Quick reference for every term used in this repo.
| Term | Definition | Why it matters |
|---|---|---|
| Agent type | An abstract role describing what kind of work gets done — not which AI model does it. Currently two: builder and validator. | Lets you swap AI providers without rewriting your process. |
| Builder | The agent type that creates work — produces deliverables, verifies quality, publishes results. | Keeps creation and review separate. |
| Committee | All of a team's personas reviewing work in sequence, each reading prior feedback first. | Produces multi-perspective review that no single reviewer can match. |
| Fresh-eyes validation | A zero-context sub-agent reads only the final spec and flags gaps. Catches assumptions the committee forgot to write down. | Ensures specs are self-contained for whoever implements them. |
| Manifest | A YAML file (manifest.yml) that defines a team's roles, pipeline stages, and vocabularies. Single source of truth. |
Change one file, update everywhere. No drift between docs and config. |
| Overlay | Optional domain-specific rules (e.g., healthcare, fintech) layered on top of the base process. Additive, never replacing. | Keeps domain compliance separate from team fundamentals. |
| Overwrite-to-consensus | After committee deliberation, members whose positions changed edit their original comments to show final positions. | Readers see clean conclusions, not debate threads. |
| Persona | A detailed character profile — title, backstory, expertise, review lens — that shapes how an AI agent thinks about a specific kind of work. | Produces deep, targeted feedback instead of shallow, generic observations. |
| Pipeline | A multi-stage workflow (e.g., define → design → implement → review → deploy → summarize). Each stage produces artifacts the next consumes. | Prevents skipping steps. Tracks progress with labels. |
| Per-stage override | An optional stages map on a manifest role that overrides the role's default agent type for specific pipeline stages. Resolution: `role.stages[stage] |
|
| Pipeline mode | How much human involvement a project requires. Autonomous: no human gates. Gated: pauses for human approval at key stages. | Lets you tune the system from fully automated to human-in-the-loop. |
| Severity level | A tag on review findings. MUST-FIX: blocks integration. SHOULD-FIX: blocks current round. NIT: suggestion only. | Shared vocabulary prevents disagreements about what's blocking. |
| Three-tier model | Configuration lives at three levels: Tier 1 (this repo — shared practices), Tier 2 (organization — optional), Tier 3 (project — tech stack and environment). | Prevents duplication. Each tier adds specificity without repeating the tier above. |
| Validator | The agent type that reviews work — audits deliverables, checks quality, flags issues. | Independent review catches blind spots the builder can't see. |