fix(proxy): ignore defer_loading when measuring a turn hook's tool delta (#4056)

## Description

A turn hook's measured tool delta counted the `defer_loading` flag as if
it were tool content. When headroom-tool-search's hot-tools hook
**un-defers** a tool, it removes `"defer_loading": true`. The serialized
array got shorter, so the hook runner booked a realized `tool_search`
saving, and the Anthropic, OpenAI chat and gateway paths added the same
amount to `turn_hook_tools_saved_tokens`, which is part of the headline.
In fact un-deferring makes the billed request larger.

Deferral is already booked through `tool_search_deferred_tokens` (see
`tool_schema_savings_policy`). The measured hook delta should therefore
ignore the flag and count only real removals and rewrites. That is what
this PR does, through one helper, `without_deferral_flags`, used at all
four measurement sites.

Found while auditing the enterprise plugins for the provider-unit
savings work (#4052 / #4055).

## Type of Change

- [x] Bug fix (non-breaking change that fixes an issue)

## Changes Made

- `tool_schema_savings_policy.without_deferral_flags(tools)`: returns
the tool array without `defer_loading` keys. Never mutates its input,
and passes non-lists through.
- Used where hook tool deltas are measured:
- `turn_hooks.run_request_hooks` (the automatic per-hook ledger entry);
  - the Anthropic handler (`_th_tok_before/_after`);
  - the OpenAI chat handler (`_th_tok_before/_after`);
  - the gateway turn (`tools_before/_after`).
- Compaction sites are unchanged; they never toggle the flag.
- **Deferral credit reconcile** (`deferred_booking` /
`reconcile_deferred_tokens`, run at the end of `run_request_hooks`, so
it covers the Anthropic handler and gateway turns):
- snapshots the booked deferred schemas before hooks run, or again when
a hook re-books;
- after all hooks, recounts the snapshot schemas of tools still
deferred;
- lowers `tool_search_deferred_tokens` / `_tools` and the latest
`tool_search` ledger entry by the released amount;
  - never raises the credit, and leaves turns with no booking alone.
- Tests: un-defer lowers both the headline and the ledger; un-defer plus
expanding another schema still lowers it; no booking is never touched;
end-to-end through the Anthropic handler (14 → 13 tools, ledger equals
headline).

## Testing

- [x] Unit tests pass (`pytest`): related suites
- [x] Linting passes (`ruff check .`)
- [x] Type checking passes (`mypy`) on changed modules
- [x] New tests added for new functionality
- [x] Manual testing performed (plugin suite against this core)

### Test Output

```text
tests/test_hook_tool_deferral_accounting.py: 10 passed
  hook deltas: un-deferring is not booked; deferring is left to the tag; a real
    removal is still measured; helper does not mutate
  deferral reconcile: un-defer lowers headline AND the deferral's ledger entry;
    growing another deferred schema cannot keep the credit; a shortened deferred
    schema is credited once; a hook's own tool_search saving is never debited;
    unbooked turns untouched; end to end through the Anthropic handler
    (14 tools deferred, hook un-defers 1 -> credit 13 tools, ledger == headline)
  each fails on the commit before its fix (e.g. 14 == 13, 246 == 123,
    0 == 235 - 136, 123 == 24)
gateway + turn_hook / tool_search / tool_schema / savings_attribution / extension /
anthropic / openai / responses / compress suites: 2094 passed, 34 skipped
plugin suites against this core (PYTHONPATH): skill-search 41, tool-search 144,
lossless-guard 237 passed
ruff check / format: clean; mypy (changed modules): Success
```

