headroom/tests/test_claude_session_mode_benchmark.py
Tejas Chopra 285176be54
fix(tokenizers): price Claude against a real BPE (tiktoken o200k) not a char estimate (#2543)
## Description

Claude's tokenizer is not public, so `get_tokenizer("claude-*")`
returned `EstimatingTokenCounter(chars_per_token=3.5)` — a
**character-ratio estimate**. The estimator's chars-per-token flips with
the *detected content type* (JSON 3.2 / code 3.5 / English 4.0 / CJK
1.5), so:

- the **same bytes** count differently before vs after compression → a
fold could appear to *increase* tokens;
- it **disagrees with the pipeline's** estimator on the same content.

Calibration vs a real BPE (tiktoken `o200k_base`) on representative
content:

| content | char-est(3.5) / o200k |
|---|---|
| English | **1.28×** (overcount 28%) |
| code | **0.83×** (undercount 17%) |
| JSON | 0.97× |
| diff | 1.02× |

i.e. the estimate is off **−17% … +28%**, and the error is
content-dependent (non-uniform) — which is exactly what produces
impossible `tok_after > tok_before` deltas.

This prices Claude against **tiktoken `o200k_base`** — a real,
deterministic, **monotone** BPE. It's not Claude's exact vocab, but it's
within ~10–20% and, crucially, **consistent before/after**, which is
what compression ratios and context-pressure gating actually need.

Closes #

## Type of Change

- [x] Bug fix (non-breaking change that fixes an issue)
- [x] Performance improvement
- [ ] Breaking change
- [ ] Documentation update
- [ ] Code refactoring (no functional changes)

## Changes Made

- `TiktokenCounter.__init__` accepts an explicit `encoding` override (so
a model can be priced against a chosen encoding regardless of name-based
resolution).
- `_create_anthropic` now returns `TiktokenCounter(model,
encoding="o200k_base")`, **failing open to the character estimator** if
the tiktoken vocab can't be loaded (reuses the existing `load_encoding`
timeout/guard).
- Updated the `MODEL_PATTERNS` comment and `test_get_anthropic_model` to
the new contract; added
`test_claude_priced_with_real_bpe_not_char_estimate`.

**Blast radius is contained:** `content_router` keeps its own
module-level estimator for routing decisions, so only the **handler's
reported counts** change. `tiktoken` is already a dependency.

**Composes with #2542** (tokenizer-consistent before/after): that PR
makes both endpoints use *one* tokenizer; this PR makes that one
tokenizer a *real BPE*. Together they fully eliminate the
inflated/phantom lines. This PR is based on `main` and is independently
valid.

## Calibration note (please review)

The one behavioral consumer of the handler's count is the
background-compression gate (`original_tokens >=
_background_compression_min_tokens`). Because o200k differs from the old
estimate by ~±20% depending on content, that threshold now trips on
slightly different requests. It's *more* accurate, but the threshold was
tuned against the estimator — worth a re-check. No other gate consumes
the handler count (routing uses `content_router`'s estimator).

## Testing

- [x] Unit tests pass (`pytest`)
- [x] Linting passes (`ruff check` + `ruff format --check`)
- [x] Type checking passes (`mypy`)
- [x] New tests added
- [x] Manual testing performed

### Test Output

```text
$ ruff check <changed files> && ruff format --check <changed files>
All checks passed!
4 files already formatted

$ mypy headroom/tokenizers/registry.py headroom/tokenizers/tiktoken_counter.py
Success: no issues found in 2 source files

$ pytest tests/test_tokenizer.py tests/test_tokenizers.py -q
55 passed, 14 skipped
```

## Real Behavior Proof

- **Unit:** `get_tokenizer("claude-opus-4-8")` →
`TiktokenCounter(encoding_name="o200k_base")`; `count_text(sample)` ==
`tiktoken.get_encoding("o200k_base").encode(sample)` exactly; a
`tool_result` shrunk 300→3 words drops 508→13 tokens (folds register).
- **Live proxy** (Anthropic path, `--proxy-extension lossless_guard`):
foldable request logs `tok_before=694 tok_after=231 tok_saved=463
transforms=turn_hook` — real o200k-scale numbers (vs the estimator's 607
for the same payload), clean fold, no inflation.
- **Not tested:** exact agreement with Anthropic's *own* count_tokens
API (o200k is a proxy; a follow-up could calibrate against it offline).
GPT paths unchanged (already tiktoken).

