fix(proxy/anthropic): repair headroom_retrieve history references the tools array cannot support (#2876)

## Description #2805 / #2807 established the mechanism: Claude Code replays one transcript across requests that carry different `tools` arrays, and Anthropic validates every history reference against the array of the request at hand. #2807 fixed it for tool-search blocks by repairing history (`strip_unsupported_tool_search_blocks`) rather than trying to predict the client's tool set. The same mechanism applies to CCR's `headroom_retrieve`, and it is tool-agnostic. A passthrough side-request (the prompt-type Stop hook evaluator, `/compact`) that the proxy forwards without declaring `headroom_retrieve` still carries a historical `tool_use` naming it, and Anthropic 400s on the dangling reference. The injection-side fixes (#2766 / #2533) decide *when to re-declare the tool*; this makes the 400 *structurally impossible* where the tool is intentionally absent. It is belt-and-braces with them, not a replacement. The fix adds the symmetric repair next to #2807's. When the outbound `tools` array does not declare `headroom_retrieve`, it replaces each `headroom_retrieve` `tool_use` and its paired `tool_result` with a text block, so no dangling reference survives. It **neutralizes** (replaces in place) rather than **drops**, which is the one deliberate difference from #2807: CCR's `tool_use` lives in an assistant turn and its `tool_result` in the next user turn, i.e. two different messages. Dropping a whole message could leave two same-role messages adjacent and break Anthropic's strict user/assistant alternation, turning one 400 into another. Replacing blocks in place keeps every message and role intact, and preserves the retrieved text the model already saw. #2807's server-tool blocks both live in the same assistant turn, so dropping was safe there. It runs after CCR tool injection, so on the main loop -- where the tool IS injected (a present marker) -- it neutralizes nothing and the prompt-cache prefix is untouched, mirroring #2807's placement and sequencing. Fixes #2814 ## Type of Change - [x] Bug fix (non-breaking change that fixes an issue) - [ ] New feature (non-breaking change that adds functionality) - [ ] Breaking change (fix or feature that would cause existing functionality to change) - [ ] Documentation update - [ ] Performance improvement - [ ] Code refactoring (no functional changes) ## Changes Made - `headroom/proxy/helpers.py`: added `strip_unsupported_ccr_retrieve_blocks(messages, tools)` (and a small `_ccr_result_as_text` helper). No-ops (returning the original object by identity) when `headroom_retrieve` is declared or no such history exists; otherwise neutralizes the `tool_use` and its paired `tool_result` to text. - `headroom/proxy/handlers/anthropic.py`: call the repair right after the tool-search history repair (which is after CCR tool injection), guarded on it actually changing anything, tagged `router:ccr_retrieve_repair:Nblocks`. - `tests/test_ccr_retrieve_history_repair.py`: 5 unit tests (no-op when declared, no-op without retrieve history, neutralize + preserve result text + keep alternation, leave foreign tool_use untouched, placeholder when the result has no text). ## Testing - [x] Unit tests pass (`pytest`) - [x] Linting passes (`ruff check .`) - [x] Type checking passes (`mypy headroom`) - [x] New tests added for new functionality - [ ] Manual testing performed ### Test Output ```text tests/test_ccr_retrieve_history_repair.py 5 passed # Broader CCR / tool-search / handler suites (unchanged behavior): tests/test_ccr_retrieve_history_repair.py tests/test_proxy/test_anthropic_ccr_deferred_injection.py tests/test_issue_746_tool_search.py 71 passed # uvx ruff@0.15.17 check -> All checks passed! # uvx mypy@1.20.2 headroom/proxy/helpers.py -> Success: no issues found in 1 source file ``` ## Real Behavior Proof - Environment: Windows 11, Python 3.12.11, project venv, pytest 9.1.1, ruff 0.15.17 and mypy 1.20.2 via uvx. - Exact command / steps: confirmed the injection point (`apply_session_sticky_ccr_tool`) and the tool-search repair placement in `handlers/anthropic.py`, confirmed `body["tools"]` reflects the CCR injection before the repair call site (`body["tools"] = tools` is written well upstream and the adjacent tool-search repair already relies on it), then drove the helper over a transcript with a `headroom_retrieve` tool_use + paired tool_result: with the tool declared it returns the original object unchanged; with the tool absent it neutralizes both blocks, preserves the result text, and keeps the message roles/count identical. - Observed result: a forwarded request that would 400 with "Tool reference 'headroom_retrieve' not found in available tools" now carries text blocks in place of the retrieve `tool_use`/`tool_result`, so there is no reference for Anthropic to reject, and user/assistant alternation is preserved. The main loop (tool present) is a no-op. - Not tested: a live multi-turn Claude Code session hitting a Stop-hook/`/compact` side-request against a real provider (no live provider here). The repair is a pure function verified directly over the exact block shapes Anthropic validates, and it mirrors the already-merged tool-search repair's mechanism and wiring. ## Review Readiness - [x] I have performed a self-review - [x] This PR is ready for human review ## Checklist - [x] My code follows the project's style guidelines - [x] I have performed a self-review of my code - [x] I have commented my code, particularly in hard-to-understand areas - [ ] I have made corresponding changes to the documentation - [x] My changes generate no new warnings - [x] I have added tests that prove my fix is effective or that my feature works - [x] New and existing unit tests pass locally with my changes - [x] I did **not** edit `CHANGELOG.md`: it is generated by release-please from my Conventional Commit PR title (a CI guard enforces this) ## Additional Notes The issue reporter noted their own logs show the tool-search variant of this 400 (61 across 11 days) but zero `headroom_retrieve` occurrences, because they run `HEADROOM_LOSSLESS=1` which disables CCR entirely. This PR fixes the CCR variant of the same, proven, tool-agnostic mechanism rather than a fresh CCR repro. The neutralize-vs-drop choice is the one place I departed from #2807, for the alternation reason above; if you would rather it drop (accepting the alternation handling that implies), I am happy to switch it. --------- Co-authored-by: Jerrett Davis <mxjerrett@gmail.com>
2026-08-14 01:35:07 +05:30
parent 6077e5a149
commit 7de35739c6
4 changed files with 327 additions and 3 deletions
@@ -2737,6 +2737,28 @@ class AnthropicHandlerMixin:
                    _ts_stripped,
                )

