[rupayon123]

Undoing an earlier hook-call monitor currently removes later monitors, and later undo calls can restore callbacks that were already removed. Keep a persistent execution chain so each undo removes only its own monitor, repeated undo stays harmless, and in-flight calls retain their original before/after pairing.

The lifetime regression invokes active callbacks before checking their release. Undo traversal now expresses the internal invariant that every active token remains in the chain; directly replacing the private _inner_hookexec while monitors are active fails that assertion. Existing public tracing/monitoring lifecycles, custom original executors, nested calls and result/exception handling remain covered. Includes the numbered changelog fragment.

Validation on macOS arm64:

• All 260 repository tests, 68 focused tests and all 10 pinned pre-commit hooks pass with the frozen lockfile.
• Every changed line and branch is covered locally; combined coverage is 99.9741%, versus 99.9727% on pristine base. The same old _caller.py branch remains uncovered locally.
• Native lifecycle checks cover all 120 removal orders for five monitors, custom original-executor identity, nested calls and callback collection. Independent review adds public normal/first-result, self-undo/reinstall and tracing/result-change scenarios.
• The hosted benchmark failure reproduces during fixture setup on both unchanged base and this branch with pytest 9.2.0 and pytest-benchmark 5.3.0: the fixture accesses the removed Function._nodeid. Local reproduction uses Python 3.10.17; the hosted job uses 3.10.22 on Linux. No benchmark, dependency, coverage threshold or configuration was changed.

Hosted checks for the latest repair still need verification. No cross-thread registration/undo guarantee is added.

[rupayon123]

I updated the lifetime test to call both monitors before checking that their callbacks are released, and simplified undo around the active-token invariant. The full 260 tests, 68 focused tests and all pinned hooks pass. Every changed line and branch is covered locally; hosted coverage for this commit still needs to run.

The benchmark failure also occurs on unchanged base with the job's pytest 9.2.0 and pytest-benchmark 5.3.0 versions: the fixture accesses the removed Function._nodeid before the benchmark body starts. I left benchmark dependencies and configuration unchanged. The PR description now records that baseline comparison and the private execution-chain invariant.

[RonnyPfannschmidt]

🤖 Written by Claude Opus 5.5 via Claude Code for the pluggy maintainers; I prompted it, it did the work, I read it.

Thanks, the undo-order bug in add_hookcall_monitoring is real. This PR looks like unattended agent output, though, and the 314-line token and chain-rebuild approach is more than the problem needs. Our AI policy asks for a contributor who engages with review themselves, so I'm closing this. We may fix it separately with a smaller list-based change.

---

Generated by Claude Code