[nguyenminhduc9988]

Hi Yasser — the PR we discussed over email. Three MIT drop-in agent templates for the LLM Gateway, one level up from dogeapi.ai.examples: typed outputs plus deterministic guardrails, and an eval harness so buyers can customize a template and verify nothing broke.

What's in dogeapi.ai.templates

support_triage — classifies inbound support messages into TriageResult (category / severity 1-4 / summary / suggested_reply / needs_human). Two guardrails beyond the prompt:

• a deterministic output validator enforces a severity floor: severity-4 results are always needs_human=True, no matter what the model said;
• a fetch_recent_tickets tool wired through SupportContext deps (plain sequence by default, so it runs with zero infrastructure — swap in your ticket-store query when you drop it in).

rag_answerer — answers over caller-retrieved chunks and refuses to emit claims without a citation: an output validator raises ModelRetry on empty citations or ids outside the retrieval set, and forces insufficient_context=true when retrieval came back empty instead of letting the model improvise. Retrieval is deliberately out of scope so the template stays storage-agnostic (pgvector, Qdrant, whatever).

evals — golden-transcript harness: GoldenCase (exact-match expected field assertions + contains substring checks), JSON loader, run_golden_cases() → EvalReport with a CI-friendly summary(). One crashed case fails that case only. Ships sample goldens for the triage agent.

Conventions followed

• Built on build_agent(); pydantic_ai stays a lazy import everywhere, so the ai extra is only required once a build_* factory is called (same discipline as agents.py/examples.py). The two tool-defining modules skip from __future__ import annotations (tool signatures must hold real types at runtime for schema building) — noted in their docstrings.
• Tests are offline respx mocks against the gateway, mirroring test_pydantic_agent.py: 9 new tests covering structured parsing, the severity floor, the citation-retry loop, empty-retrieval refusal, and the harness itself.
• Full backend suite: 123 passed locally (postgres:17-alpine + redis, uv sync --all-extras). ruff check, ruff format --check, and mypy --strict clean on the new files.

Happy to adjust naming, module placement (templates vs anything you prefer), or docs format — I kept the README delta to one line pending your preference.

🤖 Generated with Claude Code

[codecov]

Welcome to Codecov 🎉

Once you merge this PR into your default branch, you're all set! Codecov will compare coverage reports and display results in all future pull requests.

Thanks for integrating Codecov - We've got you covered ☂️