## 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
- [ ] Documentation — N/A (internal; docstrings updated)
- [x] My changes generate no new warnings
- [x] New and existing unit tests pass locally with my changes
- [x] I did **not** edit `CHANGELOG.md`

## Additional Notes

Follow-ups (not here): (1) extend the same real-BPE treatment to other
private-tokenizer providers (Gemini/Cohere/Moonshot), each with its own
calibration; (2) optional opt-in calibration against Anthropic's
`count_tokens` API to true-up the ~10–20% absolute offset.
2026-07-24 15:06:43 -07:00

542 lines
19 KiB
Python

"""Tests for Claude session mode simulation benchmark."""
from __future__ import annotations
import json
from datetime import datetime
from pathlib import Path
from types import SimpleNamespace
from benchmarks.claude_session_mode_benchmark import (
PROXY_MODE_CACHE,
PROXY_MODE_TOKEN,
ModeSummary,
ReplayTurn,
SessionReplay,
_extract_cache_stable_last_message_suffix,
_merge_appended_message_delta,
_rewrite_scope,
_write_checkpoint_by_session_id,
build_dataset_and_observed_from_files,
classify_metric_impact,
decode_project_key,
determine_winners,
load_session_replay,
resolve_checkpoint_dir,
simulate_replays,
summarize_mode_impact_vs_baseline,
summarize_observed_usage,
trim_replay_to_recent_turns,
)
def test_decode_project_key_windows_path() -> None:
assert decode_project_key("C--git-BetBlocker") == r"C:\git\BetBlocker"
def test_load_session_replay_groups_assistant_request_events(tmp_path: Path) -> None:
project_dir = tmp_path / "C--git-BetBlocker"
project_dir.mkdir()
session_file = project_dir / "sess-1.jsonl"
lines = [
{
"type": "user",
"message": {"role": "user", "content": "Hello"},
"timestamp": "2026-03-13T01:00:00Z",
},
{
"type": "assistant",
"requestId": "req-1",
"timestamp": "2026-03-13T01:00:01Z",
"message": {
"role": "assistant",
"model": "claude-sonnet-4-6",
"content": [{"type": "thinking", "thinking": "..."}],
"usage": {"output_tokens": 2},
},
},
{
"type": "assistant",
"requestId": "req-1",
"timestamp": "2026-03-13T01:00:02Z",
"message": {
"role": "assistant",
"model": "claude-sonnet-4-6",
"content": [{"type": "text", "text": "Hi"}],
"usage": {"output_tokens": 5},
},
},
{
"type": "user",
"message": {"role": "user", "content": "Next"},
"timestamp": "2026-03-13T01:01:00Z",
},
{
"type": "assistant",
"requestId": "req-2",
"timestamp": "2026-03-13T01:01:05Z",
"message": {
"role": "assistant",
"model": "claude-sonnet-4-6",
"content": [{"type": "text", "text": "Done"}],
"usage": {"output_tokens": 3},
},
},
]
session_file.write_text("\n".join(json.dumps(line) for line in lines), encoding="utf-8")
replay = load_session_replay(session_file)
assert replay is not None
assert len(replay.turns) == 2
assert replay.turns[0].request_id == "req-1"
assert replay.turns[0].output_tokens == 5
assert replay.turns[0].input_messages == [{"role": "user", "content": "Hello"}]
assert replay.turns[1].input_messages == [{"role": "user", "content": "Next"}]
assert replay.turns[1].assistant_message["content"] == [{"type": "text", "text": "Done"}]
def test_simulation_and_winner_logic() -> None:
# Realistic varied tool output. A pathologically repetitive blob (the same
# JSON object * N) is a degenerate case for a real BPE tokenizer — it merges
# the repetition to near-nothing — so a token-mode rewrite can cost MORE real
# tokens than the original, which the old character estimate masked by
# over-counting the repetition. Varied records keep the fixture representative
# of real agent tool output, where the rewrite is a genuine win.
tool_blob = json.dumps(
{
"rows": [
{"id": i, "label": f"row-{i}", "value": (i * 37) % 100, "ok": i % 3 == 0}
for i in range(150)
]
}
)
turn1 = ReplayTurn(
session_id="s1",
project_key="C--git-demo",
decoded_project_path=r"C:\git\demo",
request_id="r1",
model="claude-sonnet-4-6",
