xithead 1.5.1 → 1.5.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (85) hide show
  1. package/bin/xithead.js +486 -226
  2. package/index.d.ts +57 -0
  3. package/index.js +139 -0
  4. package/package.json +44 -63
  5. package/postinstall.js +43 -0
  6. package/LICENSE +0 -190
  7. package/README.md +0 -108
  8. package/app/api/config/route.ts +0 -22
  9. package/app/api/export/route.ts +0 -83
  10. package/app/api/models/route.ts +0 -12
  11. package/app/api/stats/route.ts +0 -46
  12. package/app/dashboard/page.tsx +0 -445
  13. package/app/globals.css +0 -35
  14. package/app/installation/page.tsx +0 -293
  15. package/app/layout.tsx +0 -35
  16. package/app/master-design-system.css +0 -2076
  17. package/app/master-design-system.source.css +0 -8
  18. package/app/page.tsx +0 -187
  19. package/app/rechtliches/agb/page.tsx +0 -165
  20. package/app/rechtliches/datenschutz/page.tsx +0 -128
  21. package/app/rechtliches/impressum/page.tsx +0 -122
  22. package/app/rechtliches/widerruf/page.tsx +0 -185
  23. package/components/layout/AppShell.tsx +0 -25
  24. package/components/layout/Footer.tsx +0 -120
  25. package/components/layout/Header.tsx +0 -145
  26. package/components/marketing/LegalHighlights.tsx +0 -45
  27. package/components/marketing/LegalPageShell.tsx +0 -49
  28. package/components/marketing/LegalPanel.tsx +0 -41
  29. package/components/marketing/index.ts +0 -3
  30. package/design-system/README.md +0 -80
  31. package/design-system/base.css +0 -127
  32. package/design-system/components.css +0 -1943
  33. package/design-system/fonts.css +0 -17
  34. package/design-system/tailwind-preset.js +0 -95
  35. package/design-system/tokens.css +0 -48
  36. package/design-system/utilities.css +0 -53
  37. package/lib/config/index.ts +0 -12
  38. package/lib/hooks/index.ts +0 -4
  39. package/lib/hooks/use-mounted.ts +0 -13
  40. package/lib/hooks/use-session.ts +0 -34
  41. package/lib/hooks/use-settings.ts +0 -49
  42. package/lib/utils/index.ts +0 -1
  43. package/lib/utils/settings.ts +0 -162
  44. package/lib/utils/utils.ts +0 -6
  45. package/middleware.ts +0 -26
  46. package/next-env.d.ts +0 -7
  47. package/next.config.ts +0 -12
  48. package/postcss.config.mjs +0 -7
  49. package/public/apple-touch-icon.png +0 -0
  50. package/public/logo-transparent.png +0 -0
  51. package/public/xithead.ps1 +0 -937
  52. package/scripts/README.md +0 -77
  53. package/scripts/audit_wheel_glibc_symbols.py +0 -217
  54. package/scripts/build-master-design-system.cjs +0 -33
  55. package/scripts/build_rust_extension.sh +0 -58
  56. package/scripts/changelog-gen.py +0 -203
  57. package/scripts/eval_output_shaper.py +0 -235
  58. package/scripts/export_kompress_v2_onnx.py +0 -267
  59. package/scripts/fixtures/anthropic_replay_body.json +0 -133
  60. package/scripts/fixtures/codex_response_create_frame.json +0 -52
  61. package/scripts/install-git-hooks.sh +0 -104
  62. package/scripts/install.ps1 +0 -1809
  63. package/scripts/install.sh +0 -1673
  64. package/scripts/pr-governance.py +0 -297
  65. package/scripts/record_fixtures.py +0 -59
  66. package/scripts/refresh_model_limits.sh +0 -48
  67. package/scripts/replay_codex_ws_load.py +0 -445
  68. package/scripts/repro_codex_replay.py +0 -699
  69. package/scripts/smoke_issue_327.py +0 -233
  70. package/scripts/sync-plugin-versions.py +0 -116
  71. package/scripts/tests/test_changelog_gen.py +0 -302
  72. package/scripts/tests/test_pr_governance.py +0 -163
  73. package/scripts/tests/test_pr_health_labels.py +0 -94
  74. package/scripts/tests/test_pr_health_workflow.py +0 -14
  75. package/scripts/tests/test_sync_plugin_versions.py +0 -131
  76. package/scripts/tests/test_version_sync.py +0 -291
  77. package/scripts/validate-workflows.sh +0 -39
  78. package/scripts/verify-versions.py +0 -106
  79. package/scripts/version-sync.py +0 -207
  80. package/tailwind.config.js +0 -31
  81. package/tailwind.master.config.cjs +0 -47
  82. package/tsconfig.json +0 -84
  83. package/types/api.ts +0 -12
  84. package/types/config.ts +0 -9
  85. package/types/index.ts +0 -2
@@ -1,235 +0,0 @@
1
- """Live before/after eval for the output shaper.
