Files
aeb09f8d18 fix(cache): keep cache mode when Claude Code re-serializes its last reply (#4003)
## Description

This PR fixes two problems that cost Headroom users money on Claude
Code. A live benchmark (`benchmarks/tool_search_vs_native/`) found both.

**1. Cache mode broke when Claude Code re-serialized its last reply.**

Headroom records the model's raw reply as the last message of the
previous turn. Claude Code sends that reply back edited:

- it strips a leading `cd <cwd> &&` from Bash commands;
- it drops `caller`.

The history then read as `relation=diverged`
(`cache_mode_prefix_mismatch`). Cache mode forwarded the raw originals
over every earlier compressed message, and the whole conversation was
written to cache again: 66.8K tokens on every long-task run.

The fix: when only the recorded reply differs, and it is the same turn,
Headroom replays the forwarded prefix and takes the client's version of
the reply. That reply was never part of a cached prefix. "Same turn"
means the same block kinds and the same `tool_use` ids.

**2. Small MCP results were compressed.**

The Jira ticket (a few hundred tokens) lost its description. The model
called `headroom_retrieve` and then re-fetched the ticket, which cost
two extra turns and about +15% on the long task. MCP results under
`HEADROOM_MCP_RESULT_MIN_CHARS` (default 4000) are now forwarded
verbatim. This covers both Anthropic `tool_result` blocks and OpenAI
`role: "tool"` messages.

No linked issue.

## 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
- [x] Performance improvement
- [ ] Code refactoring (no functional changes)

## Changes Made

- **`headroom/cache/prefix_tracker.py`:** adds
`_client_rewrote_reply_delta`, tried on `RELATION_DIVERGED` after
`_replaced_system_tail_delta`. A change earlier in the history still
diverges.
- **`headroom/transforms/content_router.py`:** adds
`_is_small_mcp_result`, applied in `_process_content_blocks` (Anthropic)
and on the `role: "tool"` branch (OpenAI). It recognises the `mcp__`,
`mcp_` and `mcp--` name prefixes. `HEADROOM_MCP_RESULT_MIN_CHARS=0`
restores the old behaviour.
- **`headroom/proxy/handlers/anthropic.py`:** adds an opt-in
`HEADROOM_DEBUG_PREFIX_MISMATCH=<dir>`. It logs the first diverging
message and dumps both histories, with the directory created `0700` and
the file `0600`, because the dump holds the full conversation.
- **`tests/test_cache_mode_delta_marker.py`:**
`test_tool_use_different_input_still_falls_back` becomes
`..._forwards_the_clients_version`, because the client's copy of the
last reply is what gets forwarded and never the recorded one. A new test
keeps the divergence for an edit earlier in the history.
- **New tests:**
  - `tests/test_prefix_tracker_client_rewrote_reply.py`;
- `tests/test_content_router_mcp_small_result.py`, which covers both
wire formats and three MCP prefixes.
- **`benchmarks/tool_search_vs_native/`:** a live A/B harness. It runs
isolated `claude -p` sessions against six fake MCP servers and a fixture
repo, and reads billed usage from Claude Code's transcripts. A run
passes only if Claude Code exits 0, the tests pass, and the four MCP
actions happened (`get_issue` counts as reading the ticket; a search
does not).

## Testing

- [x] Unit tests pass (`pytest`)
- [x] Linting passes (`ruff check .`)
- [x] Type checking passes (`mypy headroom`)
- [x] New tests added for new functionality
- [x] Manual testing performed

