fix(ccr): make --no-ccr disable server-side response handling too (#3101)

## Description

`--no-ccr` advertises **"Disable CCR entirely"**, and its help text
names the case it exists for: *"streaming / non-MCP clients that can't
resolve an injected tool."* It mapped onto only two of the three CCR
subsystems — markers and tool injection — leaving `ccr_handle_responses`
on. That field has no flag and no env var of its own, so under
`--no-ccr` it was always `True`.

That mattered because the buffered `stream: false` path keys off
`headroom_retrieve` being present in the **request's** tools, and the
client can put it there itself — the bundled OpenCode plugin registers
it unconditionally. So `--no-ccr` left the buffered path fully armed for
exactly the clients it was recommended to, and any turn whose history
still held a redeemable marker kept being flipped to buffered.

This is why the workaround handed out in #2952 / #3017 / #3079 did
nothing for `headroom wrap opencode`.

## Type of Change

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

## Changes Made

- `headroom/cli/proxy.py`: `--no-ccr` / `HEADROOM_NO_CCR` now also sets
`ccr_handle_responses=False`, so the switch covers all three CCR
subsystems rather than two.
- Rewrote the inline comment, which claimed the flag "disables both
halves at once" — there were three.
- `tests/test_no_ccr_disables_response_handling.py`: 5 tests covering
the flag→config mapping (flag, env var, and the untouched default), plus
the behaviour it buys — a client-advertised `headroom_retrieve` with a
redeemable marker no longer flips the turn to buffered.

## Testing

- [x] Unit tests pass (`pytest`)
- [x] New tests added for new functionality
- [x] Manual testing performed

### Test Output

```text
tests/test_no_ccr_disables_response_handling.py .....                    [100%]
5 passed

Full suite (both Tier 1 fixes applied):
3 failed, 11220 passed, 581 skipped in 428.90s
```

The 3 failures are pre-existing and environmental, identical to a
plain-`main` baseline run on the same machine: no `cargo` installed
(`test_no_native_tls_in_wheel_build_tree`), no `codex` CLI
(`test_learn/test_integration.py`), and
`test_run_server_installs_cancelled_error_filter`, which fails under
full-suite ordering on `main` too.

## Real Behavior Proof

- Environment: macOS (darwin 25.4.0), Python 3.12.13, worktree off
`main` @ `7ef736fb`, `HEADROOM_SKIP_UPSTREAM_CHECK=1`
- Exact command / steps: Drove one streaming `/v1/messages` turn through
`create_app()` with the `--no-ccr` posture (`ccr_inject_tool=False`,
`ccr_inject_marker=False`), a client-supplied `headroom_retrieve` in
`tools`, and a redeemable `<<ccr:…>>` marker in the message — then
recorded the `stream` value that reached the upstream stub.
- Observed result: Before — upstream received `stream=False`; the turn
was buffered despite `--no-ccr`. Only setting
`ccr_handle_responses=False` stopped it. After — `headroom proxy
--no-ccr` and `HEADROOM_NO_CCR=1` both produce
`ccr_handle_responses=False`, and the same turn keeps streaming.
Reverting just `headroom/cli/proxy.py` fails the two mapping tests and
passes them again with it restored.
- Not tested: No live OpenCode + GitHub Copilot session; the reporter's
end-to-end confirmation is still wanted. The OpenCode plugin still
registers `headroom_retrieve` unconditionally — deliberately left alone,
since with this fix an advertised tool no longer causes buffering.

## Runtime Rollout Safety

- Rollout-managed feature(s): None — no rollout channel gates this.
- Minimum rollout channel: n/a
- Stable/default behavior changed: No. Default (no flag) keeps
`ccr_handle_responses=True`, pinned by a test.
- Kill switch / disable path: This *is* the kill switch; the change
makes it work as documented.
- Unsafe override required: No.
- Qualification impact: None.
- Rollback path: Revert this commit; `--no-ccr` returns to disabling two
of three subsystems.

