Bound OpenCode review cadence and scheduler backlog

[seonghobae]

Problem

Required OpenCode Review checks are blocking merges across the org when the model pool hangs or exhausts provider capacity. The old behavior either let one model/provider consume the whole step or ended with an opaque GitHub Actions timeout, so PRs accumulated without a source-backed APPROVE/REQUEST_CHANGES result or a useful failure reason.

Current decision

This PR is the canonical OpenCode review cadence fix. It supersedes older conflicting budget PRs (#281, #348, #372).

• Put github-models/deepseek/deepseek-v3-0324 first because it has the best observed success rate in the current org queue.
• Keep compact reasoning fallbacks next, then broader catalog fallbacks, with rate-starved flagships last.
• Use a 90-minute per-model window (OPENCODE_RUN_TIMEOUT_SECONDS=5400) so deep reviews are not cut short.
• Bound the total model-pool retry budget to 300 minutes (OPENCODE_TOTAL_RETRY_BUDGET_SECONDS=18000) inside a 330-minute step, leaving time for the publish/approval gate to log the explicit reason when no model succeeds.
• Keep the approval gate running with always() so model-pool failure becomes visible review evidence instead of disappearing behind a skipped gate.
• Set scheduler dispatch defaults to process the backlog instead of only one PR per cadence loop.

Verification

• py -3 -m pytest tests/test_opencode_agent_contract.py tests/test_opencode_review_normalize_output.py tests/test_pr_review_fix_scheduler.py -> 56 passed, 1 skipped.
• git diff --check -> passed.
• codegraph sync . -> completed.
• bash -n scripts/ci/test_strix_quick_gate.sh did not complete within 120s on this Windows workspace; the edited assertions are covered by the Python contract test and static grep checks.

Queue cleanup

Closed superseded/contradictory OpenCode budget PRs:

• [codex] Restore long OpenCode review budget #281: dirty, changes requested, older long-budget branch.
• opencode-review: fail fast on hung model pool (350->30 min step timeout) #348: 30-minute fail-fast branch, contradicted by the latest review decision.
• fix(opencode-review): reach a healthy fallback model within the step budget #372: 30-minute per-model branch, contradicted by the latest review decision.

[seonghobae]

Superseded by #382. The 30-minute/10-minute concern is addressed there with DeepSeek V3 first, 90 minutes per model, and a 300-minute model-pool budget while preserving the safer merge-ref coverage checkout and least-privilege security workflow changes.

[seonghobae]

Local verification on current head 791685412186f25e6eab10f1de0e1b5a96ec85ef while GitHub-hosted Actions are queued:

• python -m pytest tests/test_opencode_agent_contract.py -q -> 12 passed, 1 skipped.
• python -m pytest tests/test_opencode_workflow_shell_syntax.py -q -> 1 passed.

No branch changes were made for this verification.

[Copilot]

Pull request overview

Note

Copilot couldn't run its full agentic review because no GitHub Actions runner was available. Make sure your repository has a runner available to run Copilot's review, or add a copilot-setup-steps.yml file specifying one with the runs-on attribute. See the docs for more details.

This PR adjusts the OpenCode review workflow to prevent org-wide merge blocking when model providers hang or exhaust capacity, by bounding model-pool retries and ensuring the publish/approval gate still runs to emit an explicit outcome.

Changes:

• Reorder the model candidate pool (DeepSeek V3 first) and add a total retry budget to keep the pool bounded.
• Allow the model-pool step to fail without aborting the job, and run the approval/publish gate unconditionally (always()).
• Change merge-scheduler dispatch defaults to process backlog (default -1) instead of dispatching a single PR per run.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

File	Description
tests/test_opencode_agent_contract.py	Updates contract expectations for model ordering, timeouts, and model-pool behavior.
scripts/ci/test_strix_quick_gate.sh	Aligns workflow assertions with the new timeouts, retry budget, and always-run publish gate.
.github/workflows/pr-review-merge-scheduler.yml	Changes scheduler input defaults to “process backlog” semantics.
.github/workflows/opencode-review.yml	Implements bounded model-pool timing/budget and ensures the publish/approval step runs even after pool failure.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.