Audit the work behind the prose
Resolve “tests pass” to the tool call, exit status, duration, and raw result instead of treating an assistant sentence as evidence.
Local-first · cross-provider · evidence-addressable
Polylogue turns AI chats, coding-agent sessions, tool calls, results, forks, costs, and reviewed notes into one local evidence system—so work can be searched, audited, and resumed without trusting a transcript summary.
nix run github:Sinity/polylogue -- demo receipts --compactEvidence pipeline
Raw evidence remains durable. Normalized views and analytics are rebuildable. Human judgment is explicit rather than blended into machine inference.
Exports, agent files, hooks, browser capture, and telemetry.
Sessions, messages, actions, results, lineage, and authoredness.
Search, costs, phases, profiles, claims, and optional vectors.
CLI, Python, MCP, daemon API, and local web reader.
What changes
Resolve “tests pass” to the tool call, exit status, duration, and raw result instead of treating an assistant sentence as evidence.
Preserve forks, resumptions, copied prefixes, and subagents while composing the logical work they represent.
Compile reviewed evidence, omissions, caveats, and user judgments into a reproducible context package for the next agent.
Separate provider-reported usage, cache lanes, reasoning tokens, catalog estimates, and subscription-credit views.
One archive, several interfaces
Polylogue ingests provider-specific evidence, but its query and reading model is provider-independent.
The committed Beads graph is rendered as a searchable roadmap with designs, acceptance criteria, dependencies, and closure records.