## Review Readiness

- [x] I have performed a self-review
- [x] This PR is ready for human review

Closes #3082

Co-authored-by: Tejas Chopra <tejas@Tejass-MacBook-Pro.local>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Tejas Chopra
2026-08-18 07:15:55 -07:00
committed by GitHub
parent 7ef736fb1a
commit 131b119c05
2 changed files with 178 additions and 4 deletions
+14 -4
View File
@@ -1330,12 +1330,22 @@ def proxy(
protect_recent=_get_env_int_optional("HEADROOM_PROTECT_RECENT"),
protect_analysis_context=_get_env_bool_optional("HEADROOM_PROTECT_ANALYSIS_CONTEXT"),
accuracy_guard=os.environ.get("HEADROOM_ACCURACY_GUARD") or None,
# CCR opt-out: --no-ccr disables both halves at once (markers in content
# AND the injected retrieve tool). Markers without a tool — or a tool
# without markers — are useless, so it is a single switch. Default keeps
# CCR fully on.
# CCR opt-out: --no-ccr disables every half at once markers in
# content, the injected retrieve tool, AND server-side response
# handling. Markers without a tool, or a tool without markers, are
# useless, so it is a single switch. Default keeps CCR fully on.
#
# Response handling has to be part of it. The buffered stream:false
# path keys off ``headroom_retrieve`` being present in the *request's*
# tools, and a client can advertise that tool on its own — the bundled
# OpenCode plugin registers it unconditionally. So with response
# handling left on, `--no-ccr` silently kept flipping streaming turns
# to buffered whenever history still held a redeemable marker, and the
# documented escape hatch for the CCR buffered-stream bugs did nothing
# for exactly the clients told to use it (#3082).
ccr_inject_tool=not no_ccr,
ccr_inject_marker=not no_ccr,
ccr_handle_responses=not no_ccr,
ccr_resolve_markers_inline=ccr_inline_resolve,
lossless=lossless,
ccr_proactive_expansion=not no_ccr_proactive_expansion,
@@ -0,0 +1,164 @@
"""`--no-ccr` has to disable server-side CCR handling too (#3082).
The flag advertises "Disable CCR entirely", and names the case it exists for:
streaming / non-MCP clients that cannot resolve an injected tool. It mapped onto
only two of the three CCR subsystems, though — markers and tool injection —
leaving ``ccr_handle_responses`` on, which has no flag and no env var of its own.
That mattered because the buffered ``stream: false`` path keys off
``headroom_retrieve`` being present in the *request's* tools, and the client can
put it there itself: the bundled OpenCode plugin registers the tool
unconditionally. So `--no-ccr` left the buffered path fully armed for exactly
the clients it was recommended to, and a turn whose history still held a
redeemable marker kept being flipped to buffered.
"""
from __future__ import annotations
import json
from unittest.mock import patch
import pytest
click = pytest.importorskip("click")
pytest.importorskip("fastapi")
httpx = pytest.importorskip("httpx")
from click.testing import CliRunner # noqa: E402
from fastapi.testclient import TestClient # noqa: E402
from headroom.cache.backends import InMemoryBackend # noqa: E402
from headroom.cache.compression_store import ( # noqa: E402
get_compression_store,
reset_compression_store,
)
from headroom.ccr.tool_injection import create_ccr_tool_definition # noqa: E402
from headroom.cli.main import main # noqa: E402
from headroom.proxy.server import ProxyConfig, create_app # noqa: E402
def _config_for(args: list[str], env: dict[str, str] | None = None) -> ProxyConfig:
"""Run `headroom proxy ...` far enough to capture the ProxyConfig it builds."""
captured: dict[str, ProxyConfig] = {}
def mock_run_server(config, **kwargs): # noqa: ANN001, ANN003
captured["config"] = config
with patch("headroom.proxy.server.run_server", mock_run_server):
result = CliRunner().invoke(main, args, env=env or {}, catch_exceptions=False)
assert result.exit_code == 0, result.output
return captured["config"]
# --------------------------------------------------------------------------- #
# The flag -> config mapping
# --------------------------------------------------------------------------- #
def test_no_ccr_flag_disables_response_handling() -> None:
config = _config_for(["proxy", "--no-ccr"])
assert config.ccr_inject_tool is False
assert config.ccr_inject_marker is False
# The half that used to survive the switch.
assert config.ccr_handle_responses is False
def test_no_ccr_env_var_disables_response_handling() -> None:
config = _config_for(["proxy"], env={"HEADROOM_NO_CCR": "1"})
assert config.ccr_inject_tool is False
assert config.ccr_inject_marker is False
assert config.ccr_handle_responses is False
def test_default_keeps_ccr_fully_on() -> None:
"""The switch must not leak into the default posture."""
config = _config_for(["proxy"])
assert config.ccr_inject_tool is True
assert config.ccr_inject_marker is True
assert config.ccr_handle_responses is True
# --------------------------------------------------------------------------- #
# What that mapping buys: no buffered flip, even when the client offers the tool
# --------------------------------------------------------------------------- #
@pytest.fixture
def _store():
reset_compression_store()
get_compression_store(backend=InMemoryBackend())
try:
yield
finally:
reset_compression_store()
def _upstream_stream_field(*, ccr_handle_responses: bool) -> object:
"""Drive one streaming turn and report the `stream` value sent upstream."""
marker = get_compression_store().store(
original=json.dumps({"earlier": "tool output"}),
compressed="{}",
original_item_count=400,
)
config = ProxyConfig(
optimize=False,
cache_enabled=False,
rate_limit_enabled=False,
memory_enabled=False,
# `--no-ccr` posture: nothing injected by us.
ccr_inject_tool=False,
ccr_inject_marker=False,
ccr_handle_responses=ccr_handle_responses,
ccr_context_tracking=False,
image_optimize=False,
)
seen: dict[str, object] = {}
app = create_app(config)
with TestClient(app) as client:
proxy = client.app.state.proxy
async def _fake_retry(method, url, headers, body, stream=False, **kwargs): # noqa: ANN001
sent = json.loads(body) if isinstance(body, (str, bytes)) else body
seen["stream"] = sent.get("stream")
return httpx.Response(
200,
json={
"id": "msg_1",
"type": "message",
"role": "assistant",
"model": "claude-sonnet-4-6",
"content": [{"type": "text", "text": "ok"}],
"stop_reason": "end_turn",
"usage": {
"input_tokens": 10,
"output_tokens": 5,
"cache_read_input_tokens": 0,
"cache_creation_input_tokens": 0,
},
},
)
proxy._retry_request = _fake_retry # type: ignore[assignment]
client.post(
"/v1/messages",
json={
"model": "claude-sonnet-4-6",
"max_tokens": 64,
"stream": True,
# The client advertises the tool itself, as the OpenCode plugin does.
"tools": [create_ccr_tool_definition("anthropic")],
"messages": [{"role": "user", "content": f"go <<ccr:{marker}>>"}],
},
headers={"x-api-key": "test-key", "anthropic-version": "2023-06-01"},
)
return seen.get("stream")
def test_client_offered_tool_still_buffers_when_handling_is_on(_store) -> None: # noqa: ANN001
"""The behaviour being switched off — pinned so the switch is meaningful."""
assert _upstream_stream_field(ccr_handle_responses=True) is False
def test_client_offered_tool_does_not_buffer_under_no_ccr(_store) -> None: # noqa: ANN001
"""With the switch honoured, the turn keeps streaming despite the tool."""
assert _upstream_stream_field(ccr_handle_responses=False) is not False