semianalysisai/cc-traces-weka-with-subagents-052726-256k

CC Traces — Weka, With Subagents, in+out ≤ 256k (May 27 2026, deduped)

A 256k-context simulation corpus derived from cc-traces-weka-with-subagents-052726. Every individual request (main-agent and sub-agent inner) where in + out > 256,000 proxy-tokenizer tokens has been dropped. After filtering, each trace's request timeline is reshifted using each surviving entry's think_time so the gaps created by removed requests collapse cleanly.

Build note (2026-05-27 rebuild): utils/proxy_to_weka.py now drops exact-duplicate rows at conversion time (proxy-side recording artifact, ~2% of raw rows). The published corpus reflects deduped counts.

Filters (on top of 052726's filters)

1. 1.Per-request in + out ≤ 256,000 tokens. Applied at request granularity, not conversation. Main-agent and sub-agent inner requests are evaluated independently. Sub-agent groups whose every inner request was filtered are dropped entirely; partial groups keep the surviving inners.

Timeline reshift policy

After filtering, each trace's request t values are recomputed:

• —First surviving entry anchored at t = 0
• —Each subsequent: new_t = prev_t + prev_api_time + this.think_time
• —Subagent inner reshift runs identically over surviving inners; group t = first surviving inner's new t; group duration_ms recomputed from the new inner span

Filter impact (vs 052726 source, both deduped)

Summary statistics

Across all surviving requests (main + sub-agent inner = 77,858):

Plots

Main-agent stream

[image] [image]

Sub-agent fan-out analysis

[image] [image]

Standalone: sub-agent invocations per session

[image]

Model composition