fix(ci): exempt draft PRs from the opencode-review current-head verdict gate

.github/workflows/opencode-review.yml

@@ -9,7 +9,7 @@ on:
   # content and never binds repository secrets. Privileged review execution is
   # isolated in opencode-review-dispatch.yml on repository_dispatch only.
   pull_request_target:
-    types: [opened, synchronize, reopened, ready_for_review, closed]
+    types: [opened, synchronize, reopened, ready_for_review, converted_to_draft, closed]
 
 concurrency:
   group: >-
@@ -264,17 +264,45 @@ jobs:
           TARGET_REPOSITORY: ${{ github.event.pull_request.base.repo.full_name || github.repository }}
           PR_NUMBER: ${{ github.event.pull_request.number }}
           HEAD_SHA: ${{ github.event.pull_request.head.sha }}
-          PR_ACTION: ${{ github.event.action }}
         run: |
           set -euo pipefail
-          if [ "$PR_ACTION" = "closed" ]; then
+          if [ -z "${PR_NUMBER:-}" ] || [ -z "${HEAD_SHA:-}" ]; then
+            echo "::error::Missing PR number or head SHA; cannot verify a current-head OpenCode verdict."
+            exit 1
+          fi
+          # Decide closed/draft from the pull request's live state, never from
+          # the triggering event's own stored payload -- a manual re-run of an
+          # old workflow run (e.g. an old converted_to_draft event) replays
+          # that event's github.event.* fields verbatim, which would let a
+          # since-ready, unreviewed PR at the same head SHA pass this required
+          # check on a stale "still draft" reading (Devin review on #1443).
+          if ! pr="$(timeout 25 gh api "repos/${TARGET_REPOSITORY}/pulls/${PR_NUMBER}")"; then
+            echo "::error::Could not fetch the pull request's live state; cannot verify a current-head OpenCode verdict."
+            exit 1
+          fi
+          pr_state="$(printf '%s' "$pr" | jq -r '.state // ""')"
+          pr_draft="$(printf '%s' "$pr" | jq -r '.draft // false')"
+          # Exposed for the dispatch step below so its repository_dispatch
+          # payload also comes from this same live fetch, not from
+          # github.event.pull_request.* -- a manual re-run of an old job
+          # would otherwise still build the dispatch payload from that stale
+          # event payload, and opencode-review-dispatch.yml's live
+          # validate-pr-metadata check hard-rejects a stale base_ref/base_sha,
+          # so the dispatch would fail even once this step correctly decides
+          # one is needed (Devin review on #1443).
+          printf 'base_ref=%s\n' "$(printf '%s' "$pr" | jq -r '.base.ref')" >>"$GITHUB_OUTPUT"
+          printf 'base_sha=%s\n' "$(printf '%s' "$pr" | jq -r '.base.sha')" >>"$GITHUB_OUTPUT"
+          printf 'head_ref=%s\n' "$(printf '%s' "$pr" | jq -r '.head.ref')" >>"$GITHUB_OUTPUT"
+          printf 'head_sha=%s\n' "$(printf '%s' "$pr" | jq -r '.head.sha')" >>"$GITHUB_OUTPUT"
+          if [ "$pr_state" = "closed" ]; then
             echo "PR closed; a current-head OpenCode verdict is not required."
             echo "verdict=CLOSED" >>"$GITHUB_OUTPUT"
             exit 0
           fi
-          if [ -z "${PR_NUMBER:-}" ] || [ -z "${HEAD_SHA:-}" ]; then
-            echo "::error::Missing PR number or head SHA; cannot verify a current-head OpenCode verdict."
-            exit 1
+          if [ "$pr_draft" = "true" ]; then
+            echo "PR is a draft; a current-head OpenCode verdict is not required until it is marked ready for review."
+            echo "verdict=DRAFT" >>"$GITHUB_OUTPUT"
+            exit 0
           fi
           if ! reviews="$(timeout 25 gh api --paginate "repos/${TARGET_REPOSITORY}/pulls/${PR_NUMBER}/reviews")"; then
             reviews="[]"
@@ -313,16 +341,28 @@ jobs:
           fi
 
