Skip to content

ROADMAP.md

Latest commit

 

History

History
63 lines (54 loc) · 3.72 KB

ROADMAP.md

File metadata and controls

63 lines (54 loc) · 3.72 KB
Raw

Roadmap

Where help is most wanted. atrophy is small on purpose, so this list stays focused on making the existing metrics more honest rather than adding surface. See CONTRIBUTING.md before opening a PR, and docs/design-v2.md for the evidence base behind the metrics.

Calibration (the ongoing gap)

  • Calibrate thresholds on more real data (#1). v0.2 moved the drift alert to a personal-percentile baseline and made the fast-accept threshold scale with output size, both checked against 30 days of real usage. More eyes and more corpora welcome: a metric is only as good as its correlation to how the day actually felt (the 1-5 rating).
  • English detection. The token sets are calibrated mostly on real French; the English pass is younger and needs tuning against real English transcripts.

Better proxies

  • Validation loop against git: do fast-accepted large outputs get reverted or fix-committed within N days? Git history is local and free; this is the cheapest ground truth available for the review metric.
  • away_summary and turn_duration: 2.1.x transcripts carry explicit away markers and per-turn durations; both could sharpen review latency and presence without any new data source.
  • Persistence proxy, scoped properly (#2).
  • gh pull requests as an output proxy for the value axis (#4).

Coverage

In priority order.

  1. Linux/Windows scheduling (#6). Since v0.2 the analyzer (report, week, recall, audit, doctor, --rate) runs anywhere Python does. What remains: presence sampling and scheduled nightly runs (systemd timers on Linux, Task Scheduler on Windows). Good first issue.
  2. More AI tools (#3). The reader is isolated in two functions (iter_session_events, scan). Known locations: Codex CLI writes JSONL to ~/.codex/sessions/YYYY/MM/DD/ (approvals and interrupts have equivalents there); Gemini CLI writes JSON to ~/.gemini/tmp/<hash>/chats/. A new source should only ship if judgment-equivalent signals exist (approvals, interruptions), so atrophy never degrades into another token counter.

Vision: from observing to improving

v0.2 shipped the first practice layer (recall, audit): retrieval practice and unassisted reps targeted by what your own transcripts show you delegate the most. Both follow the intervention literature: opt-in, rare, self-graded, unscored.

The longer arc stays the same: help you exercise and rebuild judgment, not just watch it slip. Open questions where research-minded contributions are welcome:

  • Does the recall probe's held/fuzzy/gone distribution track the felt 1-5 rating?
  • What refresher cadence actually maintains code ownership (the skill-decay literature suggests weeks, not days)?
  • Can the unassisted rep be made measurable (time-to-give-up, per Liu et al. 2026) without turning it into a score to game?

Related work in the intervention lane (atrophy deliberately stays in the passive-mirror lane): The Forge Protocol, Epistemic Protocols.

Out of scope, on purpose

  • No score to optimize, no leaderboard, no streaks. The tool stays an honest mirror, not a game.
  • No new dependency for what a few lines of standard library can do.
  • Nothing that persists a prompt, code, or client name. Aggregates only, always.
  • No workflow interception (hooks that gate or delay your agent): measurement stays passive, practice stays opt-in.