Reads your logs, finds the single root cause in your repository, proves the fix with a test, and opens a pull request. A person reviews; it never merges.
github.com/jessejohnsohn/fix-bugs-agent · v2.1.0 · declarative on the work-loop engine · source
artifacts:read artifacts:write network:api.github.com network:github.com plugins:github.com/jessejohnsohn/human-eyes-plugin state:read state:write
| field | what it is |
|---|---|
model * | The model the reasoning loop runs on. |
limits | |
repository * | The GitHub repository the agent fixes. Cloned into the run's container; the pull request targets it. Private or public makes no difference to the agent — only to what the token must be allowed to do. |
logSources * | Where the bugs come from: the logs. Each source is a command the harness runs (never the model) that prints the current FINDINGS as a JSON array — one row per distinct failure, not one row per log line. Rows: { id, title, count, firstSeen, lastSeen, samples }. `id` MUST be a stable signature of the failure (the same bug across runs yields the same id); the agent uses it to recognise a finding it has already opened a pull request for. `samples` are raw log entries, enough to locate the failure in code. What a 'distinct failure' is — grouping by exception + top frame, by route + status, by error code — is the source's decision, and the best place for it: it knows the log format, the agent does not. |