+            # CCR retrieve history repair (#2814). Run beside the tool-search
+            # repair and after both normal CCR injection and turn hooks, so it
+            # validates against the final outbound tools array. A side-request
+            # that does not declare headroom_retrieve cannot retain historical
+            # tool_use/tool_result references that Anthropic would reject.
+            from headroom.proxy.helpers import strip_unsupported_ccr_retrieve_blocks
+
+            _ccr_repaired, _ccr_neutralized = strip_unsupported_ccr_retrieve_blocks(
+                body.get("messages"), body.get("tools")
+            )
+            if _ccr_neutralized:
+                body["messages"] = _ccr_repaired
+                optimized_messages = _ccr_repaired
+                body_mutation_tracker.mark_mutated("ccr_retrieve_history_repair")
+                transforms_applied.append(f"router:ccr_retrieve_repair:{_ccr_neutralized}blocks")
+                logger.info(
+                    "[%s] CCR: neutralized %d headroom_retrieve history block(s) "
+                    "(tools array does not declare the tool this turn)",
+                    request_id,
+                    _ccr_neutralized,
+                )
+
            # Consistency: report tok_before/tok_after with ONE tokenizer. The pipeline
            # and the handler use different token estimators, and cache-mode branches
            # can leave original_tokens (handler, line ~1049) and optimized_tokens
@@ -3027,6 +3027,126 @@ def strip_unsupported_tool_search_blocks(messages: Any, tools: Any) -> tuple[Any
    return (out, removed) if changed else (messages, 0)


