@crouton-kit/humanloop 0.1.3 → 0.2.1

package/dist/cli.js

@@ -1,65 +1,124 @@
 #!/usr/bin/env node
 import { Command, Option } from 'commander';
-import { writeFileSync } from 'fs';
-import { launchTui } from './tui/app.js';
+import { writeFileSync, mkdtempSync } from 'fs';
+import { tmpdir } from 'os';
+import { resolve, join } from 'path';
 import { dispatchToTmuxPane } from './tui/tmux.js';
 import { findRecentSessionId } from './conversation/reader.js';
+import { launchReview, formatFeedbackSummary } from './editor/review.js';
+import { parseDeck } from './inbox/deck-schema.js';
+import { ask, inbox } from './api.js';
+import { display } from './surfaces/display.js';
 const program = new Command();
 program
     .name('hl')
     .description('Human-in-the-loop decision TUI.\n' +
     '\n' +
-    'Use this when you (the agent) need the human to validate decisions, choose\n' +
-    'between options, or provide freetext input before you continue. The tool\n' +
-    'blocks until the human finishes review and returns their responses as JSON.\n' +
+    'Use this when you (the agent) need the human to make a material decision\n' +
+    'before continuing — design tradeoffs, approval gates, picks between real\n' +
+    'alternatives. Blocks on an interactive TUI and returns answers as JSON.\n' +
+    'Reach for it when you have 2+ structured questions, or one decision with\n' +
+    'genuine tradeoffs; for a single freetext question, ask inline instead.\n' +
+    '\n' +
+    'AUDIENCE — the deck you write is read by a busy, technical human. Use\n' +
+    'progressive disclosure so the reader can stop at any layer:\n' +
+    '  • title    — noun-phrase topic (≤4 words). Scannable like an inbox.\n' +
+    '  • subtitle — TL;DR. One plain sentence framing the choice or stakes.\n' +
+    '  • body     — ELI12 markdown. Plain language up top, deeper material\n' +
+    '               under a heading the reader can skip.\n' +
+    'The reader can connect dots — do not write engineer-to-engineer jargon.\n' +
+    'See `hl ask --help` for content guidance and a worked example.\n' +
     '\n' +
     'Workflow:\n' +
     '  1. Write a deck file matching `hl schema` (a JSON object with interactions[]).\n' +
-    '  2. Run `hl create <file>`; it blocks and prints output JSON to stdout.\n' +
-    '  3. Parse the JSON; each response has id, optional selectedOptionId, optional freetext.\n' +
-    '\n' +
-    'Interaction options: supply options[] for choices; set allowFreetext for comment/freetext.')
+    '  2. Run `hl ask <file>`; it blocks and prints a ResolutionEnvelope JSON to stdout.\n' +
+    '  3. Parse the JSON; look up answers by `id` (humans can skip questions).')
     .version('0.1.0')
     .addHelpText('after', '\nExamples:\n' +
     '  hl schema                          # print the input JSON schema\n' +
-    '  hl create deck.json                # open TUI, block, print responses JSON\n' +
-    '  hl create deck.json --output out.json   # write result to file\n' +
-    '  hl create deck.json --no-tmux     # run in current pane even inside tmux\n');
+    '  hl schema response                 # print the resolution-envelope schema\n' +
+    '  hl ask deck.json                   # open TUI, block, print envelope JSON\n' +
+    '  hl ask deck.json --dir /tmp/ix     # store progress/response.json in /tmp/ix\n' +
+    '  hl ask deck.json --no-tmux         # run in current pane even inside tmux\n' +
+    '  hl inbox /tmp/box                  # resolve pending interactions under a root\n' +
+    '  hl display plan.md                 # live-render a file in a tmux pane\n');
 program
-    .command('create')
+    .command('ask')
     .description('Open the decisions TUI on <file> and block until the human finishes review.\n' +
-    'Prints output JSON to stdout (or to --output / --write-to).')
