Skip to content

Latest commit

 

History

History
90 lines (69 loc) · 10.2 KB

current-limitations.md

File metadata and controls

90 lines (69 loc) · 10.2 KB

Current limitations

This page describes v0.3.0-preview.144, the latest published prerelease (November 19, 2026), plus changes merged to master since; the changelog lists those under Unreleased. The standalone Email App and its Runtime are released separately: use the compatibility.json attached to the Email release you install, the current author guide and the Email limitations. The preview.144 image is built with the Apps-enabled browser profile; the API flags that turn Apps on stay off until an operator sets them.

Last reconciled on November 11, 2026.

Deft is an alpha. It is suitable for technical evaluation, internal use, and controlled pilots where an operator can tolerate breaking changes and investigate failures.

Installation and upgrades

  • Fresh installs use pnpm db:push-full to apply the Drizzle schema plus supplemental search and index SQL.
  • Supported release-to-release upgrades use pnpm db:upgrade and begin at the v0.2.0-preview.1 schema baseline.
  • Named GHCR preview images are currently amd64-only.
  • Historical databases created before v0.2.0-preview.1 are not automatically adopted. The upgrader refuses unknown or incomplete schemas rather than mutating them.
  • Raw pnpm db:migrate remains unsupported; the release upgrader owns the checksum ledger and compatibility checks.
  • Schema migrations are forward-only; successful upgrades do not have an automatic downgrade. Operators should pin an image digest, back up Postgres and uploads, and rehearse restoring both with the previous image digest before updating.
  • Content exports are partial (wiki, notes, calendar); recovery requires database, uploads, and configuration backups.

Deployment contract

  • One workspace per self-hosted deployment is the supported v1 contract.
  • Docker Compose is the primary documented deployment path.
  • The API requires a long-running container host with WebSocket support; it is not suited to a request-only serverless runtime.
  • Self-hosters operate DNS, HTTPS, reverse proxy, backups, storage, monitoring, and incident response.
  • No uptime SLA or managed support contract is included with the repository.

AI and agents

  • Core workspace features run without a Deft AI provider. Defty reasoning needs one; personal MCP clients can use their own AI provider.
  • Personal MCP clients act as the authorizing user; their writes do not automatically enter the Agent Employee approval queue. See identity and approval boundaries.
  • Knowledge summaries can be wrong or outlive a correction to their source; review and correct durable knowledge and resulting tasks.
  • Self-hosted storage does not imply local AI processing: configured providers and permitted external clients can receive workspace content. See AI data flow.
  • Model output is probabilistic. Deft validates structured drafts and applies permission and approval rules, but cannot guarantee perfect interpretation.
  • Agent employees require a separately operated compatible runtime. Deft does not host every external runtime or tool.
  • A seeded demo identity is not proof of a live Hermes integration. Preview.141 is a core release and carries no Hermes certification.
  • A healthy token or connection does not guarantee that an external runtime is online, subscribed, or able to finish assigned work.
  • Autonomous behavior remains bounded by scopes, trust, approvals, action caps, provider availability, and customer configuration.
  • Defty's chat writes always wait for approval, at every trust level. The workspace trust level governs only Defty's approved plans and background work.
  • An agent employee's own chat messages (post, reply, direct message) run without approval at every trust level, within guard rails: space membership and visibility, the reply-storm detector, the daily action cap and the unhealthy (circuit breaker) state. Everything else an employee writes follows that employee's own trust level.
  • No trust level auto-runs a call classed as destructive. That includes a connected tool that declares itself destructive, tools whose names match a destructive name in any letter case, and remove_dependency. A destructive action held for approval is stored at the full tier. Each step of a Defty plan carries a destructive flag, and a destructive step needs its own approval.
  • The action lists, receipts, dashboard activity and inbox show a Defty action only to its requester and to owners and admins (guests: their own only). Only the requester, or an owner or admin who can still see the target, can undo it, and only while the task still has the value the action set. Defty posts one reply per trigger message. An agent action is reviewed by the member who requested it. Owners and admins review when there is no active human requester; they read another active member's request in activity and history. An approver assignment makes a member a reviewer only when this rule does. The Inbox approval list and its counts, Attention and the personal MCP approval tools use this rule for the approvals they list and count. An employee's wiki page update is applied only while the fields it writes are unchanged since it was proposed, and a page creation only at the slug fixed when proposed; otherwise the approval ends and nothing is written. A plan created for an agent employee binds to an active employee of the same workspace. Connected tools are available to members, admins and owners, not to guests. The daily standup and the meeting brief are written from prepared evidence with no tools, so they appear in their evidence-only form when the draft contains a link, approval or sign-off wording, or a task key that is not in the evidence. Meeting briefs go only to active members of the workspace the meeting belongs to. Run text (employee activity entries, diagnostics, plan step text) is limited to the people who may read it, which can be fewer than the people who see the entry; other members see the entry or status without the text.
  • Deft does not record what an employee's own runtime spends, so the per-employee daily cost budget (daily_budget_cents) never triggers. The daily action cap (max_daily_actions) is enforced; use it and your runtime's own limits to bound cost.
  • Defty does not turn chat messages into tasks on its own (passive extraction is off unless an operator sets DEFT_ENABLE_BACKGROUND_TASK_CAPTURE=1). Defty does not manage agent employees, connections or triggers.
  • Realtime rooms and presence are per workspace: a person who belongs to two workspaces shows as online only in the workspace where their session is connected.
  • A public space can be browsed and joined by members (not guests), but its messages cannot be previewed before joining. Leaving a space does not hand it to anyone: a private space whose last member leaves has no reader until its creator archives it.
  • Self-hosted Deft sends no email. Invites and password recovery are links an admin generates and shares.

