><> sprat

BUILD RECORD

Milestones

Where Sprat actually is, including the parts that are not working. No dates — the phase estimates are planning figures, not commitments, and publishing them as promises would be the first dishonest thing on this site.

Phase 0

Feasibility

in progress

Prove the central bet and the risky parts before building the product.

  • Inline terminal spike done

    Raw input, live-region redraw under a fast token stream, resize, wide characters, restoration after a crash. Verdict recorded.

  • Packaging spike blocked

    Burrito 1.6.0 does not run the packaged application at all — an upstream bug, still present on main. A different single-binary route is needed.

  • Search metadata open

    Findings recorded. The choice between the provider's server-side search and a separate search API waits on the experiment below.

  • Team vs solo experiment in progress

    Ten curated tasks, three roles against one agent, blind-reviewed on evidence quality. If the team does not win clearly enough, Sprat ships as a single-agent tool and the role system is cut.

Phase 1

Foundation

in progress

One research question runs end to end in the terminal, survives a kill, and reports its cost correctly.

  • Inline terminal renderer and composer done
  • SQLite event store, state tables and recovery done
  • Session layer wired to the store done
  • Model adapter with streaming and tool use in progress
  • Usage ledger, budgets and hard money cap
  • Interrupt, resume and doctor
Phase 2

Research product

planned

Both workflows deliver a complete set of artifacts, with every material claim sourced.

  • Evidence store: sources, claims, confidence
  • Stage gates enforced by code
  • The two workflow templates and the roles
  • Opportunity scorecard and validation plan
  • Read-only repository attachment
Phase 3

Release

planned

A signed binary a stranger can install on a clean machine.

  • Signed and notarised macOS builds, Linux builds
  • Homebrew tap
  • Redaction and prompt-injection security fixtures
  • The feasibility evaluation re-run on the finished product

The one thing that could change all of this: if the team of roles does not clearly beat a single agent on evidence quality, Sprat ships as a single-agent research tool and the role system is removed. That experiment is still running.