+    'Prints a ResolutionEnvelope JSON to stdout (or to --output / --write-to).')
     .argument('<file>', 'Path to deck JSON file (see `hl schema` for format)')
+    .option('--dir <interaction-dir>', 'Interaction directory holding progress.json/response.json. Default: a managed temp dir under os.tmpdir().')
     .option('--session-id <id>', 'Claude session ID; enables per-interaction visual context from conversation history. Defaults to the most recent session in cwd.')
     .option('--no-visuals', 'Skip visual context generation (faster, no haiku calls)')
     .option('--output <path>', 'Write result JSON to <path> instead of stdout')
     .option('--no-tmux', 'Do not auto-dispatch the TUI to a new tmux pane even when $TMUX is set')
     .addOption(new Option('--write-to <path>', 'internal: tmux child mode').hideHelp())
     .addHelpText('after', '\n' +
+    'CONTENT — what to write in each interaction\n' +
+    '  The deck is read by a busy, technical human. Use progressive\n' +
+    '  disclosure so the reader can stop at any layer:\n' +
+    '\n' +
+    '    title      Noun-phrase topic (≤4 words). The *thing* being\n' +
+    '               decided, not the decision. \'Database\' not \'Use\n' +
+    '               Postgres\'. The reader scans titles like an inbox.\n' +
+    '    subtitle   TL;DR — one plain-English sentence framing the\n' +
+    '               choice or stakes. Action-ready if the call is\n' +
+    '               obvious. No jargon, no library names without\n' +
+    '               context.\n' +
+    '    body       ELI12 markdown. Audience: a smart engineer joining\n' +
+    '               the codebase — capable, but does not want a wall of\n' +
+    '               jargon. Lead with what is at stake in plain language.\n' +
+    '               Tuck anything denser (technical specifics, alternatives\n' +
+    '               considered, edge cases, related context) under a\n' +
+    '               heading like `## Details` or `## Alternatives` so the\n' +
+    '               reader can skip past it. Every layer below the TL;DR\n' +
+    '               is optional reading.\n' +
+    '\n' +
+    '  AVOID: walls of jargon; raw schema dumps or stack traces in body;\n' +
+    '  titles that bury the topic; subtitles that restate the title;\n' +
+    '  options that are not real alternatives; asking the human anything\n' +
+    '  you could decide yourself from the code.\n' +
+    '\n' +
     'INPUT FORMAT\n' +
     '  JSON file with an `interactions` array. Each interaction has an `id`,\n' +
-    '  `title`, `options[]`, and optional `allowFreetext`. Run `hl schema`\n' +
-    '  for the full schema. Example:\n' +
+    '  `title`, `options[]`, and optional `subtitle`, `body`, `allowFreetext`.\n' +
+    '  Run `hl schema` for the full schema. Example with pyramid content:\n' +
     '    {\n' +
     '      "interactions": [\n' +
-    '        {"id":"i1","title":"Use Postgres","options":[\n' +
-    '          {"id":"approve","label":"Approve"},\n' +
-    '          {"id":"reject","label":"Reject"}\n' +
-    '        ],"allowFreetext":true},\n' +
-    '        {"id":"i2","title":"Migration tool","options":[\n' +
-    '          {"id":"prisma","label":"Prisma"},\n' +
-    '          {"id":"drizzle","label":"Drizzle"}\n' +
-    '        ]},\n' +
-    '        {"id":"i3","title":"Rate limit policy","options":[],"allowFreetext":true}\n' +
+    '        {\n' +
+    '          "id": "db",\n' +
+    '          "title": "Database",\n' +
+    '          "subtitle": "Postgres or SQLite for the new capture store?",\n' +
+    '          "body": "Two services will write at the same time, which is the crux.\\n\\nPostgres handles concurrent writes natively. SQLite serializes them — fine at low traffic, but we expect bursts.\\n\\n## Details\\nSQLite WAL still serializes writers; Postgres uses MVCC.",\n' +
+    '          "options": [\n' +
+    '            {"id":"pg","label":"Postgres"},\n' +
+    '            {"id":"sqlite","label":"SQLite"}\n' +
+    '          ],\n' +
+    '          "allowFreetext": true\n' +
+    '        },\n' +
+    '        {\n' +
+    '          "id": "retry",\n' +
+    '          "title": "Retry policy",\n' +
+    '          "subtitle": "How aggressively should we retry publish failures?",\n' +