2
-
3
- Sends the SAME request to the Anthropic API twice — once as a client would
4
- send it (baseline) and once after `shape_request` rewrites it (exactly what
5
- the proxy forwards upstream) — and compares `usage.output_tokens`, which
6
- includes thinking tokens.
7
-
8
- Scenario A (verbosity steering): a complex code-review ask. Baseline vs
9
- verbosity levels 2 and 3.
10
-
11
- Scenario B (effort routing): an agentic transcript whose last message is a
12
- clean tool_result (mechanical continuation) with `output_config.effort` set
13
- to "xhigh" the way Claude Code pins it. The shaper lowers effort to "low"
14
- for this turn only.
15
-
16
- Usage:
17
- source .venv/bin/activate && python scripts/eval_output_shaper.py
18
- Requires ANTHROPIC_API_KEY in the environment or in ./.env.
19
- """
20
-
21
- from __future__ import annotations
22
-
23
- import copy
24
- import os
25
- import statistics
26
- import sys
27
- from pathlib import Path
28
- from typing import Any
29
-
30
- sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
31
-
32
- import anthropic # noqa: E402
33
-
34
- from headroom.proxy.output_shaper import OutputShaperSettings, shape_request # noqa: E402
35
-
36
- MODEL = "claude-opus-4-8"
37
- TRIALS = 2
38
-
39
- BUGGY_CODE = '''\
40
- import threading
41
- from collections import OrderedDict
42
-
43
- class TTLCache:
44
- """LRU cache with per-entry TTL."""
45
-
46
- def __init__(self, max_size=128, ttl=300):
47
- self.max_size = max_size
48
- self.ttl = ttl
49
- self._store = OrderedDict()
50
- self._lock = threading.Lock()
51
-
52
- def get(self, key, now):
53
- entry = self._store.get(key)
54
- if entry is None:
55
- return None
56
- value, expires_at = entry
57
- if now > expires_at:
58
- del self._store[key]
59
- return None
60
- self._store.move_to_end(key)
61
- return value
62
-
63
- def put(self, key, value, now):
64
- with self._lock:
65
- if key in self._store:
66
- self._store.move_to_end(key)
67
- self._store[key] = (value, now + self.ttl)
68
- if len(self._store) > self.max_size:
69
- self._store.popitem(last=True)
70
-
71
- def cleanup(self, now):
72
- for key, (_, expires_at) in self._store.items():
73
- if now > expires_at:
74
- del self._store[key]
75
- '''
76
-
77
-
78
- def load_env() -> None:
79
- env_path = Path(__file__).resolve().parent.parent / ".env"
80
- if not env_path.exists() or os.environ.get("ANTHROPIC_API_KEY"):
81
- return
82
- for line in env_path.read_text().splitlines():
83
- line = line.strip()
84
- if line and not line.startswith("#") and "=" in line:
85
- key, _, value = line.partition("=")
86
- value = value.strip().strip("'\"")
87
- os.environ.setdefault(key.strip(), value)
88
-
89
-
90
- def scenario_a_body() -> dict[str, Any]:
91
- """Complex single-turn ask — exercises verbosity steering."""
92
- return {
93
- "model": MODEL,
94
- "max_tokens": 8000,
95
- "system": "You are a senior Python engineer doing code review.",
96
- "messages": [
97
- {
98
- "role": "user",
99
- "content": (
100
- "Review this cache implementation. Identify every bug and "
101
- "thread-safety issue, then show how to fix each one:\n\n"
102
- f"```python\n{BUGGY_CODE}```"
103
- ),
104
- }
105
- ],
106
- }
107
-
108
-
109
- def scenario_b_body() -> dict[str, Any]:
110
- """Agentic mechanical continuation — exercises effort routing."""
