Your planning board,
running itself.

Sully watches your board and spawns headless Claude Code sessions that plan, implement, open PRs for, and review each ticket — every ticket in its own git worktree, with results posted back to your board and GitHub.

Coming soonSee how it works

2 runningweek38%

Board

watching your board columns and acting on them

10 tickets tracked Deploy New ticket

Uncategorized

NOVA-118nova-web3h ago

Audit unused feature flags and delete the dead branches

⋯

NOVA-121nova-api1d ago

Bump Node to 22 in CI

⋯

Planning

NOVA-119nova-api$0.061m ago

Retry failed webhook deliveries with exponential backoff

reading src/billing/invoice.ts — tracing how delivery failures surface today

working

Stop⋯

NOVA-122nova-web$0.116m ago

Split the checkout form into its own route

? Should the retry live in the worker or the API?

has questions

⋯

Answer questions (2)

Plan ready

NOVA-115nova-billing$1.2412m ago

Move invoice PDF rendering off the request path

⋯

Review plan

In progress

NOVA-124nova-web$0.41just now

Inline card validation on the checkout form

editing components/CheckoutForm.tsx — wiring the field-level error state

working

Stop⋯

NOVA-112nova-api$0.6322m ago

Seed script for the staging dataset

http://localhost:3000

⋯

In review

NOVA-117nova-web$0.2231m ago

Dark mode pass on the settings panel

✓ CI○ review◆ figma 2

⋯

NOVA-108nova-billing$0.8748m ago

Cache the plan comparison table at the edge

✕ CI± changes

⋯

View GitHub review (3)

NOVA-104localnova-web$0.552h ago

Add an audit log to the admin console

✓ CI✓ approved

⋯

agent chatNOVA-119 Chat Terminal

Before you plan this — how are we retrying today?

▸Grep "webhook_failures" src/▸Read src/api/webhooks/send.ts

We are not. A failed delivery is logged and dropped: send() catches the error, writes a row to webhook_failures, and returns. Nothing reads that table.

Two ways to retry: an in-process timer in the API, or a durable job on the worker queue. The worker already has at-least-once delivery and a dead-letter table, so I would reuse it rather than add a second retry path.

Before I write the plan — should failed deliveries keep their original order per endpoint?

Reuse the worker. Order does not matter, but cap it at 6 attempts.

Worker queue, 6 attempts, no ordering guarantee. Writing that into .sully/plan.md now.

.sully/plan.md

+## Retry policy

+- Reuse the worker queue — at-least-once, existing dead-letter table

+- Cap at 6 attempts with exponential backoff (30s → 16m)

+- No per-endpoint ordering guarantee

thinking…

the agent is working — Enter queues your message, Esc interrupts and sends the queue Stop

How it works

Five steps from ticket to pull request.

1
A ticket lands in a mapped column
You map your board columns to phases — planning, in progress, in review. Use Linear, or Sully’s own local tickets if you would rather not connect anything. Moving a card is the only trigger Sully needs.
2
Sully cuts a worktree and a branch
Every ticket gets a dedicated git worktree and branch, so parallel work never collides and your main checkout stays untouched.
3
A headless agent runs the phase
A Claude Code (or Codex) session runs the phase the column implies: writing the plan, implementing the change, creating the PR, or reviewing it.
4
Plans land in .sully/plan.md for review
You read the plan before a line of code is written. Agents can ask blocking questions, and you answer them in the in-app ticket chat.
5
Sully advances it to a merged PR
Coding, PR creation, and review follow as the ticket moves, with progress posted back to your board and GitHub along the way.

Sessions

Nothing runs off in the dark.

Every headless session Sully spawns shows up in one list with its phase, agent, model, elapsed time and spend. Open any of them to watch the normalized stream — tool calls, agent prose, test results — as it arrives, and stop it mid-flight if it is heading the wrong way.

