Process agents

The process layer ships eight focused agents under process/agents/. Each agent has a narrow job, a small context budget, and declared safety boundaries in its YAML frontmatter.

Agent inventory

Agent	Stage	Safety tier	Mutation	Context budget
Pipeline Orchestrator	(all)	tier1	read-only	~20k tokens
Business Requirements Analyst	1	tier1	read-only	~50k tokens
Service Designer	2	tier1	read-only	~35k tokens
Data Architect	3	tier1	read-only	~30k tokens
API Architect	4	tier1	read-only	~40k tokens
UI Architect	5	tier1	read-only	~30k tokens
Scaffolding Orchestrator	6	tier1	read-only	~20k tokens
Client Documentation Orchestrator	CD	tier1	scoped-write	~25k tokens

Frontmatter fields

Every process agent declares the following in its YAML frontmatter:

• stage: the pipeline stage this agent serves (omitted for the pipeline orchestrator, which coordinates all stages).
• safety_tier: the maximum privilege tier. All process agents are tier1.
• mutation: what the agent may write. Seven agents are read-only; the client-documentation orchestrator is scoped-write with mutation_scope: ["requirements/client/**"].
• context_budget: the approximate token budget for the agent's working context.

Governance envelope ceilings

The process governance envelope (process/governance-envelope.yaml) declares aggregate ceilings that must bound the per-agent frontmatter:

• max_tier: tier1 (all agents are tier1).
• max_mutation: scoped-write (the ceiling is set by the single scoped-write agent).

OAP independently recomputes the aggregate from the agent frontmatter and confirms the declared ceilings bound it. This is a two-sided, fail-closed validation.

Pipeline Orchestrator

Coordinates the end-to-end pipeline. It invokes each stage agent in order, runs the verification gate after each stage, persists pipeline state, and hands off to the scaffolding orchestrator. It holds no domain content of its own; it sequences and checks.

Key responsibilities:

• Run stages in order; never start stage N+1 until stage N's exit gate passes.
• Pause for human confirmation with deterministic facts only.
• Write pipeline state after every successful step.
• Resume from artifacts on disk rather than memory.

Scaffolding Orchestrator

Sequences the adapter's code-generation agents against the frozen Build Specification at Stage 6. It never writes application code itself; it invokes the adapter's agents (data scaffolder, API scaffolder, UI scaffolder, configurer, trimmer) and runs the verification harness after each feature.

Client Documentation Orchestrator

The only process agent with write permissions. It runs off the critical path (optional, never blocks the build), consumes Stage 1 and 2 outputs, and produces client-facing documentation under requirements/client/. Its scoped-write mutation is bounded to that directory only.