111
- return {
112
- "model": MODEL,
113
- "max_tokens": 8000,
114
- "thinking": {"type": "adaptive"},
115
- "output_config": {"effort": "xhigh"},
116
- "system": (
117
- "You are a coding agent. Use the Read tool to inspect files, then "
118
- "report findings concisely."
119
- ),
120
- "tools": [
121
- {
122
- "name": "Read",
123
- "description": "Read a file from the repository.",
124
- "input_schema": {
125
- "type": "object",
126
- "properties": {"path": {"type": "string"}},
127
- "required": ["path"],
128
- },
129
- }
130
- ],
131
- "messages": [
132
- {
133
- "role": "user",
134
- "content": "Check whether cache.py has thread-safety issues.",
135
- },
136
- {
137
- "role": "assistant",
138
- "content": [
139
- {"type": "text", "text": "Reading cache.py first."},
140
- {
141
- "type": "tool_use",
142
- "id": "toolu_eval_01",
143
- "name": "Read",
144
- "input": {"path": "cache.py"},
145
- },
146
- ],
147
- },
148
- {
149
- "role": "user",
150
- "content": [
151
- {
152
- "type": "tool_result",
153
- "tool_use_id": "toolu_eval_01",
154
- "content": BUGGY_CODE,
155
- }
156
- ],
157
- },
158
- ],
159
- }
160
-
161
-
162
- def run(client: anthropic.Anthropic, body: dict[str, Any]) -> dict[str, int]:
163
- # The installed SDK may predate output_config as a typed kwarg; the API
164
- # accepts it either way, so pass it through extra_body.
165
- body = dict(body)
166
- extra_body = None
167
- if "output_config" in body:
168
- extra_body = {"output_config": body.pop("output_config")}
169
- response = client.messages.create(**body, extra_body=extra_body)
170
- if response.stop_reason == "refusal":
171
- raise RuntimeError("request was refused by safety classifiers")
172
- return {
173
- "input_tokens": response.usage.input_tokens,
174
- "output_tokens": response.usage.output_tokens,
175
- }
176
-
177
-
178
- def main() -> int:
179
- load_env()
180
- if not os.environ.get("ANTHROPIC_API_KEY"):
181
- print("ANTHROPIC_API_KEY not found (env or .env)", file=sys.stderr)
182
- return 1
183
- client = anthropic.Anthropic()
184
- which = sys.argv[1].upper() if len(sys.argv) > 1 else "ALL"
185
-
186
- conditions: list[tuple[str, str, dict[str, Any]]] = []
187
-
188
- if which in ("A", "ALL"):
189
- # Scenario A: baseline vs steered.
190
- conditions.append(("A:verbosity", "baseline", scenario_a_body()))
191
- for level in (2, 3):
192
- body = scenario_a_body()
193
- shape_request(body, OutputShaperSettings(enabled=True, verbosity_level=level))
194
- conditions.append(("A:verbosity", f"shaped L{level}", body))
195
-
196
- if which in ("B", "ALL"):
197
- # Scenario B: baseline (effort=xhigh) vs shaped (effort routed to low).
198
- conditions.append(("B:effort-routing", "baseline xhigh", scenario_b_body()))
199
- body = scenario_b_body()
200
- result = shape_request(body, OutputShaperSettings(enabled=True, verbosity_level=0))
201
- assert body["output_config"]["effort"] == "low", result.labels
202
- conditions.append(("B:effort-routing", "shaped low", body))
203
-
204
- print(f"model={MODEL} trials={TRIALS}\n")
205
- print(f"{'scenario':<18} {'condition':<16} {'trial':<6} {'in_tok':>7} {'out_tok':>8}")
206
- print("-" * 60)
207
-
208
- results: dict[tuple[str, str], list[int]] = {}
209
- for scenario, condition, body in conditions:
210
- for trial in range(1, TRIALS + 1):
211
- usage = run(client, copy.deepcopy(body))
212
- results.setdefault((scenario, condition), []).append(usage["output_tokens"])
213
- print(
214
- f"{scenario:<18} {condition:<16} {trial:<6} "
215
- f"{usage['input_tokens']:>7} {usage['output_tokens']:>8}"
216
- )
217
-
218
- print("\n=== Summary (mean output tokens, reduction vs baseline) ===")
219
- baselines: dict[str, float] = {}
220
- for (scenario, condition), outs in results.items():
221
- if condition.startswith("baseline"):
222
- baselines[scenario] = statistics.mean(outs)
223
- for (scenario, condition), outs in results.items():
224
- mean = statistics.mean(outs)
225
- base = baselines.get(scenario, 0)
226
- if condition.startswith("baseline") or not base:
227
- print(f"{scenario:<18} {condition:<16} {mean:>8.0f} (baseline)")
228
- else:
229
- pct = (base - mean) / base * 100
230
- print(f"{scenario:<18} {condition:<16} {mean:>8.0f} ({pct:+.1f}% vs baseline)")
231
- return 0
232
-
233
-
234
- if __name__ == "__main__":
235
- sys.exit(main())
@@ -1,267 +0,0 @@
1
- #!/usr/bin/env python
2
- """Export a Kompress PyTorch checkpoint to ONNX INT8 for Headroom's light path.
