fix(function_manager): check names and dependencies against the whole store in add_functions

[breken-ai]

Summary

FunctionManager.add_functions read existing names through list_functions(). That call applies the instance's filter_scope and exclude_compositional_ids, and the actor sets both at runtime. As a result, a stored function hidden from discovery was treated as if it did not exist during a write:

• Dependency detection misses it. A new function that calls a hidden function is stored with depends_on == []. _inject_dependencies then never injects the dependency when the function is executed later, and delete_function(delete_dependents=True) no longer cascades to it.
• overwrite=True inserts instead of updating. Adding a function with the same name as a hidden one runs an INSERT, which fails on UNIQUE constraint failed: functions.name. Because the inserts share one transaction, every other new function in that batch is rolled back with it.

The docstring of tests/function_manager/core/test_filter_scope.py states that "Write paths are unaffected" by filter_scope. list_function_name_to_ids() is documented as the unscoped catalogue "so a reference resolves even when its function is hidden from discovery". This PR makes add_functions use that catalogue for its name and dependency checks.

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

Two new regression tests in tests/function_manager/core/test_filter_scope.py:

• test_scoped_add_records_dependency_on_hidden_function
• test_scoped_overwrite_updates_hidden_function

# on main (fa912bec): both fail
AssertionError: depends_on [] != ['hello_world']
ValueError: Failed to add function(s): hello_world: error: Failed to create log - UNIQUE constraint failed: functions.name
2 failed, 9 passed

# on this branch
.venv/bin/python -m pytest tests/function_manager/ tests/guidance_manager/ tests/actor/code_act/test_directly_callable.py tests/actor/environments -m "not llm_call"

The broader run gives the same result on main and on this branch, except for the new tests. The only failures are 8 tests in test_live_steering_eval.py and test_default_env_steering.py, which need a live model key that I don't have. They fail the same way on main. 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 tests, and submitted the PR from the breken-ai account.

🤖 Generated with Claude Code