Skip to content

Gap analysis: spec against the code

Reviewed against main at 8cd34d5 (the conversation workspace, #64) on October 24, 2026. Verdicts: Built, Partial, Not built, Diverges (built differently on purpose). Paths are relative to the repository root.

Built as specified

Spec item Evidence
Centered conversation title; project left; Context, People, Settings right; title on its own row on a phone apps/web/src/ui.tsx (Page conversation layout), apps/web/src/ConversationShell.tsx, apps/web/src/styles.css
Conversations grouped by project with the agents that took part apps/web/src/Sidebar.tsx
Activity replaces execution history, opens from conversation settings, shows handoffs, subagents and tools, and has its own route apps/web/src/Activity.tsx, apps/web/src/route.ts
Clarifications separate from approvals packages/components/src/components.tsx (ClarificationCard, ApprovalCard)
Versioned deliverables with previews, presentation and revision requests apps/web/src/Deliverables.tsx, apps/web/src/artifactPreview.ts
Context provenance: origin and real import time, honestly labelled Indexed apps/web/src/ConversationShell.tsx
Project tree with kinds and sub-projects; instructions, notes, agents, skills, knowledge and experts inherited down pkg/projects, internal/platform/projects.go
Tighten-only limits, usage ledger, admission, budgets pkg/policy, pkg/usage, apps/web/src/Billing.tsx
Memory approval queue with provenance and forget pkg/memory, apps/web/src/AgentDetail.tsx
Slack channel bound to a project channel, two-way, Slack users mapped to people pkg/projects/channels.go, pkg/clients
One way external systems wrap in: one surface per scheme, tenant from the client record, every run through Invoke internal/platform/connectors.go
Create-agent form; personal credentials apps/web/src/AgentDetail.tsx, internal/platform/credentials.go
Monitor and alert rules apps/web/src/Monitor.tsx, pkg/alerts

Partial

Spec item What exists Missing
Needs you Inbox with approvals, questions, failures, alerts, budget, memory, mentions One item per cause (failures are per run); stale-context items
Agent roster Working, waiting, idle; home project; run tape Paused and failed states; every project that enables the agent; last run time
Agent autonomy Per-tool: listed, in approvals, or absent; callee call policy Workspace or project Allow, Ask, Never by action category, inherited tighten-only; reminders; stop rules; manager field
Handoffs pkg/call modes, depth and fan-out limits, typed inputs and validated output Due date, contract states (Blocked, Needs a person), cycle detection
Skills Workspace and project skills inherited down Personal scope, per-skill versions, provenance, promotion, console editor
Memory Agent and project notes, approval, provenance Expiry, personal memory, per-note scope
Audit Policy, project and agent histories with before and after; break-glass reads One cross-domain log; memory and call-policy changes
Sharing and People Share inspector; read-only participant list Who can see this; roles; removal
Evidence chips Bounded citation projection on run records and streaming events, including memory/Postgres restart Source lifecycle health is tracked separately
Thread branches Regenerate creates sibling answers with a switcher Fork into a new conversation; branches in rooms
GitHub Installations, webhooks, reviews, labels Repo-to-project links; GitHub user mapping
Live context Person-scoped MCP tool providers First-party live search, OAuth, hybrid mode
Composer One-line initial height, bounded growth, wrapping controls Existing text-control styling is retained
Workspace message presentation Shared persistent Copy/Reply footer for people and agents; recorded agent timestamps with focus, hover and tap detail; answer before expandable work Structured outcomes and Redirect

Not built

Spec item Note
Structured outcomes Built through shared primitives and the explicit record_outcome tool; Redirect prepares editable human intent through the normal conversation path
Moving a conversation after agents ran Refused by design (ErrStaysInProject in internal/platform/threads.go); see company-workspace.md
Durable assignments Listed as next work in conversation-workspace.md
Pause an agent, pause all Console Stop Turn calls the existing stop API and waits for the run ledger, with bounded polling and retryable errors
Context health and source lifecycle pkg/knowledge sources have no state, attempts or last good copy
Agent network view No component
Dry run for new agents Evals run only after saving
Initiatives with cross-child agents; status rollups Counts and unread are per project only
Project fork No route
Integration auto-linking rules and suggestions Links are manual (Slack) or the client's home (GitHub); no Slack binding UI in the console
Google Drive and Notion connectors Knowledge arrives as text through the API; retrieval is lexical
Quick switcher No Cmd-K
Notification preferences, personal agents, privacy defaults Not in pkg/prefs

Diverges on purpose

Spec item Repo rule
Agents as initials in rounded squares AGENTS.md: agents are told apart by color (AgentMark), not letters in boxes
Composer chips as pills AGENTS.md: controls under the composer are text
Outcome sections as web-only markup Shared components are built from primitives and render in Slack too

Suggested order

  1. Answer before work steps, then outcome structure with Redirect, through the shared primitives.
  2. Pause and stop: agent pause, pause all, the console stop control.
  3. Source lifecycle records: the base for Context health, stale-context items and Drive ingestion.
  4. Durable assignments.
  5. Conversation re-homing after runs, per company-workspace.md.
  6. Repo-to-project links and a Slack binding screen.