fix: release the token counter when CompactionHook is closed

[nanhanq1]

Related Issues

• fixes CompactionHook.close() releases the compactor but never the token counter #12984

Proposed Changes:

CompactionHook owns two objects that can hold resources - the token_counter and the compactor - but only the compactor had a full lifecycle. warm_up() and warm_up_async() warmed up both, while close() and close_async() released only the compactor, and nothing else in haystack/ closes a token counter. A counter that had been warmed up was therefore never released.

OpenAITokenCounter is where this shows: warm_up() builds an openai.OpenAI client, and with it an HTTP connection pool, and its own close() drops it again. Agent.close() -> close_hooks() -> CompactionHook.close() never reached that close(), so the client outlived the hook that created it.

The hook now treats both resources the same way in all four lifecycle methods:

• close() and close_async() release the token counter, then the compactor.
• warm_up_async() and close_async() await a resource's warm_up_async / close_async when it defines one, and fall back to the sync method otherwise. That is the contract haystack/hooks/utils.py already uses for hooks; warm_up_async() previously called a token counter's sync warm_up() unconditionally.

The two getattr/hasattr pairs are factored into the private _warm_up_async / _close_async helpers so both resources go through the same path instead of repeating the pattern.

How did you test it?

Three new unit tests in test/hooks/compaction/test_hooks.py:

• test_lifecycle_also_closes_the_token_counter and test_lifecycle_also_closes_the_token_counter_async use a counter that records the lifecycle calls it receives, and assert it sees the same calls as the compactor does.
• test_lifecycle_closes_a_real_openai_token_counter uses the real OpenAITokenCounter and asserts counter.client is built by warm_up() and dropped by close(). No network call is made - warm_up() only constructs the client.

Checks run locally:

• A/B on the fix: with hooks.py reverted to main, exactly those three tests fail (3 failed, 38 passed); with the fix, 41 passed.
• test/hooks + test/components/agents: 682 passed, 5 skipped.
• ruff check, ruff format --check, mypy and codespell clean on the changed files.

Notes for the reviewer

The warm_up_async() change is the one deliberate behaviour change beyond the reported leak: a token counter that defines warm_up_async is now awaited rather than having its sync warm_up() called. No counter in the repository defines one, so nothing changes for the built-in counters, and it makes the token counter match how the compactor is already handled. Happy to drop it if you would rather keep the diff to the close path.

The token counter is duck-typed the same way the compactor already is - via hasattr, not a change to the TokenCounter protocol - so ApproximateTokenCounter and TiktokenCounter, which have no close(), are unaffected.

Checklist

• I have read the contributors guidelines and the code of conduct.
• I have updated the related issue with new insights and changes.
• I have added unit tests and updated the docstrings.
• I've used one of the conventional commit types for my PR title: fix:.
• I have documented my code.
• I have added a release note file, following the contributors guidelines.
• I have run pre-commit hooks and fixed any issue.

This PR was written with the help of an AI assistant. I have reviewed the changes and run the relevant tests locally.

[vercel]

@nanhanq1 is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[anakin87]

Thank you!

I simplified a bit the implementation

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
haystack/hooks/compaction
hooks.py					296
Project Total

This report was generated by python-coverage-comment-action