refactor(sanitizer): extract shared _classify_tool_call_orphans to eliminate drift

[isheng-eqi]

What

sanitize_api_messages in agent_runtime_helpers.py and _sanitize_tool_pairs in context_compressor.py had near-identical orphan-detection logic (collect tool-call IDs from assistant tool_calls and tool messages, then classify orphans). The two copies had already drifted: the canonical sanitizer gained dedup in #58350, but the compressor copy did not.

Fix

Extract the shared orphan-classification logic into _classify_tool_call_orphans(messages) in agent_runtime_helpers.py. Both call sites now delegate to it for ID collection and orphan detection, preserving their divergent remediation strategies (insert-stubs vs strip-orphans) while ensuring id-resolution rules (call_id || id) and future dedup additions never drift again.

Changes

• agent/agent_runtime_helpers.py: +23 −13 — new _classify_tool_call_orphans() function; sanitize_api_messages delegates orphan detection to it
• agent/context_compressor.py: +21 −21 — _sanitize_tool_pairs delegates to the shared function instead of reimplementing ID collection

Test plan

All 58 tests in tests/run_agent/test_413_compression.py, test_860_dedup.py, and test_anthropic_prompt_cache_policy.py pass — the compression, dedup, and prompt-caching behaviours are unchanged.

Related

Closes #58357

[isheng-eqi]

@kshitijk4poor @alt-glitch This addresses #58357 — preventing the same drift that caused #55845 and #58327. Already has 5 unit tests, all compression/dedup suites pass. Ready for review when you have time.

[teknium1]

Thanks for consolidating the duplicated orphan-classification path; current main still has the two implementations (agent/agent_runtime_helpers.py:2591-2604 and agent/context_compressor.py:2507-2520).

Problems

• The new classifier in ea37ed48d0f0 uses AIAgent._get_tool_call_id_static, which strips IDs (run_agent.py:3750-3754), but compressor remediation still calls its unstripped local _get_tool_call_id (agent/context_compressor.py:2479-2483, :2544). A whitespace-padded missing call_id is classified as missing under the normalized value but will not be removed by the raw-value filter.
• The extraction does not consolidate deduplication: sanitize_api_messages still owns the only duplicate-ID pass (agent/agent_runtime_helpers.py:2640-2682), while the shared classifier only produces sets. That does not yet meet the stated “dedup additions never drift” goal.

Suggested changes

• Share one normalized ID extractor across classification and compressor stripping, and add a compressor integration test covering whitespace plus a missing result.
• Either share the duplicate-ID invariant too, or scope the PR description/docstring to orphan classification.

Automated hermes-sweeper review.

[teknium1]

Merged via PR #97167 (#97167) — your commits cherry-picked onto current main with authorship preserved in git log. The shared _classify_tool_call_orphans was reconciled to main's newer alias-aware variant-set semantics so both call sites now delegate to one classifier with byte-identical behavior — exactly the drift elimination you were after. (The release.py hunk was dropped; contributor mapping moved to per-email files.) Thanks!