### Test Output

```text
$ pytest -q tests/test_prefix_tracker_client_rewrote_reply.py tests/test_content_router_mcp_small_result.py tests/test_cache_mode_delta_marker.py
============================== 20 passed in 3.42s ==============================

$ pytest -q tests -k "content_router or router or prefix or cache_mode or delta"
========= 1 failed, 983 passed, 15147 deselected in 222.61s (0:03:42) ==========
# the 1 failure is test_content_router_single_item_deadline.py::test_single_cache_miss_fails_open_at_deadline,
# a wall-clock assertion (0.126s vs < 0.12s). It fails identically on origin/main on this machine.

$ ruff check .            (ruff 0.16.8, CI's pinned version)
All checks passed!
$ ruff format --check .
1757 files already formatted
$ mypy headroom/proxy/handlers/anthropic.py headroom/transforms/content_router.py headroom/cache/prefix_tracker.py --ignore-missing-imports
Success: no issues found
```

## Real Behavior Proof

- Environment: macOS 26.6 (arm64), Python 3.12.13 (the repo's `.venv`),
Claude Code 2.1.290, `claude-sonnet-5-5`, real Anthropic API calls, with
the proxy started from this branch.
- Exact command / steps: `python3
benchmarks/tool_search_vs_native/run.py --arms native,hr_hot,hr_client
--reps 3 --warmup`, then the same with `--context-kb 200` for the long
task.
- Observed result: median of 3 runs after a warmup; every run passed
(table below). Before the fix, `HEADROOM_DEBUG_PREFIX_MISMATCH` showed
`relation=diverged first_diff_index=8`: the Bash `tool_use` with `cd …
&&` stripped and `caller` dropped.
- Not tested: harnesses other than Claude Code (the OpenAI-format MCP
path is unit-tested only), Bedrock and Vertex upstreams, real (larger)
MCP server catalogs, models other than Sonnet 5.5.

| | native Claude Code | Headroom + CC ToolSearch | Headroom, routine
tools resident |
|---|---|---|---|
| long task, before | $0.283 | $0.323 | $0.426 (66.8K-token cache
re-write every run) |
| long task, after | $0.284 | **$0.284** | **$0.255** (6 API calls vs
native's 7) |
| short task, after | $0.061 | $0.062 | **$0.046** |

## Runtime Rollout Safety

- Rollout-managed feature(s): none; both changes are in the always-on
cache-mode and compression paths.
- Minimum rollout channel: stable.
- Stable/default behavior changed: yes. A re-serialized last reply no
longer drops cache mode to raw forwarding, and MCP results under 4000
characters are no longer compressed, giving up a few tokens of savings
on those results.
- Kill switch / disable path: `HEADROOM_MCP_RESULT_MIN_CHARS=0` for the
MCP floor. The prefix fix has no flag, because it only applies where the
old code forwarded raw.
- Unsafe override required: no.
- Qualification impact: less compression on small MCP results; fewer
cache writes and extra turns.
- 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
- [x] 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)

## Screenshots (if applicable)

N/A

## Additional Notes

- **Devin review follow-ups are in `d043bd5e`:**
  - the OpenAI-format MCP floor;
  - owner-only debug dumps;
  - exit-code and `get_issue` scoring;
  - the environment and commands in the README.
- **The two `npm audit` failures** (docs, e2e/wrap/tools) also fail on
`main`.
- **`tests/test_cli_proxy_improvements.py` hangs** when the whole file
runs locally, also on `main`. That is pre-existing and not touched here.

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

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: JD Davis <mxjerrett@gmail.com>
2026-10-07 10:58:23 -05:00
..

Claude Code tool search: native vs Headroom

A live A/B benchmark. Each run is a fresh, isolated Claude Code session with its own throwaway HOME, a new copy of fixture/, and six fake enterprise MCP servers (mcp_server.py: GitHub, Jira, Slack, Datadog, PagerDuty and Confluence, 63 tools). Every session does the same task:

  1. Read Jira BENCH-42.
  2. Fix the pricing bug.
  3. Run the tests.
  4. Search the Datadog logs.
  5. Comment on the ticket.
  6. Post to Slack.

