remediate disposition. Diagnosis and remediation share one cycle key, but remain distinct stages: diagnosis is read-only and short-lived; remediation may prepare code.
Typical flow
- Squasher receives an error and produces an evidence-backed diagnosis.
- The diagnosis either requests remediation, asks for a human, or records that no safe fix exists.
- For
remediate, project policy and repository settings are checked deterministically. - Squasher prepares a candidate patch and a focused regression test when the failure can be reproduced safely.
- The sandbox independently reruns the reported test commands and blocks failed validation from being pushed.
- A separate verification pass reviews the production root cause, diff, and executable test evidence before Squasher opens a pull request.
- Squasher waits for the pull request’s GitHub checks. Failed CI is returned to the same sandboxed agent for a bounded repair, test, push, and recheck loop.
- The PR result is sent to configured Slack or webhook review destinations only after CI passes.
Recommended setup
- Connect GitHub.
- Add a review channel with Slack if your team wants notifications.
- Pair with Linear if you want issue tracking alongside the PR.
Notes
- Auto-fix is review-oriented. Your team decides whether and when to merge.
- Better error context generally leads to better candidate fixes.
- Pull requests require independently passing test commands before they open. After opening, Squasher records bounded GitHub check evidence and up to three CI repair attempts; a run fails if CI never becomes green.
- Squasher records terminal cycles, safe refusals, PR creation, tests, verification, and result delivery against the prompt and model version that produced them.
Agent handoff
Use this prompt when a coding agent is deciding whether to start a fix run:squasher agent-runs ..., or MCP Code Mode after search("agentRuns"). Creating a fix run is a write action and should be confirmed by the user.