Make AgentCiv practical to try with local collaboration evidence

[blisspixel]

AgentCiv's README now explains what the framework provides, why retained shared work is useful, and what works today in its first three paragraphs. A diagram makes the handoff visible; setup, security, research, and validation details live in linked guides.

A small fictional stock experiment exercises real HTTP publication, source corrections, host restart, revoked author credentials, and a fresh newcomer. Participants can use installed local models, publish their own checklists, object, decline, or stop. A separate oracle checks arithmetic and exact current source references; invalid decisions are never replaced by scripted answers. The task and its English labels are replaceable example choices, not protocol requirements.

Retained evidence covers six stock model launches with USD 0 external spend: three completed, including one preliminary run, and three stopped during decision validation. Completed tasks exist on both hosts. Only the preliminary newcomer declared peer-artifact derivation; the other completions reconstructed from original shared sources. All failures, source conditions, budgets, and limits remain visible. Two earlier guide trials and both scripted stock baselines are retained separately.

The history client now validates page/world consistency, order, known authorship fields, duplicate JSON members, and finite JSON numbers before model use. It preserves schema-permitted integer representations such as 1.0. Failure tests cover incorrect quantities, stale sources, forged claims, source drift, changed originals, provider failures, uncertain publication, and shared budgets. Fixture tests preserve Spanish and Chinese text through real HTTP and a child process with a non-UTF-8 stdout locale.

Validation: formatting, workspace Clippy, repository/schema checks, strict mypy, all local suites, raw HTTP walks, restart exercises, and the visibility/lifecycle matrix passed. The matrix recorded 350 passes, zero failures, and ten optional skips. Rust line coverage was 89.66%; Python statement coverage was 90.73%. CI adds scripted stock runs on the existing supported platforms. Full profile coverage, independent interoperability, general model reliability, autonomous project selection, and a benefit from collaboration remain unclaimed.