AI-assisted triage for GitHub Issues & Pull Requests. AutoTriage summarizes items, applies / removes labels, posts comments, suggests better titles, and can optionally change issue state. It is driven by a project-specific prompt plus two Gemini model passes (fast first, then a higher quality review only if needed). A lightweight JSON DB (optional) preserves cumulative reasoning so model decisions stay explainable and append-only.
- Two-stage analysis (fast pass, conditional review pass) to save tokens.
- Deterministic prompt template you fully control (checked into your repo).
- Cumulative reasoning log: every run appends instead of overwriting context.
- Dry-run mode for safe experimentation (no writes).
- Explicit op budget (max operations) to keep runs predictable.
- Artifacts (prompt + model raw output + planned operations) for debugging.
| Input | Description | Default |
|---|---|---|
issue-number |
Single issue or PR number to triage (defaults to event context). | - |
issue-numbers |
Space or comma separated list of issue/PR numbers to triage explicitly. | - |
prompt-path |
Repo-relative path to the prompt template. | .github/AutoTriage.prompt |
enabled |
true = apply changes, false = dry-run only. |
true |
db-path |
Optional JSON file storing per-issue summary + reasoning history. | - |
model-fast |
Gemini model for the first (fast) pass. | gemini-2.5-flash |
model-pro |
Gemini model for the second (review) pass. | gemini-2.5-pro |
model-temperature |
Sampling temperature (0-2). Lower = more deterministic. | 1.0 |
max-timeline-events |
Max most-recent timeline events included in prompt. | 50 |
max-operations |
Cap on actions (label/comment/title/state) per run. | 10 |
| Secret | Purpose |
|---|---|
GITHUB_TOKEN |
Standard Actions token (provided automatically in most workflows). |
GEMINI_API_KEY |
API key for Google Gemini models. |
Create a project prompt (default path: .github/AutoTriage.prompt). Keep: rules, labeling conventions, closure criteria, tone for comments, title style guidelines. The action injects issue body, metadata, timeline, prior reasoning, and a strict JSON output schema section.
See ready-to-use workflow files in examples/workflows:
autotriage-issues.yml– run on issue events.autotriage-prs.yml– run on pull request events.autotriage-backlog.yml– scheduled/backlog sweep.
Copy one into .github/workflows/ and adjust enabled, schedules, or permissions as needed.
To triage specific items manually (e.g. via a workflow_dispatch input or a one-off run), set issue-numbers or issue-number in the job step with: block. Spaces or commas are both accepted, e.g. issue-numbers: "123 456,789". Combine with enabled: "false" for a safe dry-run.
Set enabled: "false" to log planned operations with an 🧪 [dry-run] prefix. No labels, comments, titles, or closures are changed.
If db-path is provided and enabled is true, each run appends to the reasoning log + maintains a canonical summary. This powers duplicate detection or clustering later (out of scope here, but data is retained).
For each item the action writes (in ./artifacts):
ISSUEID-gemini-input-<model>.md- exact prompt sent to the model.ISSUEID-gemini-output-<model>.json- raw Gemini response.ISSUEID-analysis-<model>.json- parsed structured analysis.ISSUEID-operations.json- final planned ops.
MIT