3
-
4
- Why this exists
5
- ---------------
6
- Headroom's ``[proxy]`` extra ships ``onnxruntime`` but **not** torch — the
7
- proxy runs Kompress text compression on ONNX Runtime alone. The loader
8
- (``headroom/transforms/kompress_compressor.py``) downloads
9
- ``onnx/kompress-int8.onnx`` from the model repo and runs it through
10
- ``_OnnxModel``, which expects a single graph output named ``final_scores``
11
- (per-token importance in ``[0, 1]``, kept when ``> 0.5``).
12
-
13
- ``chopratejas/kompress-v2-base`` ships only PyTorch weights
14
- (``model.safetensors`` / ``merged.pt``) — no ONNX. So pointing Headroom at v2
15
- without an ONNX export would silently force the heavier ``[ml]`` (torch) path
16
- on every proxy install. This script reproduces v1's exact ONNX contract from
17
- the v2 PyTorch checkpoint, so a default swap stays zero-cost for light installs.
18
-
19
- The model is a *custom* dual-head ModernBERT (token classifier + span CNN), not
20
- a standard HF architecture, so ``optimum-cli export onnx`` does not apply — we
21
- trace the real module from ``kompress_compressor._get_model_class()``.
22
-
23
- Requires
24
- --------
25
- pip install xtream-head[ml] onnxruntime # torch + transformers + onnxruntime
26
-
27
- Usage
28
- -----
29
- # Convert + verify locally (writes onnx/kompress-int8.onnx):
30
- python scripts/export_kompress_v2_onnx.py --model-id chopratejas/kompress-v2-base
31
-
32
- # Convert, verify, and upload back to the HF repo (needs `huggingface-cli login`):
33
- python scripts/export_kompress_v2_onnx.py --model-id chopratejas/kompress-v2-base --upload
34
- """
35
-
36
- from __future__ import annotations
37
-
38
- import argparse
39
- import logging
40
- import sys
41
- from pathlib import Path
42
-
43
- logging.basicConfig(level=logging.INFO, format="%(levelname)s %(message)s")
44
- logger = logging.getLogger("export_kompress_v2_onnx")
45
-
46
- # ModernBERT encoder + tokenizer base (must match training and the loader).
47
- BASE_MODEL = "answerdotai/ModernBERT-base"
48
- DEFAULT_MODEL_ID = "chopratejas/kompress-v2-base"
49
-
50
-
51
- def _build_core(model_id: str):
52
- """Instantiate HeadroomCompressorModel and load the merged v2 weights.
53
-
54
- The v2 repo's ``model.safetensors`` is the *unmerged* PEFT structure
55
- (``encoder.base_model.model...`` with separate ``base_layer`` + LoRA
56
- adapters), which does not map onto ``HeadroomCompressorModel``. The
57
- canonical artifact is ``merged.pt`` — a structured checkpoint with already
58
- LoRA-merged sub-state-dicts:
59
-
60
- {"encoder_state_dict", "token_head_state_dict",
61
- "span_conv_state_dict", "config", "checkpoint_kind"}
62
-
63
- Each loads cleanly (0 missing / 0 unexpected) into the encoder + heads.
