fix(actor): resolve prompt_guidance titles that contain a quote

[breken-ai]

Summary

_resolve_prompt_guidance builds title = '{identifier}' without escaping. When a guidance title contains an apostrophe (for example prompt_guidance=["Client's Refund Policy"]), SQLite rejects the clause. GuidanceManager.filter then returns its invalid_filter error payload, which is a dict, instead of a list. The next line iterates that dict's keys, so primitives.actor.act(...) fails before the inner actor starts, with AttributeError: 'str' object has no attribute 'guidance_id'.

The fix escapes single quotes in the literal by doubling them, which is the standard SQL escape. The rest of the function is unchanged.

Type of change

• Bug fix (non-breaking change that fixes incorrect behavior)
• Feature (non-breaking change that adds functionality)
• Refactor (no behavior change)
• Breaking change (API or data-model change — Unify has zero-backward-compat policy, but please call it out)
• Test-only (no source changes)
• Docs / chore

Areas touched

• Actor / CodeAct
• ConversationManager / slow brain
• A specific state manager (Contact / Knowledge / Transcript / Guidance / Function / File / Ingestion / Image / Web / Secret / Data / Memory)
• Async tool loop (unify/common/_async_tool/)
• Event bus / observability
• The local store (unify/db/)
• Tests / test infra (tests/, conftest.py, parallel_run.sh)
• Build / packaging

Test plan

New regression test tests/guidance_manager/test_gm_environment_exclusion.py::test_resolve_prompt_guidance_title_with_apostrophe.

# on main (fa912bec): fails
unify/actor/environments/actor.py:160: AttributeError: 'str' object has no attribute 'guidance_id'
1 failed, 16 passed

# on this branch
.venv/bin/python -m pytest tests/guidance_manager -m "not llm_call"
82 passed

black and autoflake are clean on the changed files.

• All relevant tests pass locally
• If this is a bug fix, I added a regression test (or explained why one isn't feasible)

Behavior / migration notes

None.

Checklist

• Followed conventional commit style (feat(scope):, fix(scope):, refactor(scope):, chore(scope):, etc.)
• No try/except added defensively — only around specific, recoverable errors
• No "new" / "updated" / "TODO from chat" temporal comments (see .agents/rules/no-temporal-comments.md)
• No test-specific shortcuts in production code (see .agents/rules/no-test-info-in-production-code.md)
• Updated AGENTS.md / ARCHITECTURE.md if I changed architectural conventions (not applicable)

An AI agent (Breken, using Claude) found this bug, wrote the fix and the test, and submitted the PR from the breken-ai account.

🤖 Generated with Claude Code