Background review fork tools cache fragmentation fix

Fixes on record

Merged to main. A stamp counts the issues a fix is linked to.

• #29704 fix(background_review): inherit parent toolset config for tools[] cache parity (salvage #29568)

  merged to main

  files: agent/background_review.py, scripts/release.py, tests/run_agent/test_background_review_cache_parity.py, tests/run_agent/test_background_review_toolset_restriction.py

  Show the diff

  • agent/background_review.py

    @@ -390,6 +390,9 @@ def _bg_review_auto_deny(command, description, **kwargs):
                 # parent below so memory(action="add") writes from
                 # the review still land on disk; the review just
                 # has zero side effects on external providers.
    +            # Match parent's toolset config so ``tools[]`` is byte-identical
    +            # in the request body — Anthropic's cache key includes it.
    +            # (The runtime whitelist below still restricts dispatch.)
                 review_agent = AIAgent(
                     model=agent.model,
                     max_iterations=16,
    @@ -401,6 +404,8 @@ def _bg_review_auto_deny(command, description, **kwargs):
                     api_key=_parent_runtime.get("api_key") or None,
                     credential_pool=[redacted:secret-assignment] "_credential_pool", None),
                     parent_session_id=agent.session_id,
    +                enabled_toolsets=getattr(agent, "enabled_toolsets", None),
    +                disabled_toolsets=getattr(agent, "disabled_toolsets", None),
                     skip_memory=True,
                 )
                 review_agent._memory_write_origin = "background_review"

  • scripts/release.py

    @@ -718,6 +718,7 @@
         "[email protected]": "cresslank",
         "[redacted:email]": "trevthefoolish",
         "[email protected]": "ziliangpeng",
    +    "[redacted:email]": "ziliangpeng",
         "[email protected]": "centripetal-star",
         "[email protected]": "LeonSGP43",
         "[email protected]": "LeonSGP43",

  • tests/run_agent/test_background_review_cache_parity.py

    @@ -38,6 +38,9 @@ def _make_agent_stub(agent_cls):
         agent._MEMORY_REVIEW_PROMPT = "review memory"
         agent._SKILL_REVIEW_PROMPT = "review skills"
         agent._COMBINED_REVIEW_PROMPT = "review both"
    +    # Non-None so the test catches a missing-kwarg regression.
    +    agent.enabled_toolsets = ["memory", "skills", "terminal"]
    +    agent.disabled_toolsets = ["spotify", "feishu_doc"]
         return agent
     
     
    @@ -183,3 +186,54 @@ def close(self):
             "Review fork did not inherit parent's session_id — "
             "system-prompt rebuild paths would diverge."
         )
    +
    +
    +def test_review_fork_inherits_parent_toolset_config():
    +    """``tools[]`` byte-stability: fork must inherit parent's toolset config."""
    +    import run_agent
    +
    +    agent = _make_agent_stub(run_agent.AIAgent)
    +
    +    captured = {}
    +
    +    class _Recorder:
    +        def __init__(self, *args, **kwargs):
    +            captured["enabled_toolsets"] = kwargs.get("enabled_toolsets")
    +            captured["disabled_toolsets"] = kwargs.get("disabled_toolsets")
    +            self._cached_system_prompt = None
    +            self._memory_write_origin = None
    +            self._memory_write_context = None
    +            self._memory_store = None
    +            self._memory_enabled = None
    +            self._user_profile_enabled = None
    +            self._memory_nudge_interval = None
    +            self._skill_nudge_interval = None
    +            self.suppress_status_output = None
    +            self.session_start = None
    +            self.session_id = None
    +
    +        def run_conversation(self, *args, **kwargs):
    +            raise RuntimeError("stop after recording — don't actually call the API")
    +
    +        def shutdown_memory_provider(self):
    +            pass
    +
    +        def close(self):
    +            pass
    +
    +    with patch.object(run_agent, "AIAgent", _Recorder), \
    +         patch("threading.Thread", _SyncThread):
    +        agent._spawn_background_review(
    +            messages_snapshot=[],
    +            review_memory=True,
    +            review_skills=False,
    +        )
    +
    +    assert captured.get("enabled_toolsets") == agent.enabled_toolsets, (
    +        f"enabled_toolsets mismatch: {captured.get('enabled_toolsets')!r} "
    +        f"vs expected {agent.enabled_toolsets!r}"
    +    )
    +    assert captured.get("disabled_toolsets") == agent.disabled_toolsets, (
    +        f"disabled_toolsets mismatch: {captured.get('disabled_toolsets')!r} "
    +        f"vs expected {agent.disabled_toolsets!r}"
    +    )

  • tests/run_agent/test_background_review_toolset_restriction.py

    @@ -38,6 +38,9 @@ def _make_agent_stub(agent_cls):
         agent._MEMORY_REVIEW_PROMPT = "review memory"
         agent._SKILL_REVIEW_PROMPT = "review skills"
         agent._COMBINED_REVIEW_PROMPT = "review both"
    +    # Non-None so the test catches a missing-kwarg regression.
    +    agent.enabled_toolsets = ["memory", "skills", "terminal"]
    +    agent.disabled_toolsets = ["spotify", "feishu_doc"]
         return agent
     
     
    @@ -52,13 +55,8 @@ def start(self):
                 self._target()
     
     
    -def test_background_review_does_not_narrow_toolset_schema():
    -    """The review fork must NOT pass enabled_toolsets to AIAgent.
    -
    -    Narrowing the schema diverges the ``tools`` cache key from the parent's,
    -    which sits above ``system`` in Anthropic's cache hierarchy and forces a
    -    full prefix-cache miss on every review (see #25322, PR #17276).
    -    """
    +def test_background_review_matches_parent_toolset_config():
    +    """Fork must receive parent's toolset config so ``tools[]`` cache key matches."""
         import run_agent
     
         agent = _make_agent_stub(run_agent.AIAgent)
    @@ -66,6 +64,7 @@ def test_background_review_does_not_narrow_toolset_schema():
     
         def _capture_init(self, *args, **kwargs):
             captured["enabled_toolsets"] = kwargs.get("enabled_toolsets", "UNSET")
    +        captured["disabled_toolsets"] = kwargs.get("disabled_toolsets", "UNSET")
             raise RuntimeError("stop after capturing init args")
     
         with patch.object(run_agent.AIAgent, "__init__", _capture_init), \
    @@ -77,11 +76,13 @@ def _capture_init(self, *args, **kwargs):
             )
     
         assert "enabled_toolsets" in captured, "AIAgent.__init__ was not called"
    -    # The kwarg must be absent — letting AIAgent inherit the default full
    -    # toolset so the schema bytes match the parent's.
    -    assert captured["enabled_toolsets"] == "UNSET", (
    -        f"Review fork narrowed the toolset schema (got {captured['enabled_toolsets']!r}), "
    -        "which breaks prefix-cache parity with the parent."
    +    assert captured["enabled_toolsets"] == agent.enabled_toolsets, (
    +        f"enabled_toolsets mismatch: {captured['enabled_toolsets']!r} "
    +        f"vs expected {agent.enabled_toolsets!r}"
    +    )
    +    assert captured["disabled_toolsets"] == agent.disabled_toolsets, (
    +        f"disabled_toolsets mismatch: {captured['disabled_toolsets']!r} "
    +        f"vs expected {agent.disabled_toolsets!r}"
         )