Glossary

The vocabulary the rest of the docs use, defined once. Each entry is one or two sentences and links to the page that owns the concept in full. This page is for integrators who hit a term mid-page and want the precise meaning without re-reading the chapter it came from. Names in backticks are verbatim: the exact config key, event string, or identifier the engine uses.

A few internal identifiers and _meta keys still carry a symphony/ prefix for historical reasons, even though the user-facing surface is lorenz.

A

ACP bridge - the external subprocess that adapts a coding agent (Codex or Claude) to the Agent Client Protocol, spawned and driven by the built-in acp executor over @agentclientprotocol/sdk. The bridge is a shell command (bridge_command, defaulting to codex-acp or claude-agent-acp) that runs locally under bash -lc or over SSH on a worker host, where the bridge binary must be installed. See ../agents/acp-bridges.md.

Agent kind - a named per-agent configuration block under agents.<kind> (for example agents.codex, agents.claude) selecting which executor runs and carrying that executor's options. The active kind comes from agent.kind (default codex); AgentKind is an open string alias, not a closed set. See ../agents/index.md and configuration.md.

C

Continuation turn - the retry Lorenz schedules after a clean agent exit, distinct from a failure retry. finishAsync(normal) always writes a continuation retry (attempt 1, fixed backoff near 1000ms); reconciliation later prunes it if the issue is no longer active. See ../agent-orchestrator.md and workflow-prompt.md for the continuation prompt.

D

Dispatch - the act of turning an eligible issue into a running agent. The runtime dispatches each eligible issue as a detached per-run promise: the static or local path claims a slot and runs immediately, the pool path reserves a slot, acquires a worker, then binds. See ../dispatch.md.

Dispatch block - a reason a candidate issue is held back this tick instead of dispatched. The DispatchBlockReason values are global_concurrency_cap, local_concurrency_cap, and worker_host_capacity; blocked issues surface in the snapshot's blocked lane. See ../dispatch.md and events.md.

E

Eligibility - the orchestrator's decision about which fetched issues may be dispatched now. eligibleIssuesAsync sweeps expired reservations, prunes retry attempts for inactive issues, resets the blocked list, then sorts and filters: issues with retries not yet due are dropped. See ../dispatch.md.

Ensemble - running more than one agent on the same issue concurrently, one per slot. The count comes from an ensemble:<n> label on the issue or falls back to agent.ensemble_size. See ../features/context-ensembles.md.

Executor - the runtime contract (AgentExecutor) that starts a session and runs turns for one agent. Each executor is registered by an AgentExecutorProvider under the agents.<kind>.executor selector; the only built-in value is acp. See ../extensions/agent-executor.md.

H

Hook - an operator script run at a point in the workspace lifecycle: after_create, before_run, after_run, or before_remove, under hooks.* with a shared hooks.timeout_ms. Hooks run arbitrary code in the daemon's trust boundary. See ../workspace.md.

L

Ledger - the write-ahead JSON file the worker pool keeps for cloud drivers so it can recover leased machines after a crash. It is inert (zero filesystem I/O) unless the driver's capabilities.usesLedger is true and a ledger path is supplied; a sibling spend.json tracks daily worker-seconds. See ../workers/worker-pool.md.

Lease - one agent run's hold on a pooled worker. acquire() yields a WorkerLease (on a leased result) whose release(outcome) (healthy or poison), fail(reason), and heartbeat() settle it exactly once; a poison outcome recycles the machine on its last lease return. See ../workers/worker-pool.md.

M

MCP endpoint - the HTTP POST /mcp JSON-RPC surface that exposes agent-callable tools, mounted by @lorenz/mcp and leased per agent run as a local server, an SSH reverse tunnel, or a per-run tunnel. The ACP server name shown to an agent is lorenz_<kind>. See http-api.md and jira-tools.md.

P

Provider - the extension that backs a tracker. A TrackerProvider is registered under a kind, creates the dispatch client, and declares the tool packs it owns through defaultToolPacks; selected by trackers.<name>.provider (or the flat tracker.kind). See ../extensions/tracker-provider.md and ../trackers/index.md.

