Summary

Combined salvage of three community tool-pair sanitizer PRs, applied fixes-first so the final refactor deduplicates the already-fixed logic:

1. fix(compression): strip whitespace from tool_call_id in _sanitize_tool_pairs #55845 by @srojk34 — strip whitespace from tool_call_id in ContextCompressor._sanitize_tool_pairs, closing the sibling gap of fa3ab2ffd: IDs differing only by stray whitespace were misclassified as orphaned and replaced with [Result unavailable] stubs on every compression cycle.
2. fix(agent): drop stale empty tool_calls on repair_message_sequence merge (#77921) #78063 by @JoaoMarcos44 — fix three stale-field bugs in agent/agent_runtime_helpers.py: drop stale empty tool_calls surviving the consecutive-assistant merge (fixes Still reproducing in v0.19.1: empty tool_calls after repair_message_sequence (follow-up to #58755) #77921), drop the stale api_content sidecar when the merge rewrites content (with the wz-heng review's negative controls), and drop tool results with a missing/empty tool_call_id at the pre-API chokepoint (sanitize_api_messages does not catch tool messages with missing/empty tool_call_id #78071).
3. refactor(sanitizer): extract shared _classify_tool_call_orphans to eliminate drift #59434 by @isheng-eqi — extract a shared _classify_tool_call_orphans(messages) used by both sanitize_api_messages and _sanitize_tool_pairs, so orphan-detection/id-resolution rules can never drift between the two copies again (refactor(compression): consolidate _sanitize_tool_pairs with the canonical agent_runtime_helpers sanitizers (dedup/whitespace/orphan drift) #58357).

All three authors' commits are cherry-picked so authorship is preserved; conflict-resolution glue is in a separate commit.

Fixes #77921. Closes #58357. Closes #78071 (chokepoint half).

Changes

#55845 (whitespace ids) — agent/context_compressor.py, tests/agent/test_context_compressor.py (+3 tests)

• Main has since adopted variant-set matching (tool_call_id_variants / tool_result_id_variants from agent/message_sanitization.py, Tool results silently dropped when provider returns composite tool_call ids (call_xxx|fc_yyy) — model never sees output and confabulates #63000/fix(agent): generic 'Unexpected item type in content' 400s now strip images and retry (#57948, salvage #58120) #97075 churn), whose _expand_tool_id_variants already .strip()s — so the behavioral whitespace fix was partially superseded. Salvaged: the .strip() on _get_tool_call_id (still raw on main) and the author's 3 regression tests, which now pin the whitespace-pairing contract against future drift.

#78063 (stale fields on merge) — agent/agent_runtime_helpers.py, tests/run_agent/test_message_sequence_repair.py (+7 tests)

• Fix 1 (empty tool_calls dropped in repair_message_sequence merge) landed on main independently as f316f7d086 — redundant, not re-applied.
• Fix 2 salvaged: content_rewritten tracking in the consecutive-assistant merge; drop the api_content sidecar only when the merged content value actually changed, so prompt-cache replay bytes stay stable when nothing moved (empty-string / None / multimodal negative controls included).
• Fix 3 salvaged: explicit pre-filter in sanitize_api_messages dropping role=tool messages with a missing/empty tool_call_id, which previously slipped through the orphan sweep (truthy-only set membership) and reached strict providers as a schema violation.

#59434 (dedup refactor) — agent/agent_runtime_helpers.py, agent/context_compressor.py, tests/run_agent/test_message_sequence_repair.py (+5 tests)

• New _classify_tool_call_orphans(messages) in agent_runtime_helpers.py is the single owner of orphan detection; both call sites delegate to it while keeping their divergent remediation (insert-stubs vs strip-orphans, incl. the [Bug]: context compression can drop an in-flight tool chain — side effect completes, result never reaches the agent, agent replays #79278 in-flight-chain protection).
• Reconciled with main's post-PR variant-set semantics: the classifier registers EVERY id alias (id/call_id/response_item_id/composite call|item, [Bug]: mnemosyne_recall and cronjob list consistently return [Result unavailable] — Hermes drops tool results #55626/Tool results silently dropped when provider returns composite tool_call ids (call_xxx|fc_yyy) — model never sees output and confabulates #63000) and returns orphaned message objects / missing tool_call entries (identity-matched), so the refactor is behavior-preserving against today's alias-aware code rather than regressing to single-id matching. The author's 5 classifier unit tests were adapted to the alias-aware return shape.

Validation

Live A/B on the real modules (.venv python + worktree sys.path), BEFORE = current main, AFTER = this branch:

Invariants held in every probe: role alternation preserved, no orphaned tool results, deterministic output.

Live repro: BEFORE on main — api_content sidecar survives the assistant merge (stale bytes win at API-build time) and id-less tool results pass sanitize_api_messages untouched (tool ids on wire: ['call_Z', '', None]); AFTER — sidecar dropped exactly when content changed, wire ids ['call_Z'], and both sanitizers classify orphans through the one shared function.

Tests: tests/run_agent/test_message_sequence_repair.py + tests/agent/test_context_compressor.py → 194 passed (memory-capped run), including the 15 new tests from the three PRs. ruff check clean on all touched files.

Note on #59434's scripts/release.py hunk: it appended to the frozen AUTHOR_MAP — dropped per current attribution policy (contributor mapping lives in contributors/emails/; [email protected], [email protected], and srojk34 noreply are all already mapped — scripts/audit_pr_attribution.py --fix reports clean).

Infographic