       - name: Request current-head OpenCode review execution
-        if: github.event.action != 'closed' && steps.verdict.outputs.verdict == ''
+        # steps.verdict.outputs.verdict alone is the live, authoritative
+        # signal now: the verdict step resolves closed/draft from a live
+        # gh api call, not this event's own stored payload, so it is already
+        # empty only when a real dispatch is actually needed. The former
+        # extra github.event.action/github.event.pull_request.draft conjuncts
+        # were themselves stale-payload reads -- a manual re-run of an old
+        # closed/draft-era job could still suppress this dispatch on a since-
+        # reopened/ready PR even after the verdict step's own fix, leaving the
+        # required check red with no review ever requested (Devin review on
+        # #1443).
+        if: steps.verdict.outputs.verdict == ''
         env:
           OIDC_AUDIENCE: opencode-github-action
           OPENCODE_API_BASE_URL: https://api.opencode.ai
           TARGET_REPOSITORY: ${{ github.event.pull_request.base.repo.full_name || github.repository }}
           PR_NUMBER: ${{ github.event.pull_request.number }}
-          BASE_BRANCH: ${{ github.event.pull_request.base.ref }}
-          BASE_SHA: ${{ github.event.pull_request.base.sha }}
-          HEAD_BRANCH: ${{ github.event.pull_request.head.ref }}
-          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+          # Sourced from the verdict step's own live gh api fetch, not
+          # github.event.pull_request.* -- see the comment on that step.
+          BASE_BRANCH: ${{ steps.verdict.outputs.base_ref }}
+          BASE_SHA: ${{ steps.verdict.outputs.base_sha }}
+          HEAD_BRANCH: ${{ steps.verdict.outputs.head_ref }}
+          HEAD_SHA: ${{ steps.verdict.outputs.head_sha }}
         run: |
           set -euo pipefail
           if [ -z "${ACTIONS_ID_TOKEN_REQUEST_TOKEN:-}" ] || [ -z "${ACTIONS_ID_TOKEN_REQUEST_URL:-}" ]; then
@@ -361,6 +401,9 @@ jobs:
           if [ "$VERDICT" = "CLOSED" ]; then
             exit 0
           fi
+          if [ "$VERDICT" = "DRAFT" ]; then
+            exit 0
+          fi
           if [ -z "$VERDICT" ]; then
             echo "::error::No APPROVED or CHANGES_REQUESTED from opencode-agent on the current head. This required check is not a review and must not succeed until the authenticated dispatch posts a current-head verdict."
             exit 1

CHANGELOG.md

@@ -5,6 +5,71 @@ this file. The format follows Keep a Changelog, and versioned releases follow
 Semantic Versioning where the repository publishes a release.
 