R

Reaper - the single serial recurring pass that reconciles the worker pool against driver.list(): it destroys unknown pool-owned machines, reaps over-TTL and idle workers above the floor, demotes failing probes, and tops up warm capacity. Cadence is worker.worker_pool.reap_interval_ms (default 15000). See ../workers/worker-pool.md.

Reconciliation - the per-poll passes that bring in-memory state back in line with truth. Stall reconciliation finishes runs idle past their stall timeout; tracked-issue reconciliation refetches in-flight issues and stops, refreshes, or cleans up each one based on its current tracker state. Reasons are terminal, unrouted, blocked, and inactive. See ../agent-orchestrator.md.

Route label - the tracker label that marks an issue for a specific worker, matched against the route_label_prefix (default Lorenz:). With accept_unrouted (default true) and only_routes, the prefix governs which issues a worker claims. See ../features/dispatch-routing.md.

RuntimeSnapshot - the single object that feeds every UI. It carries appStatus, poll, running[], reserving?[], retrying[], blocked[], runHistory[], usageTotals, rateLimits, logFile, and recentEvents[], assembled from orchestrator state plus the projection ring buffers (last 20 events, last 50 history entries). See ../observability.md.

S

Skill - a directory of agent instructions a tool pack overlays into the workspace when mounted, so enabling a tool ships its companion guidance. The mounted set is the union of every active pack's skills. See ../agents/skills.md.

Slot - one unit of agent concurrency on an issue, keyed by (issueId, slotIndex). A slot moves through claimed or reserved, running, finished, then retry; reserved (in-acquire) slots count toward every concurrency cap so dispatch cannot exceed agent.max_concurrent_agents. See ../dispatch.md.

Stall timeout - the inactivity timer that cancels a turn after no agent update for agents.<kind>.stall_timeout_ms (default 300000); a value of 0 or less disables stall detection. A stall-finished run records a stalled outcome and poisons its worker. See ../agents/index.md.

Status override - a per-state config block under status_overrides[state] that adjusts behavior for issues in a given tracker state, for example a per-state agent.max_concurrent_agents. See configuration.md.

T

Tool pack - a named bundle of agent-callable tools (a ToolProvider) registered under a name, mounted into the flat MCP namespace, and optionally shipping skills. The Jira extension's jira pack serves the seven jira_* tools; provider-specific packs (linear, local, slack) ship in the tracker extensions. See ../extensions/tool-pack.md and jira-tools.md.

Tracker - the issue backend that drives dispatch: Linear, Jira, a local board, Slack, or the in-memory fixture. Lorenz polls the tracker for candidate issues and writes status and comments back through it. See ../trackers/index.md.

TraceViz - the dashboard's per-issue trace view, fed by the HTTP API and a WebSocket stream of events and events_append messages. See ../observability.md.

Turn timeout - the hard timer that cancels a turn after agents.<kind>.turn_timeout_ms (default 3600000) regardless of activity, distinct from the stall timeout. On fire the ACP executor rejects the active and queued turns, terminates the session, and rejects the run. See ../agents/index.md.

W

Warm pool - the standing set of idle workers the pool keeps ready so an acquire need not wait on provisioning. The reaper tops up toward max(min, warm) within budget; worker.worker_pool.warm defaults to 1. See ../workers/worker-pool.md.

Workflow - the WORKFLOW.md file: YAML front matter (config) plus a Liquid prompt body. The runtime reloads it transactionally before each poll and falls back to the last-known-good copy if the reload fails. See ../workflows.md and ../features/workflow-hot-reload.md.

Worker - the SSH-addressable host an agent run executes on, identified by its workerHost. A run reaches a worker either through the legacy static worker.ssh_hosts list or through the warm worker pool. See ../workers/index.md.

Worker driver - the swappable backend (WorkerDriver) the pool calls to provision, probe, destroy, and list machines, selected by worker.worker_pool.driver. Built-ins are fake, static-ssh, and docker; drivers can also load out-of-tree by module specifier with an sdkVersion handshake. See ../extensions/worker-driver.md and ../extensions/out-of-tree.md.

Glossary

A

C

D

E

H

L

M

P

R

S

T

W

See also