64
- """
65
- import torch
66
- from huggingface_hub import hf_hub_download
67
-
68
- from headroom.transforms.kompress_compressor import _get_model_class
69
-
70
- ckpt_path = hf_hub_download(model_id, "merged.pt")
71
- ckpt = torch.load(ckpt_path, map_location="cpu")
72
- for key in ("encoder_state_dict", "token_head_state_dict", "span_conv_state_dict"):
73
- if key not in ckpt:
74
- raise RuntimeError(
75
- f"merged.pt missing '{key}'. Found: {sorted(ckpt)}. "
76
- "This script targets the v2 'merged' checkpoint format."
77
- )
78
-
79
- core = _get_model_class()(model_name=BASE_MODEL)
80
-
81
- def _strict_load(module, sd, label: str) -> None:
82
- missing, unexpected = module.load_state_dict(sd, strict=False)
83
- if missing or unexpected:
84
- raise RuntimeError(
85
- f"{label}: state_dict mismatch (missing={list(missing)[:5]}, "
86
- f"unexpected={list(unexpected)[:5]}). Architecture drifted from the checkpoint."
87
- )
88
- logger.info(" %s loaded (%d tensors, exact match)", label, len(sd))
89
-
90
- logger.info("Loading merged.pt (checkpoint_kind=%s)", ckpt.get("checkpoint_kind"))
91
- _strict_load(core.encoder, ckpt["encoder_state_dict"], "encoder")
92
- _strict_load(core.token_head, ckpt["token_head_state_dict"], "token_head")
93
- _strict_load(core.span_conv, ckpt["span_conv_state_dict"], "span_conv")
94
-
95
- core.eval()
96
- return core
97
-
98
-
99
- def _export_wrapper(core):
100
- """Wrap the dual head so forward() returns `final_scores` (== get_scores)."""
101
- import torch
102
- import torch.nn as nn
103
-
104
- class ExportWrapper(nn.Module):
105
- def __init__(self, inner):
106
- super().__init__()
107
- self.inner = inner
108
-
109
- def forward(self, input_ids, attention_mask): # noqa: ANN001
110
- hidden = self.inner.encoder(input_ids, attention_mask=attention_mask).last_hidden_state
111
- token_probs = torch.softmax(self.inner.token_head(hidden), dim=-1)[:, :, 1]
112
- span_scores = self.inner.span_conv(hidden.transpose(1, 2)).squeeze(1)
113
- return token_probs * (0.5 + 0.5 * span_scores)
114
-
115
- return ExportWrapper(core).eval()
116
-
117
-
118
- def export(model_id: str, out_path: Path, opset: int, precision: str) -> None:
119
- import numpy as np
120
- import torch
121
-
122
- core = _build_core(model_id)
123
- wrapper = _export_wrapper(core)
124
-
125
- out_path.parent.mkdir(parents=True, exist_ok=True)
126
-
127
- # fp32 path: trace straight to the final artifact (lossless — verified 100%
128
- # keep-decision agreement with PyTorch). int8 path: trace to a temp fp32
129
- # graph, then dynamically quantize into the final artifact.
130
- trace_target = out_path if precision == "fp32" else out_path.with_name("kompress-fp32-tmp.onnx")
131
-
132
- dummy_ids = torch.randint(0, 1000, (1, 64), dtype=torch.long)
133
- dummy_mask = torch.ones((1, 64), dtype=torch.long)
134
-
135
- logger.info("Tracing → ONNX (opset %d, precision=%s) ...", opset, precision)
136
- with torch.no_grad():
137
- torch.onnx.export(
138
- wrapper,
139
- (dummy_ids, dummy_mask),
140
- str(trace_target),
141
- input_names=["input_ids", "attention_mask"],
142
- output_names=["final_scores"],
143
- dynamic_axes={
144
- "input_ids": {0: "batch", 1: "seq"},
145
- "attention_mask": {0: "batch", 1: "seq"},
146
- "final_scores": {0: "batch", 1: "seq"},
147
- },
148
- opset_version=opset,
149
- do_constant_folding=True,
150
- dynamo=False,
151
- )
152
-
153
- if precision == "int8":
154
- from onnxruntime.quantization import QuantType, quantize_dynamic
155
-
156
- logger.info("INT8 dynamic quantization (MatMul only) → %s", out_path)
157
- # Restrict to MatMul: the encoder's linear layers carry ~all the weight
158
- # mass and ORT's CPU provider implements MatMulInteger. Quantizing the
159
- # tiny span_conv Conv1d layers would emit ConvInteger, which ORT CPU
160
- # cannot run. per_channel recovers transformer accuracy at the 0.5 boundary.
