The plan your coding agent writes, and the engine that runs it
One plain-English request becomes an ordered pipeline: investigate, plan, file, task, execute, commit. Every artifact is a markdown file in your repository, written before the code it describes.
brew install FlowChargeApp/flowcharge/flowchargeTested in Claude Code, OpenCode and OpenAI Codex. It runs in any coding agent that loads skill folders.
Live · updated 14:02:11The reports page shows each report as a table, with no way to take the rows out. Add an Export CSV button that downloads the rows the current filter shows.
reports×Summary
Each report renders from GET /v1/reports/:id. A new endpoint streams the same rows as CSV, and a button on the reports page calls it with the active filter.
Stages
- Serialise report rows to CSV.
- Add the export endpoint.
- Add the Export CSV button to the reports page.
Acceptance criteria
- The export holds exactly the rows the current filter shows, in the same order.
- A value that contains a comma, a quote or a line break survives a round trip.
- A report of 100,000 rows exports without loading every row into memory.
- The button is disabled while an export is in progress.
- The file name carries the report name and the export date.
Plan stage 1. Add the encoder that turns report rows into CSV text, with the escaping the acceptance criteria require.
The work moves. The board moves with it.
FlowCharge reads the flowcharge/ folder in any project you point it at, and shows every workstream as a card carrying its plan, its issues and its task progress.
Nothing on the board is typed in. It is a view of what your agent wrote, so it changes when the work changes. You never refresh it.
Every project carries its own record
Every workstream, plan, issue list and task list is a markdown file with YAML frontmatter, under flowcharge/ in your project. You open them, read them and edit them like any other file, and they commit with the code they describe.
They belong to the repository rather than to the tool. Clone the project on another machine and the whole record comes with it. Remove FlowCharge and the files stay exactly where they are.
You say it once. It writes the whole run.
Type /flowcharge once to load the orchestrator, then type your request in plain English. FlowCharge runs it as an ordered pipeline. The investigate stage returns its findings, and each stage after it writes a file into your repository before the next one starts.
Create a workstream to add a CSV export to the reports page.
Write up a plan.
Open tasks.
Execute all tasks, committing after each task.- 1. Investigate. The agent reads the code and returns what it found.
- 2. Plan. The findings become a plan with its own ID.
- 3. File. Anything that needs its own record becomes an issue.
- 4. Task. The plan becomes a task list, one task per unit of work.
- 5. Execute. The same agent session runs each task in turn.
- 6. Commit. Each task lands in its own commit, with the record beside the code.
The files are the prompt
Those artifacts are not a record you read afterwards. They are what the agent works from. Context is loaded before the run.
A task carries nine fields: what to do, the exact anchor to edit, the pattern to follow, the imports it needs, how it stays compatible with what already exists, the trap it must avoid, the commands that prove it worked, a checklist to answer before it moves on, and its own assessment. That is one task, for one small change.
Prompting well is a skill. FlowCharge does that work for you and leaves the result on disk.
One key decides how much it asks you
One key in flowcharge/agents.md decides how often a run comes back to you.
prompts: manual | assist | cruisemanual is the default. Every open question reaches you, and the run checks with you before it executes and again before it commits. assist settles the routine questions on your behalf and brings you the rest. cruise settles almost everything and writes down the risky decisions.
Under all three sits a floor you do not set. A question with no recommendation reaches you. Anything that could lose work, data or history reaches you. A task that aborts, a checklist item that stays failed or a stage that cannot complete stops the run and says why. A failing test run is the one exception: the run tries two rounds of fixes before it stops.
Turn the key as far as you like. Some things still come to you.
Four more things it does quietly
In the middle of an ordinary conversation, ask for a plan, an issue list or a task list, and get one. Nothing to launch and nothing to switch into.
Asked to record a commit SHA in a folder that was not a git repository, it wrote down that it could not, and why. A value it cannot verify is a value it will not invent.
Each workstream, plan, issue list, task list and issue carries an ID with a random six-character suffix. Two clones of the same project never collide, and no ID is ever reused.
FlowCharge writes the Core skill files into Claude Code, Cursor, Windsurf, OpenCode and Codex, and tracks which version each one is running.
The engine is open. FlowCharge is free.
FlowCharge Core is the engine: the skill and script files your coding agent loads. It is open source under the MIT license, it needs git and Node 16 or newer, and there is nothing else to install.
FlowCharge is the application that installs Core, tracks its versions and shows the board. It is free to use, and it runs on macOS and Linux (Windows version coming soon).
It sits under your tracker rather than across from it. Keep Jira, Linear or GitHub Issues for the work the whole company can see, and let FlowCharge hold the detail your agent works from.
Questions
What it needs, what it writes, and how it is licensed.
Install FlowCharge
brew install FlowChargeApp/flowcharge/flowchargeThen give it its first run.