The agent starts from three inputs. Failure logs carry the error lines, seeds reproduce the run, and recent commits bound the suspect set.
Input | What it holds | Why the agent needs it |
|---|---|---|
Failure logs | Error lines plus surrounding context | Signatures for clustering. |
Seeds | Random seeds per failing run | Exact reruns of a failure. |
Recent commits | Changed files since the last green run | The first suspect list. |
The agent reads through three narrow tools. Each tool has one job and returns plain data the agent can quote.
Tool | Does | Returns |
|---|---|---|
Log search | Finds lines matching a pattern | Matching lines with file offsets. |
Rerun a seed | Replays one failing run | Pass or fail plus the new log tail. |
Bisect | Runs seeds across a commit range | First bad commit in the range. |
The output is one card per failure group. Each card lists the cluster size, the ranked suspects, and one evidence line per suspect.
Field | Example content |
|---|---|
Group | 3 runs sharing one signature. |
Suspects | Commits ordered c1, c3, c2. |
Evidence | Touched file, rerun result, bisect vote. |
The listing below is schematic. It shows the shape of one pass and is not output from any real system.
schematic illustration of one triage pass, not output from any system
1. normalize: strip timestamps, seeds, and addresses from each log line
2. cluster: runs sharing one signature form one group
3. rank: score each suspect commit against the group evidence
4. report: write one card per group with suspects and evidence linesThe recent-commits input. Only commits since the last green run belong in the suspect set.
One card. The signature matches, so they form one group; the two seeds become rerun evidence.