Glossary

One-sentence definitions for the terms a first-reader meets in Alfred docs. Entries are alphabetical, with cross-links to the page that covers each term in depth.

• agent:authored label: Marks a PR opened by an Alfred agent, used by Ra’s al Ghul, Nightwing, and automerge to know which PRs are theirs to act on. See also: Issue claim state machine.
• agent:bundle:<slug> label: Groups GitHub issues that belong to one cross-repo feature; Batman resolves the bundle from this label. See also: Worked example.
• agent-cleanup: Housekeeping agent that sweeps stale worktrees, expired spend files, stuck locks, and stale agent:in-flight claims. See also: The agent fleet.
• agent:done label: Terminal state for an issue whose PR has merged; set by release_issue(transition_to="agent:done"). See also: Issue claim state machine.
• agent:implement label: Marks an issue ready for Lucius (or another feature-dev agent) to claim and implement. See also: Issue claim state machine.
• agent:in-flight label: Marks an issue currently claimed by an agent; another agent will not pick it up until the claim is released. See also: Issue claim state machine.
• agent:large-feature label: Marks an issue Batman should plan as a multi-repo rollout rather than a single-repo implementation. See also: Worked example.
• agent:pr-open label: Set on the issue when the implementing agent opens its PR; cleared when the PR merges or closes. See also: Issue claim state machine.
• AGENT_RUNNER: The shared Python package at lib/agent_runner/ that every agent script imports for preflight, locking, spend, gh, and Slack. See also: agent_runner API.
• AgentResult: The dataclass an engine returns to the runner, containing success flag, subtype, turn count, cost, session id, and result text. See also: How it works.
• ALFRED_HOME: Operator-overridable directory holding runtime state, prompts, worktrees, and logs; defaults to ~/.alfred. See also: Install.
• Bane: Test-coverage agent that picks the lowest-coverage actively changed file and opens a tests-only PR for it. See also: The agent fleet.
• Bat-signal: The Slack alert raised when an agent prints [BLOCKED] or when a fleet-wide spend or rate-limit cap trips. See also: Output samples.
• Batman: Cross-repo architect that turns agent:large-feature issues into rollout plans and, on the parent-issue path, approved child agent:implement issues for the normal fleet queue. See also: Worked example.
• claude -p: Claude Code’s non-interactive subprocess mode, the surface Alfred uses to invoke Claude with a prompt and capture an AgentResult. See also: Claude Code and Codex.
• code-map: JSON snapshot of every watched repo’s source files, symbols, imports, API calls, server routes, and contract drift, written to $ALFRED_HOME/state/code-map.json by code-map-refresh. See also: Alfred on a monorepo.
• Codex: OpenAI’s local coding agent, supported as an engine alongside Claude Code and selectable per agent via ALFRED_<AGENT>_ENGINE. See also: Claude Code and Codex.
• doctor.sh: Script at bin/doctor.sh that runs every enabled agent in doctor mode (no LLM spend) and reports preflight status. See also: Output samples.
• Drake: Planner agent that reads specs, roadmap, and code-reality and files the next well-scoped agent:implement issue for the fleet to work. See also: Specs-driven development.
• dry-run: Mode toggled by --dry-run or ALFRED_DRY_RUN=1 that narrates a full firing lifecycle without LLM calls or side effects. See also: Dry-run mode.
• engine: The coding backend an agent invokes; Claude Code, Codex, or a hybrid that tries Claude first and falls back to Codex on rate limits. See also: Claude Code and Codex.
• engine routing: Per-agent assignment of which engine to use, set via ALFRED_<AGENT>_ENGINE (claude, codex, or hybrid). See also: Claude Code and Codex.
• fast-cleanup: agent-cleanup’s sub-pass that runs after every Lucius firing to delete just-closed worktrees without waiting for the nightly sweep. See also: The agent fleet.
• firing: One run of one agent triggered by the host scheduler; bounded by lock, preflight, spend caps, and a hard timeout. See also: How it works.
• GH_ORG: The GitHub org or user that owns the repos Alfred operates against; agents refuse to act on repos outside it. See also: Install.
• hybrid fallback: Engine routing mode where the runner tries Claude first and falls back to Codex when Claude hits a rate limit or budget cap. See also: Claude Code and Codex.
• IAM-per-agent: AWS pattern where each agent gets its own IAM identity and Secrets Manager scope, so a compromised agent can only reach its own keys. See also: AWS.
• launchd macOS host scheduler that owns the firing cadence on Mac; agents ship as .plist files in ~/Library/LaunchAgents/. See also: Architecture.
• Lucius: Feature-dev agent that claims an agent:implement issue, opens a worktree, invokes the engine, and pushes a PR labelled agent:authored. See also: How it works.
• Nightwing: Review-fix agent that lands P0/P1 reviewer comments on open agent:authored PRs without re-litigating design. See also: The agent fleet.
• plist macOS launchd unit file; one per agent, generated by the renderer from launchd/agents.conf and launchd/template.plist. See also: launchd plist template.
• preflight: Cheap pre-firing check that the required CLIs, gh auth, and workspace checkouts exist before any LLM turn is spent. See also: How it works.
• Ra’s al Ghul: Code-review agent that posts a multi-axis review (correctness, security, performance, maintainability) on every fresh agent:authored PR. See also: The agent fleet.
• role runner: A Python script under bin/ that implements one agent role; codenames map to role runners via AGENT_CODENAME. See also: The agent fleet.
• sentinel string: Bracketed marker like [OK], [BLOCKED], or [SILENT] that an agent prints to stdout to signal its exit path. See also: Output samples.
• Slack post: One outbound message Alfred sends to a configured Slack channel via webhook or bot token, at info, warn, or alert severity. See also: Slack.
• starter fleet: The recommended first roster (Drake, Lucius, Ra’s al Ghul, agent-cleanup) the installer enables when you pass --agents starter. See also: Install.
• state machine: The label transitions on an issue (agent:implement → agent:in-flight → agent:pr-open → agent:done) that coordinate agent handoffs. See also: State machine.
• systemd --user: Linux host scheduler that owns the firing cadence on Debian/Ubuntu; the equivalent of launchd on macOS. See also: Linux.
• turn budget: Per-firing cap on LLM turns (max_turns) plus daily rolling caps on turns and cost, enforced before and during a firing. See also: Architecture.
• worktree: Throwaway git worktree under $ALFRED_HOME/worktrees/ branched from a fresh origin/main, where the engine writes code for one firing. See also: Alfred on a monorepo.
• WORKSPACE_ROOT: Operator-set directory whose product/ subdirectory contains the local checkouts of every repo Alfred operates against. See also: Workspace patterns.