Siftbench

Changelog

What just shipped

Linear-style shipping notes, authored as Markdown in content/changelog.

Agents with mandatory citations

agentsfindings

M6 lands cited findings, human accept/reject, and API rejection of fabricated pointers.

Agent runs are first-class jobs with durable Postgres transcripts, SSE events, re-steering, and a hard citation contract. Findings cross one public API boundary that resolves every pointer — a fabricated locator fails closed with a 422 and never lands.

The workbench adds live run transcripts, citation chips with evidence peeks, and keyboard findings triage. Human accept/reject records identity and reason; acceptance never launders AI content into evidence.

Unified timeline and explorers

explorerstimeline

M5 ships the dense workbench — registry, SQLite, EVTX, plist, text/hex, and a case timeline.

Purpose-built explorers land for Tier 1 artifacts alongside a unified timeline scoped to the case. Keyboard-first navigation, command palette, and virtualized grids treat the workbench like an instrument — information-dense, calm, and fast.

Evidence stays monospace and neutral. The case overview becomes a command center for parse jobs and upcoming agent work.

Cloud parse farm for triage packages

ingestparsing

M4 unpacks collections, maps hosts, and fans parse work across workers.

Triage archives unpack into a case topology: hosts, companion files, and artifact identity. Parse jobs fan across the worker plane so a multi-host collection returns browsable records in minutes instead of a day of local tooling.