## Real Behavior Proof

- Environment: macOS, Python 3.12, `LITELLM_LOCAL_MODEL_COST_MAP=True`.
The real proxy app (`create_app`) is driven through FastAPI's
TestClient, with only the upstream Anthropic call mocked (respx).
- Exact command / steps:
1. `HEADROOM_TOOL_SEARCH=1 pytest
tests/test_hook_tool_deferral_accounting.py -k anthropic_headline`. This
POSTs `/v1/messages` with 14 MCP tools. Core deferral books all 14, then
a registered hook shaped like tool search's hot tools un-defers
`mcp__srv__tool_0`.
2. `pytest tests/test_hook_tool_deferral_accounting.py`. The
runner-level cases cover hook deltas and deferral reconcile: un-defer,
un-defer plus schema growth, a shortened deferred schema, and a
`tool_search` hook with its own saving.
3. headroom-tool-search's suite run with `PYTHONPATH` set to this
branch.
- Observed result:
- **Hook deltas:** a hook that only removes `defer_loading` books
nothing; a hook that removes a tool is credited exactly the removed
size.
- **Deferral reconcile:** after hooks, `tool_search_deferred_tools` is
13 and `tool_search_deferred_tokens` equals the Claude counter's count
of the 13 still-deferred booked schemas. The deferral's own
`tool_search` ledger entry equals the tag, while a hook's own measured
saving is left intact. Growth in another schema doesn't restore the
credit, and a shortened deferred schema is credited once.
- **Plugin:** headroom-tool-search's `HotToolsHook._rebook`, which
already lowered the credit itself, still passes (144 passed). Core's
reconcile then finds nothing more to release.
- Not tested: a live Claude Code session with `HEADROOM_TOOL_SEARCH_HOT`
set against the real API. The change is in counting only, so requests
are byte-identical.

## Runtime Rollout Safety

- Rollout-managed feature(s): none.
- Minimum rollout channel: n/a
- Stable/default behavior changed: yes, savings accounting only.
Requests are byte-identical.
  - The phantom hot-tools hook saving disappears.
- The deferral credit (headline and `/stats`
`savings.by_source.tool_search`) now covers only booked tools still
deferred after hooks, each at the smaller of its booked and final size.
- Kill switch / disable path: n/a. This is accounting only.
- Unsafe override required: no
- Qualification impact: none.
- Rollback path: revert this PR.

## 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`

## Additional Notes

- **Docs:** no change needed. The rule is documented in the
`tool_schema_savings_policy` module docstring and the helper's
docstring.
- **Native deferral:** headroom-tool-search's NativeDeferralHook keeps
booking deferral through its tag. Before this PR, flagging a tool made
the array longer and clamped to zero, so the behaviour is unchanged; now
it is zero by construction.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

<!-- devin-review-badge-begin -->

---

<a href="https://app.devin.ai/review/headroomlabs-ai/headroom/pull/4056"
target="_blank"><picture><source media="(prefers-color-scheme: dark)"
srcset="https://static.devin.ai/assets/gh-devin-review-dark.svg?v=4"><img
src="https://static.devin.ai/assets/gh-devin-review-light.svg?v=4"
alt="Devin Review"></picture></a>
<!-- devin-review-badge-end -->

