Field evidence · updated 2026-07-15
Pluribus is an evidence layer for agent context boundaries. The strongest support for that direction is not another document in this repository; it is a falsifiable boundary that an independent maintainer ships or protects in tests.
A synthetic canary showed that Observer returned raw historical tool arguments to the agent and created a world-readable database. The maintainer shipped metadata-only, session-scoped defaults, raw opt-in, pre-store redaction, and 0600 storage; then requested and merged the hermetic regression suite.
A failover test asked whether backup agent B inspects work left by quota-limited agent A. Usher's maintainer turned that first external feedback into a default continuation guard, machine-readable marker, opt-out, end-to-end test, and tagged release in 37 minutes. The remaining boundary is explicit: warning delivered does not prove workspace inspected.
skill-graveyard accepted a privacy-safe receipt that separates inventory from observed use, with CLI and test coverage.
halo-record accepted runtime authority snapshots and a follow-up that deduplicates unchanged evidence instead of pretending every observation is new.
agent-context-economy accepted metadata that makes a generated repository map's staleness inspectable instead of treating existence as current authority.
mcpm accepted sync dry-run receipts followed by rollback snapshots, separating previewed intent from applied and recoverable state.
claude-handoff-revive accepted a pre-edit resume receipt so a handoff can be checked before a new session mutates the repository.
promptblock accepted hidden-comment indexes in warnings, making invisible instruction locations reviewable instead of reporting only a generic finding.
An independent maintainer listed Pluribus in awesome-ai-agents-2026 under the observability category.
The context-engineering resource repository accepted Pluribus as a context-management tool.
awesome-gemini-cli accepted the Pluribus listing, exposing the source-to-native-file workflow to Gemini CLI users.
gaia-skill-tree accepted the privacy-safe evidence-attestation recipe into its named registry. After an unrelated staging branch contaminated the first attribution patch, the maintainer recreated it as a one-commit, one-file change with all checks green. This proves independent curation and accurate contribution provenance—not installs or runtime use.
The browser demo shows the narrowest useful Pluribus workflow: one reviewed source, native agent-file previews, and hashes for what was generated. The receipt remains honest that generation does not prove a client loaded the output.
Run the 30-second context sync demo →
For the latest CLI release, use GitHub Releases. npm may lag during release windows.