atrislabs/improve

/improve

Runs one improvement tick on the workspace. Calls POST /api/improve on the backend, which plans one task, builds it, verifies it, and scores it. Returns what shipped + the reward. Writes the scorecard locally.

This is the product. The thing the user pays for. One call, one verifiable result.

How it works

/improve
  → POST /api/improve { workspace: ".", mode: "full" }
  → backend picks a task, plans, builds, reviews, verifies
  → returns { task, reward, files_changed, verify_pass, summary }
  → CLI writes scorecard to .atris/presidio/scorecards.md
  → CLI reports result to user

The inference is Claude Code (or whatever model the backend uses). The environment is the folder. The endpoint is the bridge.

On invoke

1. Read ~/.atris/credentials.json for auth token
2. Read .atris/business.json for the API base URL (or default to http://localhost:8000)
3. Call POST /api/improve with:

   {
     "workspace": "<current working directory>",
     "mode": "full",
     "model": "sonnet"
   }
   

4. Wait for response (may take 1-5 minutes)
5. On success:
   • Show what shipped (task name, files changed, verify result)
   • Show the reward score
   • Write scorecard to .atris/presidio/scorecards.md
   • Append tick to today's journal
6. On failure:
   • Show the error
   • Write a lesson to atris/lessons.md
   • Do not write a scorecard

Modes

• full — plan, build, review, verify (default)
• plan — just pick the task and show what it would do
• dry_run — run everything but don't commit

Fallback

If the backend is unreachable (no auth, no network, localhost not running), fall back to local mode: run atris autopilot --auto --iterations=1 instead. Same loop, just local inference via claude -p subprocess. Report that it ran locally.

Output

improved.

  task:    fixed the stale wiki ref in auth-flow.md
  verify:  pass (npm test, 143/143)
  reward:  +4
  files:   atris/wiki/briefs/auth-flow.md
  time:    47s

  scorecard updated.

Rules

• One tick only. Never batch.
• Always verify. No reward without a check.
• Show what shipped, not what was attempted.
• Write the scorecard. This is the receipt.
• If verify fails, halt honestly and write a lesson.
• Fallback to local if backend is unreachable. Never error silently.
• The user pays because something real happened. Never fake it.