## TL;DR
Check the draft flag before doing anything: if the PR is a draft, post informational comments only (or skip entirely) and never submit a formal verdict. Run the full review automatically when the PR is marked ready. Requesting changes on a draft is noise - the author already knows it is unfinished.

## The problem

```text
review agent failed to detect a draft PR: requested changes too early: how to fix
```

## When to use
- The bot submits REQUEST_CHANGES or APPROVE on draft PRs
- Draft authors complain about premature formal reviews
- Reviewers get change-request notifications for unfinished work
- Your workflow uses drafts heavily and the bot spams them

## When not to use
- The bot only left informational comments on a draft (acceptable; leave it)
- An operator explicitly asked for a full review of a draft (an override, not a bug)
- The PR was not a draft when the review ran (check timing before blaming detection)

## Steps

### 1. Fetch the draft flag first, before any review work
```bash
gh pr view 123 --json isDraft --jq .isDraft
```

Expected output: `true` or `false`. This one call decides the whole shape of the run.

### 2. Branch on the flag: drafts get comments only, never verdicts
If draft: run the analysis but post findings as plain informational comments, or skip the review entirely per team policy. Never submit REQUEST_CHANGES or APPROVE on a draft.

Expected output: draft PRs get at most helpful notes; no formal verdicts, no notification noise.

### 3. Subscribe to the ready_for_review event
Listen for the PR moving out of draft and trigger the full review then. Expected output: the real review runs automatically at the moment it becomes useful, with no manual re-trigger.

### 4. Re-check the flag immediately before submitting the verdict
A PR can leave draft status mid-run. Re-fetch `isDraft` right before the submit call; if it flipped to ready, run the full verdict path.

Expected output: no missed reviews for PRs that went ready during the run, and no verdicts on PRs still in draft.

### 5. Log the draft decision explicitly
Record "skipped formal review: PR is draft" (or "comments-only mode") in the run log. Expected output: draft-skips are distinguishable from failures in monitoring.

## Tool compatibility
- GitHub REST and GraphQL: the `draft` field on pull requests; stable on github.com and GHES.
- GitHub `ready_for_review` webhook event: the trigger for the post-draft review.
- gh CLI 2.x: `pr view --json isDraft` for the flag check.

## Variant phrasings

### bot reviewed a work-in-progress PR
Same fix. Decide the team policy first (step 2): comments-only versus full skip, then encode it.

### change request on draft pull request
The specific harm is the REQUEST_CHANGES verdict on unmergeable code. Gate verdict submission on the draft flag and the complaint disappears.

### skip review for draft PRs
If the team prefers silence over comments on drafts, step 2's skip branch does it; keep step 3 so the review still happens on ready.

## Why it happens
Review triggers usually fire on PR open and on synchronize, and neither distinguishes drafts. The agent does exactly what it was told - review the PR - without knowing the PR is explicitly marked not-ready. The draft flag exists precisely to signal this; the pipeline just never checked it.

## Edge cases
- Repos that use drafts as the default workflow: comments-only mode is usually the right default there, not full skip.
- Draft converted mid-run: step 4 handles it; without the recheck the review either misfires or never runs.
- Operator override: support an explicit "review this draft fully" input for the rare cases it is wanted, and log when it is used.
- Fork PRs: the draft flag works the same; no special handling needed.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_F-Jou9R_di6cCh5py-le-Q