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Tejas Chopra
2026-10-08 22:29:58 -05:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 17abee2721
commit f8212a39dc
6 changed files with 444 additions and 8 deletions
+3 -2
View File
@@ -51,6 +51,7 @@ from typing import Any
from headroom.proxy.gateway_responses import VIEW_MARKER as GATEWAY_RESPONSES_VIEW_MARKER
from headroom.proxy.outcome import RequestOutcome
from headroom.proxy.tool_schema_savings_policy import without_deferral_flags
from headroom.proxy.turn_hooks import (
TurnContext,
merge_provider_headers,
@@ -893,7 +894,7 @@ class RequestTransformer:
)
if hooks:
msg_before = _safe_count(count_messages, messages)
tools_before = _safe_count(count_tools, tools)
tools_before = _safe_count(count_tools, without_deferral_flags(tools))
deferred_before = self.tags.get("tool_search_deferred_tools")
run_request_hooks(ctx, stream_safe_only=not self.caps.redrive_allowed)
messages = ctx.messages
@@ -911,7 +912,7 @@ class RequestTransformer:
if msg_before is not None and msg_after is not None and msg_after < msg_before:
transforms.append("turn_hook")
messages_rewritten = True
tools_after = _safe_count(count_tools, tools)
tools_after = _safe_count(count_tools, without_deferral_flags(tools))
if tools_before is not None and tools_after is not None:
saved = max(0, tools_before - tools_after)
if saved > 0:
+11 -2
View File
@@ -20,6 +20,7 @@ from typing import TYPE_CHECKING, Any
from urllib.parse import urlsplit
from headroom.proxy.stage_timer import StageTimer, emit_stage_timings_log
from headroom.proxy.tool_schema_savings_policy import without_deferral_flags
if TYPE_CHECKING:
from fastapi import Request
@@ -3364,7 +3365,11 @@ class AnthropicHandlerMixin:
except Exception:
_pre_hook_tokens = None
_th_tools_before = body.get("tools")
_th_tok_before = _count_tool_tokens(_th_tools_before) if _th_tools_before else 0
_th_tok_before = (
_count_tool_tokens(without_deferral_flags(_th_tools_before))
if _th_tools_before
else 0
)
run_request_hooks(_req_ctx, stream_safe_only=bool(stream))
if _req_ctx.messages is not optimized_messages:
optimized_messages = _req_ctx.messages
@@ -3376,7 +3381,11 @@ class AnthropicHandlerMixin:
# so measure the FINAL tools object. Deferral-shaped (removes schemas
# count_messages never saw), hence a tag rather than a fold — mirrors
# the OpenAI chat path so a turn-hook extension is credited on both.
_th_tok_after = _count_tool_tokens(_req_ctx.tools) if _req_ctx.tools else 0
_th_tok_after = (
_count_tool_tokens(without_deferral_flags(_req_ctx.tools))
if _req_ctx.tools
else 0
)
_th_saved = max(0, _th_tok_before - _th_tok_after)
if _th_saved > 0:
tags["turn_hook_tools_saved_tokens"] = (
+7 -2
View File
@@ -39,6 +39,7 @@ from headroom.proxy.modes import is_cache_mode
from headroom.proxy.rate_limit_identity import rate_limit_identity
from headroom.proxy.semantic_cache_key_policy import compute_request_cache_partition
from headroom.proxy.stage_timer import StageTimer, emit_stage_timings_log
from headroom.proxy.tool_schema_savings_policy import without_deferral_flags
from headroom.proxy.upstream_guard import is_safe_upstream_url
from headroom.proxy.ws_headers import WS_HOP_BY_HOP_HEADERS
from headroom.proxy.ws_session_registry import (
@@ -4833,7 +4834,9 @@ class OpenAIHandlerMixin:
if registered_turn_hooks():
_th_tools_before = body.get("tools")
_th_tok_before = (
tokenizer.count_text(json.dumps(_th_tools_before, default=str))
tokenizer.count_text(
json.dumps(without_deferral_flags(_th_tools_before), default=str)
)
if _th_tools_before
else 0
)
@@ -4882,7 +4885,9 @@ class OpenAIHandlerMixin:
except Exception:
logger.debug("turn-hook token re-count skipped", exc_info=True)
_th_tok_after = (
tokenizer.count_text(json.dumps(_th_ctx.tools, default=str)) if _th_ctx.tools else 0
tokenizer.count_text(json.dumps(without_deferral_flags(_th_ctx.tools), default=str))
if _th_ctx.tools
else 0
)
_th_saved = max(0, _th_tok_before - _th_tok_after)