+def _ccr_result_as_text(block: dict[str, Any]) -> str:
+    """Flatten a ``tool_result`` block's content to plain text, preserving what
+    the model already saw. Falls back to a short placeholder when there is no
+    textual content to keep."""
+    content = block.get("content")
+    if isinstance(content, str) and content.strip():
+        return content
+    if isinstance(content, list):
+        parts = [
+            b.get("text", "")
+            for b in content
+            if isinstance(b, dict) and b.get("type") == "text" and b.get("text")
+        ]
+        joined = "\n".join(part for part in parts if part)
+        if joined.strip():
+            return joined
+    return "[headroom_retrieve result omitted]"
+
+
+def strip_unsupported_ccr_retrieve_blocks(messages: Any, tools: Any) -> tuple[Any, int]:
+    """Neutralize ``headroom_retrieve`` history references the outbound ``tools``
+    array cannot support.
+
+    Claude Code replays one transcript across requests that carry different
+    ``tools`` arrays, and Anthropic validates every history ``tool_use`` against
+    the array of the request at hand. A passthrough side-request (the prompt-type
+    Stop hook evaluator, ``/compact``) that the proxy forwards without declaring
+    ``headroom_retrieve`` then 400s on a historical ``tool_use`` that names it --
+    the CCR sibling of the tool-search history repair (#2814 / #2807). This is
+    belt-and-braces with the injection-side fixes: they keep the tool available
+    where it belongs; this makes the 400 structurally impossible where it cannot.
+
+    When the request does NOT declare ``headroom_retrieve``, replace each
+    ``headroom_retrieve`` ``tool_use`` block and its paired ``tool_result`` with a
+    text block, so no dangling reference survives. Neutralize rather than drop:
+    CCR's ``tool_use`` (an assistant turn) and its ``tool_result`` (the next user
+    turn) live in DIFFERENT messages, so removing a message could leave two
+    same-role messages adjacent and break Anthropic's user/assistant alternation.
+    Replacing blocks in place keeps every message and role intact, and preserves
+    the retrieved text the model already saw.
+
+    Returns ``(messages, blocks_neutralized)``, and the ORIGINAL ``messages``
+    object when nothing changed -- callers rely on identity to skip the write-back.
+    """
+    from headroom.ccr.tool_injection import CCR_TOOL_NAME
+
+    if not isinstance(messages, list):
+        return messages, 0
+
+    tool_list = tools if isinstance(tools, list) else []
+    available = {str(t["name"]) for t in tool_list if isinstance(t, dict) and t.get("name")}
+    # The tool is declared this turn (e.g. the main loop, or sticky re-injection),
+    # so its history references resolve. Nothing to repair.
+    if CCR_TOOL_NAME in available:
+        return messages, 0
+
+    # First pass: collect the ids of headroom_retrieve tool_use blocks so their
+    # paired tool_result blocks (in a later user turn) can be matched.
+    retrieve_ids: set[str] = set()
+    for message in messages:
+        content = message.get("content") if isinstance(message, dict) else None
+        if not isinstance(content, list):
+            continue
+        for block in content:
+            if (
+                isinstance(block, dict)
+                and block.get("type") == "tool_use"
+                and block.get("name") == CCR_TOOL_NAME
+            ):
+                use_id = block.get("id")
+                if use_id:
+                    retrieve_ids.add(str(use_id))
+
+    if not retrieve_ids:
+        return messages, 0
+
+    out: list[Any] = []
+    neutralized = 0
+    changed = False
+    for message in messages:
+        content = message.get("content") if isinstance(message, dict) else None
+        if not isinstance(content, list):
+            out.append(message)
+            continue
+
+        new_content: list[Any] = []
+        touched = False
+        for block in content:
+            if isinstance(block, dict):
+                if block.get("type") == "tool_use" and block.get("name") == CCR_TOOL_NAME:
+                    new_content.append(
+                        {
+                            "type": "text",
+                            "text": "[headroom_retrieve call omitted: tool not available this turn]",
+                        }
+                    )
+                    neutralized += 1
+                    touched = True
+                    continue
+                if (
+                    block.get("type") == "tool_result"
+                    and str(block.get("tool_use_id", "")) in retrieve_ids
+                ):
+                    new_content.append({"type": "text", "text": _ccr_result_as_text(block)})
+                    neutralized += 1
+                    touched = True
+                    continue
+            new_content.append(block)
+
+        if touched:
