WeaverTools

The suite-level repository of WeaverTools: what belongs to the suite as a whole rather than to one of its code repositories. WeaverTools is a local-first agent framework whose primary artifact is the trace, split on 2026-09-30 into repositories along its network boundaries, one per party a consumer meets across a socket:

Repository	What it holds
WeaverAgent	The agent: harness, SPU, gate, admin, state, trace and the crates beneath them, with their charters, Specs and contracts.
WeaverAnalysis	The diagnostic consumer: the tools that read the record the agent's admin hands off, derive and preload a replay, and read what comes back.
WeaverWeb	The frontend: the human's surface over running agents and over what the analysis emits.
WeaverTools, this repository	The suite: experiments, and the documents that bind more than one repository.

The three code repositories came out of one monorepo, kept as WeaverTools-old2, whose history is the readable record of how each subsystem reached the state it left in. Each new repository opens at its own root commit.

What is here

experiments/: the suite's pre-registered measurement acts, each an experiment directory holding a charter (README.md), one directory per arm, and under each arm one per probe with its Spec, code/ and results/. They live here rather than in WeaverAgent because their code reads WeaverAnalysis as much as it reads the agent: the determinism matrix under deployment-tuple/baseline/ drives an installed agent stack and checks the declaration WeaverAnalysis derives, and its grammar test reads a WeaverAnalysis fixture by path.

• experiments/deployment-tuple/: the hypothesis that a deployment's output is reproducible exactly when six declared fields are held. Arms baseline/, device/, kernel-stack/. Its charter is the README a reader opens first.
• experiments/trace-content/: the trace-content classifier work, its label split, HeroBench agents and classifier evaluation.

A run's evidence lives in its deposit on the shared bulk store, and the repository carries only the result note and the scripts that produced its figures. A result note names its deposits and copies no data from them.

Still to come, on the operator's word: the cross-repository contracts (weaver-gate-world, weaver-admin-operator, weaver-analysis-state, weaver-analysis-web), the suite vision, the four process documents, and the paper material. Until they land, the contracts stand in WeaverAgent under docs/crates/contracts/ and the process documents under process/.

Running the experiments' code

Stdlib Python, no dependencies. The harness code expects the suite's repositories as siblings in one directory, which is how the workshop is laid out:

<workshop>/
  WeaverAgent/
  WeaverAnalysis/
  WeaverWeb/
  WeaverTools/        this repository

Where they are not siblings, WEAVER_ANALYSIS_DIR names the WeaverAnalysis checkout. Each probe's code/README.md says how to run it on a box, and each harness's test_*.py files are plain scripts that exit non-zero on the first failure:

cd experiments/deployment-tuple/baseline/determinism-matrix/code
for t in test_*.py; do python3 "$t" || break; done

Editorial

ASCII only; absolute dates; each election's reasoning stated where it is made; superseded text removed, since git is the archive; commit subjects code:, docs:, process:. The repository is public: no commercial or strategy material enters it.