if _th_saved > 0:
@@ -25,6 +25,113 @@ name to :data:`TOOL_SCHEMA_SAVINGS_TAGS` and every surface picks it up.
from __future__ import annotations
import copy
from typing import Any
def without_deferral_flags(tools: object) -> object:
"""The tool array as it is MEASURED for a hook's tool delta.
Deferral is reported through ``tool_search_deferred_tokens`` (it removes
schemas from the billed context while leaving them in the array, flagged
``defer_loading``). Measuring the raw array would see only the flag
appear or disappear: un-deferring a tool would shorten the JSON by the
flag text and be booked as a small saving although the bill goes UP, and
deferring one would add a phantom cost. So the flag is ignored here, and a
hook's measured tool delta covers only real removals and rewrites. Never
mutates its input.
"""
if not isinstance(tools, list):
return tools
return [
{k: v for k, v in tool.items() if k != "defer_loading"}
if isinstance(tool, dict) and "defer_loading" in tool
else tool
for tool in tools
]
DeferredBooking = tuple[int, dict[str, Any], "dict[str, Any] | None"]
def deferred_booking(tags: object, tools: object) -> DeferredBooking | None:
"""Snapshot the deferral currently booked: the tag, the tools it covers and
its own ledger entry.
``None`` when nothing is booked. The schemas are copied as they were when
booked, so a later in-place edit cannot change what the credit stands for.
The ledger entry is the latest ``tool_search`` one marked ``estimated``:
both the core deferral and the native deferral hook book that way, and a
hook's own measured saving never is, so it is never the one debited.
"""
if not isinstance(tags, dict) or "tool_search_deferred_tokens" not in tags:
return None
try:
booked = int(tags.get("tool_search_deferred_tokens") or 0)
except (TypeError, ValueError):
return None
snapshot = {
str(t["name"]): copy.deepcopy(t)
for t in (tools if isinstance(tools, list) else [])
if isinstance(t, dict) and t.get("defer_loading") and t.get("name")
}
from headroom.proxy.savings_attribution import SAVINGS_ATTRIBUTION_TAG
ledger = tags.get(SAVINGS_ATTRIBUTION_TAG)
entry = next(
(
item
for item in reversed(ledger if isinstance(ledger, list) else [])
if isinstance(item, dict)
and item.get("source") == "tool_search"
and item.get("estimated")
),
None,
)
return booked, snapshot, entry
def reconcile_deferred_tokens(
tags: object, booking: DeferredBooking | None, tools: object, count_tools: object
) -> None:
"""Cap the deferral credit at the booked tools still deferred after hooks.
A hook may un-defer tools the deferral booked (tool search's hot tools);
those are billed, so their share of the credit goes. Each tool still
deferred is credited at the SMALLER of its booked and its final schema:
growing another schema cannot keep an un-deferred tool's credit alive, and
a schema a hook shortened (and booked as its own saving) is not credited
twice. Only ever lowers the tag, and lowers the deferral's own ledger entry
by the same amount so ``/stats`` by-source agrees with the headline.
Never raises.
"""
if booking is None or not isinstance(tags, dict) or not callable(count_tools):
return
booked, snapshot, entry = booking
try:
final = {
str(t["name"]): t
for t in (tools if isinstance(tools, list) else [])
if isinstance(t, dict) and t.get("defer_loading") and t.get("name")
}
kept = []
for name, schema in snapshot.items():
if name not in final:
continue
now = final[name]
kept.append(now if int(count_tools([now])) < int(count_tools([schema])) else schema)
measured = int(count_tools(kept)) if kept else 0
released = booked - measured
if released <= 0:
return
tags["tool_search_deferred_tokens"] = measured
tags["tool_search_deferred_tools"] = len(kept)
if isinstance(entry, dict):
entry["tokens"] = max(0, int(entry.get("tokens") or 0) - released)
except Exception: # accounting must never break a request
return
TOOL_SCHEMA_SAVINGS_TAGS: tuple[str, ...] = (
"tool_search_deferred_tokens",