timestamp=datetime.fromisoformat("2026-03-13T01:00:00+00:00"),
input_messages=[
{"role": "user", "content": "Summarize this JSON"},
{
"role": "user",
"content": [
{
"type": "tool_result",
"tool_use_id": "tool-1",
"content": tool_blob,
}
],
},
],
assistant_message={"role": "assistant", "content": "ok"},
output_tokens=20,
)
turn2 = ReplayTurn(
session_id="s1",
project_key="C--git-demo",
decoded_project_path=r"C:\git\demo",
request_id="r2",
model="claude-sonnet-4-6",
timestamp=datetime.fromisoformat("2026-03-13T01:03:00+00:00"),
input_messages=[
{"role": "user", "content": "Now tell me the anomalies again"},
],
assistant_message={"role": "assistant", "content": "ok2"},
output_tokens=25,
)
replay = SessionReplay(
session_id="s1",
project_key="C--git-demo",
decoded_project_path=r"C:\git\demo",
turns=[turn1, turn2],
)
dataset, summaries = simulate_replays([replay], cache_ttl_minutes=5)
assert dataset.requests == 2
assert summaries["baseline"].raw_input_tokens > 0
assert (
summaries[PROXY_MODE_TOKEN].forwarded_input_tokens
<= summaries["baseline"].forwarded_input_tokens
)
assert summaries[PROXY_MODE_CACHE].cache_read_tokens >= 0
assert summaries["baseline"].cache_bust_turns == 0
assert summaries[PROXY_MODE_CACHE].cache_bust_turns == 0
assert summaries[PROXY_MODE_TOKEN].cache_bust_turns >= 0
assert summaries[PROXY_MODE_TOKEN].rewrite_turns >= 0
assert summaries[PROXY_MODE_CACHE].rewrite_turns >= 0
winners = determine_winners(summaries)
assert winners["total_cost"] in {"baseline", PROXY_MODE_TOKEN, PROXY_MODE_CACHE}
assert winners["window_with_cache"] in {"baseline", PROXY_MODE_TOKEN, PROXY_MODE_CACHE}
def test_observed_usage_summary_tracks_cache_patterns() -> None:
turns = [
ReplayTurn(
session_id="s1",
project_key="C--git-demo",
decoded_project_path=r"C:\git\demo",
request_id="r1",
model="claude-sonnet-4-6",
timestamp=datetime.fromisoformat("2026-03-13T01:00:00+00:00"),
input_messages=[{"role": "user", "content": "a"}],
assistant_message={"role": "assistant", "content": "x"},
output_tokens=5,
observed_input_tokens=10,
observed_cache_read_tokens=0,
observed_cache_write_tokens=100,
),
ReplayTurn(
session_id="s1",
project_key="C--git-demo",
decoded_project_path=r"C:\git\demo",
request_id="r2",
model="claude-sonnet-4-6",
timestamp=datetime.fromisoformat("2026-03-13T01:01:00+00:00"),
input_messages=[{"role": "user", "content": "b"}],
assistant_message={"role": "assistant", "content": "y"},
output_tokens=6,
observed_input_tokens=9,
observed_cache_read_tokens=80,
observed_cache_write_tokens=90,
),
ReplayTurn(
session_id="s1",
project_key="C--git-demo",
decoded_project_path=r"C:\git\demo",
request_id="r3",
model="claude-sonnet-4-6",
timestamp=datetime.fromisoformat("2026-03-13T01:02:00+00:00"),
input_messages=[{"role": "user", "content": "c"}],
assistant_message={"role": "assistant", "content": "z"},
output_tokens=7,
observed_input_tokens=9,
observed_cache_read_tokens=80,
observed_cache_write_tokens=120,
),
]
replay = SessionReplay(
session_id="s1",
project_key="C--git-demo",
decoded_project_path=r"C:\git\demo",
turns=turns,
)
observed = summarize_observed_usage([replay])
assert observed.requests == 3
assert observed.cache_read_tokens == 160
assert observed.cache_write_tokens == 310
assert observed.healthy_growth_turns == 1
assert observed.broken_prefix_turns == 2
def test_checkpoint_write_omits_per_turn_payload(tmp_path: Path) -> None:
summary = ModeSummary(
mode=PROXY_MODE_TOKEN,
sessions=1,
requests=1,
turns=[],
)
_write_checkpoint_by_session_id(tmp_path, PROXY_MODE_TOKEN, "session-1", summary)
payload = json.loads((tmp_path / f"{PROXY_MODE_TOKEN}--session-1.json").read_text())
assert payload["turns"] == []
def test_trim_replay_to_recent_turns_keeps_latest_slice() -> None:
replay = SessionReplay(
session_id="s1",
project_key="C--git-demo",
decoded_project_path=r"C:\git\demo",
turns=[
ReplayTurn(
session_id="s1",
project_key="C--git-demo",
