Prompt Pipeline

Every conversation turn dynamically assembles agent instructions before calling the LLM via the @openai/agents SDK. This page describes how the harness builds per-turn context.

Source files:

• packages/harness/src/runtime/skill-resolver.ts — per-turn skill resolution
• packages/harness/src/runtime/runner.ts — turn orchestration
• packages/web/api/src/functions/converse.ts — HTTP handler

Assembly Order

POST /api/converse arrives
    │
    ├─ 1. Session lookup or creation
    ├─ 2. Guardrail input check (core/token-budget, etc.)
    │
    ├─ 3. Runner selects active Agent
    │      session.activeAgent ?? "core.triage"
    │
    ├─ 4. Dynamic instructions assembled per turn
    │      = agent.body (base .agent.md text)
    │      + resolvedSkills(agent.name, recentTurns)  ← skill resolver
    │      + catalog snapshot (component type list)
    │
    ├─ 5. Agent streams via SDK — text chunks + tool calls
    │      core.emit_ui → a2ui SSE events
    │      core.write_file, azure.arm_get, etc.
    │
    ├─ 6. UserAction encountered?
    │      → pause, emit user_action_required SSE
    │      → browser acts (MSAL popup, GitHub OAuth, etc.)
    │      → POST /api/converse/resume with typed result
    │      → Runner resumes
    │
    ├─ 7. Guardrail output check
    │
    └─ 8. AgentOutput { message, intent } → SSE done

Skill Resolution

resolveSkills(agentName, context) in packages/harness/src/runtime/skill-resolver.ts:

1. Match appliesTo — glob each skill's appliesTo field against the current agent name.
2. Keyword scoring — score matched skills against the recent conversation turns.
3. Priority ordering — sort by priority (higher first), apply token budget cap (2000 tokens default).
4. Inject — append skill text to the agent's dynamic instructions.

Skills are authored as SKILL.md files inside packs. They are pure text — no code, no execution.

Agent Instructions Structure

{agent.body}                     ← base .agent.md text (persona + rules)

## Active Skills
{skill1 text}
---
{skill2 text}

## Component Catalog
{registered component list with schemas}

Guardrails

Guardrails contributed by packs run at three stages:

Stage	When
input	Before the agent receives the user message
tool	Before each tool call result is returned to the agent
output	After AgentOutput is produced

Built-in guardrails: core/token-budget, core/content-safety. Pack guardrails: azure/no-hardcoded-creds, aks/no-privileged-containers, etc.