Pin MCP SDK to v1 and fix stale prompt instructions

[Brandon Werner (brandwe)]

Summary

Three commits:

11b2b32 fix: pin MCP Python SDK to v1

• Adds an upper bound on mcp in pyproject.toml, with a constraint test in tests/test_dependency_constraints.py.

3fd0eb9 test: align suite with canonical agent instructions

• Updates test_prompt_doctrine.py (and one script test) to follow the instruction consolidation in f878cea: CLAUDE.md now delegates to AGENTS.md, and .github/copilot-instructions.md is gone.

8a67b89 fix(prompts): remove contradictory and stale prompt instructions
Comes from a prompt audit against Claude Opus 5.5. Most of it fixes instructions that contradicted each other or sent the model to the wrong tool:

• read_interactions told the model to scan history before every send, which the body prompt forbids.
• wait_for_sponsor_dm told the model to loop on itself, which makes it wait twice now that send_teams_message waits on its own. It also said Claude Code has no channel push, which is wrong.
• read_file and tools/files.py pointed to read_workbook_range, which doesn't exist.
• The SessionStart hook told the model to Read the persisted-output file, but the require_body_prompt gate doesn't accept that file. The hook now names prompts/agent_system.md and the anatomy/ files.
• Removed the PostToolUse hook that added a stale watch_teams_replies reminder after every send.
• identity-and-tools.md:
  • removed the stale workflow steps and the partial prose tool list (11 of 37 tools)
  • fixed memory routing to match AGENTS.md
  • cut the Python hygiene section down to deps this repo actually has
• channel-discipline.md: dropped "when in doubt, spawn" and the reference to the missing superpowers: skill.
• AGENTS.md: corrected the line that contradicted body-first prompt loading, and cut "Active Work" down to pointers to docs/project/status.md.
• refresh-persona skill: now reloads from persona-sati instead of the local memory dir.
• send_teams_message / create_chat: removed examples that contradicted always-HTML. The six Agent 365 tools now have full descriptions.

security.md was not changed.

Test plan

• pytest -v && ruff check .: 1688 passed, 1 skipped, lint clean
• Check live Teams behavior after the changes most likely to affect it: the removed after-send reminder, "when in doubt, spawn", and the read_interactions wording

🤖 Generated with Claude Code