[ada-cl25]

Summary

Build on the backend infrastructure merged in #3, within packages/dsh-triton-riscv:

• Add versioned parent/child memory records and asynchronous embedding indexing through the existing MySQL Outbox and RabbitMQ workers. Support Qdrant backfill, activation, retirement and cleanup.
• Add opt-in Qdrant retrieval with workspace/version/checksum validation. Expand matching evidence with required background, stage-specific evidence and bounded adjacent fragments within the existing context budget.
• Add an experimental main/semantic/implementation/review workflow. Roles produce validated proposals in isolated Harness sessions without executable tools; the host retains approval, write and validation authority.
• Add workbench workflow controls, revision-bound approvals, asynchronous validation collection and bounded repair continuation. Repairs preserve their frozen acceptance tests.
• Fix numerical-failure diagnosis when sandbox metadata happens to contain compiler paths, and distinguish original source from proposed source during independent review.

Related Issue

Follow-up to #3. No issue is closed by this PR.

Validation

Vitest is the single automated test entry point. Existing backend assertions remain behind the established Vitest-to-pytest bridge; direct Vitest coverage checks proposal-role tool denial, SDK credential forwarding, and includes the frontend suite in repository CI. The former Python YAML/JavaScript configuration check is migrated to Vitest with an explicit JS parser dependency, removing the undeclared PyYAML dependency found in the first CI run. This PR does not add a second CI test runner.

With disposable MySQL, Redis, RabbitMQ and Qdrant services configured as described in the plugin README:

pre-commit run --all-files
pnpm test:all --reporter=verbose
cd packages/dsh-triton-riscv
npm test
npm run test:frontend
npm run build
npm run check:package

Local results:

• Repository Vitest: 212 passed across 22 test files.
• Plugin Vitest: 166 passed, including the backend and frontend integration wrappers.
• Backend regressions driven by Vitest: 804 passed; frontend Vitest: 16 passed. One former Python check now runs directly in Vitest. These are nested results, not additional top-level Vitest tests.
• Build, package audit and repository pre-commit checks passed. Package audit excludes runtime state and private artifacts.
• Separately authorized live smoke testing on RISC-V/Triton 3.4.0: a new operator passed 9 tests; a controlled faulty operator went from 8 failures to 8 passes without changing its tests. Five intermediate proposal-stage blocks were retained during development. This is not a first-attempt success-rate or load-test claim.

CI adds a disposable Qdrant service. Automated CI tests require neither model credentials nor access to the private RISC-V server.

Latest GitHub CI run for 5b8754d: both pre-commit and vitest passed.

Compatibility and Limits

• Apply additive schema migrations 0003 through 0005 after backing up the database.
• Vector indexing/retrieval and proposal workflows are opt-in. Normal chat and legacy retrieval defaults remain unchanged. Backfill and generation activation are explicit operations.
• Live role calls require configured model credentials and queue.proposalWorkflows; proposals never bypass write or validation approvals.
• Vector reads currently bypass Redis. The evidence budget is measured in Unicode characters, not model tokens. Live smoke tests do not establish retrieval-quality gains or production concurrency guarantees.
• No credentials, private reports, model weights, runtime databases or generated operator experiments are included.

Checklist

• Changes are limited to the plugin and its CI service configuration.
• Followed the repository's Vitest-based testing convention.
• Preserved backend regression coverage and added direct Vitest checks.
• Updated user-facing setup and configuration instructions.
• Ran formatting, build, packaging and relevant test checks locally.
• Documented migration requirements, opt-in behavior and known limitations.

[shirohasuki]

Sorry for the late review, and thank you so much for the quick update!

That said, I would suggest separating vector retrieval and the multi-agent framework into different plugins, rather than integrating both into triton-riscv.

Decoupling application logic from the framework is, after all, the whole point of a plugin-based architecture.