Automated fix to PR design

Design reference for a hosted fix-to-open-PR flow where aislop scores every iteration and only keeps changes that improve deterministic quality.

This is a design reference for a hosted fix -> open PR flow. The current CLI already supports local aislop agent --pr; the hosted flow needs GitHub App installation auth, a server-side sandbox, and job orchestration.

Principle

The score is the gate. An LLM may propose edits, but aislop scores them. A change is only kept when the deterministic score improves. The model is never the judge of its own work.

User flow

1. Trigger

   A user starts a fix run from the hosted dashboard or a future CLI/API command against a repository and ref.

2. Queue

   The platform queues a job with repository, ref, installation, and policy context.

3. Run in sandbox

   The worker clones the repository, runs the loop, and records score deltas and findings.

4. Open PR or hand off

   If the final score improves, the GitHub App opens a PR. If not, the run returns a summary without a PR.

Repair loop

clone repo @ ref -> scan baseline score S0
repeat up to N iterations:
  run aislop fix
  if findings remain:
    ask model for bounded patch
  re-scan score Si
  if Si > S(i-1):
    keep changes and commit
  else:
    revert this iteration
stop when clean, no score gain, or budget hit
if final score Sf > S0:
  open PR
else:
  hand off summary

aislop fix runs first because it is deterministic. The model only handles findings that need context. Every kept iteration is committed so the PR history shows the improvement path.

Branch naming

aislop/fix/<job-id>

One branch per run, created from the target ref. The hosted App never pushes directly to the default branch.

PR body

The generated PR body should include:

• Score delta, such as 72 -> 91.
• Per-engine breakdown.
• Findings resolved, grouped by engine and rule.
• Files changed.
• Anything left unresolved and why.
• aislop version and config hash for reproducibility.

Auth model

• A GitHub App is installed on the repository.
• The server mints a short-lived installation token for the job.
• Token scopes are limited to repository contents and pull requests.
• Tokens are injected as environment secrets, never persisted to disk or logs.
• Branch protection is respected because output is always a PR.

Safety rails

Rail	Requirement
Sandbox	Ephemeral, network-restricted worker with only the cloned repo writable
Command allowlist	Only git, aislop, and safe project install/build/test commands
Patch application	Model output is applied as patches, not arbitrary shell
Time and budget caps	Hard limits on wall-clock time, iterations, and token spend
Score-gated writes	Revert any iteration that does not improve the deterministic score
Config protection	.aislop/config.yml and .aislop/rules.yml are treated as read-only policy inputs

Relation to the CLI

Use aislop agent --pr today for local, developer-controlled repair PRs. The hosted fix-to-PR design applies the same scoring model to a server-side workflow with GitHub App identity and dashboard orchestration.