This glossary defines each Kiln term in one or two sentences and links to the page that covers it.
Factories and work
| Term | Definition |
|---|---|
| Factory | A team of agents attached to the repositories that ship together, defined by a tree rooted at factory.yaml. See Definitions as code. |
| Alias | The factory's handle, such as webapp. People reach the foreman with @webapp in Slack, GitHub, Linear or Jira. Lowercase letters, digits and dashes, up to 60 characters. |
| Work item | One request, from intake to completion. It keeps its source, conversation, branch, pull request, cost and the definition SHA it runs under. See How Kiln works. |
| Foreman | The agent with agentType: FOREMAN. It coordinates work items and chooses the path when no pipeline is set. A factory has exactly one. |
| Agent | A role in the factory, defined in agents/<name>/agent.md with an agentType, a harness or model, and instructions. See Factory agents. |
| Stage | One step of a work item's path: triage, spec, build, verify, preview, review or hand-off. |
| Pipeline | An optional, fixed list of stages in pipelines/<name>.yaml. See Pipelines and gates. |
| Gate | A pause where a person decides: spec approval, preview acceptance, a question or merge. Open gates make up the inbox. |
| Hand-off | The last stage. The pull request is marked ready and a person is asked to merge it. |
| Project | A group of factories for planning and reporting. See Portfolio and projects. |
| Roadmap | Epics and issues the planning agent produces from a goal or PRD. See Planning. |
Configuration
| Term | Definition |
|---|---|
| Definition tree | All files that define a factory: factory.yaml, agents/, automations/, pipelines/, gates/, policies/, decisions/, runners/, routers/, scorers/, benchmarks/, skills/, webhooks/ and memories/. |
| Definition SHA | A 40-character content hash of the whole tree. Kiln stores every applied tree by this SHA, and every run records it. |
| Applied definition | The tree new work items use. Kiln applies a tree whole or not at all; an invalid tree leaves the previous one active. |
| Config repo | A GitHub repository that holds the definition tree. Merges to its branch apply the new tree. |
| Automation | A trigger file that starts work from an event. See Automations. |
| Skill | A SKILL.md file with know-how that Kiln hands to agents. See Skills. |
| Memory | A learned convention stored as memories/<slug>.md and added to later runs. See Skills. |
| Runner | Compute settings for an agent's sandbox: image, backend, shape, setup commands and egress. See Sandboxes. |
| Router | A file in routers/ that picks a model per task, by complexity or by prompt. Agents use it with model: router:<name>. |
| Policy | Rules in policies/<name>.yaml, such as paths self-improvement may not change. See Policies and budgets. |
| Decision | A typed question set with thresholds, in decisions/<name>.yaml. See Typed decisions. |
| Expression | The small language used by skip_if, gate rules and decision policies. See Typed decisions. |
Runs and safety
| Term | Definition |
|---|---|
| Run | One agent session in one sandbox. It records the agent, harness, model, definition SHA, start state, events, cost and outcome. |
| Harness | The coding agent program a run uses: claude-code, codex or opencode. See Harnesses. |
| Sandbox | The disposable environment a run works in: local directory, Docker container or Kubernetes pod. |
| Start state | What a run started from: prompt, instructions, skills and the exact commit of each repository. Replays start from it. |
| Result file | .kiln/result.json, which the agent writes at the end of a run to report structured results. See Harnesses. |
| Run token | A short-lived token for one run. It can push only kiln/* branches of the factory's repositories, open pull requests, comment and label. It can never approve or merge. |
| Command guard | The check that rates shell commands before they run and allows or blocks them. It never asks a person: a risky command runs in the run's isolated sandbox and is recorded. |
| Steering | A message sent to a running agent that it reads before its next step. |
| Scope | A permission on an API or MCP token, granted per factory. See Admin, RBAC and audit. |
Measure and improve
| Term | Definition |
|---|---|
| Scorer | A rubric with labels that grades a share of an agent's runs. See Scorers. |
| Score | One scorer's label for one run, with a pass or fail against the scorer's passingScore. |
| Self-improvement | The loop that turns unreviewed failures into a proposed definition change, proven by replay. See Self-improvement. |
| Proposal | A self-improvement change with its replay report, opened as a pull request or kept for review in Kiln. |
| Replay | A new run from a saved start state, optionally under another definition. |
| Canary | A candidate definition that only a share of new work items use, so you can compare it to the applied one. |
| Benchmark | A suite of tasks pinned to exact commits, run under up to six configurations. See Benchmarks. |
| Autonomy | The share of merged pull requests whose only human touch was the original request. See Metrics. |