"turn_hook_tools_saved_tokens",
+16 -2
View File
@@ -26,6 +26,12 @@ from collections.abc import Awaitable, Callable
from dataclasses import dataclass, field
from typing import Any, Protocol, runtime_checkable
from headroom.proxy.tool_schema_savings_policy import (
deferred_booking,
reconcile_deferred_tokens,
without_deferral_flags,
)
log = logging.getLogger(__name__)
# Re-drive the model with a message list; returns the provider's response JSON.
@@ -198,18 +204,22 @@ def run_request_hooks(ctx: TurnContext, *, stream_safe_only: bool = False) -> No
working on streamed OpenAI-compatible traffic. Default off ⇒ conservative:
a hook is treated as buffered-only unless it declares itself stream-safe.
"""
# Deferral booked before or by a hook; reconciled once every hook has run,
# since a later hook may un-defer booked tools (tool search's hot tools).
booking = deferred_booking(ctx.tags, ctx.tools)
for hook in registered_turn_hooks():
if stream_safe_only and not getattr(hook, "stream_safe", False):
continue
fn = getattr(hook, "on_request", None)
if fn is None:
continue
booked_tag = ctx.tags.get("tool_search_deferred_tokens")
before_messages = before_tools = None
try:
if ctx.count_messages is not None:
before_messages = ctx.count_messages(ctx.messages)
if ctx.count_tools is not None:
before_tools = ctx.count_tools(ctx.tools)
before_tools = ctx.count_tools(without_deferral_flags(ctx.tools))
fn(ctx)
message_saved = (
max(0, before_messages - ctx.count_messages(ctx.messages))
@@ -217,7 +227,7 @@ def run_request_hooks(ctx: TurnContext, *, stream_safe_only: bool = False) -> No
else 0
)
tool_saved = (
max(0, before_tools - ctx.count_tools(ctx.tools))
max(0, before_tools - ctx.count_tools(without_deferral_flags(ctx.tools)))
if before_tools is not None and ctx.count_tools is not None
else 0
)
@@ -232,6 +242,10 @@ def run_request_hooks(ctx: TurnContext, *, stream_safe_only: bool = False) -> No
)
except Exception: # a hook must never break the proxy
log.exception("turn hook %r on_request failed", getattr(hook, "name", hook))
if ctx.tags.get("tool_search_deferred_tokens") != booked_tag:
# This hook booked (or re-booked) the deferral: it is the new baseline.
booking = deferred_booking(ctx.tags, ctx.tools)
reconcile_deferred_tokens(ctx.tags, booking, ctx.tools, ctx.count_tools)
async def run_response_hooks(
+300
View File
@@ -0,0 +1,300 @@
"""A turn hook's measured tool delta ignores the ``defer_loading`` flag.
Deferral is booked through ``tool_search_deferred_tokens``. Measuring the raw
tool array would see only the flag text appear or disappear, so un-deferring a
tool (tool search's hot-tools hook) was booked as a realized saving although
it makes the request larger.
"""
from __future__ import annotations
import json
import pytest
from headroom.proxy.savings_attribution import from_tags, record_savings
from headroom.proxy.tool_schema_savings_policy import without_deferral_flags
from headroom.proxy.turn_hooks import (
TurnContext,
clear_turn_hooks,
register_turn_hook,
run_request_hooks,
)
@pytest.fixture(autouse=True)
def _clean_registry():
clear_turn_hooks()
yield
clear_turn_hooks()
def _tool(name: str, size: int = 400) -> dict:
return {"name": name, "description": "d" * size, "input_schema": {"type": "object"}}
def _count(value: object) -> int:
# Stand-in tokenizer: one token per 4 characters of the serialised array.
return len(json.dumps(value, default=str)) // 4 if value else 0
class _UndeferHook:
"""Shaped like headroom-tool-search's HotToolsHook: drops the flag in place."""
name = savings_source = "tool_search"
stream_safe = True
def on_request(self, ctx: TurnContext) -> None:
for tool in ctx.tools:
tool.pop("defer_loading", None)
class _DeferHook:
"""Shaped like NativeDeferralHook: flags tools, books through the tag."""
name = savings_source = "tool_search"
stream_safe = True
def on_request(self, ctx: TurnContext) -> None:
ctx.tools = [{**t, "defer_loading": True} for t in ctx.tools]