+    '          "body": "Affects the reliability budget. Too aggressive and we hammer downstream during outages; too lax and transient blips become user-visible.",\n' +
+    '          "options": [],\n' +
+    '          "allowFreetext": true\n' +
+    '        }\n' +
     '      ]\n' +
     '    }\n' +
     '\n' +
-    'OUTPUT FORMAT (stdout on success, JSON)\n' +
+    'OUTPUT FORMAT (stdout on success, JSON — a ResolutionEnvelope)\n' +
     '  {\n' +
+    '    "summary": "<one line per answered interaction>",\n' +
+    '    "responsePath": "/abs/path/to/<dir>/response.json",\n' +
+    '    "schema": "humanloop.response/v2",\n' +
     '    "responses": [ ... ],\n' +
     '    "completedAt": "2026-04-20T15:23:00.000Z"\n' +
     '  }\n' +
     '\n' +
+    '  On-disk <dir>/response.json holds only { responses, completedAt }.\n' +
+    '  Run `hl schema response` for the envelope schema.\n' +
+    '\n' +
     '  Response shape:\n' +
     '    { id: string, selectedOptionId?: string, freetext?: string }\n' +
     '\n' +
@@ -69,10 +128,10 @@ program
     'BEHAVIOR\n' +
     '  tmux       When $TMUX is set, the TUI auto-splits into a new pane to the\n' +
     '             right (-d keeps focus on the caller). Disable with --no-tmux.\n' +
-    '  progress   Responses are persisted atomically to <file>.progress.json after\n' +
-    '             every change. If the process is killed, the next run resumes\n' +
-    '             from where the human left off. The file is removed on full\n' +
-    '             completion; partial-response files are preserved.\n' +
+    '  storage    progress.json/response.json live in --dir (default: a managed\n' +
+    '             temp dir). Responses persist atomically after every change;\n' +
+    '             a hard kill resumes from progress.json on the next run.\n' +
+    '             response.json is written when the human finishes.\n' +
     '  visuals    With --session-id (or auto-detected) haiku generates a short\n' +
     '             ANSI context block per interaction from recent conversation turns.\n' +
     '\n' +
@@ -84,6 +143,7 @@ program
     const sessionId = opts.visuals
         ? (opts.sessionId || findRecentSessionId(process.cwd()) || findRecentSessionId() || undefined)
         : undefined;
+    const dir = opts.dir ? resolve(opts.dir) : mkdtempSync(join(tmpdir(), 'hl-ix-'));
     const emit = (result) => {
         const json = JSON.stringify(result, null, 2) + '\n';
         if (opts.writeTo) {
@@ -99,7 +159,7 @@ program
     try {
         if (process.env.TMUX && opts.tmux && !opts.writeTo) {
             try {
-                const result = await dispatchToTmuxPane(file, { sessionId, visuals: opts.visuals });
+                const result = await dispatchToTmuxPane(file, { sessionId, visuals: opts.visuals, dir });
                 emit(result);
                 process.exit(0);
             }
@@ -107,98 +167,289 @@ program
                 process.stderr.write(`tmux dispatch failed, running locally: ${err instanceof Error ? err.message : String(err)}\n`);
             }
         }
-        const result = await launchTui(file, sessionId);
+        const deck = parseDeck(resolve(file));
+        const result = await ask(deck, { dir, sessionId });
         emit(result);
         process.exit(0);
     }
     catch (err) {
         const msg = err instanceof Error ? err.message : String(err);
         process.stderr.write(`ERROR: ${msg}\n`);
-        if (msg.includes('not found')) {
+        if (msg.includes('ENOENT') || msg.includes('no such file')) {
             process.stderr.write('\nFix: pass a path to an existing deck JSON file.\nSee format: hl schema\n');
         }
-        else if (msg.includes('No interactions')) {
-            process.stderr.write('\nFix: the file must contain a non-empty `interactions` array.\nSee format: hl schema\n');
+        else if (msg.includes('not valid JSON') || msg.includes('JSON')) {
+            process.stderr.write('\nFix: the deck file must be valid JSON matching `hl schema`.\n');
         }
         else if (msg.includes('TTY')) {
             process.stderr.write('\nFix: hl needs an interactive terminal. If the caller captures stdin,\nrun inside tmux so hl can auto-dispatch the TUI to a new pane, or pipe\nstdin from /dev/tty.\n');
         }
-        else if (msg.includes('JSON')) {
-            process.stderr.write('\nFix: the deck file must be valid JSON matching `hl schema`.\n');