decoded_project_path=r"C:\git\demo",
request_id=f"r{i}",
model="claude-sonnet-4-6",
timestamp=datetime.fromisoformat(f"2026-03-13T01:0{i}:00+00:00"),
input_messages=[{"role": "user", "content": str(i)}],
assistant_message={"role": "assistant", "content": str(i)},
output_tokens=i,
)
for i in range(4)
],
)
trimmed = trim_replay_to_recent_turns(replay, 2)
assert [turn.request_id for turn in trimmed.turns] == ["r2", "r3"]
def test_build_dataset_and_observed_from_files_applies_recent_turn_sampling(
tmp_path: Path,
) -> None:
project_dir = tmp_path / "C--git-BetBlocker"
project_dir.mkdir()
session_file = project_dir / "sess-1.jsonl"
lines = []
for i in range(3):
lines.append(
{
"type": "user",
"message": {"role": "user", "content": f"Hello {i}"},
"timestamp": f"2026-03-13T01:0{i}:00Z",
}
)
lines.append(
{
"type": "assistant",
"requestId": f"req-{i}",
"timestamp": f"2026-03-13T01:0{i}:01Z",
"message": {
"role": "assistant",
"model": "claude-sonnet-4-6",
"content": [{"type": "text", "text": f"Hi {i}"}],
"usage": {
"output_tokens": 3,
"input_tokens": 10,
"cache_read_input_tokens": 20,
"cache_creation_input_tokens": 5,
},
},
}
)
session_file.write_text("\n".join(json.dumps(line) for line in lines), encoding="utf-8")
dataset, observed = build_dataset_and_observed_from_files(
[session_file],
recent_turns_per_session=2,
)
assert dataset.requests == 2
assert dataset.sampled_requests == 2
assert dataset.sampling_note == "Most recent 2 turns per session"
assert observed.requests == 2
def test_determine_winners_includes_no_cache_counterfactual() -> None:
summaries = {
"baseline": ModeSummary(
mode="baseline",
paid_input_cost_usd=1.0,
cache_read_cost_usd=0.2,
paid_output_cost_usd=0.5,
),
PROXY_MODE_TOKEN: ModeSummary(
mode=PROXY_MODE_TOKEN,
paid_input_cost_usd=0.8,
cache_read_cost_usd=0.1,
paid_output_cost_usd=0.5,
),
PROXY_MODE_CACHE: ModeSummary(
mode=PROXY_MODE_CACHE,
paid_input_cost_usd=0.7,
cache_read_cost_usd=0.3,
paid_output_cost_usd=0.5,
),
}
winners = determine_winners(summaries)
assert winners["no_cache_total_cost"] == PROXY_MODE_TOKEN
def test_resolve_checkpoint_dir_namespaces_sampling_mode() -> None:
base = Path("benchmark_results") / "checkpoints"
assert resolve_checkpoint_dir(base).name == "v5__ttl_5m__full"
assert (
resolve_checkpoint_dir(base, recent_turns_per_session=200).name == "v5__ttl_5m__recent_200"
)
def test_cache_suffix_helpers_support_append_only_text_growth() -> None:
suffix_delta = _extract_cache_stable_last_message_suffix(
[{"role": "user", "content": "prefix + raw suffix"}],
[{"role": "user", "content": "prefix"}],
[{"role": "user", "content": "COMPRESSED_PREFIX"}],
)
assert suffix_delta is not None
stable_prefix, stable_last_message, delta_messages = suffix_delta
assert stable_prefix == []
assert stable_last_message == {"role": "user", "content": "COMPRESSED_PREFIX"}
assert delta_messages == [{"role": "user", "content": " + raw suffix"}]
merged = _merge_appended_message_delta(
stable_last_message,
{"role": "user", "content": " + COMPRESSED_SUFFIX"},
)
assert merged == {"role": "user", "content": "COMPRESSED_PREFIX + COMPRESSED_SUFFIX"}
def test_mode_impact_classification_marks_assist_harm_and_no_change() -> None:
baseline = ModeSummary(
mode="baseline",
forwarded_input_tokens=100,
cache_read_tokens=50,
cache_write_tokens=10,
regular_input_tokens=40,
output_tokens=5,
total_cost_usd=1.0,
)
token = ModeSummary(
mode=PROXY_MODE_TOKEN,
forwarded_input_tokens=80,
cache_read_tokens=70,
cache_write_tokens=8,
regular_input_tokens=30,
output_tokens=5,
total_cost_usd=0.8,
)
cache = ModeSummary(
mode=PROXY_MODE_CACHE,
forwarded_input_tokens=120,
cache_read_tokens=45,
cache_write_tokens=15,
regular_input_tokens=60,
output_tokens=5,
total_cost_usd=1.2,
)
assert classify_metric_impact(baseline, token, "forwarded_input_tokens")["impact"] == "assist"