+            changed = True
+            repaired = dict(message)
+            repaired["content"] = new_content
+            out.append(repaired)
+        else:
+            out.append(message)
+
+    return (out, neutralized) if changed else (messages, 0)
+
+
 # ---------------------------------------------------------------------------
 # Server-side Tool Search injection — OpenAI Responses API (gpt-5.4+).
 #
@@ -0,0 +1,119 @@
+"""Regression tests for the CCR headroom_retrieve history repair (#2814).
+
+A passthrough side-request (prompt-type Stop hook evaluator, /compact) replays a
+transcript whose history contains a ``headroom_retrieve`` tool_use, but the
+request's own ``tools`` array does not declare that tool. Anthropic 400s on the
+dangling reference. The repair neutralizes those history blocks so the 400 is
+structurally impossible, without breaking user/assistant alternation.
+"""
+
+from __future__ import annotations
+
+from headroom.ccr.tool_injection import CCR_TOOL_NAME
+from headroom.proxy.helpers import strip_unsupported_ccr_retrieve_blocks
+
+
+def _transcript_with_retrieve() -> list[dict]:
+    return [
+        {"role": "user", "content": [{"type": "text", "text": "read config"}]},
+        {
+            "role": "assistant",
+            "content": [
+                {"type": "text", "text": "Let me expand that."},
+                {
+                    "type": "tool_use",
+                    "id": "toolu_ccr_1",
+                    "name": CCR_TOOL_NAME,
+                    "input": {"hash": "abc123def456abc123def456"},
+                },
+            ],
+        },
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "tool_result",
+                    "tool_use_id": "toolu_ccr_1",
+                    "content": "the full expanded config text",
+                }
+            ],
+        },
+        {"role": "assistant", "content": [{"type": "text", "text": "Done."}]},
+    ]
+
+
+def test_noop_when_tool_is_declared() -> None:
+    messages = _transcript_with_retrieve()
+    tools = [{"name": CCR_TOOL_NAME}]
+    out, n = strip_unsupported_ccr_retrieve_blocks(messages, tools)
+    assert n == 0
+    assert out is messages  # identity: caller skips the write-back
+
+
+def test_noop_when_no_retrieve_history() -> None:
+    messages = [
+        {"role": "user", "content": [{"type": "text", "text": "hi"}]},
+        {"role": "assistant", "content": [{"type": "text", "text": "hello"}]},
+    ]
+    out, n = strip_unsupported_ccr_retrieve_blocks(messages, tools=[])
+    assert n == 0
+    assert out is messages
+
+
+def test_neutralizes_retrieve_reference_when_tool_absent() -> None:
+    messages = _transcript_with_retrieve()
+    out, n = strip_unsupported_ccr_retrieve_blocks(messages, tools=[{"name": "Bash"}])
+
+    assert n == 2  # tool_use + its paired tool_result
+    # No dangling references survive anywhere.
+    for message in out:
+        for block in message["content"]:
+            assert block.get("type") != "tool_use" or block.get("name") != CCR_TOOL_NAME
+            assert not (
+                block.get("type") == "tool_result" and block.get("tool_use_id") == "toolu_ccr_1"
+            )
+
+    # Message count and roles are unchanged (alternation intact).
+    assert [m["role"] for m in out] == [m["role"] for m in messages]
+    # The assistant's own text survives; the tool_use became a text block.
+    assistant = out[1]["content"]
+    assert assistant[0] == {"type": "text", "text": "Let me expand that."}
+    assert assistant[1]["type"] == "text"
+    # The retrieved result text is preserved, not dropped.
+    assert out[2]["content"][0] == {
+        "type": "text",
+        "text": "the full expanded config text",
+    }
+
+
+def test_leaves_foreign_tool_use_untouched() -> None:
+    messages = [
+        {
+            "role": "assistant",
+            "content": [
+                {"type": "tool_use", "id": "toolu_bash_1", "name": "Bash", "input": {"cmd": "ls"}}
+            ],
+        },
+        {
+            "role": "user",
+            "content": [
+                {"type": "tool_result", "tool_use_id": "toolu_bash_1", "content": "file.txt"}
+            ],
+        },
+    ]
+    out, n = strip_unsupported_ccr_retrieve_blocks(messages, tools=[{"name": "Bash"}])
+    assert n == 0
+    assert out is messages
+
+
+def test_result_falls_back_to_placeholder_without_text() -> None:
+    messages = [
+        {
+            "role": "assistant",
+            "content": [{"type": "tool_use", "id": "t1", "name": CCR_TOOL_NAME, "input": {}}],
+        },
+        {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "t1", "content": []}]},
+    ]
+    out, n = strip_unsupported_ccr_retrieve_blocks(messages, tools=[])
+    assert n == 2
+    assert out[1]["content"][0] == {"type": "text", "text": "[headroom_retrieve result omitted]"}
@@ -21,12 +21,18 @@ pytest.importorskip("fastapi")

