filter to CC >=2.1.139 (header-keyed grouping); +2 traces (98 total)

README.md

@@ -22,94 +22,54 @@ configs:
 
 # CC Traces — Weka, No-Subagents, v5 only (May 18 2026)
 
-A collection of **96 multi-turn agentic traces** (≈ 21.6 k individual
+A collection of **98 multi-turn agentic traces** (≈ 22.8 k individual
 model requests) drawn from real production traffic against the Claude
 Code CLI ≥ 2.1.139. Each trace captures the full request/response
 sequence of a single agent session, including per-request KV block
 hashes, so the dataset can be replayed against an inference engine or
 used to simulate prefix-cache behavior offline.
 
-**No-subagents, v5-only variant.** This dataset is a derivative of the
-source weka traces with two filters applied:
+**No-subagents, v5-only, CC ≥ 2.1.139 variant.** Three filters:
 
 1. **No subagents.** All `WekaSubagentEntry` blocks have been stripped —
    only top-level main-agent turns remain.
 2. **v5 only.** Every replayable request in every included session has
-   `trace_version = 5` (the latest proxy schema). Sessions containing
-   any earlier-format request are excluded entirely.
-
-The combined filter is stricter than the
-[cc-traces-weka-no-subagents-051226](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-no-subagents-051226)
-variant: that one mixes v1–v4 traces (older qwen3-tokenizer rows)
-alongside v5. Use **this dataset** when the token-count accounting
-needs to be uniform across the corpus; use 051226 when you want a
-larger pool with mixed tokenizers.
-
-Also filtered: only sessions whose post-strip main-agent stream has
-**≥ 20 turns**. Sessions with shorter streams are excluded so every
-trace exercises the multi-turn replay path meaningfully.
-
-- **Traces:** 96
-- **Requests:** 21,566 total, mean **224.6** per trace, max **2,171**
+   `trace_version = 5` (latest proxy schema, o200k_base tokenizer).
+3. **Claude Code CLI ≥ 2.1.139.** Every successful request in every
+   included session is on a CC build that ships the
+   `x-claude-code-agent-id` HTTP header — the canonical signal sub-agent
+   grouping depends on. Sessions where any row predates this header
+   are excluded entirely because we can't reliably distinguish a
+   Task-tool sub-agent from a utility call without it.
+
+Each remaining trace has **≥ 20 main-agent turns** post-strip so every
+trace exercises multi-turn replay meaningfully.
+
+Companion: [`cc-traces-weka-051826`](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-051826)
+is the same 98 traces with sub-agent fan-out preserved.
+
+- **Traces:** 98
+- **Requests:** 22,822 total, mean **232.9** per trace, max **2,171**
 - **Models:** `claude-opus-4-7`, `claude-opus-4-6`, `claude-sonnet-4-6`
 - **KV block size:** 64 tokens
-- **Hash scope:** `local` — block hash IDs are only comparable *within*
-  a single trace; they are not a global content-addressable identity.
-
-## Important: tokenizer caveat
-
-The `in` field on each request and the `hash_ids` array are both
-measured in **the proxy's tokenizer**. v5 rows use `o200k_base`
-(GPT-4o family) for byte-pair encoding. Anthropic typically reports
-~60 % of the o200k token count for the same content. So the ISL
-numbers below are larger than what the Anthropic API would have
-billed for the same prompt — but they're self-consistent between
-`in` and `hash_ids`, which is what matters for KV-cache replay
-simulation.
-
-If you replay these traces against a real Claude (or Claude-like)
-server, the server will re-tokenize the text and see ~Claude-equivalent
-token counts, so cache-hit-rate measurements remain accurate.
-
-## What's in each trace
-
-Top-level trace fields:
-
-| field           | type             | description                                         |
-|-----------------|------------------|-----------------------------------------------------|
-| `id`            | str              | Trace identifier (the proxy session id)             |
-| `models`        | list[str]        | Models used by the trace                            |
-| `block_size`    | int              | KV block size used to derive `hash_ids` (64)        |
-| `hash_id_scope` | str              | `local` — hash IDs are per-trace, not global        |
-| `requests`      | list[object]     | Ordered list of per-request records (main-agent only) |
-
-Each entry in `requests` is:
-
-| field         | type      | description                                                              |
-|---------------|-----------|--------------------------------------------------------------------------|
-| `t`           | float     | Seconds since start of trace                                             |
-| `type`        | str       | `n` (non-streaming) or `s` (streaming)                                   |
-| `model`       | str       | Target model for this request                                            |
-| `in`          | int       | Effective prompt tokens covered by `hash_ids` (proxy tokenizer)          |
-| `out`         | int       | Output tokens (Anthropic-reported)                                       |
-| `hash_ids`    | list[int] | Per-trace local block-hash IDs for the input, one per 64-token block     |
-| `api_time`    | float     | End-to-end server time for this call (seconds)                           |
-| `think_time`  | float     | Wall-clock gap from previous request's end (seconds)                     |
-| `ttft`        | float?    | Time to first token (streaming requests only)                            |
-
-`hash_ids` make the dataset unusually useful for KV-cache work: the
-contiguous common prefix between turn *t* and turn *t − 1* exactly
-measures the portion of the input that a local prefix cache would be
-able to reuse.
+- **Hash scope:** `local`
+
+## Tokenizer caveat
+
+The `in` field and `hash_ids` are measured in the proxy's tokenizer
+(`o200k_base`, GPT-4o family). Anthropic typically reports ~60 % of the
+o200k token count for the same content — so ISL numbers below are
+larger than what Anthropic would bill for the same prompt, but
+self-consistent between `in` and `hash_ids` for cache-hit simulation.
 
