github-triage

Files (1)

• SKILL.md

SKILL.md (first lines)

# GitHub triage: issue and PR processor

## Tool restrictions

Read-only GitHub and repository analysis. Do not modify repo files or GitHub state. Local report writes are allowed only under `.omca/scratch/github-triage-{datetime}/` in the project root. MCP tools: `notepad_write`, `evidence_log`, `ast_search`.

Fetch open issue/PR metadata, classify each, spawn 1 background executor per item. Each subagent fetches full details for its item and writes a report under `.omca/scratch/github-triage-{datetime}/` in the project root. Never take destructive action.

## Zero-action policy

This skill reads and reports; a human maintainer decides what happens to every item, and a mutation made during triage takes that decision away. Run only these read-only commands:

- `gh issue list`, `gh issue view`
- `gh pr list`, `gh pr view`
- `gh api --method GET repos/{REPO}/pulls/{number}/files`
- `gh repo view`

Never run a `gh` command that merges, closes, edits, comments on, labels, or reviews an item (`gh pr merge`, `gh pr close`, `gh issue close`, `gh issue edit`, `gh pr edit`, `gh pr review`), and never call `gh api` with `POST`, `PUT`, `PATCH`, or `DELETE`.

## Evidence rule (mandatory)

Every factual claim MUST cite a GitHub permalink containing a commit SHA. Branch permalinks (`blob/main`, `blob/master`, branch names) are forbidden.

Format:
```
CLAIM: "The handler for X is in Y"
EVIDENCE: https://github.com/{REPO}/blob/{COMMIT_SHA}/path/to/file.py#L42
```

Without a commit-SHA permalink, do not make the claim. Write "UNVERIFIED" instead.

Applies to: bug root cause (file + line), "feature exists" (cite where), "fix correct" (cite what), any code reference.

## Architecture

```
1 issue or PR  =  1 Agent(subagent_type="oh-my-claudeagent:executor")
```

| Rule | Value |
|------|-------|
| Agent type for ALL items | `oh-my-claudeagent:executor` |
| Execution mode | Background by default, with the result in a `<task-notification>`; a subagent the platform runs in the foreground returns its result directly. Accept either |
| Parallelism | Bounded batches, max 5 concurrent agents |
| Total items per run | No platform cap. Bound the run yourself and record the overflow (see below) |
| Result storage | `issue-{number}.md` or `pr-{number}.md` under `.omca/scratch/github-triage-{datetime}/` in the project root |
| Final collection | Orchestrator reads all reports and writes `SUMMARY.md` |

---

## Phase 1: setup output directory

Set `{OUTDIR}` to the absolute path `<project root>/.omca/scratch/github-triage-<datetime>`, where `<datetime>` is the current date and time as `YYYYMMDD-HHMMSS`. When the session context gives only the date, append a short word of your own so two runs on one day get different directories. OMCA's server writes `.omca/.gitignore` when it starts, so reports stay out of commits.

Run no shell command to create the directory: the Write tool creates it when the first report is written. Tell the user the path now: `Reports will be written to: {OUTDIR}`.

---

## Phase 2: fetch open item metadata only

When no repo argument was given, run `gh repo view --json nameWithOwner -q .nameWithOwner` and use the `owner/name` it prints as `{REPO}` from here on.

```bash
# Issues: all open metadata only. Do not request body/comments here; control characters can break batching.
gh issue list --repo {REPO} --state open --limit 500 --json number,title,state,createdAt,updatedAt,labels,author

# PRs: all open metadata only. Subagents fetch body/comments/reviews/files per item.
gh pr list --repo {REPO} --state open --limit 500 --json number,title,state,createdAt,updatedAt,labels,author,headRefName,baseRefName,isDraft,mergeable,reviewDecision,statusCheckRollup
```

If either returns exactly 500 results, paginate using `--search "created:<LAST_CREATED_AT"` until exhausted.

---

## Phase 3: classify each item

For each item, determine its type from metadata only: title, labels, author, and PR state fields. Do not fetch body/comments during classification.

### Issues

Title prefixes and labels are strong signals. A `?` in a title is weak, because bug reports are often phrased as questions: an item that reports broken or unexpected behavior is `ISSUE_BUG` however it is phrased.

| Type | Signals |
|------|-----------|
| `ISSUE_BUG` | Title contains `[Bug]` or `Bug:`, labels indicate bug, or the title describes broken or unexpected behavior |
| `ISSUE_FEATURE` | Title contains `[Feature]`, `[RFE]`, `[Enhancement]`, `Feature Request`, `Proposal`, or labels indicate enhancement |
| `ISSUE_QUESTION` | Title contains `[Question]` or `[Discussion]`, or labels indicate question/discussion, and no defect is reported |
| `ISSUE_OTHER` | Anything else |

### PRs

| Type | Detection |
|------|-----------|
| `PR_BUGFIX` | Title starts with `fix`, `fix:`, `fix(`, branch contains `fix/` or `bugfix/`, or labels include `bug` |
| `PR_OTHER` | Everything else (feat, refactor, docs, chore, etc.) |

---

## Phase 4: spawn 1 background agent per item

### Run-size bound (decide before spawning anything)

No platform ceiling limits how many subagents a session spawns in total. The only
spawn limit that applies here is the concurrent one: the Agent tool refuses a spawn
with `Concurrent subagent limit reached` once the session's running count reaches
`CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS` (default 20), and it starts succeeding again
as soon as running agents finish. Batches of 5 stay under that and under the
tool-call concurrency default of 10, so a batched run never trips it. A session
running with ultracode on is exempt from the concurrent limit entirely.

What actually bounds a run is wall-clock time and the orchestrator's own context:
every finished agent's result lands inline in this conversation. So the total is a
judgment call, not a constant. Pick a bound for the run, from the user's stated scope
if they gave one, otherwise from how large the classified list is against how much
room is left in the conversation. Then, if the classified list is longer than the
bound:

1. Sort by `updatedAt` descending and take the first N. Recently-touched items are
   the ones a triage pass is for.

From UtsavBalar1231/oh-my-claudeagent, MIT.

More from UtsavBalar1231/oh-my-claudeagent