 ## [Unreleased]
+- Fix `opencode-review.yml`'s required `opencode-review-target` check
+  reporting a hard `exit 1` failure on every push to a draft PR, forever,
+  until the PR is marked ready for review. Root cause:
+  `scripts/ci/pr_review_merge_scheduler.py` deliberately never dispatches an
+  OpenCode review request for a draft PR (`if pr.get("isDraft"): return
+  Decision(number, "skip", "draft PR")`), but the required check had no
+  draft handling and unconditionally demanded a current-head verdict on
+  every `opened`/`synchronize`/`reopened` event regardless. The
+  `Resolve current-head formal OpenCode verdict` step now exits early with
+  `verdict=DRAFT` for a draft PR (mirroring its existing `closed` early
+  exit), the `Request current-head OpenCode review execution` dispatch
+  step's own `if:` also skips drafts (so a transient OIDC/dispatch failure
+  can't turn a draft PR's required check red before the exemption even
+  runs), and the trivial `Fail closed without a current-head OpenCode
+  verdict` step treats `VERDICT=DRAFT` the same as `VERDICT=CLOSED`. Also
+  adds `converted_to_draft` to the workflow's `pull_request_target` trigger
+  types, so a ready PR converted back to draft with no new commit still
+  gets a fresh required-workflow run that can apply the exemption (without
+  it, its previously failed check would show failure indefinitely). This
+  was re-derived from scratch against the event-driven
+  `opencode-review-target` design `#1507`/`#1532` landed on `main` (the
+  325-minute synchronous poll loop this fix originally targeted no longer
+  exists); see `tests/test_opencode_required_verdict_regression.py` for
+  shell-level regression coverage executing each step's actual production
+  body, including the new `_run_verdict_step`/`_run_fail_closed_step`
+  helpers matching the three-step split.
+- Fix a review-flagged gap in the draft-gate fix above (Devin review on
+  `#1443`): the `Resolve current-head formal OpenCode verdict` step decided
+  closed/draft purely from the triggering event's own stored payload
+  (`github.event.action`/`github.event.pull_request.draft`). A manual
+  re-run of an old workflow run (e.g. a stale `converted_to_draft` run)
+  replays that event's payload verbatim, so a since-ready, unreviewed PR at
+  the same head SHA could pass this required check on a stale "still draft"
+  reading — a real required-review bypass, not merely a false alarm. The
+  step now fetches the pull request's live state (`gh api
+  repos/.../pulls/<number>`) and decides closed/draft from that instead,
+  failing closed if the lookup itself fails; the two payload-derived `env:`
+  vars (`PR_ACTION`/`PR_DRAFT`) are removed entirely from this step.
+- Fix a second review-flagged gap in the same lineage (Devin review on
+  `#1443`): the `Request current-head OpenCode review execution` dispatch
+  step's own `if:` still gated on `github.event.action`/
+  `github.event.pull_request.draft` alongside `steps.verdict.outputs.verdict
+  == ''`. Once the verdict step above was fixed to resolve closed/draft from
+  live state, that made the dispatch step's stale-payload conjuncts a
+  liability rather than a safety net: a manual re-run of an old closed/draft
+  job could still suppress the dispatch for a since-reopened/ready PR at the
+  same head SHA, leaving the required check red with no review ever
+  requested. The dispatch step now gates purely on
+  `steps.verdict.outputs.verdict == ''`, the one live-computed signal that
+  already means "a real dispatch is needed."
+- Fix a third review-flagged gap in the same lineage (Devin review on
+  `#1443`): even with the dispatch step correctly enabled from the live
+  verdict, it still built its `repository_dispatch` payload's
+  `pr_base_ref`/`pr_base_sha`/`pr_head_ref`/`pr_head_sha` fields from
+  `github.event.pull_request.*` — the same stale-payload source the two
+  fixes above removed from the pass/fail decision. On a manual re-run of an
+  old job whose base branch has since advanced (head SHA unchanged),
+  `opencode-review-dispatch.yml`'s live `validate-pr-metadata` check
+  hard-rejects that stale `base_sha`, so the dispatch would fail even though
+  the verdict step correctly decided one was needed — leaving the required
+  check red with no review ever requested. The verdict step now exposes
+  `base_ref`/`base_sha`/`head_ref`/`head_sha` as step outputs from the same
+  live `gh api repos/.../pulls/<number>` response it already uses for
+  closed/draft, and the dispatch step builds its payload from
+  `steps.verdict.outputs.*` instead of the event payload.
 - Fail closed when the first top-level Noema JSON candidate is malformed,
   preventing a later approval object from overriding malformed preface data;
   multiple-object output remains supported when its first object is valid.
@@ -671,7 +736,8 @@ Semantic Versioning where the repository publishes a release.
   Informational, no change: the gap-baseline's repeated review-round
   narrative is this repo's own documented, intentional convention
   (ADR-0002: the baseline is "an operational snapshot," not a duplicate of
-  the ADR's design record), not accidental redundancy.- Raise `contextual_orchestrator_review_sidecar.sh`'s
+  the ADR's design record), not accidental redundancy.
+- Raise `contextual_orchestrator_review_sidecar.sh`'s
   `ORCHESTRATOR_CATALOG_FAMILY_CAP` default from 4 to 8: root-caused the
   live "no provider route passed the Strix plain-chat preflight" outage
   blocking `noema-review`/`opencode-review`/`strix` org-wide to

docs/product-technical-gap-baseline.md

@@ -2344,6 +2344,47 @@ contract assertion, and `docs/adr/0003-contextual-orchestrator-vendored-free-zdr
 "today" reference. Landed in the same PR (`#1463`) as the streaming revert,
 not split out, since the revert is unsafe without it.
 
