Opening the context black box between agents and models

Research reports

1. Oct 2026Measurement note

   Where an Agent Turn's Tokens Go

   1 to 7%of a turn's tokens is the answer a person reads

   Tool calls with their results were 42 to 95% of a turn and thinking 4 to 53%. On Claude Code and DeepSeek Harness, the prompt that opened the next turn still held the previous turn's thinking. The shares reflect one person's use; the answer stayed a small minority in every runtime.

   3,237 turns · Claude Code, Codex, DeepSeek Harness, Pi · 30 days of the author's own logs

   Read → 中文 Script ↗ Aggregate data ↗

2. Aug 2026Pilot v2

   Context repair: how one error contaminates the rest of a conversation

   152 → 162of 162 derailed outcomes repaired (18 cases × 9 endpoints): deleting the source, then removing the subgraph

   Removing an error at its source does not take it out of the downstream context. Across 162 derailed outcomes (18 cases on 9 endpoints), deleting only the source repaired 152, recomputing downstream nodes 161, and removing the contaminated subgraph all 162. With reasoning switched off, source-only repair fell to 2 of 18.

   9 model endpoints · 1,215 captured conditions · 0 capture failures

   Read → 中文 Case study Benchmark ↗

3. Aug 2026Pilot v1 · superseded by v2

   Context Repair for LLMs, first pilot

   The first run of the benchmark on four endpoints. Kept as an archived snapshot; v2 extends it to nine endpoints and adds the reasoning ablation.

   4 model endpoints · 540 captured conditions

   Read → 中文

Collaboration

1. Since Oct 2026Collaboration

   TraceVista

   TraceVista visualizes agent workflows as maps: from the overview down to each action and its original record, with answers to questions linked to the evidence. Its reading layer builds on ThoughtDAG's zoom summaries.

   With @Superone77 and contributors · reading layer and map Q&A by Xia Chen

   Live demo ↗ GitHub ↗

How these studies are published

• Every number comes from a script in the repository, with its aggregate output next to it.
• Where a study reads personal session logs, the logs stay on that machine and only aggregate counts are published.
• Each report ends with what it does not establish.
• Reports are versioned. A newer version links back to the one it replaces.

Cite

@software{chen_thoughtdag_2026,
  author  = {Chen, Xia},
  title   = {ThoughtDAG: an instrument for legible human-AI collaboration},
  year    = {2026},
  url     = {https://github.com/chenxiachan/thoughtdag},
  license = {MIT}
}

GitHub's “Cite this repository” button reads the same entry from CITATION.cff. CITATION.cff ↗