161
- quantize_dynamic(
162
- str(trace_target),
163
- str(out_path),
164
- weight_type=QuantType.QInt8,
165
- op_types_to_quantize=["MatMul"],
166
- per_channel=True,
167
- )
168
- trace_target.unlink(missing_ok=True)
169
-
170
- _verify(model_id, core, out_path, np, torch)
171
-
172
-
173
- def _verify(model_id: str, core, out_path: Path, np, torch) -> None:
174
- """Compare ONNX scores against PyTorch get_scores on a real tokenized sample."""
175
- import onnxruntime as ort
176
- from transformers import AutoTokenizer
177
-
178
- tok = AutoTokenizer.from_pretrained(BASE_MODEL)
179
- sample = (
180
- "The proxy compresses tool outputs before they reach the model. "
181
- "Errors and stack traces should survive; boilerplate should not. "
182
- ) * 6
183
- words = sample.split()
184
- enc = tok(
185
- words,
186
- is_split_into_words=True,
187
- truncation=True,
188
- max_length=512,
189
- padding=True,
190
- return_tensors="pt",
191
- )
192
-
193
- with torch.no_grad():
194
- torch_scores = core.get_scores(enc["input_ids"], enc["attention_mask"])[0].cpu().numpy()
195
-
196
- sess = ort.InferenceSession(str(out_path), providers=["CPUExecutionProvider"])
197
- onnx_scores = sess.run(
198
- ["final_scores"],
199
- {
200
- "input_ids": enc["input_ids"].numpy().astype(np.int64),
201
- "attention_mask": enc["attention_mask"].numpy().astype(np.int64),
202
- },
203
- )[0][0]
204
-
205
- max_abs = float(np.max(np.abs(torch_scores - onnx_scores)))
206
- keep_torch = torch_scores > 0.5
207
- keep_onnx = onnx_scores > 0.5
208
- agree = float((keep_torch == keep_onnx).mean())
209
- logger.info(
210
- "Verify: max|Δscore|=%.4f keep-decision agreement=%.1f%% (fp32 ~100%%, int8 ~98-100%%)",
211
- max_abs,
212
- agree * 100,
213
- )
214
- if agree < 0.98:
215
- logger.warning(
216
- "Keep-decision agreement below 98%% — for fp32 this means a tracing "
217
- "problem; for int8 consider per_channel/fp32. Inspect before publishing."
218
- )
219
-
220
-
221
- def upload(model_id: str, out_path: Path) -> None:
222
- from huggingface_hub import upload_file
223
-
224
- # Publish under onnx/<artifact filename> so int8 and fp32 can coexist.
225
- repo_path = f"onnx/{out_path.name}"
226
- logger.info("Uploading %s → %s:%s", out_path, model_id, repo_path)
227
- upload_file(
228
- path_or_fileobj=str(out_path),
229
- path_in_repo=repo_path,
230
- repo_id=model_id,
231
- commit_message="Add ONNX export for Headroom lightweight (no-torch) path",
232
- )
233
- logger.info("Uploaded. Headroom's ONNX loader will now find it on next cold start.")