-        }
-        else if (msg.startsWith('interactions[') || msg.includes('Duplicate interaction id') || msg.includes('must be')) {
+        else {
             process.stderr.write('\nFix: the deck file must match `hl schema`. Run `hl schema` to see the required shape.\n');
         }
         process.exit(1);
     }
 });
 program
-    .command('schema')
-    .description('Print the v2 Interaction[] deck schema to stdout')
+    .command('inbox')
+    .description('Resolve pending interactions across one or more root directories.\n' +
+    'Each root\'s immediate subdirs are treated as interaction dirs (a dir\n' +
+    'with deck.json and no response.json is pending). Lists them, lets the\n' +
+    'human pick and resolve one at a time, writing each response.json.')
+    .argument('<roots...>', 'Root dir(s) whose immediate subdirs are interaction dirs')
     .addHelpText('after', '\n' +
-    'Use this to learn the exact input format for `hl create`.\n' +
-    '\n' +
-    'Example:\n' +
-    '  hl schema > deck.schema.json   # save for reference\n' +
-    '  hl schema | jq                 # pretty-print\n')
-    .action(() => {
-    const schema = {
-        $schema: 'https://json-schema.org/draft/2020-12/schema',
-        description: 'Input schema for hl create (v2)',
-        type: 'object',
-        required: ['interactions'],
-        properties: {
-            title: {
-                type: 'string',
-                description: 'Optional deck title shown in the TUI header',
-            },
-            interactions: {
-                type: 'array',
-                minItems: 1,
-                items: {
-                    type: 'object',
-                    required: ['id', 'title', 'options'],
-                    properties: {
-                        id: { type: 'string', description: 'Unique identifier' },
-                        title: { type: 'string', description: 'Short display label (≤4 words). Required.' },
-                        subtitle: { type: 'string', description: 'One-line "why this matters"' },
-                        body: { type: 'string', description: 'Markdown body shown in item-review' },
-                        bodyPath: { type: 'string', description: 'Path to body file; sisyphus inlines before mount' },
-                        options: {
-                            type: 'array',
-                            description: 'Selectable choices. Empty for freetext-only interactions.',
-                            items: {
-                                type: 'object',
-                                required: ['id', 'label'],
-                                properties: {
-                                    id: { type: 'string' },
-                                    label: { type: 'string' },
-                                    description: { type: 'string' },
-                                    shortcut: {
-                                        type: 'string',
-                                        description: 'Single char shortcut. Auto-assigned if absent. Avoid: c r n p q j k space',
-                                    },
+    'BEHAVIOR\n' +
+    '  Runs in the current terminal (requires a TTY). Up/down (or j/k) to\n' +
+    '  navigate, enter to resolve the selected interaction, q/esc to quit.\n' +
+    '  After each resolution the list rescans — resolved items drop out.\n' +
+    '\n' +
+    'EXIT CODES\n' +
+    '  0  finished (human quit, or nothing pending)\n' +
+    '  1  error (no TTY, unreadable root, etc.)\n' +
+    '\n' +
+    'Examples:\n' +
+    '  hl inbox /tmp/box\n' +
+    '  hl inbox ~/.sisyphus/asks ~/.crtr/pending\n')
+    .action(async (roots) => {
+    try {
+        await inbox(roots.map((r) => resolve(r)));
+        process.exit(0);
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        process.stderr.write(`ERROR: ${msg}\n`);
+        if (msg.includes('TTY')) {
+            process.stderr.write('\nFix: hl inbox needs an interactive terminal.\n');
+        }
+        process.exit(1);
+    }
+});
+program
+    .command('display')
+    .description('Render <path> live in a tmux pane (via the managed termrender). The pane\n' +
+    'live-updates as the file changes. Auto-splits, or opens a new window when\n' +
+    'the current window is at its pane budget.')
