Sync Sakura AI Engine models automatically

[dceoy]

Summary

• add a shell-based Sakura AI Engine model discovery/sync script
• probe /v1/chat/completions to keep only OpenCode-usable chat models
• add a daily/manual workflow that opens or updates a PR when the model set changes
• use the repository-provided GITHUB_TOKEN with contents: write and pull-requests: write to create sync PRs
• fail closed on invalid model listings, authentication/rate-limit errors, transport errors, and server errors
• add Bats coverage for model filtering/order and empty-list safety

Configuration

The scheduled workflow requires only the repository secret SAKURA_AI_ENGINE_API_KEY. It uses the workflow GITHUB_TOKEN for repository writes and pull request creation; no GH_TOKEN or personal access token is required.

Because GitHub suppresses new workflow runs for events triggered by GITHUB_TOKEN, sync PRs created by this workflow do not automatically trigger pull_request workflows.

Validation

• bash -n scripts/sync-sakura-models.sh
• mocked normal sync updates only the Sakura models block and preserves the rest of opencode.jsonc
• mocked empty /v1/models response exits non-zero without modifying the config

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit:e7b2ae35ef

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[claude]

Consolidated review from two automated passes (pr-review-toolkit: code/comments/tests/errors/types, run in parallel, plus Claude Code's security-review with adversarial false-positive filtering). All findings below were independently verified against the current PR diff and surrounding code (including cross-checking .github/workflows/opencode-bot.yml and run-opencode.sh). No high-confidence security vulnerabilities survived the adversarial filtering pass — two candidates (an awk brace-depth desync from crafted model IDs, and unsanitized model IDs in log output) were both excluded as speculative / not concrete security issues (the latter also falls under standard "log spoofing is not a vulnerability" guidance), though the underlying brace-depth fragility is noted below as a correctness issue. Remaining items are correctness/robustness and test-coverage gaps, left as inline comments. Nice to see the fail-closed design intent (empty-listing guard, set -euo pipefail, pinned action SHAs, scoped workflow permissions, no pull_request trigger) — the comments below are refinements on top of that, not blockers to the overall approach.

[opencode-agent]

The trusted-context validation gate has failed:

::error::Pinned review context is unavailable or invalid, or the PR head changed.

Per the pr-review skill's fail-closed rules ("Once context succeeds, any later metadata, diff, or validation failure must abort the review"; "If validation fails, stop"), I am aborting this review run.

Review status: aborted — no findings published.

• Discovery and independent validation completed across all 3 changed files (.github/workflows/sync-sakura-models.yml, scripts/sync-sakura-models.sh, tests/test-sync-sakura-models.bats) against pinned head e7b2ae35efe8dbe0387c559c79ac99cac71d5bcd.
• 5 candidate findings passed independent validation (2 important: string-blind brace counting in the awk JSONC splice, and curl --retry corrupting captured status/body on recovered transients; 3 suggestions: multi-"models"-block replacement scope, wasted auth-failure retries, and an uncovered empty-model-list guard).
• Because the pre-submission validation gate now reports the pinned review context as unavailable/invalid or the head SHA as changed, submitting inline findings can no longer be safely anchored to a verified head commit, so nothing was written to the pull request.

opencode session | github run

[dceoy]

Reviewed the current head (cdb79441a1b96f1b0b40f832aa34de7d16622898). The earlier provider-scoping, partial-extraction, default-model, and failure-path coverage issues are addressed. One correctness issue remains in the probe classification below.

[dceoy]

Found one merge-blocking functional issue in the production model-classification path. CI is green, but the normal mock does not reproduce Sakura's mixed model listing.

[dceoy]

Reviewed current head 52cc515aee909b7e6457fefc9e6d88159bd5ccd6. The previously reported provider-scoping, partial-extraction, default-model, ambiguous-4xx, and mixed non-chat-model issues are addressed, and CI is green. I found one remaining operational correctness issue: the documented minimal configuration can create sync PRs whose CI does not run automatically.

[dceoy]

Found one configuration issue that should be fixed before relying on the scheduled sync. CI is currently green, and the previously reported model-filtering/default-model issues are addressed in the current head.

[dceoy]

Reviewed current head 145d1164a0be5ccb4087d6676d00ca1d3c026620. CI is green and the previously reported functional issues are addressed. I found one remaining least-privilege issue in the scheduled workflow.

[dceoy]

Reviewed current head ceb40d809d8e74cbf513d97dd8266bed6826e01a. The previously reported provider scoping, partial extraction, default-model consistency, ambiguous probe handling, mixed non-chat model handling, dedicated-token configuration, and least-privilege issues are addressed. CI is green, including Bats, shell lint, and GitHub Actions lint/scan. I found no additional high-confidence merge blockers in the current diff.