Instruction file imported from rodriguesvali/recruitment-assistant (.github/instructions/adapter-crewai.instructions.md). Copyright stays with the author.

CrewAI Adapter Rules

Purpose

• Define runtime-specific guidance when AAMAD_TARGET_RUNTIME=crewai.
• This adapter governs how implementation personas scaffold and validate the generated MVP runtime.

Setup

• Install CrewAI and dependencies in the generated backend environment.
• Required runtime files should include:
  • config/agents.yaml
  • config/tasks.yaml
  • crew.py (or equivalent runtime entrypoint)
• Load secrets from environment variables and provide names in .env.example.
• Record resolved llm, temperature, and max token controls in Audit.

Mapping

• All CrewAI agent and task definitions MUST be externalized to YAML files under a config/ directory (e.g., config/agents.yaml, config/tasks.yaml).
• Use explicit task context chaining (Task.context) for deterministic dependency flow.
• Define expected_output with required headings and target artifact paths.
• Prefer allow_delegation=false unless the SAD justifies delegated manager patterns.

Execution

• Prefer sequential process mode for reproducible MVP builds.
• Use hierarchical process mode only when justified in SAD and documented in Audit.
• Baseline controls:
  • max_iter <= 12 for MVP tasks unless explicitly justified.
  • max_execution_time tuned per epic.
  • max_retry_limit >= 2.
  • max_rpm set at crew level for budget stability.
• If using function-calling agents, do not switch mode mid-run; record chosen mode in Audit.
• Use kickoff_for_each only for truly independent batch items and document deterministic merge keys.

Tools

• Bind only the minimum required tool set per agent/task.
• Validate YAML-referenced tools before kickoff to avoid runtime binding errors.
• Keep tool configs JSON-serializable; load secret parameters via env vars.
• Permit web/API tools only where explicitly justified by persona scope.

Logging

• Capture rendered system/user prompts in Prompt Trace before execution.
• Record lifecycle events (task start/stop, retries, guardrail outcomes) in Trace Log.
• Keep Prompt Trace and Trace Log in project-context artifacts, not inline with generated code.
• If using step callbacks/event listeners, redact secrets and persist logs under project-context/2.build/logs.

Quality Gates

• Validate required template headings before final artifact write.
• Use Task.guardrail for high-risk outputs (schema limits, content rules, size checks).
• Require Task.id and explicit output path for traceability.
• For machine-ingested output sections, enforce plain markdown/JSON without code fences.
• For high-risk outputs, require a human review gate (human_input=true or explicit review task).

Failure Policy

• On missing runtime prerequisites, unresolved tools, heading mismatches, or guardrail failure: halt and write Diagnostic.
• On context overflow or budget breach: halt and report remediation notes in the artifact.

Memory

• Default memory=False for reproducibility; if memory=True, constrain to current epic, redact secrets, and persist logs to project-context/2.build/logs.
• If memory=True, set CREWAI_STORAGE_DIR to a project-scoped path and record scope/retention in Audit.