# supi-agent

Runs user-requested delegation through managed Agent Runs while keeping profiles, workspace policy, and results explicit.

## Language

**Agent Profile**:
A named user or trusted-project policy that supplies an Agent Run's instructions, model preferences, and allowed tools.
_Avoid_: Agent type, role, preset

**Profile Directory**:
A Profile Directory definition containing a partial `profile.json` settings manifest and optional Markdown prompt assets. Package, global, and trusted-project Profile Directories are separate sources rather than entries in shared SuPi configuration.
_Avoid_: Profile file, agents config entry, profile bundle

**Profile ID**:
The stable name of a Profile Directory and the identifier used by Delegation Tasks. Project, global, and package sources are resolved per field; omitted fields fall through to lower-precedence sources while lists remain whole-field values.
_Avoid_: Profile name field, agent name, source-qualified profile

**Profile Catalogue**:
The immutable, maximum-32 set of Profile IDs, source entries, and Profile Diagnostics retained for one extension runtime at session start or reload. Complete manifests are resolved during task preflight; the first 32 sorted IDs enter the model-facing schema and additional IDs are unavailable with diagnostics.
_Avoid_: Live profile registry, discovered agents

**Profile Diagnostic**:
A bounded configuration error for one unavailable Profile Directory source or incomplete resolved Profile ID. Invalid sources are skipped as a whole, lower-precedence fields may supply the effective profile, and diagnostics appear in settings, startup warnings, and SuPi debug events without entering model-facing guidance.
_Avoid_: Profile warning, fallback notice, Agent Run failure

**Agent Capability Set**:
The fixed package-owned set of tools that Agent Profiles may select for Agent Runs. It includes PI built-ins and explicitly supported child-capable extensions, never ambient extensions.
_Avoid_: Inherited tools, parent tool set, extension allowlist

**Agent Prompt Policy**:
An Agent Profile's choice of one complete base prompt: PI-native, package-owned by Prompt ID, or a Custom System Prompt from the Profile Directory's `SYSTEM.md`.
_Avoid_: Prompt mode, system prompt type, layered prompt

**Custom System Prompt**:
The complete system prompt stored in a Profile Directory's `SYSTEM.md` and selected by `systemPrompt: "custom"`.
_Avoid_: Profile Prompt, Markdown body, appended instructions

**Agent Instruction Scope**:
One of the explicit `global` or `project` AGENTS.md/CLAUDE.md sources selected by an Agent Profile's required `instructionScopes` set. `project` uses PI's normal cwd-ancestry context discovery; PI SYSTEM.md and APPEND_SYSTEM.md files are not instruction-scope inputs.
_Avoid_: Applied context, prompt inheritance, resource mode

**Delegation Task**:
One caller-defined objective assigned to an Agent Profile and producing its own Agent Run Outcome independently of sibling tasks.
_Avoid_: Job, step, chain link

**Delegation Batch**:
An ordered set of independent Delegation Tasks accepted only when every selected profile and policy resolves successfully. Every task in a valid read-only multi-task batch starts concurrently without a queue; canceling the batch cancels every active Agent Run in it.
_Avoid_: Workflow, chain, DAG

**Delegation Batch Result**:
The ordered, attributed parent-facing result of a Delegation Batch. It has one aggregate model-output bound; each task remains represented, and complete bounded task text stays available through human details and a temporary Markdown spill when needed.
_Avoid_: unbounded joined output, Agent transcript, continuation artifact

**Mutation-Capable Profile**:
An Agent Profile whose allowed tools can change workspace state. It is eligible only for a single-task Delegation Batch.
_Avoid_: Writer agent, unsafe profile

**Read-Only Profile**:
An Agent Profile restricted to tools that cannot intentionally change workspace state.
_Avoid_: Safe profile, inspection-only reviewer