Create a single project-level slash command .claude/commands/test-e2e.md that uses agent-browser CLI to intelligently test dashboard pages affected by the current PR/branch. It auto-starts the dev server if needed, authenticates as the test user, maps changed files to routes, and tests each affected page.
feat: add /test-e2e command for agent-browser E2E testing
| # E2E Shard 2/4 Flaky Failure: Drizzle v1 "decoder" Bug | |
| ## Problem | |
| E2E Shard 2/4 fails intermittently across multiple PRs (confirmed on #6498 and #6548). | |
| The root cause is the drizzle-orm v1 (beta.15) "decoder" bug affecting E2E setup scripts. | |
| ## Error | |
| ``` | |
| ❌ Failed to grant feature flag: Unknown relational filter field: "decoder" | |
| ❌ Failed to set user context: Unknown relational filter field: "decoder" |
Add support for first-party connectors that bypass Unified.to, starting with a generic OAuth framework, then a provider-agnostic task management tool, then Todoist as the first native provider, and finally wiring Jira/Asana/etc through Unified.to's existing task API.
Each PR builds on the previous one. Each section contains enough context for an independent agent session to implement it.
These decisions were made during planning. Revisit if any feel wrong.
Credit source is per-workspace, user's choice — Each user picks "workspace credits" or "personal credits" per workspace. This is a toggle on their membership, not a global setting.
No fallback between sources — If a user picks "personal credits" and runs out, they get blocked (CREDITS_EXHAUSTED). They must buy more or switch back to workspace credits. No silent fallback.
After Phases 0-2, the codebase is on drizzle-orm@1.0.0-beta.15-859cf75 with:
db.query → v2 relational query API (new syntax)db._query → v1 compat shim (old syntax, currently used by all 54 files)Phase 3 migrates each file from db._query (v1) → db.query (v2), batch by batch.
Plans are hardcoded as a TypeScript union type and a static PLAN_METADATA object. Adding a new plan requires a code change and deploy. We need to support:
| # OBV-1846: Message Pagination — Remaining Work | |
| ## Problem | |
| Threads with >200 messages have older messages permanently inaccessible. | |
| Mirror hydration loads 200 messages per thread. There is no UI to load more. | |
| ## Architecture | |
| ### Current data flow | |
| 1. **Hydration** (`/hydrate/project/:scopeId`) loads 200 most recent messages per thread |
Completed in PR #6396: Steps 1–4 (WorkbookSessionContext, useSimpleWorkbookData decomposition, ColumnOperationsShared, orchestrator extraction).
Rename simple-workbook-table/ → workbook-table/ and reorganize hooks into subdirectories.
Why deferred: Touches 29+ files for mechanical import renaming. High conflict risk, no behavioral change. Best done during a quiet period with no other active workbook PRs.
Add workspace-scoped chat as a new capability alongside the existing project-scoped chat. This is purely additive -- project chat continues to work exactly as it does today. Workspace chat introduces a general-purpose assistant that operates without a fixed project context.
| Project Chat (existing) | Workspace Chat (new) |
|---|
| # Mirror Convenience Hooks Proposal | |
| ## Overview | |
| This proposal introduces three convenience hooks to reduce unnecessary re-renders and improve developer experience when working with Mirror data. These hooks cover the most common patterns found in the codebase that currently cause suboptimal re-render behavior. | |
| **Current state:** | |
| - `useMirrorList` and `useGlobalMirrorList` now use derived atoms with shallow array equality ✅ | |
| - Selector hooks (PR #3754) were reverted due to complexity concerns ❌ |