234
-
235
-
236
- def main() -> int:
237
- ap = argparse.ArgumentParser(description=__doc__)
238
- ap.add_argument("--model-id", default=DEFAULT_MODEL_ID)
239
- ap.add_argument(
240
- "--precision",
241
- choices=["fp32", "int8"],
242
- default="fp32",
243
- help="fp32 = lossless, larger artifact. int8 = ~2x smaller, tiny accuracy cost.",
244
- )
245
- ap.add_argument(
246
- "--out",
247
- type=Path,
248
- default=None,
249
- help="Local output path. Defaults to onnx/kompress-<precision>.onnx.",
250
- )
251
- ap.add_argument("--opset", type=int, default=17)
252
- ap.add_argument(
253
- "--upload",
254
- action="store_true",
255
- help="Upload to the HF repo under onnx/<filename> (needs HF write auth).",
256
- )
257
- args = ap.parse_args()
258
-
259
- out_path = args.out or Path(f"onnx/kompress-{args.precision}.onnx")
260
- export(args.model_id, out_path, args.opset, args.precision)
261
- if args.upload:
262
- upload(args.model_id, out_path)
263
- return 0
264
-
265
-
266
- if __name__ == "__main__":
267
- sys.exit(main())
@@ -1,133 +0,0 @@
1
- {
2
- "model": "claude-opus-4-7",
3
- "max_tokens": 4096,
4
- "stream": true,
5
- "system": "You are Claude Code, Anthropic's official CLI for disciplined software engineering. You are helping the operator triage a resilience bug in a Python/FastAPI LLM proxy that handles both Anthropic /v1/messages traffic and OpenAI Codex /v1/responses WebSocket traffic on the same process.\n\nWorking style: Read code with symbolic tools when available; prefer small, reviewable diffs; never bypass compression or add fast paths that skip transforms; keep /livez trivial and IO-free; treat the WebSocket relay lifecycle as load-bearing. When you are uncertain about upstream behavior, instrument first and mitigate second. Multi-agent reconnect storms arrive as a burst of /v1/messages?beta=true POSTs immediately after process restart; these are the failure mode under investigation.\n\nInvariants: compression stays on; /livez must respond under 100ms even during cold-start replay storms; memory-context lookups are wrapped in wait_for with a bounded timeout; upstream WS handshake has retry/open-timeout hardening; debug endpoints are loopback-only.",
6
- "tools": [
7
- {
8
- "name": "read_file",
9
- "description": "Read a file from the local filesystem. Returns file contents as a string. Use when you need to inspect exact text of a source file before editing.",
10
- "input_schema": {
11
- "type": "object",
12
- "properties": {
13
- "path": {"type": "string", "description": "Absolute file path."},
14
- "start_line": {"type": "integer", "description": "Optional 1-indexed start line."},
15
- "end_line": {"type": "integer", "description": "Optional inclusive end line."}
16
- },
17
- "required": ["path"]
18
- }
19
- },
20
- {
21
- "name": "search_code",
22
- "description": "Search the codebase for a regex pattern. Returns matching file paths and surrounding lines. Prefer over read_file when you don't know the file location yet.",
23
- "input_schema": {
24
- "type": "object",
25
- "properties": {
26
- "pattern": {"type": "string"},
27
- "glob": {"type": "string", "description": "Optional glob filter, e.g. '**/*.py'."},
28
- "max_results": {"type": "integer", "default": 50}
29
- },
30
- "required": ["pattern"]
31
- }
32
- },
33
- {
34
- "name": "run_tests",
35
- "description": "Run the pytest suite or a subset. Returns summary output including passed/failed/errored counts and tracebacks for failures.",
36
- "input_schema": {
37
- "type": "object",
38
- "properties": {
39
- "expression": {"type": "string", "description": "pytest -k expression."},
40
- "paths": {"type": "array", "items": {"type": "string"}, "description": "Specific test files or directories."}
41
- }
42
- }
43
- },
44
- {
45
- "name": "edit_file",
46
- "description": "Apply an exact-string replacement to a file. The old_string must be unique within the file; otherwise use replace_all. Prefer over write_file for modifications.",
47
- "input_schema": {
48
- "type": "object",
49
- "properties": {
50
- "path": {"type": "string"},
51
- "old_string": {"type": "string"},
52
- "new_string": {"type": "string"},
53
- "replace_all": {"type": "boolean", "default": false}
54
- },
55
- "required": ["path", "old_string", "new_string"]
56
- }
57
- },
58
- {
59
- "name": "run_command",
60
- "description": "Execute a shell command and return stdout, stderr, and exit code. Use for git, build, and verification commands.",
61
- "input_schema": {
62
- "type": "object",
63
- "properties": {
64
- "command": {"type": "string"},
65
- "timeout_seconds": {"type": "integer", "default": 120}
66
- },
67
- "required": ["command"]
68
- }
69
- }
70
- ],
71
- "messages": [
72
- {
73
- "role": "user",
74
- "content": [
75
- {
76
- "type": "text",
77
- "text": "The shared proxy on 127.0.0.1:8787 started acting like a zombie this morning after a restart. New /livez probes time out, new /v1/responses WebSocket handshakes hang for ~20 seconds before the client gives up, but the existing agent sessions that were already streaming when the restart happened keep working fine. Two of us were connected at the time so there was a reconnect burst of Anthropic /v1/messages?beta=true POSTs carrying multi-turn history (~30 KB each). I suspect the replay storm is interacting badly with cold-start ONNX/embedder loads. Start by reading headroom/proxy/handlers/anthropic.py:handle_anthropic_messages and headroom/proxy/server.py:startup — I want to know whether the Anthropic pre-upstream stages (deep_copy, first compression pass, memory-context lookup) run before or after the Kompress/Magika preload settles, and whether there is any concurrency limit in front of that path. Don't propose a fix yet; just map the stages and tell me where the event loop would plausibly starve."
