Research documentation on how IDE-embedded agentic assistants actually reason — process, control flow, tool selection, and observability — plus practical notes for operating agent runtimes economically.
Part of the AIML Solutions public work; see also the security-first flagship triage-mesh and the evaluation harness multiclaw-harness.
| File | What it is |
|---|---|
agentic-reasoning-framework.md |
The primary document: a dense, arXiv-style treatment (abstract, 10 sections, appendices) of agentic reasoning in IDE assistants — request interpretation, decomposition, tool invocation (search/read/edit/shell/MCP), iteration, and trace capture. |
AGENTIC-THOUGHT-PROCESS.md |
Condensed summary: the framework, decision points, and capture methods. |
session-reasoning-template.md |
Per-session capture template: user request → reasoning trace → tool sequence → outcome. |
OPENCLAW_FREE_TIER.md |
Operating notes: keeping agent-runtime token usage disciplined (batching, in-repo summaries, CI offload). |
build-pdf.sh · md-to-print-html.py · print.css · pdf-header.tex |
Toolchain to render the paper as a print-quality PDF (Pandoc+LaTeX, or pure-Python HTML route). |
./build-pdf.sh # Pandoc + LaTeX (best quality)
python3 md-to-print-html.py # no-LaTeX route: HTML → browser → print to PDFEvaluating and operating agents well requires a precise model of how they reason. This repo is the written-down model — the same understanding that underpins the trajectory assertions in multiclaw-harness and the harness-enforced boundaries in triage-mesh.