status

Show where the work is, and carry on from it.

Start with catch_up. It returns two blocks: the story so far, and whatever comes next. Paste them in that order and stop. That is the whole reply on the common path, which is somebody reopening after a gap and needing to know where they were.

Unless there is no second block. An empty next_block does not mean there is nothing to do, it means nothing is waiting on the user — the step is decided and the next move is Forge's. Do not stop there. Carry straight on in the same turn: current_step, then plan_files, then build it. This ended a real session on the line "No question is open; nothing is blocking you", which reads as a finished report and is actually a loop that forgot to take its turn. Follow the next field that catch_up returns; it says which of the two endings applies.

The summary is assembled from the records every time rather than kept in a log, so there is no second version of the history to drift from the first (decision 011: the repository is the memory). It carries the idea in the user's own words, how many questions are answered against the estimate, how many decisions are recorded, phases finished, steps built, which step is open now, and the last three decisions with what was chosen.

If it reports needs_repair, call check_history, show what is damaged, and offer repair_history, which is the tool that fixes it: it quarantines the damaged records before it writes anything, so nothing is destroyed and the original is still on disk afterwards. Say that, then ask, then run it if they say yes. Telling somebody their notes are broken and naming no repair is the shape of an error message, not of a product. Nothing else on this screen is worth reading until the notes are sound.

If the user asks whether they ever asked for something — "did I say I wanted dark mode", "what did I ask about the database" — call what_did_i_ask_for with their words. It searches the records and returns what they asked, when, and what was decided. Answering from memory here is guessing about the one thing this product exists to have written down.

Everything below is the longer report, for when the user asks for more than a glance.

Steps:

1. Call next_step for the stage, the subagent, the model, and whether the next step needs the user.
2. Call current_state for what is decided and what is open.
3. Call usage_report for tokens used so far. If it reports available: false, say usage is unavailable in one line and carry on — nothing else depends on it.
4. Call check_history to confirm the decision records are intact.

Then check the version — a status report from a plugin that is two weeks behind describes a project state that may already have moved:

python "${CLAUDE_PLUGIN_ROOT}/scripts/forge_update.py"

It prints nothing when the copy is current. Print what it returns, add nothing.

Then print, using the Forge visual identity:

• the banner, then the progress bar with the count of questions answered against the estimate (rule R4 — always show how many remain)
• the current stage in plain words, and what happens next
• the open question, if there is one, and that code is blocked until it is answered — put the question itself through render_action, because it is the one thing on this screen the user can do something about, and everything around it is a report
• which model does the next step and why (rule R2 — name it: "Fable 5", not "the teaching model")
• usage, and any threshold warning
• anything the review still has open

Keep it short. This is a glance, not a report.