AgentSkillsCN

team-prompt-contract-testing

当您修改 src/agent_loom/team/prompts.py（或团队运行时中面向提示的行为）时，应确保提示保持确定性，并通过 tests/test_team_prompts.py 进行充分测试。

SKILL.md
--- frontmatter
name: team-prompt-contract-testing
description: Use when modifying src/agent_loom/team/prompts.py (or prompt-facing behavior in team runtime) to keep prompts deterministic and covered by tests/test_team_prompts.py.
license: MIT
compatibility: opencode,claude
metadata:
  created_at: "2026-01-30T20:09:25.393Z"
  updated_at: "2026-01-30T20:09:25.393Z"
  version: "1"
<!-- BEGIN:compound:skill-managed -->

When to use

  • •You changed prompt assembly, templates, or any code in src/agent_loom/team/prompts.py.
  • •You changed prompt-facing behavior in src/agent_loom/team/core.py or src/agent_loom/team/cli.py.
  • •You introduced/modified team runtime modules that influence prompt inputs (commonly src/agent_loom/team/inbox.py, src/agent_loom/team/models.py, src/agent_loom/team/constants.py).

Goal

Keep prompt outputs stable, reviewable, and regression-tested.

Steps

  1. •

    Identify the prompt contract you changed

    • •Inputs: what variables/context are interpolated?
    • •Outputs: what exact text format or invariants must hold?
  2. •

    Add/update tests

    • •Prefer updating tests/test_team_prompts.py.
    • •Assert on stable invariants (sections/headers/required phrases) instead of brittle full snapshots unless a full snapshot is the contract.
    • •Add a regression test for any bug fix.
  3. •

    Run diagnostics and checks

    • •Run LSP diagnostics on touched files via lsp_diagnostics.
    • •Run uv run ruff check ..
    • •Run uv run pytest tests/test_team_prompts.py (or the smallest relevant subset).
  4. •

    Review prompt diffs like API diffs

    • •Read the final prompt text as if you are the model.
    • •Confirm: no accidental leakage (paths, secrets), no contradictory instructions, no duplicated sections, and consistent terminology.

Common failure modes

  • •Tests assert full prompt text but the prompt contains nondeterministic content (timestamps, paths, order-dependent dict iteration).
  • •Prompt changes land without corresponding test coverage, causing silent regressions.
  • •“Small” refactors subtly reorder or drop constraints (role, tool rules, JSON-only requirements).
<!-- END:compound:skill-managed -->

Manual notes

This section is preserved when the skill is updated. Put human notes, caveats, and exceptions here.