[ada-cl25]

Summary

Follow up on #5 with approval-gated parallel analysis, optional model routing, and workspace artifact access. Changes stay within dsh-multi-agent and dsh-triton-riscv; the existing plugin separation is preserved.

• Add reusable bounded parallel execution and fast/strong routing primitives to dsh-multi-agent.
• Let the main agent propose read-only replicas. Show goals, context and models for approval before starting; declining parallel investigation continues the structured workflow sequentially. Code changes and validation still require their own lifecycle approvals.
• Route bounded information tasks to an explicitly configured fast model; retain the strong/current model for implementation, review and ambiguous work. Routing is off by default, with no automatic paid retry on errors.
• Add account-scoped links to operator code, tests, patches, receipts and logs. File previews are read-only, size-bounded and protected against traversal, symlinks and cross-account tab reuse.
• Fix approval-card/revision handling, bounded scalar-shape discovery, provider-error reporting, and proposal publication before worker-lease release.
• Keep native cancellation alive until replicas are reaped, use the effective host model after switching/restoring sessions, and require fresh approval when expired or configuration-stale plans are replaced.
• Align registration, password changes and the frontend on a 6-character minimum. This intentionally relaxed policy is for the trusted workbench, not a security improvement or a public-deployment recommendation.

Related Issue

Follow-up to #5. No separate issue.

Validation

GitHub CI run 42586974576 passed both pre-commit and Vitest on 68c5536. Linux CI also reports 80 files passed, 584 tests passed and 5 skipped. The results below describe the separate local run.

• pre-commit run --all-files --show-diff-on-failure: passed.
• Root Vitest (the pnpm test:all entry point): 80 files passed; 584 tests passed, 5 skipped with disposable MySQL, Redis, RabbitMQ and Qdrant.
• The root run includes the existing Python bridge: 66 modules, 800 Python tests passed, 3 conditional skips, and the frontend Vitest wrapper (147 tests passed). These nested counts are not additional root tests.
• npm run build and npm run check:package in packages/dsh-triton-riscv: passed.
• Targeted tests cover cancellation/reaping and quota release, effective native model selection, renewed approval after expiry/config changes, sequential fallback, routing, password boundaries and account-bound artifact access.

The five root skips require an optional historical prompt-comparison baseline. The three Python skips require Redis Bloom commands, a separate Redis restart executable, or private broker restart controls. No tests were removed or newly disabled to obtain a pass.

Scope And Limitations

• New regression cases use Vitest. Existing Python tests continue through the repository's Vitest wrapper; no second test entry point is introduced.
• Local verification uses disposable services, stubbed model responses and harmless process fixtures. It does not call a paid model or sg2044, or establish real-operator quality improvements.
• Replica concurrency is bounded locally, not by a cluster-wide scheduler. Replica findings do not count as successful compilation or validation.
• This does not add automatic repair continuation to the ordinary chat path or claim to resolve all failures from prior operator experiments.
• No credentials, personal databases, generated operator artifacts, local experiment reports or submodule updates are included.

Checklist

• The change is focused and contains no unrelated modifications.
• I have followed the target repository's contribution guidelines.
• I have added or updated tests where applicable.
• I have updated related documentation where applicable.
• I have run the relevant formatting, lint, build, and test checks.
• I have described known limitations or compatibility impact.