The TasksAI journey
A transparent view of what has shipped, what is being refined, and where the copilot is heading next. All sixteen milestones (M00–M15) landed as reviewed pull requests — the current build-out is the “creative & useful copilot” epic.
Milestones shipped
M00–M15 + the epic-#244 build-out · @asafarim/tasks-ai
Product charter, ADRs & taxonomy
ICP, jobs-to-be-done, KPI dictionary, event taxonomy, five architecture decision records, and the compliance / license gate.
Platform foundation
Next.js 16 shell, shared SSO via Hub, an isolated PostgreSQL database and Prisma client, a BullMQ worker skeleton, CI and test harnesses.
Multi-tenant work graph + /api/v1
Workspace / project / task / dependency model with row-level tenant scoping, RBAC, a versioned REST API with OpenAPI, idempotency and optimistic concurrency.
- Every mutation writes an ActivityEvent + OutboxEvent in one transaction
- IDOR / privilege-escalation / cross-tenant isolation tests
Fast task experience
Subtasks, dependencies, estimates, dates, labels, custom fields, recurrence and templates across Inbox, List, Board, Calendar and Timeline — no AI required.
Collaboration & realtime
Invitations and membership lifecycle, comments, @mentions, reactions, attachments, an in-app notification inbox and SSE-based presence.
Capture, search & portability
Global Postgres full-text search scoped by authorization, per-workspace inbound email capture, deterministic CSV/JSON import-export with dry-run and a deletion manifest.
AI safety & provider boundary
Provider-neutral lib/ai with a deterministic fixture provider for CI, redaction, prompt-injection fencing, an operation allowlist, per-workspace quotas, a kill switch and an offline eval suite.
- AI never mutates domain data — it only drafts a Proposal
- Disabling AI leaves core task management fully functional
AI copilot — intent to approved plan
Pasted notes / briefs / threads become an editable proposal diff grouped by create / update / link, with citations, partial accept, regenerate, undo and typed feedback.
Focus, risk & workload intelligence
Explainable, deterministic focus ranking and risk signals with per-factor evidence, plus an anti-surveillance guard enforced in code on every payload.
Automations & integration platform
A trigger-condition-action rules engine with dry run and execution log, scoped API tokens, signed webhooks, and the first (read-only) GitHub issue integration.
Goals, cycles, time & analytics
OKRs, cycles, capacity, time entries, portfolio dashboards and versioned forecasting with stated assumptions and drill-down to the underlying tasks.
PWA, accessibility, i18n & performance
WCAG 2.2 AA on critical journeys, an installable PWA with an offline mutation queue, en / nl / fr foundations and route-level performance budgets.
Security, privacy, admin & reliability
Threat model, CSP and security headers, workspace admin with audit export and break-glass, data-subject request workflows, and backup / restore rehearsals.
Design-partner beta tooling
Beta onboarding under explicit consent, a segmented metrics dashboard, a feedback triage board and the decision-report scaffold.
- Field work — recruiting design-partner teams — is still open
Billing & packaging
Free / Pro / Business / Enterprise plans, subscriptions, entitlements that fail safe, metered AI allowances and Stripe test-mode — inert until the commercial license is signed.
Enterprise readiness
SCIM provisioning, domain claiming, configurable retention, legal hold and audit streaming, plus a scale-validation and integration-catalog plan.
What's next
The “creative & useful copilot” epic (#244)
Cross-project duplicate detection
Surface duplicate and near-duplicate tasks across projects, built on the grounded retrieval that #232 landed.
Wider — but still bounded — proposal ops
Let the copilot propose labels, a status suggestion, a dependency edge and a due-date suggestion. Still confirm + undo + audit; still never assigns people or writes committed dates.
Auto-changelog on completion
Turn a closed cycle's completed tasks into grounded, editable release notes before publishing — the same page you are reading, generated.
A command palette that acts
Wire natural-language queries into real saved views, and add a read-only “ask” mode answered from retrieval with citations.