@nicknisi/pi-answer

Extracts unanswered questions from the last completed assistant message and presents them in an interactive Q&A overlay, then sends your answers back into the session as a single formatted user message. It exists for the common case where the agent ends a turn with several questions scattered through its output — instead of re-reading and composing a reply by hand, you tab through a focused form and fire the answers back in one shot.

What it adds

• Slash command: /answer
• Keybinding (shortcut): ctrl+. — same handler as the command
• Overlays (via ctx.ui.custom):
  • A BorderedLoader while questions are being extracted by an LLM (abortable with Esc)
  • An AnswerComponent — a bordered, multi-question form with navigation, per-question status, and a submit confirmation step
• No tools, no widgets, no custom message/entry types. The submitted answers go out as a regular user message via pi.sendUserMessage.

How it works

1. Walks ctx.sessionManager.getBranch() backwards to find the last assistant message with stopReason === "stop" and non-empty text. Incomplete (e.g. aborted or tool-loop) assistant messages are skipped, with a warning if any were skipped.

2. Sends that message's text to an extraction model with a system prompt instructing it to return {"questions": [{question, context?}]} JSON — only questions that require user input, each self-contained.

3. Parses the response leniently: tries the raw text, ```json fenced blocks, and the first {..} span. If parsing fails or the response is empty, falls back to a regex-based extractor that pulls ?-terminated sentences out of the original message. Questions are trimmed and case-insensitively deduped.

4. Opens the answer form. On submit, builds a message:

   Here are my answers to your questions:
   
   Q: <question>
   Context: <context, if any>
   A: <answer, or "(no answer)">
   

5. If the session is idle (ctx.isIdle()), sends via pi.sendUserMessage(answers). If the agent is still running, sends with { deliverAs: "followUp" } and notifies that it was queued.

Usage

/answer

or press ctrl+. at the prompt.

Answer form keybindings

The form renders up to 120 columns wide and shows: question counter (n/N), answered count, a per-question status strip (current highlighted, answered in success color, unanswered dim), the question text with optional Context: line, and the answer editor. Unanswered questions submit as (no answer) — the confirmation step tells you how many.

The extraction loader can be aborted with Esc, which cancels the whole flow.

Configuration

Optional global config: ~/.pi/agent/configs/answer.json. Changes take effect on the next session or after /reload.

{
  "extractionModels": ["anthropic/claude-fable-5", "anthropic/claude-opus-5"]
}

extractionModels is a non-empty ordered list of provider/model-id strings. The first model that both exists in ctx.modelRegistry and passes getApiKeyAndHeaders() is used. Model IDs may contain additional slashes; only the first slash separates the provider. Missing or invalid config falls back to the list above and invalid config emits a warning. See answer.example.json.

The current session model (ctx.model) is not used for extraction—only checked for presence as a guard. API keys come from Pi's normal model authentication.

Dependencies

• @earendil-works/pi-coding-agent (peer, >=0.86.0) — ExtensionAPI, ExtensionContext, BorderedLoader, Theme; extraction runs through ctx.modelRegistry.complete(model, { systemPrompt, messages }, { signal }), which applies models.json overrides, extension-registered providers, and request-time auth
• @earendil-works/pi-tui — Component, Focusable, Editor, EditorTheme, Key, matchesKey, TUI, truncateToWidth, visibleWidth, wrapTextWithAnsi
• @earendil-works/pi-ai — types only (Api, Model, UserMessage)

No third-party npm dependencies.

Caveats

• Extraction-model availability. If none of the configured candidates exist with working authentication, the command errors out listing every model it checked.
• Pi internals: depends on ctx.sessionManager.getBranch() entry shapes (entry.type === "message", message.role, message.stopReason === "stop", content part { type: "text", text }), ctx.modelRegistry.find/getApiKeyAndHeaders/complete, ctx.ui.custom component contract, and pi.sendUserMessage(..., { deliverAs: "followUp" }). Any of these changing across pi versions will break it.
• Editor render slicing. AnswerComponent.render strips the first and last lines of the embedded Editor.render() output (editorLines[1..len-2]) assuming the Editor wraps its content in a border frame. If pi-tui's Editor render shape changes, answers will render wrong.
• Extraction is heuristic at the edges. The LLM prompt tries to keep questions self-contained, but the regex fallback (fallbackExtractQuestions) is naive — it splits on sentence boundaries ending in ? and can produce context-free or false-positive questions when extraction JSON parsing fails.
• Multi-part assistant messages are flattened by joining all text parts with \n; non-text parts (tool calls, thinking) are ignored.
• Requires interactive UI (ctx.hasUI); errors out in headless mode.

Install

pi install /Users/nicknisi/Developer/pi-extensions/packages/answer