Current capabilities
This page separates what ships today from what is planned. Cully's other pages describe only the first two groups.
Shipped means it works as described and has tests. Partial means it works with limits listed here. Planned means it does not exist yet; see product direction.
Observe
| Capability | Status | Notes |
|---|---|---|
| Cully terminal for Claude Code, Codex, Cursor CLI and any executable | Shipped | macOS and Linux. Not available on Windows. |
| Agent health bar | Shipped | Agent, project and branch, time, context, loops, unchecked edits. Segments with no data are omitted. |
| Tool events from hooks | Shipped | Claude Code and Codex hooks, and Cursor shell, file-edit and MCP hooks. Hooks do nothing outside the Cully terminal. |
| Cursor shell exit codes | Partial | Cursor's shell hook reports no exit code, so a failed Cursor command is not marked failed. Loop detection is weaker for Cursor. |
| Native footer (model, tokens, quota) | Partial | Codex only. |
| Context pressure | Partial | Codex and Claude Code. Other agents show none. |
| Cost and cache values | Planned | Shown only if an agent reports them. |
Understand
| Capability | Status | Notes |
|---|---|---|
| Session journal | Shipped | Local, bounded, no contents. See privacy. |
| Loop detection | Shipped | Rule based. Reports what repeated, never a cause. |
| Verification state | Shipped | Whether a check passed since the last edit. |
| Session health and a rule-based risk label | Shipped | The label names its reason. It is not a score. |
| Session task and progress | Partial | The task comes from cully task and the linked session record; a progress percentage is still planned because the journal has no measured source for it. |
Remember
| Capability | Status | Notes |
|---|---|---|
| Project memory through MCP | Shipped | PostgreSQL source records, Mem0 semantic recall, owner-scoped. |
cully_context bounded recall | Shipped | Up to five short previews. |
| Continuity hooks | Shipped | Claude Code and Codex start hooks; Claude Code and Cursor turn-end reminders. |
| Shared project learning | Partial | Private drafts, explicit publish/edit/unshare/delete, current-source lookup and maintainer adoption; no team-wide visibility, shared semantic indexing or learning worker. |
Advise
| Capability | Status | Notes |
|---|---|---|
| Advisor with context, search, failure and verification warnings | Shipped | Rule warnings work offline. |
| Model-backed advisor analysis | Partial | Uses the agent's own CLI in a headless run. Needs that CLI and its Cully MCP connection. |
| Clickable advisor in the terminal | Shipped | Click the panel, or Ctrl+] / F6. |
| Project workflow hint | Shipped | After three earlier sessions with edits where checks usually followed, it warns when the current session has unchecked edits. |
| Guardrails that block or approve actions | Planned |
Continue and review
| Capability | Status | Notes |
|---|---|---|
cully timeline | Shipped | Grouped history of a session. |
cully replay | Shipped | Step player, file activity, git reconcile, JSON and offline HTML export. File paths come from structured file tools. The interactive view and the HTML page have been exercised in a pseudo-terminal and a stubbed DOM, not yet in a range of terminals and browsers. |
cully handoff | Shipped | A structured handoff you can print or start another agent with. |
cully rescue | Shipped | Evidence and recovery steps; an optional headless advisor adds a probable cause. |
cully status session health | Shipped | |
| Session forensics | Shipped | Flight-recorder postmortems, audit-grade session review and replayable playbooks, composed of timeline, replay, reconcile and handoff. |
Coordinate project work
| Capability | Status | Notes |
|---|---|---|
| Team/project access and stable tasks | Shipped | OAuth principals; explicit project roles; additive migration 006. Private memory stays separate. |
| Board and decision inbox | Partial | CLI/MCP project views for tasks, blockers, reviews and stale work. No web view. |
| Atomic claims and attempt leases | Shipped | Current-version writes; one claim winner; 30-minute lease and explicit reclaim. Manual agent launch. |
| Portable team checkpoint | Shipped | Selected context, next step and branch/artifact references; no automatic journal uploads or process transfer. |
| Evidence-backed task acceptance | Partial | Reported evidence for each criterion, exact revision/version and noncontributing maintainer approval. No independent CI attestation or configurable completion policy. |
| Maintained project playbooks | Partial | Explicit adoption from completed tasks; withdrawal invalidates retrieval. Editing/archive controls remain planned. |
| Audit events | Partial | Transactional database events for operators; no user-facing audit query/export. |
| Managed agent execution | Planned | No autonomous dispatch, enforced tool budgets or automatic cancellation. |
See team workflows for inputs, limits and roles.
Improve
| Capability | Status | Notes |
|---|---|---|
| Workflow hint from past sessions | Shipped | Narrow on purpose. |
| Learned team workflow | Planned | |
| Agent performance comparison | Planned | Needs measured outcomes first. See product direction. |
What Cully does not claim
- It has not measured token or time savings. A benchmark is on the roadmap.
- It does not read transcripts or send them anywhere.
- It does not see inside an agent's reasoning. It sees tool calls and results.
- A missing signal is shown as missing, never guessed.