Interactive Q&A bridge for the Pi coding agent that converts unstructured questions from LLMs into structured TUI forms.

When an LLM sends a message containing multiple questions, traditional chat requires manual text replies that are error-prone and tedious. pi-answr extracts those questions and renders an interactive terminal form, ensuring the LLM receives exactly the data format it needs.

• Question Extraction — Uses remote LLMs to parse unstructured text into structured question schemas
• Interactive TUI Forms — Terminal UI with keyboard navigation for radio, checkbox, and text inputs
• Draft Autosave — Responses are persisted automatically to prevent data loss during long forms
• Questionnaire Restoration — Reopen the last questionnaire with /answer:again
• Tool Integration — LLMs can trigger forms directly via the ask_user_question tool
• Template System — Configurable answer formatting templates

pi install git:https://github.com/k0valik/pi-answr

Restart Pi if already running.

Extracts questions from the last assistant message and displays an interactive form.

1. Finds the most recent assistant message in the session branch
2. Sends the message to an extraction model (configured in settings)
3. Parses the JSON response into a question schema
4. Renders the TUI form
5. On submit, formats answers and sends back to the chat

Implemented in index.ts:

• answerHandler() — Main command handler (lines ~295-380)
• selectExtractionModel() from extraction.ts — Model selection logic

Reopens the last questionnaire without re-extracting. Prompts to restore previous answers if drafts exist.

1. Attempts to reconstruct cache from session entries
2. Prompts user to restore saved answers
3. Re-renders the form with prior responses filled in

Implemented in index.ts:

• answerAgainHandler() — Lines ~420-510

LLMs can invoke this tool directly to request structured user input. Registered only if toolEnabled: true in settings.

{
  "value": "postgres",
  "label": "PostgreSQL",
  "description": "Best for relational data"
}

Answers are returned as formatted text:

Q1: Database: postgres
Q2: Languages: typescript, rust
Q3: Project Name: my-cool-app

Or if cancelled: User cancelled the form

Implemented in index.ts:

• Tool registration — pi.registerTool() block (lines ~175-290)
• execute() function — Form rendering and result formatting
• Input schema defined via AskUserQuestionParams (lines ~140-155)

• See example.config.json

Add an answer block to your settings.json:

{
  "answer": {
    "toolEnabled": true,
    "extractionModels": [
      { "provider": "openai-codex", "id": "gpt-5.4-mini" },
      { "provider": "github-copilot", "id": "gpt-5.4-mini" },
      { "provider": "anthropic", "id": "claude-haiku-4-5" }
    ],
    "extractionTimeoutMs": 30000,
    "debugNotifications": false,
    "answerTemplates": [
      { "label": "Q&A", "template": "Q{{index}}: {{question}}\nA: {{answer}}" },
      { "label": "Concise", "template": "{{question}}: {{answer}}" }
    ],
    "drafts": {
      "enabled": true,
      "autosaveMs": 1000,
      "promptOnRestore": true
    }
  }
}

Default extraction models (DEFAULT_MODEL_PREFERENCES in extraction.ts):

[
  { provider: "openai-codex", id: "gpt-5.4-mini" },
  { provider: "github-copilot", "id": "gpt-5.4-mini" },
  { provider: "anthropic", id: "claude-haiku-4-5" }
]

Settings are read from (in order of precedence):

1. Project: ./.pi/agent/settings.json
2. Global: ~/.pi/agent/settings.json

Implemented in index.ts:

• loadAnswerSettings() — Settings loading with caching (lines ~95-130)
• getAnswerSettingsPaths() — Path resolution

1. User runs /answer
2. answerHandler() finds the last assistant message
3. selectExtractionModel() picks an available model from preferences
4. Message sent to model with DEFAULT_SYSTEM_PROMPT (extraction.ts)
5. Model returns JSON with question schema
6. parseExtractionResult() parses JSON (schema.ts)
7. Questions normalized via normalizeQuestions() (schema.ts)
8. TUI form rendered
9. Answers formatted and sent as chat message

schema.ts provides the core data model:

• UnifiedQuestion — Raw question from extraction or tool params
• NormalizedQuestion — Fully typed question for TUI
• normalizeQuestions() — Converts raw input to normalized form
• parseExtractionResult() — Parses LLM JSON response

Implemented in drafts.ts:

• createDraftStore() — Creates a draft persistence store
• getLatestDraftForEntry() — Retrieves saved draft
• getInitialResponses() — Restores responses from draft
• deriveAnswersFromResponses() — Converts TUI responses to answer strings

Drafts are stored as session entries with type answer:draft.

When a user completes all questions and presses Tab once more, they reach the Summary tab (the final position in the tab order).

The Summary page displays:

1. All answers reviewed — Shows every question with its answer (or "(no answer)" if unanswered)
2. Two selection options:
   • "Confirm All" — Submit the form
   • "Revisit Questions" — Go back to a specific unanswered question

Before submission, the form checks for unanswered questions with getUnansweredQuestions() (tui.ts). If questions are marked required: true and left blank:

1. First press of Enter shows a warning banner: ⚠ X question(s) not answered
2. Second press of Enter proceeds with submission anyway
3. Pressing Up/Down to select "Revisit Questions" navigates directly to the first unanswered question

The progress indicator shows:

• [●] — Current question
• [✓] — Answered and confirmed (pressed Enter)
• [○] — Answered (selected but not confirmed)
• [ ] — Unanswered

Navigation in Summary:

• ↑/↓ — Toggle between Confirm All and Revisit Questions
• Enter — Confirm selection
• Esc — Go back to last question
• Esc Esc — Direct jump to first unanswered question
• Tab — Cycle back to first question

Implemented in tui.ts:

• showingConfirmation state variable
• confirmPageSelection — "confirm" or "revisit"
• confirmWarningShown — controls warning display
• getUnansweredQuestions() (lines ~170-177) — scans for required-but-empty questions
• submit() (lines ~310-330) — final submission handler

Answer templates use variable substitution:

Example template:

{{index}}. {{question}}
   Answer: {{answer}}

Renders as:

1. Which color do you prefer?
   Answer: blue

Implemented in templates.ts:

• normalizeTemplates() — Parses template configuration
• applyTemplate() — Variable substitution

• Pi coding agent (latest version)
• At least one model configured in ~/.pi/agent/models.json