where-am-i

where-am-i

Render the task's decisions as an ASCII tree. The point is orientation: show the path already settled, mark the active fork, and end with the exact questions left to answer.

How to render

Read the current state from the conversation, then draw a single tree.

• Root = the task in one short phrase.
• [N] numbers top-level decisions only, in the order they arose; sub-decisions stay unnumbered. The closing list references these [N] tags.
• Q: lines at each fork, stating the question being decided there.
• Indent sub-decisions under the option that produced them.

Markers

Attach a marker after the node with a ◄── connector (node ◄── ✅ chosen), as the example shows. Use this one form for every marker.

• ✅ — a decided node; append a terse why/result in parentheses.
• 📍 YOU ARE HERE — the single active fork, the decision in question right now. Use it once per tree.
• not taken — an option considered and rejected.
• (not started) — a decision that exists but hasn't been reached yet.
• Keep settled branches visible (collapsed to one line, not deleted) so the user sees the path behind them, not only what is ahead.

Always end with

A numbered "Open questions, in order" list below the tree. Include only the undecided nodes, ordered the way they need answering. Reference each by its [N] tag so the user can point back at the tree.

Example

deploy pipeline rework
│
├─ [1] runner image          Q: pin or float?
│   ├─ pin to :1.4    ◄── ✅ chosen (reproducible CI)
│   └─ float :latest  ◄── not taken
│
├─ [2] cache strategy   ◄── 📍 YOU ARE HERE
│   │   Q: per-branch or shared?
│   ├─ shared ──┐  Q: evict by age or size?
│   │           ├─ age
│   │           └─ size
│   └─ per-branch
│
└─ [3] rollout   ◄── (not started)
        Q: canary or all-at-once?

Open questions, in order:
1. [2] cache strategy — shared or per-branch?
2. [3] rollout — canary or all-at-once?

Rules

• One tree per render. Redraw the whole thing each time rather than patching it.
• Alignment is approximate — don't chase perfect columns; readability beats pixel-perfect spacing.
• Keep node labels short; detail goes in the parenthetical or the questions list.
• Do not invent decisions. Only show forks that actually came up in the task.
• If every decision is settled, render the full tree (all ✅) and write Open questions: none. If no decision has come up yet, say so in one line instead of drawing a tree.
• No files, no state. The conversation is the source of truth.