+    .argument('<path>', 'Path to the markdown/text file to render')
+    .option('--no-watch', 'Render once instead of live-watching the file')
+    .option('--new-window', 'Open in a new tmux window instead of splitting the current one')
+    .addHelpText('after', '\n' +
+    'EXIT CODES\n' +
+    '  0  spawned (prints the pane id) — or no-op when not in tmux / renderer\n' +
+    '     unavailable (message on stderr)\n' +
+    '\n' +
+    'Examples:\n' +
+    '  hl display plan.md\n' +
+    '  hl display plan.md --no-watch\n' +
+    '  hl display plan.md --new-window\n')
+    .action((path, opts) => {
+    const res = display(resolve(path), {
+        watch: opts.watch,
+        window: opts.newWindow ? 'new' : 'auto',
+    });
+    if (res.paneId) {
+        process.stdout.write(`opened in tmux pane ${res.paneId} (live — edits to the file refresh the view)\n`);
+    }
+    else {
+        process.stderr.write('display: no pane opened (not in tmux, or termrender unavailable)\n');
+    }
+    process.exit(0);
+});
+program
+    .command('propose')
+    .description('Open a markdown document in a read-only editor review session and block\n' +
+    'until the human finishes and quits. Use this when you have produced a\n' +
+    'document (plan, design doc, spec, draft) and need targeted human review\n' +
+    'before continuing — not a structured decision (use `hl ask` for that).')
+    .argument('<file>', 'Path to the markdown (.md) file to get feedback on')
+    .option('--output <path>', 'Path for the answers JSON (live autosave + finalized on exit). Default: <file>.feedback.json')
+    .option('--editor <bin>', 'Editor binary to use. Default: first of nvim, vim on PATH')
+    .option('--no-tmux', 'Run the editor in the current terminal instead of a tmux split pane')
+    .addHelpText('after', '\n' +
+    'WHAT THIS IS FOR\n' +
+    '  Freeform review of a markdown doc you wrote. The human leaves comments\n' +
+    '  anchored to real source lines or visual selections using native vim\n' +
+    '  keybindings — you do not predefine questions; the human tells you what\n' +
+    '  is wrong and where.\n' +
+    '\n' +
+    'BEHAVIOR\n' +
+    '  • Opens the file READ-ONLY in a CLEAN editor (nvim -u NONE: no\n' +
+    '    init.lua / LazyVim / plugins / keymaps). Look/feel is ONLY the\n' +
+    '    user\'s gloam colorscheme + built-in treesitter markdown\n' +
+    '    highlighting. Review keys are buffer-scoped on <Space>,\n' +
+    '    plus :HL* commands.\n' +
+    '  • When $TMUX is set, opens in a split pane (disable with --no-tmux);\n' +
+    '    otherwise takes over the current terminal.\n' +
+    '  • Comments autosave to the answers JSON continuously. A killed/closed\n' +
+    '    session RESUMES from the autosave on next run.\n' +
+    '  • ANY quit submits. On editor exit the JSON is finalized (submitted:true)\n' +
+    '    and echoed to stdout, then the command exits 0.\n' +
+    '  • Quitting with zero comments sets approved:true ("looks good").\n' +
+    '\n' +
+    '  In-editor (<Space> maps; or use the :HL* commands):\n' +
+    '    <Space>c / :HLComment   Comment on the visual selection or current line\n' +
+    '    <Space>l / :HLList      Toggle comments list\n' +
+    '    <Space>u / :HLUndo      Undo last comment\n' +
+    '    <Space>s / :HLSubmit    Submit & quit\n' +
+    '\n' +
+    'OUTPUT\n' +
+    '  stdout is a COMPACT listing (kept small so it does not clog context):\n' +
+    '  a header with the source path, then per comment:\n' +
+    '\n' +
+    '     1. L46:5-35\n' +
+    '        text:    <the original text in that span>\n' +
+    '        comment: <the human\'s note>\n' +
+    '\n' +
+    '  Range is L<line> (whole line), L<line>:<colStart>-<colEnd> (partial\n' +
+    '  selection), or L<l1>-<l2> / L<l1>:<c1>-<l2>:<c2> across lines. text\n' +
+    '  is the exact selected quote, or the whole line(s) when there was no\n' +
