Learn more.
Sparse n-gram index: Space-efficient character n-gram postings for fast fuzzy prefilter before verification. Learn more.
Studio workspace: Primary UI pane for intent-to-release: orchestration, gates, evidence, contracts, and releases. Learn more.
Surveillance track: Research track for monitoring-style studies with drift detection and alerting. Learn more.
Survey track: Research track for structured surveys, cohorts, and human-in-the-loop protocols. Learn more.
Telemetry: Streaming state, health, and sensor summaries from devices or agents for dashboards and alarms. Learn more.
Tier: Delivery level (prototype, pilot, production) that determines which gates apply. Learn more.
Vibe coding: Flow where you describe intent and the agent implements with gates and evidence. Learn more.
Webhook secret: Shared HMAC key used to verify inbound webhook signatures before executing automation handlers. Learn more.
Worktree voting: Running multiple agent candidates in separate git worktrees and merging winning patches. Learn more.

How Midcore works · Gates & evidence · Outcome Compiler