+## 2026-09-01 opencode-review.yml draft-gate fix (`.github#1443`): three-round live-state hardening, tied to `#1531` queue pressure
+
+**Buyer/control-plane effect.** Before this fix, every draft PR org-wide showed the required
+`opencode-review` check as a hard, permanent `exit 1` failure: `scripts/ci/pr_review_merge_scheduler.py`
+deliberately never dispatches an OpenCode review request for a draft PR
+(`if pr.get("isDraft"): return Decision(number, "skip", "draft PR")`), but the required check had no
+draft handling and unconditionally demanded a current-head verdict. Net effect: draft PRs allocated
+required-workflow OpenCode review dispatch work they could never complete or merge from — the exact
+kind of wasted required-workflow queue allocation `#1531` tracks — while presenting a false-alarm
+failure with no actionable next step for the PR author.
+
+**Fix, in three review-driven rounds, all on `.github#1443`:**
+1. The `Resolve current-head formal OpenCode verdict` step now exits early (`verdict=DRAFT`) for a
+   draft PR, mirroring its existing `closed` early exit, and the `Request current-head OpenCode review
+   execution` dispatch step's own `if:` matches — so a draft PR no longer dispatches OpenCode review
+   work it can never merge from. `converted_to_draft` was added to the workflow's `pull_request_target`
+   trigger types so a ready-to-draft conversion with no new commit still gets a fresh required-workflow
+   run that can apply the exemption.
+2. Devin review found the draft/closed decision, and later the dispatch step's own `if:`, were still
+   reading `github.event.action`/`github.event.pull_request.draft` — the *triggering event's own stored
+   payload*. A manual re-run of an old workflow run (e.g. a stale `converted_to_draft` run) replays that
+   payload verbatim, so a since-ready, unreviewed PR at the same head SHA could pass the required check
+   on a stale "still draft" reading, or a since-ready PR's real dispatch could stay wrongly suppressed.
+   Both steps now decide from the pull request's *live* state (one `gh api repos/.../pulls/<number>`
+   fetch per run), failing closed if the lookup itself fails.
+3. Devin review (third round) found that even with the dispatch step correctly gated on the live
+   verdict, it still built its `repository_dispatch` payload's `pr_base_ref`/`pr_base_sha`/
+   `pr_head_ref`/`pr_head_sha` from that same stale event payload — so a re-run after the base branch
+   advanced could still have its dispatch rejected by `opencode-review-dispatch.yml`'s live
+   `validate-pr-metadata` check, leaving the required check red with no review ever requested. The
+   verdict step now exposes `base_ref`/`base_sha`/`head_ref`/`head_sha` as step outputs from that same
+   live fetch, and the dispatch step builds its payload from those outputs instead.
+
+**Net result.** Stale workflow reruns can no longer grant a draft/closed exemption or dispatch stale
+base/head metadata — both the pass/fail decision and the dispatch payload are sourced from one live
+PR-state fetch per run, never from the frozen triggering-event payload. Full regression coverage in
+`tests/test_opencode_required_verdict_regression.py` executes each step's actual production bash body
+against fake `gh` fixtures. This closes one confirmed, now-eliminated source of required-workflow queue
+waste (draft PRs); it does not by itself resolve the separate org-wide runner-capacity congestion
+`#1531` also tracks.
+
 ## 5. 실행 루프와 고객의 다음 행동
 
 각 hourly pass는 아래 순서를 유지한다.

scripts/ci/test_strix_quick_gate.sh

@@ -512,7 +512,7 @@ assert_opencode_review_uses_codegraph_and_contextual_orchestrator() {
 	local opencode_config="$REPO_ROOT/opencode.jsonc"
 
 	assert_file_contains "$bootstrap_file" "pull_request_target:" "opencode required workflow loads its metadata-only bootstrap from the protected base ref"
-	assert_file_contains "$bootstrap_file" "types: [opened, synchronize, reopened, ready_for_review, closed]" "opencode required workflow reacts to current PR head changes and closed-PR cleanup"
+	assert_file_contains "$bootstrap_file" "types: [opened, synchronize, reopened, ready_for_review, converted_to_draft, closed]" "opencode required workflow reacts to current PR head changes, draft reconversion, and closed-PR cleanup"
 	assert_file_contains "$bootstrap_file" "required-workflow-bootstrap:" "opencode required workflow materializes at least one job for pull_request ruleset runs"
 	assert_file_contains "$bootstrap_file" "Required OpenCode workflow materialized without checking out or" "opencode required workflow bootstrap documents its data-only trust boundary"
 	assert_file_contains "$bootstrap_file" "coverage-source-tree:" "opencode required workflow preserves the stable coverage-source-tree branch-protection context"