assert classify_metric_impact(baseline, token, "cache_read_tokens")["impact"] == "assist"
assert classify_metric_impact(baseline, cache, "total_cost_usd")["impact"] == "harm"
assert classify_metric_impact(baseline, token, "output_tokens")["impact"] == "no_change"
impacts = summarize_mode_impact_vs_baseline(
{"baseline": baseline, PROXY_MODE_TOKEN: token, PROXY_MODE_CACHE: cache}
)
assert impacts[PROXY_MODE_TOKEN]["total_cost_usd"]["impact"] == "assist"
assert impacts[PROXY_MODE_CACHE]["cache_write_tokens"]["impact"] == "harm"
def test_rewrite_scope_distinguishes_retroactive_from_latest_turn_only() -> None:
rewrite, retroactive = _rewrite_scope(
[{"role": "user", "content": "prefix"}, {"role": "user", "content": "new raw"}],
[{"role": "user", "content": "prefix"}, {"role": "user", "content": "new compressed"}],
stable_prefix_message_count=1,
)
assert rewrite is True
assert retroactive is False
rewrite, retroactive = _rewrite_scope(
[{"role": "user", "content": "prefix"}, {"role": "user", "content": "new raw"}],
[
{"role": "user", "content": "compressed prefix"},
{"role": "user", "content": "new compressed"},
],
stable_prefix_message_count=1,
)
assert rewrite is True
assert retroactive is True
def test_synthetic_token_mode_busts_cache_while_cache_mode_stays_stable(monkeypatch) -> None:
class _FakeProvider:
@staticmethod
def get_context_limit(model: str) -> int:
return 200_000
class _FakePipeline:
@staticmethod
def apply(messages, **kwargs): # noqa: ANN001
rewritten = []
should_rewrite_history = len(messages) > 2
for message in messages:
content = message.get("content")
if (
should_rewrite_history
and isinstance(content, list)
and any(
isinstance(block, dict) and block.get("type") == "tool_result"
for block in content
)
):
new_blocks = []
for block in content:
if isinstance(block, dict) and block.get("type") == "tool_result":
new_blocks.append({**block, "content": "[compressed-tool-result]"})
else:
new_blocks.append(block)
rewritten.append({**message, "content": new_blocks})
else:
rewritten.append(message)
return SimpleNamespace(messages=rewritten)
class _FakeProxy:
def __init__(self) -> None:
self.config = SimpleNamespace(image_optimize=False)
self.anthropic_provider = _FakeProvider()
self.anthropic_pipeline = _FakePipeline()
monkeypatch.setattr(
"benchmarks.claude_session_mode_benchmark._make_proxy",
lambda mode: _FakeProxy(),
)
tool_blob = "X" * 800
replay = SessionReplay(
session_id="synth-bust",
project_key="C--git-synth",
decoded_project_path=r"C:\git\synth",
turns=[
ReplayTurn(
session_id="synth-bust",
project_key="C--git-synth",
decoded_project_path=r"C:\git\synth",
request_id="r1",
model="claude-sonnet-4-6",
timestamp=datetime.fromisoformat("2026-03-13T01:00:00+00:00"),
input_messages=[
{"role": "user", "content": "Summarize this tool output"},
{
"role": "user",
"content": [
{
"type": "tool_result",
"tool_use_id": "tool-1",
"content": tool_blob,
}
],
},
],
assistant_message={"role": "assistant", "content": "ok"},
output_tokens=10,
),
ReplayTurn(
session_id="synth-bust",
project_key="C--git-synth",
decoded_project_path=r"C:\git\synth",
request_id="r2",
model="claude-sonnet-4-6",
timestamp=datetime.fromisoformat("2026-03-13T01:02:00+00:00"),
input_messages=[{"role": "user", "content": "What changed?"}],
assistant_message={"role": "assistant", "content": "done"},
output_tokens=12,
),
],
)
_, summaries = simulate_replays([replay], cache_ttl_minutes=5)
token = summaries[PROXY_MODE_TOKEN]
cache = summaries[PROXY_MODE_CACHE]
assert token.cache_bust_turns == 1
assert token.rewrite_turns >= 1
assert token.busting_rewrite_turns >= 1
assert token.non_cache_eligible_rewrite_turns == 0
assert token.stable_replay_rewrite_turns == 0
assert token.retroactive_rewrite_turns >= 1
assert cache.cache_bust_turns == 0
assert cache.busting_rewrite_turns == 0
assert cache.non_cache_eligible_rewrite_turns == 0
assert cache.retroactive_rewrite_turns == 0