+    '  partial selection. Zero comments => "approved" (looks good, proceed).\n' +
+    '\n' +
+    '  The FULL record is written to --output (default <file>.feedback.json);\n' +
+    '  stdout ends with its path + this schema (read the file only if you\n' +
+    '  actually need the verbose fields — usually you do not):\n' +
+    '    {file, submitted, approved,\n' +
+    '     comments:[{id, line, endLine, quote?, colStart?, colEnd?,\n' +
+    '                lineText, comment, createdAt}],\n' +
+    '     submittedAt, savedAt}\n' +
+    '  cols are 0-based byte offsets, colEnd exclusive. Act on each comment\n' +
+    '  via the source path + range (quote/cols when present).\n' +
+    '\n' +
+    'EXIT CODES\n' +
+    '  0  finished — feedback summary emitted\n' +
+    '  1  error (file missing, no nvim/vim found)\n' +
+    '\n' +
+    'Examples:\n' +
+    '  hl propose plan.md\n' +
+    '  hl propose plan.md --output /tmp/fb.json\n' +
+    '  hl propose plan.md --no-tmux\n' +
+    '  hl propose plan.md --editor vim\n')
+    .action(async (file, opts) => {
+    try {
+        const output = opts.output ?? `${file}.feedback.json`;
+        const result = await launchReview(file, { output, editor: opts.editor, noTmux: !opts.tmux });
+        process.stdout.write(formatFeedbackSummary(result, resolve(output)) + '\n');
+        process.exit(0);
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        process.stderr.write(`ERROR: ${msg}\n`);
+        if (msg.startsWith('Markdown file not found')) {
+            process.stderr.write('\nFix: pass a path to an existing markdown file.\n');
+        }
+        else if (msg.startsWith('Editor not found') || msg.startsWith('No editor found')) {
+            process.stderr.write('\nFix: install Neovim (`brew install neovim`) or pass --editor <path>.\n');
+        }
+        process.exit(1);
+    }
+});
+const REQUEST_SCHEMA = {
+    $schema: 'https://json-schema.org/draft/2020-12/schema',
+    description: 'Input schema for hl ask (v2)',
+    type: 'object',
+    required: ['interactions'],
+    properties: {
+        title: {
+            type: 'string',
+            description: 'Optional deck title shown in the TUI header',
+        },
+        interactions: {
+            type: 'array',
+            minItems: 1,
+            items: {
+                type: 'object',
+                required: ['id', 'title', 'options'],
+                properties: {
+                    id: { type: 'string', description: 'Unique identifier. Used to look up answers in the output — never assume index alignment.' },
+                    title: { type: 'string', description: 'Noun-phrase topic (≤4 words) — the *thing* being decided, not the decision. \'Database\' not \'Use Postgres\'. The reader scans titles like an inbox.' },
+                    subtitle: { type: 'string', description: 'TL;DR — one plain-English sentence framing the choice or stakes. Action-ready if the call is obvious. No jargon, no library names without context.' },
+                    body: { type: 'string', description: 'ELI12 markdown. Audience: a smart engineer joining the codebase — capable, but does not want a wall of jargon. Lead with what is at stake in plain language. Tuck anything denser (technical specifics, alternatives considered, edge cases) under a heading like `## Details` or `## Alternatives` so the reader can skip past. Every layer below the TL;DR is optional reading.' },
+                    bodyPath: { type: 'string', description: 'Path to a markdown file used in place of `body`; inlined before mount, resolved relative to the deck JSON\'s directory and confined to it. Same content guidance as `body`.' },
+                    options: {
+                        type: 'array',
+                        description: 'Selectable choices. Empty for freetext-only interactions.',
+                        items: {
+                            type: 'object',
+                            required: ['id', 'label'],