Sessions

every headless command Sully runs, live and stoppable

running (2)

NOVA-124codingclaude · opus

editing components/CheckoutForm.tsx — wiring the field-level error state

running

2m 14s · $0.41

NOVA-119planningclaude · sonnet

reading src/billing/invoice.ts — tracing how delivery failures surface today

running

48s · $0.06

history

✓

NOVA-117pr reviewclaude · opus

posted 4 review comments — 2 blocking, 2 nits

done

1m 02s · $0.22

✓

NOVA-115create prclaude · haiku

opened nova-billing#412 — Move invoice PDF rendering off the request path

done

34s · $0.03

NOVA-108repromptcodex · gpt-5

timed out after 20m waiting on the failing Actions log

error

20m 00s · $0.19

NOVA-124coding

claude -p --output-format stream-json --permission-mode acceptEdits --model opus

running · 2m 14s Stop Raw

session started · claude opus · cwd nova-web-worktrees/nova-124

Implement the approved plan at .sully/plan.md for ticket NOVA-124. Verify your work.

▸ Read .sully/plan.md

▸ Read components/CheckoutForm.tsx

▸ Read lib/validation/card.ts

The plan has three steps: field-level validation, a shared error component, and a test for each branch. Starting with the validator so the form has something to call.

▸ Bash npm run lint

✓ no lint errors

The form only validates on submit, so a bad card number surfaces after a round trip. The plan moves validation to the field level and keeps the submit guard as a backstop.

▸ Grep "validateCard" src/

▸ Edit components/CheckoutForm.tsx

▸ Edit lib/validation/card.ts

▸ Bash npm run test -- checkout

✓ 14 passed · 0 failed (2.8s)

Wiring the field-level error state into the shared input compone

Phase configuration

Every phase gets its own agent, model, and budget.

Plan with Opus and a planning skill, implement with Sonnet, open the PR with Haiku, review with Codex. Each phase carries its own injected prompt, permission mode, timeout, and MCP server subset, so you are not paying frontier-model prices to write a commit message.

Planning
Coding
Create PR
Commit & Push
Address Comments
Figma Comments
PR Review
Error Investigation

Sully — Settings

Phase configuration

Planning

drafts the implementation plan

Agentclaude (default)

ModelOpus (latest)

Skill (blank = built-in prompt)/plan:create

Injected prompt (prepended to every session)Sibling repos live under ~/dev — read the API contract before planning.

Permissionsaccept edits

Timeout (min)30

MCP serverschoose servers

figmalinearposthog

Coding

implements the approved plan

Agentclaude (default)

ModelSonnet (latest)

Skill (blank = built-in prompt)built-in prompt

Injected prompt (prepended to every session)extra context for this phase, e.g. sibling repo paths

Permissionsbypass all (yolo)

Timeout (min)45

MCP serversnone (save tokens)

Create PR

commits, pushes, and opens the PR

Agentclaude (default)

ModelHaiku (latest)

Skill (blank = built-in prompt)/gh:create-pr

Injected prompt (prepended to every session)extra context for this phase, e.g. sibling repo paths

Permissionsaccept edits

Timeout (min)10

MCP serversnone (save tokens)

PR Review

reviews teammates’ PRs

Agentcodex

Modelgpt-5.1-codex-max

Skillnot supported by codex

Injected prompt (prepended to every session)extra context for this phase, e.g. sibling repo paths

Permissionsaccept edits

Timeout (min)20

MCP serversmanaged by codex

Watchers

It also watches the things you forget to.

Two background watchers run whether or not a ticket is moving. One picks up every PR where you are a requested reviewer and posts a real review on GitHub. The other stays on the PRs Sully opened: when a check goes red or a reviewer leaves comments, it reads the failing job log, reprompts the agent, and pushes the fix.

