atrislabs

improve

"Run one verified, scored improvement tick and write its receipt. Use when the user asks to improve, make this better, ship one thing, run a tick, or get smarter. `atris improve` alone shows metabolism vitals; use `atris improve tick` to ship work."

atrislabs 67 4 Updated 1mo ago
GitHub

Install

npx skillscat add atrislabs/atris/improve

Install via the SkillsCat registry.

SKILL.md

/improve

Run one improvement tick on the workspace. A successful tick ships one result, passes a real verifier, receives a score, and writes a machine-readable scorecard.

One call means one verifiable result. Never turn one invocation into a hidden batch.

Command contract

atris improve alone shows the self-improvement metabolism vitals. It does not ship a change.

Use the explicit tick subcommand when the user asks for improvement work:

atris improve tick                    # one full tick: plan, build, verify, score
atris improve tick --json             # the same tick as machine-readable output
atris improve tick plan               # plan only; no change or receipt
atris improve tick --dry-run          # execute without committing; no shipping receipt
atris improve tick --no-fallback      # report API failure instead of running locally
atris improve history                 # reward trend, credits, and pass rate
atris improve                         # vitals only
atris improve --json                  # vitals only as JSON

Run one tick

  1. Run atris improve tick, adding --json only when structured output is useful.
  2. Inspect the returned source, task summary, verifier result, reward, files, and receipt path.
  3. Count the tick only when ok is true, verification passed, and the scorecard was written.
  4. Report what shipped and the exact verifier result. If the tick fails, report the failure and stop.

For a paid API tick, the CLI:

  1. Loads credentials with utils/auth.loadCredentials.
  2. Calls POST /api/improve with the workspace, mode, model, and dry-run setting.
  3. Lets the backend plan, build, verify, score, and bill the successful tick.
  4. Writes the normalized receipt to .atris/state/scorecards.jsonl and the human trail to the daily Atris journal.

The full response may omit the billed credit count. In that case the CLI reports that billing happened server-side instead of inventing a number.

Local fallback

Local fallback is allowed when there is no login, the backend is unreachable, or the hosted backend cannot access the local workspace path. The CLI runs this exact bounded command internally:

atris mission run --due --headless --max-ticks 1 --complete-on-pass --json

The local result succeeds only when all of these are true:

  1. Exactly one mission tick ran.
  2. A headless worker actually ran; caller-session and no-worker placeholders are rejected.
  3. The mission verifier passed.
  4. The scorecard and journal receipt were written.

A verified local tick earns the conservative local reward of +1 and deducts no credits. No due headless mission, a skipped worker, a failed or missing verifier, or a receipt write failure makes the improve command fail without a reward.

Do not silently fall back for answerable API failures such as insufficient credits, unrelated authorization failures, or server errors.

Expected output

User says: "Improve this once."

Action: run atris improve tick.

Expected shape:

improved.
  task:    fixed the stale wiki reference
  verify:  pass
  reward:  4
  files:   atris/wiki/auth-flow.md
  scorecard: .atris/state/scorecards.jsonl

For local fallback, the first line is improved (local fallback). and the report still names the task, passing verifier, reward, and scorecard.

Failure handling

  • If verification fails or is missing, halt honestly and write a durable lesson. Do not claim improvement.
  • If the command reports no due headless mission, stop; do not manufacture busywork or a scorecard.
  • If the API returns insufficient credits or another answerable error, report it without a local retry.
  • If the workspace is dirty, preserve existing changes. Use an isolated worktree for manual follow-up fixes.
  • If a worker lands work but the outer receipt fails, report the landed commit separately and do not count the improve tick.

Rules

  • One tick only. Never batch.
  • No reward without a passing verifier.
  • Show what shipped, not what was attempted.
  • A scorecard is required for success.
  • Preserve user work and unrelated changes.
  • The user pays because something real happened. Never fake it.