class _RemoveHook:
"""Shaped like a router: really removes a tool from the array."""
name = savings_source = "tool_router"
stream_safe = True
def on_request(self, ctx: TurnContext) -> None:
ctx.tools = ctx.tools[:1]
def _run(hook: object, tools: list) -> list[dict]:
register_turn_hook(hook)
ctx = TurnContext(
provider="anthropic",
model="claude-sonnet-4-6",
messages=[{"role": "user", "content": "hi"}],
tools=tools,
count_messages=_count,
count_tools=_count,
)
run_request_hooks(ctx)
return from_tags(ctx.tags)
def test_undeferring_a_tool_is_not_booked_as_a_saving() -> None:
tools = [{**_tool("a"), "defer_loading": True}, {**_tool("b"), "defer_loading": True}]
assert _run(_UndeferHook(), tools) == []
def test_deferring_is_left_to_the_deferral_tag() -> None:
# Only the flag changes, so the measured delta is zero and the hook's own
# tool_search_deferred_tokens tag stays the single place deferral is booked.
assert _run(_DeferHook(), [_tool("a"), _tool("b")]) == []
def test_a_real_removal_is_still_measured() -> None:
(entry,) = _run(_RemoveHook(), [_tool("a"), _tool("b")])
assert entry["source"] == "tool_router"
assert entry["tokens"] == _count([_tool("a"), _tool("b")]) - _count([_tool("a")])
def test_without_deferral_flags_never_mutates_and_passes_non_lists() -> None:
tools = [{**_tool("a"), "defer_loading": True}, "not-a-dict"]
out = without_deferral_flags(tools)
assert "defer_loading" in tools[0]
assert out[0] == _tool("a")
assert out[1] == "not-a-dict"
assert without_deferral_flags(None) is None
# ── deferral credit is capped at what is still deferred after hooks ──────────
class _BookDeferral:
"""Shaped like the native deferral hook: flags tools and books them."""
name = savings_source = "tool_search"
stream_safe = True
def on_request(self, ctx: TurnContext) -> None:
ctx.tools = [{**t, "defer_loading": True} for t in ctx.tools]
ctx.tags["tool_search_deferred_tokens"] = _count(ctx.tools)
ctx.tags["tool_search_deferred_tools"] = len(ctx.tools)
record_savings(ctx.tags, "tool_search", tokens=_count(ctx.tools), estimated=True)
class _UndeferFirstGrowSecond:
name = savings_source = "other"
stream_safe = True
def __init__(self, grow: int = 0) -> None:
self.grow = grow
def on_request(self, ctx: TurnContext) -> None:
ctx.tools[0].pop("defer_loading", None)
ctx.tools[1]["description"] += "g" * self.grow
def _booked(*hooks: object) -> tuple[dict, list]:
for hook in hooks:
register_turn_hook(hook)
tags: dict = {}
ctx = TurnContext(
provider="anthropic",
model="claude-sonnet-4-6",
messages=[{"role": "user", "content": "hi"}],
tools=[_tool("a"), _tool("b")],
count_messages=_count,
count_tools=_count,
tags=tags,
)
run_request_hooks(ctx)
return tags, [e for e in from_tags(tags) if e["source"] == "tool_search"]
def test_undeferring_a_booked_tool_lowers_the_headline_and_the_ledger() -> None:
tags, (entry,) = _booked(_BookDeferral(), _UndeferFirstGrowSecond())
expected = _count([{**_tool("b"), "defer_loading": True}])
assert tags["tool_search_deferred_tokens"] == expected
assert tags["tool_search_deferred_tools"] == 1
assert entry["tokens"] == expected # /stats by-source agrees with the headline
def test_growing_another_deferred_schema_cannot_keep_the_credit() -> None:
# b grows by far more than a's share: the final deferred set outweighs the
# booking, yet a is billed now, so its share must still go.
tags, (entry,) = _booked(_BookDeferral(), _UndeferFirstGrowSecond(grow=4_000))
expected = _count([{**_tool("b"), "defer_loading": True}])
assert tags["tool_search_deferred_tokens"] == expected
assert entry["tokens"] == expected
def test_reconcile_never_raises_the_credit_or_touches_unbooked_turns() -> None:
tags, (entry,) = _booked(_BookDeferral())
both = _count([{**_tool("a"), "defer_loading": True}, {**_tool("b"), "defer_loading": True}])
assert tags["tool_search_deferred_tokens"] == both == entry["tokens"]
clear_turn_hooks()
untouched, _ = _booked(_UndeferFirstGrowSecond())