MCP and external tools

  • Deft supports streamable HTTP MCP with OAuth or personal bearer-token access, depending on the client.
  • Client support varies by vendor, application, account tier, and connector policy.
  • Native Slack, Gmail, GitHub, Google Calendar OAuth, Linear, and Notion connectors are not part of the current self-hosted v1 promise.
  • External tools should be connected through a customer-owned agent runtime or MCP server.
  • Calendar subscriptions through ICS are read-only and are fetched only from public internet addresses on ports 80 and 443, except exact feed URLs a self-hosted operator lists in DEFT_CALENDAR_PRIVATE_URL_ALLOWLIST; native Deft calendar events are writable inside Deft.

Apps and Modules

Modules and Apps use different packaging and installation paths. Standalone Modules do not require the experimental Apps flags. Module v1 records are organization-wide for owners, admins, and members; guests are denied, and private rows, collections, and fields are not supported.

  • Apps are disabled by default. The API requires DEFT_APPS_ENABLED=true, and the web build requires NEXT_PUBLIC_FEATURE_APPS=true.
  • Connected App execution and bounded automations add further flags and operator review. They are included in preview.144, experimental, and disabled by default.
  • Internal Apps provide declarative records and native views. The Apps developer preview adds bounded trusted UI, external Runtime and governed source-sync contracts that need reviewed setup. Arbitrary custom UI, public portals and self-granted execution are not provided.
  • A package's requested permissions are review input. The host owns installation, grants, connector binding, activation, and execution authority.

Scale and performance

  • The project includes synthetic 60-person certification and substantial local/public-demo dogfooding.
  • That evidence is not a production benchmark, capacity guarantee, or substitute for load testing on the operator's hardware.
  • Sustained concurrency, large file volume, long retention, WebSocket fanout, notification bursts, and job backlog should be tested for each deployment.

Security and compliance

  • Deft has ongoing auth, org-isolation, space-membership, MCP, upload, and rich-content security tests.
  • Deft has not completed an independent security audit.
  • Deft does not claim SOC 2, ISO 27001, HIPAA, GDPR certification, DLP, eDiscovery, or an uptime SLA.
  • Operators remain responsible for regional privacy, retention, backup, access, and compliance obligations.

Product boundaries

  • Web application only; no native desktop or mobile app.
  • No offline-first mode.
  • Huddles are browser-based and are not a substitute for a dedicated large-group meeting platform.
  • Knowledge is link- and graph-oriented, not a full Notion-style block database system.
  • Task management does not currently target portfolio planning, sprint analytics, OKRs, billing, or a custom report builder.
  • Some lower-frequency settings and administrative flows have less dogfood depth than chat, tasks, knowledge, MCP, and agent workflows.

These limitations should shrink over time. Track direction in ROADMAP.md and release-specific changes in CHANGELOG.md.