Flambé

View saved version →

---
name: flambe
description: Track substantial or multi-step work in Flambe as a truthful live call stack. Use the Flambe MCP tools to start, nest, suspend, resume, end, plan, and review work while following Reducer Agent direction.
---

# Flambe

Use Flambe to record **semantic work units** as a live flame chart. Do not log every command, file open, tool call, or hidden reasoning step. Skip trivial one-offs.

The MCP server is the execution layer. This skill is the workflow layer.

## Establish context

The bundled MCP connection accesses **Cloud** at `https://flambe.fly.dev/mcp`.
It cannot access **Single Burner** or **Development**. A trace id selects data
inside one instance; it does not switch instances. Discussing or developing
Single Burner does not authorize recording that work in Cloud.

On the first Flambe interaction in a conversation, and after an explicit switch:

- Resolve the user's intended instance: **Cloud**, **Single Burner**, or **Development**. Use a separately connected local instance only when its tools are available and the user has selected it. If intent is ambiguous and more than one connection is available, ask which instance before reading activity data or writing.
- Call `flambe_traces` on the intended connection, even if a trace id was mentioned earlier. Verify the returned `instance` matches the requested destination. If that connection is unavailable, explain that it is not connected; never substitute Cloud or another instance.
- Briefly name the instance when establishing or switching context, for example: "Using Single Burner for this chat."
- Keep the connection, `instance.id`, and `trace_id` together for this conversation. Reuse a trace id only when it belongs to that verified instance. If the user names a trace, match it inside that instance; otherwise use `default_trace_id`.
- Pass the exact returned `instance.id` as `expected_instance_id` on every other Flambe tool call, including status reads. Never infer it from a label or copy it between connections.
- Check `instance` in every result, including errors. On `INSTANCE_REQUIRED` or `INSTANCE_MISMATCH`, stop the attempted operation and rediscover the intended connection. Do not replace the expected id with a different instance's id just to make the request succeed.
- Stay with the selected instance until the user explicitly switches. On a switch, discard the previous instance's trace/activity ids and rediscover on the new connection.
- Create one stable conversation-local `agent_id` such as `chatgpt:<8 random lowercase hex characters>` and reuse it for every Flambe tool call in this conversation.
- Pass `platform: "ChatGPT"`.
- Do not invent an `agent_name`; let the Reducer Agent name the agent.

Do not list threads or categories merely to decide placement. The reducer owns placement.

## Think in a stack

Treat attention as a **call stack**, not a flat to-do list.

When work on a parent goal exposes a specific bug, prerequisite, design question, test failure, or verification step, that is a **push**: call `flambe_start` for a nested activity and leave the parent running.

When the nested work is complete, **pop** it with `flambe_end` and continue the parent.

- Leave parents running while going deeper.
- Put nested work under the workstream it serves, not beside the root.
- When moving to a true sibling, end the current leaf first, then start the sibling.
- Use `parent_id: null` only for a genuinely new top-level workstream.
- Omit `parent_id` to let the reducer infer this agent's active parent.
- Keep the active stack truthful as attention changes. Do not pre-plan a detailed flame or reconstruct one after the fact.

## Tool mapping

Use the MCP tools directly:

- `flambe_start`: begin meaningful work. Keep names concrete and under 255 characters.
- `flambe_end`: finish work. Prefer an outcome message over a generic "done".
- `flambe_suspend`: explicitly table work that is being paused.
- `flambe_resume`: resume suspended work.
- `flambe

Preview of the first 4,000 characters. Read full instructions →