semianalysisai/cc-traces-weka-051826

CC Traces — Weka, With Subagents, v5 only (May 18 2026)

A collection of 98 multi-turn agentic traces drawn from real production traffic against 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 AND the original sub-agent fan-out structure (Task-tool spawned sub-agents grouped into WekaSubagentEntry blocks).

With-subagents, v5-only, CC ≥ 2.1.139 variant. Companion to cc-traces-weka-no-subagents-051826, which is the same 98 traces with all WekaSubagentEntry blocks stripped.

Three filters:

1. v5 only. Every replayable request in every included session has trace_version = 5 (latest proxy schema, o200k_base tokenizer).
2. 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 without it would have unreliable grouping (legacy stretch-based fallback can mis-group concurrent same-label sub-agents) and are excluded entirely.
3. ≥ 20 main-agent turns post-strip. Sub-agent fan-out is on top of that.

• Traces: 98
• Top-level entries: 23,443 (22,822 main turns + 621 subagent groups)
• Sub-agent inner requests: 16,332
• Total individual model requests: 39,154
• Models: claude-opus-4-7, claude-haiku-4-5-20251001, claude-opus-4-6, claude-sonnet-4-6
• KV block size: 64 tokens
• Hash scope: local

Sub-agent grouping

Identical algorithm to the proxy dashboard's buildRequestRuns. Two passes over chronologically-sorted DB rows:

1. Pre-collect header-keyed groups: for every row with x-claude-code-agent-id, gather ALL rows with the same id into one group across the entire session (not just contiguous stretches). Background sub-agents that overlap with main-agent turns collapse correctly.

2. Emit chronologically:

   • subagent_label IS NULL → main turn at position
   • has agent-id, first sighting → emit the pre-collected group
   • has agent-id, already emitted → skip
   • has label but no agent-id → fallback contiguous-stretch grouping

For CC ≥ 2.1.139 all groups go through the header-keyed path; the fallback isn't exercised. That's why we require CC ≥ 2.1.139 — the fallback can mis-group concurrent same-label sub-agents.

Tokenizer caveat

in and hash_ids are in proxy tokenizer (o200k_base). Anthropic reports ~60 % of the o200k count for the same content. ISL numbers are bigger than what Anthropic would bill, but self-consistent between in and hash_ids for cache-hit simulation.

What's in each trace

Top-level: id, models, block_size: 64, hash_id_scope: "local", requests: list.

Main-agent request (type: "n" or "s")

field	type	description
t	float	seconds since trace start
type	str	n (non-streaming) or s (streaming)
model	str	target model
in	int	input tokens (proxy tokenizer)
out	int	output tokens
hash_ids	list[int]	per-trace local block IDs
api_time	float	end-to-end server seconds
think_time	float?	gap from previous request's end
ttft	float?	streaming only

Sub-agent group (type: "subagent")

field	type	description
t	float	start of the group
type	str	"subagent"
agent_id	str	stable id (slug + 8-char suffix from x-claude-code-agent-id)
subagent_type	str	"Subagent" for CC ≥ 2.1.139
duration_ms	int	first→last inner wall-clock
total_tokens	int	Σ(in+out) across inners
tool_use_count	null	not tracked
status	str	"completed"
requests	list	inner Shape-A entries (all type: "n")
models	list[str]	distinct models inside

Inner requests carry absolute t values, so flattening + chronological re-sort works directly: sorted(main + flatten(inner), key=lambda x: x['t']).

Summary statistics

Across all requests (main + sub-agent inner = 39,154):

	p50	p75	p90	p95	p99	mean
ISL (tokens)	103,026	231,708	440,033	573,767	799,819	173,073
OSL (tokens)	223	476	1,152	1,907	5,413	527
Top entries/trace	126	304	571	733	1,120	239

Plots

Main-agent stream

Histograms across all 22,822 main-agent turns (sub-agent groups skipped). Identical to the no-subagents sibling.

Sub-agent fan-out analysis

Histograms across all 621 sub-agent groups and their 16,332 inner requests.

Six panels: groups per trace · inner requests per group · group wall-clock duration · group total tokens · inner-request ISL · intra-group cache hit rate.

Model composition

model	requests (main + subagent inner)
claude-opus-4-7	38,058
claude-haiku-4-5-20251001	702
claude-opus-4-6	343
claude-sonnet-4-6	51

Source

Same proxy → weka pipeline as the no-subagents sibling, without the post-step that strips WekaSubagentEntry blocks.

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 (subagent grouping per proxy lib/subagent-runs.ts)
3. Concatenate the resulting weka JSONs into traces.jsonl — one trace per line.

The same-98-trace contract with the no-subagents sibling makes the two variants safe to A/B compare for any benchmark that cares about sub-agent fan-out's impact on cache-hit-rate or throughput.