 ## Summary statistics
 
 |                              |   p50    |   p75    |   p90    |   p95    |   p99    |   mean   |
 |------------------------------|---------:|---------:|---------:|---------:|---------:|---------:|
-| ISL (tokens)                 |  182,115 |  345,997 |  537,070 |  675,840 |  840,000 |  243,765 |
-| OSL (tokens)                 |      216 |      576 |    1,480 |    2,555 |    6,653 |      633 |
-| `think_time` (s)             |     1.00 |     2.75 |    12.25 |    72.31 |   472.53 |    96.41 |
-| Turn depth (requests/trace)  |       96 |      300 |      568 |      676 |    1,140 |      225 |
+| ISL (tokens)                 |  181,474 |  338,879 |  525,701 |  664,749 |  838,501 |  240,502 |
+| OSL (tokens)                 |      212 |      569 |    1,455 |    2,499 |    6,528 |      623 |
+| `think_time` (s)             |     0.98 |     2.71 |    12.45 |    71.64 |   448.24 |    92.32 |
+| Turn depth (requests/trace)  |      110 |      300 |      568 |      730 |    1,119 |      233 |
 
 ![Distributions — log x](plots/distributions_log.png)
 ![Distributions — linear x](plots/distributions_linear.png)
@@ -118,20 +78,18 @@ able to reuse.
 
 | model | requests |
 |-------|---------:|
-| `claude-opus-4-7` | 21,195 |
+| `claude-opus-4-7` | 22,484 |
 | `claude-opus-4-6` | 302 |
-| `claude-sonnet-4-6` | 69 |
+| `claude-sonnet-4-6` | 36 |
 
 ## Source
 
 Derived from the SemiAnalysis Claude Code traffic proxy. Build pipeline:
 
-  1. `utils/sample_proxy_traces.py --min-trace-version 5 --min-main-turns 20 --privacy-mode anon`
-     pulls v5-only Anthropic sessions out of the proxy DB.
+  1. `utils/sample_proxy_traces.py --min-trace-version 5 --min-main-turns 20 --require-cli-min 2.1.139 --privacy-mode anon`
   2. `utils/proxy_to_weka.py` converts per-session JSONL into weka-format
-     trace JSON (with subagent grouping per the proxy dashboard's
-     algorithm).
-  3. A post-step strips all `WekaSubagentEntry` blocks and drops any
-     trace with fewer than 20 surviving main-agent turns.
-  4. The remaining traces are concatenated into `traces.jsonl`
-     (one trace per line).
+     trace JSON. Sub-agent grouping is identical to the
+     [proxy dashboard's `buildRequestRuns`](https://github.com/SemiAnalysisAI/semianalysis-claude-code-proxy/blob/main/packages/app/src/lib/subagent-runs.ts).
+  3. Post-step strips all `WekaSubagentEntry` blocks; traces with <20
+     surviving main turns are dropped (none dropped at this build).
+  4. Remaining traces concatenated into `traces.jsonl` (one trace per line).

traces.jsonl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fe0ae6f7b987261ea05af0f328a1213a04cfdf3bf0ac7bc2808c5f0af1685178
-size 516326432
+oid sha256:b2a7698312e97f52e5eaa01d9ebd7dafcd17fc8e7d1fa55c9a3aa30749a440ec
+size 538929960