PR Reviews

auto-reviews PRs where you are requested or assigned

watching

Cache the plan comparison table at the edge

nova-billing #409 · by priya · 4m ago

reviewing

✓

Move invoice PDF rendering off the request path

nova-billing #412 · by sully · 26m ago

changeschanges requested

✓

Add an audit log to the admin console

nova-web #401 · by marcus · 2h ago

approved

✓

Drop the legacy /v1 webhook receiver

nova-api #397 · by dana · 5h ago

commented

reads failing checks and review comments, then reprompts the agent

auto-fix

Move invoice PDF rendering off the request path

nova-billing #412 · sully/nova-412-invoice-pdf-worker

1 failing

✓lint41s

✓typecheck1m 12s

✕unit tests3 failing · billing/tax.test.ts2m 04s

·e2e—

fixing unit tests + 3 review comments

read 214 log lines · editing billing/tax.ts

attempt 1 of 3

Features

Everything the loop needs, in one window.

Board-driven orchestration

Column-to-phase mapping drives the whole workflow, with per-ticket opt-in via labels so nothing runs that you did not ask for.

Git worktree isolation

Each ticket gets its own worktree and branch.

Plan review and ticket chat

Review agent plans and answer agent questions without leaving the app.

Automated PR creation and review

Creates PRs, and watches for PRs where you are a requested reviewer — spawning review sessions automatically.

CI auto-fix

Reads failing GitHub Actions logs and PR review comments, then reprompts the agent to fix them.

Production error investigation

Pulls errors from PostHog error tracking and spawns investigation sessions.

Embedded terminals and browser

Split and dockable terminals, a browser panel, and a dev server runner.

One-click deploys

Run your own release commands per repo.

Usage tracking and doctor

Per-session budgets, a live usage bar, and diagnostics for your CLIs, tokens, and MCP servers.

In the loop

Answer the agent. Drive the rest from the keyboard.

When an agent hits a decision it should not make alone, it asks — in a thread attached to the ticket, resumable across app restarts. Everything else is one keystroke away: ⌘K searches tickets, sessions, reviews, terminals and browser tabs at once.

NOVA-119agent chat

Before you plan this — how are we retrying today?

We are not. A failed delivery is logged and dropped: send() catches the error, writes a row to webhook_failures, and returns. Nothing reads that table.

Before I write the plan — should failed deliveries keep their original order per endpoint?

Reuse the worker. Order does not matter, but cap it at 6 attempts.

Worker queue, 6 attempts, no ordering guarantee. Writing that into .sully/plan.md now.

Edit .sully/plan.md

the agent is working… Send

Command palette — search anything, run actions: ⌘K
Collapse or expand the sidebar: ⌘B
Toggle the browser panel: ⌘⇧B
Split the terminal pane side by side: ⌘D

Integrations

It plugs into the tools already on your machine.

Linear: Polls tickets by column, moves states, and reads and posts comments through the GraphQL API. Optional — Sully also ships local tickets that live in your repo.
GitHubgh: Creates PRs, checks CI status, fetches failing Actions logs, and reads and posts PR review comments.
Claude Codeclaude: The main agent for planning, coding, PR creation, review, and error investigation.
OpenAI Codexcodex: Optional alternative agent, configurable per phase.
MCP servers: Per-phase server selection from ~/.claude.json, with OAuth re-login support.
Figma: Injects a Figma token into sessions so agents can read design comments.
PostHog: Queries error tracking to feed production error investigation sessions.
Git: Worktrees, branch sync, and mapping repos to local paths.
CI providers: GitHub Actions log fetching, plus CircleCI and Vercel status through the PR check rollup.

Stop babysitting tickets.

Sully is still in private development. The first builds go out soon — watch a ticket go from plan to PR without babysitting it.

Coming soon

Your planning board,running itself.

Five steps from ticket to pull request.

A ticket lands in a mapped column

Sully cuts a worktree and a branch

A headless agent runs the phase

Plans land in .sully/plan.md for review

Sully advances it to a merged PR

Nothing runs off in the dark.

Every phase gets its own agent, model, and budget.

It also watches the things you forget to.

Everything the loop needs, in one window.

Board-driven orchestration

Git worktree isolation

Plan review and ticket chat

Automated PR creation and review

CI auto-fix

Production error investigation

Embedded terminals and browser

One-click deploys

Usage tracking and doctor

Answer the agent. Drive the rest from the keyboard.

It plugs into the tools already on your machine.

Stop babysitting tickets.

Your planning board,
running itself.