 from fastapi.testclient import TestClient

+from headroom.ccr.tool_injection import CCR_TOOL_NAME
 from headroom.proxy.server import ProxyConfig, create_app
 from headroom.proxy.turn_hooks import clear_turn_hooks, register_turn_hook

 _SEARCH_TOOL = {"type": "tool_search_tool_20250917", "name": "tool_search"}
 _GREP = {"name": "Grep", "description": "search files", "input_schema": {"type": "object"}}
 _READ = {"name": "Read", "description": "read a file", "input_schema": {"type": "object"}}
+_CCR_TOOL = {
+    "name": CCR_TOOL_NAME,
+    "description": "retrieve compressed content",
+    "input_schema": {"type": "object"},
+}

 # A transcript that already carries a resolved tool-search round trip for `Grep`.
 _POISONED_MESSAGES = [
@@ -53,6 +59,31 @@ _POISONED_MESSAGES = [
    {"role": "user", "content": "now use it"},
 ]

+_CCR_POISONED_MESSAGES = [
+    {"role": "user", "content": "expand the saved result"},
+    {
+        "role": "assistant",
+        "content": [
+            {
+                "type": "tool_use",
+                "id": "toolu_ccr_1",
+                "name": CCR_TOOL_NAME,
+                "input": {"hash": "abc123def456abc123def456"},
+            }
+        ],
+    },
+    {
+        "role": "user",
+        "content": [
+            {
+                "type": "tool_result",
+                "tool_use_id": "toolu_ccr_1",
+                "content": "the expanded result",
+            }
+        ],
+    },
+]
+

@pytest.fixture(autouse=True)
 def _clean_registry():
@@ -77,7 +108,12 @@ class _InertHook:
        return None


-def _run(hook) -> dict:  # noqa: ANN001
+def _run(
+    hook,  # noqa: ANN001
+    *,
+    messages: list[dict] = _POISONED_MESSAGES,
+    tools: list[dict] | None = None,
+) -> dict:
    """POST a poisoned transcript through the handler, return the forwarded body."""
    captured: dict[str, object] = {}
    register_turn_hook(hook)
@@ -123,8 +159,8 @@ def _run(hook) -> dict:  # noqa: ANN001
            json={
                "model": "claude-sonnet-4-6",
                "max_tokens": 64,
-                "messages": _POISONED_MESSAGES,
-                "tools": [_SEARCH_TOOL, _GREP, _READ],
+                "messages": messages,
+                "tools": tools or [_SEARCH_TOOL, _GREP, _READ],
            },
        )
        assert response.status_code == 200
@@ -174,3 +210,30 @@ def test_repair_leaves_resolvable_history_alone_when_the_hook_keeps_the_tool() -

    assert _referenced_tool_names(forwarded) == ["Grep"]
    assert "tool_search_tool_result" in _block_types(forwarded)
+
+
+def test_ccr_repair_sees_the_tools_array_the_hook_left_behind() -> None:
+    """CCR repair also runs after a hook removes headroom_retrieve."""
+    forwarded = _run(
+        _DropToolHook(CCR_TOOL_NAME),
+        messages=_CCR_POISONED_MESSAGES,
+        tools=[_CCR_TOOL, _READ],
+    )
+
+    assert CCR_TOOL_NAME not in [t.get("name") for t in forwarded["tools"]]
+    assert "tool_use" not in _block_types(forwarded)
+    assert "tool_result" not in _block_types(forwarded)
+    assert forwarded["messages"][2]["content"] == [{"type": "text", "text": "the expanded result"}]
+
+
+def test_ccr_repair_leaves_resolvable_history_when_hook_keeps_tool() -> None:
+    """CCR history remains structured when the final tools array declares it."""
+    forwarded = _run(
+        _InertHook(),
+        messages=_CCR_POISONED_MESSAGES,
+        tools=[_CCR_TOOL, _READ],
+    )
+
+    assert CCR_TOOL_NAME in [t.get("name") for t in forwarded["tools"]]
+    assert "tool_use" in _block_types(forwarded)
+    assert "tool_result" in _block_types(forwarded)