78
- }
79
- ]
80
- },
81
- {
82
- "role": "assistant",
83
- "content": [
84
- {
85
- "type": "text",
86
- "text": "I've read both files. Mapping the pre-upstream stages on the Anthropic path: (1) handle_anthropic_messages reads the raw request bytes, (2) _read_request_json decodes and json.loads — this is synchronous and runs on the event loop thread, (3) a deep copy of the body is taken for logging, (4) the first compression stage fires through anthropic_pipeline.transforms, which on cold start is the moment Kompress ONNX and Magika get lazy-loaded, (5) memory-context lookup runs via MemoryHandler.search_and_format_context inside asyncio.wait_for, (6) only then does the handler reach http_client.stream() for upstream.\n\nStages 3 and 4 are the choke: deep_copy is synchronous Python on 30KB bodies (fine individually, but N concurrent replays mean N deep copies on one event loop), and the first compression stage blocks on Kompress load when the eager preload hasn't finished yet. There is currently no asyncio.Semaphore in front of stages 1-5. /livez does not go through this path, but the event loop is shared — if stages 3-5 hog it long enough, /livez's response coroutine doesn't get scheduled until they yield.\n\nI have not instrumented timings yet. Do you want me to add a StageTimer pass first, or jump straight to proposing a pre-upstream semaphore?"
87
- }
88
- ]
89
- },
90
- {
91
- "role": "user",
92
- "content": [
93
- {
94
- "type": "text",
95
- "text": "Instrument first — that's the whole point of not guessing. Add per-stage timings for read_request_json, deep_copy, compression_first_stage, memory_context, upstream_connect, upstream_first_byte on the Anthropic HTTP path, and accept -> first_client_frame -> upstream_connect -> upstream_first_event -> total_session on the Codex WS path. Both should share a single StageTimer utility, log one structured line per request with request_id + session_id + all stage fields, and emit Prometheus histograms. Don't change any behavior. Once that's in, we can let the harness run and read the real numbers before touching concurrency."
96
- }
97
- ]
98
- },
99
- {
100
- "role": "assistant",
101
- "content": [
102
- {
103
- "type": "tool_use",
104
- "id": "toolu_anthropic_replay_01",
105
- "name": "search_code",
106
- "input": {
107
- "pattern": "class PrometheusMetrics",
108
- "glob": "headroom/proxy/**/*.py"
109
- }
110
- }
111
- ]
112
- },
113
- {
114
- "role": "user",
115
- "content": [
116
- {
117
- "type": "tool_result",
118
- "tool_use_id": "toolu_anthropic_replay_01",
119
- "content": "headroom/proxy/prometheus_metrics.py:41:class PrometheusMetrics:\n Holds in-process counters/gauges/histograms plus async record_* helpers. Backend-agnostic; OTLP export is deferred. Pattern: register each histogram once in __init__, update via record_request / record_stage_timings."
120
- }
121
- ]
122
- },
123
- {
124
- "role": "assistant",
125
- "content": [
126
- {
127
- "type": "text",
128
- "text": "Plan: create headroom/proxy/stage_timer.py with a StageTimer context manager exposing .measure(stage_name) as both a sync and async context manager, accumulating into a dict[str, float] keyed by stage with millisecond values. Extend PrometheusMetrics with one histogram per stage, keyed by path ('anthropic' | 'codex_ws'). Extend request_logger with a stage_timings field. Then plumb stage_timer through handle_anthropic_messages and handle_openai_responses_ws without changing any existing control flow — the timer is a pure observer. I'll land the util + unit tests first, then thread it through the two handlers in separate commits so the diff stays reviewable."
129
- }
130
- ]
131
- }
132
- ]
133
- }