fix(repair): Pass 1.5 must match tool results on call_id, not just id

[Kyzcreig]

Symptom

AIAgent._repair_message_sequence deletes a valid assistant(tool_calls) turn whose tool_call carries only call_id (no id), even though the immediately-following tool message answers it correctly.

Reproduced on fork/main (0b2a95966f) before any change:

messages = [user,
            assistant(tool_calls=[{call_id: "call_XYZ", ...}]),
            tool(tool_call_id="call_XYZ"),
            user]

repairs      -> 1                          (expected 0)
roles after  -> ['user', 'tool', 'user']   # assistant turn destroyed

The result is a stray tool message with no preceding tool_calls — precisely the HTTP 400 shape this repair exists to prevent ("Messages with role 'tool' must be a response to a preceding message with 'tool_calls'" on DeepSeek/Kimi; "tool_use ids were found without tool_result blocks" on Anthropic).

Root cause

agent/agent_runtime_helpers.py:827 (pre-fix):

for tc in calls:
    _cid = tc.get("id") if isinstance(tc, dict) else None   # <-- id ONLY
    if _cid and _remaining.get(_cid, 0) > 0:

Pass 1 registers the answer budget against the superset of both keys (for key in ("id", "call_id"), per NousResearch#58168) — but Pass 1.5 consumes that budget against tc.get("id") alone. A tool_call carrying only call_id (or a Codex-Responses call whose id = fc_... and call_id = call_... differ) therefore matches nothing, answered_flags is all-False, and the "NONE answered — the whole tool-call turn is an orphan" branch drops the turn.

This is pre-existing on fork/main, not parity-merge damage. The 2026-08-07 parity merge only surfaced it by restoring the upstream tests that catch it.

Fix

Minimal and surgical — Pass 1.5 now consumes the budget on id OR call_id, mirroring Pass 1. No other pass touched:

for tc in calls:
    _matched = False
    if isinstance(tc, dict):
        for _key in ("id", "call_id"):
            _cid = tc.get(_key)
            if _cid and _remaining.get(_cid, 0) > 0:
                _remaining[_cid] -= 1
                _matched = True
                break
    answered_flags.append(_matched)

The multiset/count semantics are preserved exactly (each entry consumes at most one result slot, so a [X, X] turn answered once still keeps exactly one X), and the break keeps a single tool_call from consuming two slots when id and call_id are both present and both answered.

RED -> GREEN proof

RED first. New focused regression test test_repair_pass_1_5_keeps_call_id_only_assistant_turn written and run before any source change:

>       assert repairs == 0
E       assert 1 == 0
1 failed, 25 deselected

GREEN after the fix — scripts/run_tests.sh (canonical parallel runner), exact observed counts:

Command	Result
scripts/run_tests.sh tests/run_agent/test_message_sequence_repair.py -q	26 passed, 0 failed
scripts/run_tests.sh tests/run_agent -q	177 files, 1622 passed, 0 failed (66.9s)
scripts/run_tests.sh tests/agent -q	422 files, 4958 passed, 0 failed (110.7s)
ruff check (both changed files)	All checks passed!

MUTATION PROOF

Explicitly performed. Reverting only the source change (git checkout HEAD~1 -- agent/agent_runtime_helpers.py, leaving the tests in place) and re-running the file:

=== Summary: 1 files, 23 tests passed, 3 failed ===
FAILED ...::test_repair_keeps_tool_matching_codex_call_id
FAILED ...::test_repair_keeps_tool_matching_only_call_id
FAILED ...::test_repair_pass_1_5_keeps_call_id_only_assistant_turn

Restoring the source (git checkout HEAD -- agent/agent_runtime_helpers.py) returns the file to 26 passed, 0 failed. The tests fail without the fix and pass with it — they are real gates on this exact behavior, not vacuous greens.

Removed xfail markers

Both @pytest.mark.xfail(strict=False) markers in tests/run_agent/test_message_sequence_repair.py — parked on the two upstream tests pending exactly this fix — are deleted, so those tests now gate for real:

• test_repair_keeps_tool_matching_codex_call_id (distinct id + call_id)
• test_repair_keeps_tool_matching_only_call_id (call_id only)

The now-unused import pytest was dropped with them (ruff-clean). Note both were strict=False and had begun xpassing for the codex_call_id case in some orderings — as markers they proved nothing; the mutation run above shows they genuinely fail against the unfixed source.

Files changed

• agent/agent_runtime_helpers.py (+18 / -6) — Pass 1.5 id/call_id superset match + explanatory comment
• tests/run_agent/test_message_sequence_repair.py (+30 / -39) — new regression test; two xfail markers removed