quadrant: reference audience: both see-also: - architecture.md - cli-reference.md
Terminology¶
Canonical terms used across the codebase, CLI, and docs. Alphabetical.
| Term | Meaning |
|---|---|
| Action | A unit emitted by the engine to drive the orchestration loop. One of: DISPATCH, GATE, APPROVAL, COMPLETE, FAILED, WAIT, FEEDBACK, INTERACT. See ActionType in agent_baton/models/execution.py. |
| Agent | A distributable specialist defined in agents/<name>.md (frontmatter + prompt body). Dispatched by the orchestrator or invoked directly via Claude Code's Agent tool. |
| Approval | Explicit human (or designated reviewer) sign-off required for HIGH-risk plans and certain phase transitions. Driven by the APPROVAL action and baton execute approve. |
| Assurance pack | A governance bundle in .claude/packs/<pack>/ with a pack.json manifest, policy/signals/rubric files, and optional gates/evidence. Distinct from knowledge packs: assurance packs enforce checks and risk controls. |
| Bead | A persistent incident or follow-up record stored by the external bd tool (gastownhall/beads) in a per-project .beads/ workspace. Created with baton beads create. Used for autonomous bug filing, regression beads, audit trails. |
| Baton | The project name (capitalized in prose). |
baton |
The CLI binary (lowercase, monospace). |
agent_baton |
The Python package (snake_case, importable). |
| Budget | The token/cost ceiling for an execution. Tiers are tight, standard, generous. Set per phase by the planner. Run-level cap: BATON_RUN_TOKEN_CEILING. |
| Central DB | ~/.config/agent-baton/central.db — federated registry of all projects' executions. Distinct from each project's baton.db. |
| Classification | Risk tier (LOW / MEDIUM / HIGH / REGULATED) assigned by the planner using core/govern/classifier.py. |
| Daemon | The optional long-running process (baton daemon) that watches for triggers and drives executions without explicit invocation. |
| Dispatch | The act of spawning a specialist agent for a step. Surfaced as the DISPATCH action. |
| Engine | The Python orchestration code in agent_baton/core/engine/. Owns state machine, persistence, gating, tracing. |
| Federated sync | Bidirectional sync between per-project baton.db files and the central database. Driven by baton sync. |
| Forge | The plan-generation subsystem. May invoke a headless Claude Code subprocess for plan synthesis. |
| Gate | An automated check between phases (typically pytest, lint, type check, link check). Emitted as the GATE action. Recorded with baton execute gate. |
| Guardrail preset | A bundle of risk-tier-specific rules: required reviewers, mandatory gates, allowed tools. See references/guardrail-presets.md. |
| Immune system | Background immune-* agents that proactively scan for drift (deprecated APIs, doc drift, stale comments, TODO rot, untested edges). |
| INTERACT | A multi-turn step primitive that pauses the loop for back-and-forth with a user or another agent. |
| Knowledge pack | A discoverable bundle of reference docs in .claude/knowledge/<pack>/ with a knowledge.yaml manifest. Resolved by the knowledge router during planning. |
| Orchestrator | The agent (agents/orchestrator.md) that drives baton plan → baton execute for complex tasks. |
| Phase | A named segment of a plan. Composed of one or more steps and (optionally) a gate. Phases are sequential by default; some can run in parallel. |
| Plan | The output of baton plan. Two artifacts: machine-readable plan.json and human-readable plan.md, both in .claude/team-context/. |
| PMO | The optional REST API + React UI (served at /pmo/) that visualizes plans, executions, traces, retrospectives, and learning data across projects. |
| Reference procedure | A reusable, agent-readable doc in references/<name>.md that codifies a workflow contract (engine protocol, routing logic, guardrails, patterns). |
| Retrospective | The post-execution analysis written by the engine (and optionally augmented by learning-analyst). Stored in .claude/team-context/retros/. |
| Gate retry | A single automated retry triggered when BATON_GATE_RETRY=1 and a gate fails for the first time. The engine re-dispatches the failing step once with the gate output appended to the prompt. Second failure is terminal. |
| Step | A unit of work within a phase. Either a DISPATCH (agent), a GATE (check), or an INTERACT (dialogue). |
| Subagent | An agent dispatched by another agent (e.g., the orchestrator dispatches backend-engineer). |
| Trace | The full event log of an execution. Generated automatically; inspected via baton trace. |
For the architectural concepts and how they relate, see architecture.md.