Usage is read from Claude Code's own transcript, which records what the API billed, subagents included.

python3 benchmarks/tool_search_vs_native/run.py --arms native,hr_client,hr_server,hr_hot --reps 3 --warmup
python3 benchmarks/tool_search_vs_native/run.py ... --context-kb 200   # lookups deep in a long session

The script needs ANTHROPIC_API_KEY (or ~/env.txt) and the repo's .venv. A full run of four arms with three reps and a warmup costs about $1 short and $5 long, on Sonnet 5.5.

Arm Setup
native Claude Code straight to api.anthropic.com, with its own ToolSearch (on by default).
eager Claude Code straight to Anthropic with ENABLE_TOOL_SEARCH=false, so nothing is deferred.
hr_client Claude Code through Headroom with ENABLE_TOOL_SEARCH=true (what headroom wrap sets today).
hr_server Claude Code through Headroom with ENABLE_TOOL_SEARCH=false. Headroom defers tools on Anthropic's side.
hr_hot As hr_server, but the MCP tools a developer uses routinely stay loaded (HEADROOM_TOOL_SEARCH_CORE_TOOLS).

Results (2026-10-06, Claude Code 2.1.290, Sonnet 5.5, n=3 after warmup, median)

Run on macOS 26.6 (arm64) with Python 3.12.13 (the repo's .venv) and real Anthropic API calls. Commands:

python3 benchmarks/tool_search_vs_native/run.py --arms native,hr_hot,hr_client --reps 3 --warmup
python3 benchmarks/tool_search_vs_native/run.py --arms native,hr_hot,hr_client --reps 3 --warmup --context-kb 200

A run passes only if Claude Code exits 0, the fixture tests pass, and the session called Jira get_issue, Datadog search_logs, Jira add_comment and Slack post_message. All 73 recorded runs pass under that rule.

native hr_client hr_server hr_hot
short task, before fixes $0.059 $0.062 $0.081 $0.049
long task, before fixes $0.283 $0.323 $0.350 $0.426
long task, after fixes $0.284 $0.284 — $0.255 (−10%)
short task, after fixes $0.061 $0.062 — $0.046 (−25%)

The "after" rows held in every run: on the long task hr_hot cost $0.254–0.256 and native $0.282–0.284.

What the runs showed

  • A tool lookup costs a re-read of the whole conversation. That holds for Claude Code's ToolSearch, which takes an extra turn, and for Anthropic's server-side search, which takes an extra sampling pass inside the same call (usage sums the iterations). A schema that stays loaded costs about 0.1× its size per turn. So keeping a routinely used tool loaded is far cheaper than looking it up.
  • Server-side search without names searches blindly. Claude Code tells the model the deferred tool names, so it loads everything it needs in one select:. Headroom's server-side deferral gives no names, and the model ran 2–5 regex searches. Each search pulled in about 4.7K tokens of schemas, uncached.
  • Bug: cache mode broke when Claude Code re-serialized its last reply. Headroom records the model's raw reply as the end of the previous turn. Claude Code sends that reply back edited: it strips a leading cd <cwd> && from Bash commands and drops caller. The history then read as diverged, Headroom forwarded the raw originals over its earlier compressed messages, and the whole conversation was re-written to cache: 66.8K tokens on every long hr_hot run. Fixed in headroom/cache/prefix_tracker.py (_client_rewrote_reply_delta). Set HEADROOM_DEBUG_PREFIX_MISMATCH=<dir> to log and dump where a history diverges.
  • Bug: compressing small MCP results cost turns. The Jira ticket (a few hundred tokens) was compressed and lost its description, so the model called headroom_retrieve and then re-fetched the ticket: two extra turns, about +15% on the long task. MCP results under HEADROOM_MCP_RESULT_MIN_CHARS (default 4000) are now forwarded verbatim, in both Anthropic tool_result blocks and OpenAI role: "tool" messages.

The fake MCP schemas are terse compared with real servers such as GitHub's official one, so absolute savings here understate a real catalog.