assert "tool_search_deferred_tokens" not in untouched
def test_anthropic_headline_drops_the_credit_of_a_tool_a_hook_undefers(monkeypatch) -> None:
"""End to end: the built-in deferral books 14 MCP tools, then a hook
(shaped like tool search's hot tools) un-defers one. The headline credit
must cover only the 13 still deferred."""
import httpx
import respx
pytest.importorskip("fastapi")
from fastapi.testclient import TestClient
from headroom.proxy.loopback_guard import require_loopback
from headroom.proxy.server import ProxyConfig, create_app
from headroom.tokenizers import get_tokenizer
monkeypatch.setenv("HEADROOM_TOOL_SEARCH", "1")
class _UndeferOne:
name = savings_source = "tool_search"
stream_safe = True
def on_request(self, ctx: TurnContext) -> None:
for tool in ctx.tools:
if isinstance(tool, dict) and tool.get("name") == "mcp__srv__tool_0":
tool.pop("defer_loading", None)
register_turn_hook(_UndeferOne())
app = create_app(
ProxyConfig(
optimize=False,
cache_enabled=False,
rate_limit_enabled=False,
cost_tracking_enabled=False,
log_requests=False,
)
)
app.dependency_overrides[require_loopback] = lambda: None
outcomes: list = []
async def _spy(_self, outcome, *a, **kw): # noqa: ANN001, ANN002, ANN003, ANN202
outcomes.append(outcome)
monkeypatch.setattr(type(app.state.proxy), "_record_request_outcome", _spy, raising=True)
tools = [
{
"name": f"mcp__srv__tool_{i}",
"description": "Look things up. " * 30,
"input_schema": {"type": "object"},
}
for i in range(14)
]
with respx.mock:
respx.post("https://api.anthropic.com/v1/messages").mock(
return_value=httpx.Response(
200,
json={
"id": "a",
"type": "message",
"role": "assistant",
"model": "claude-sonnet-4-5",
"content": [{"type": "text", "text": "ok"}],
"stop_reason": "end_turn",
"usage": {"input_tokens": 100, "output_tokens": 1},
},
)
)
with TestClient(app) as client:
result = client.post(
"/v1/messages",
json={
"model": "claude-sonnet-4-5",
"max_tokens": 16,
"messages": [{"role": "user", "content": "hi"}],
"tools": tools,
},
headers={"x-api-key": "sk-ant-test", "anthropic-version": "2023-06-01"},
)
assert result.status_code == 200
tags = outcomes[-1].tags
assert tags.get("tool_search_mode") == "headroom"
still = [{**t, "defer_loading": True} for t in tools[1:]]
tok = get_tokenizer("claude-sonnet-4-5")
assert tags["tool_search_deferred_tools"] == 13
assert tags["tool_search_deferred_tokens"] == tok.count_text(json.dumps(still, default=str))
(entry,) = [e for e in from_tags(tags) if e["source"] == "tool_search"]
assert entry["tokens"] == tags["tool_search_deferred_tokens"]
class _ShrinkSecondUndeferFirst:
"""A tool_search-sourced hook: shortens b (still deferred), un-defers a."""
name = savings_source = "tool_search"
stream_safe = True
def on_request(self, ctx: TurnContext) -> None:
ctx.tools[0].pop("defer_loading", None)
ctx.tools[1]["description"] = "short"
def test_release_debits_the_deferral_entry_not_a_hooks_own_saving() -> None:
tags, entries = _booked(_BookDeferral(), _ShrinkSecondUndeferFirst())
deferral = [e for e in entries if e.get("estimated")]
measured = [e for e in entries if not e.get("estimated")]
(deferral_entry,) = deferral
(hook_entry,) = measured
shrunk = {**_tool("b"), "description": "short"}
# The hook keeps its own measured saving (b's shrink, flags ignored)...
assert hook_entry["tokens"] == _count([_tool("a"), _tool("b")]) - _count([_tool("a"), shrunk])
# ...and only the deferral entry is debited, down to what is still deferred.
assert deferral_entry["tokens"] == tags["tool_search_deferred_tokens"]
def test_a_shortened_deferred_schema_is_credited_once() -> None:
# b's shrink is the hook's saving; the deferral credits b at its final
# (smaller) size, so those tokens are not counted twice.
tags, _ = _booked(_BookDeferral(), _ShrinkSecondUndeferFirst())
shrunk = {**_tool("b"), "description": "short", "defer_loading": True}
assert tags["tool_search_deferred_tokens"] == _count([shrunk])
assert tags["tool_search_deferred_tools"] == 1