+                            properties: {
+                                id: { type: 'string' },
+                                label: { type: 'string' },
+                                description: { type: 'string' },
+                                shortcut: {
+                                    type: 'string',
+                                    description: 'Single char shortcut. Auto-assigned if absent. Avoid: c r n p q j k space',
                                 },
                             },
                         },
-                        allowFreetext: {
-                            type: 'boolean',
-                            description: 'If true, user can add a freetext comment (or respond freely if options is empty)',
-                        },
-                        freetextLabel: {
-                            type: 'string',
-                            description: 'Prompt shown above the freetext input. Default: "Comment" or "Response"',
-                        },
-                        kind: {
-                            type: 'string',
-                            enum: ['notify', 'validation', 'decision', 'context', 'error'],
-                            description: 'Display hint — opaque to humanloop, used by sisyphus for inbox icons',
-                        },
+                    },
+                    allowFreetext: {
+                        type: 'boolean',
+                        description: 'If true, user can add a freetext comment (or respond freely if options is empty)',
+                    },
+                    freetextLabel: {
+                        type: 'string',
+                        description: 'Prompt shown above the freetext input. Default: "Comment" or "Response"',
+                    },
+                    kind: {
+                        type: 'string',
+                        enum: ['notify', 'validation', 'decision', 'context', 'error'],
+                        description: 'Display hint — opaque to humanloop, used by consumers for inbox icons',
                     },
                 },
             },
         },
-    };
+    },
+};
+const RESPONSE_SCHEMA = {
+    $schema: 'https://json-schema.org/draft/2020-12/schema',
+    $id: 'humanloop.response/v2',
+    description: 'Resolution envelope emitted by `hl ask` / returned by ask(). The on-disk <dir>/response.json holds only { responses, completedAt }; responsePath points at it.',
+    type: 'object',
+    required: ['summary', 'responsePath', 'schema', 'responses', 'completedAt'],
+    properties: {
+        summary: {
+            type: 'string',
+            description: 'Deterministic, no-LLM. One line per answered interaction: "<title>: <option label>[ — <freetext>]".',
+        },
+        responsePath: {
+            type: 'string',
+            description: 'Absolute path to the on-disk response.json ({ responses, completedAt }).',
+        },
+        schema: {
+            const: 'humanloop.response/v2',
+            description: 'Identifies this response contract.',
+        },
+        responses: {
+            type: 'array',
+            description: 'Inline answers. May have FEWER entries than input interactions — humans can skip. Look up by id.',
+            items: {
+                type: 'object',
+                required: ['id'],
+                properties: {
+                    id: { type: 'string' },
+                    selectedOptionId: { type: 'string' },
+                    freetext: { type: 'string' },
+                },
+            },
+        },
+        completedAt: {
+            type: 'string',
+            description: 'ISO 8601 timestamp when the human finished.',
+        },
+    },
+};
+program
+    .command('schema')
+    .description('Print a JSON Schema. `request` (default) = the `hl ask` deck input; `response` = the resolution envelope.')
+    .argument('[kind]', 'request | response', 'request')
+    .addHelpText('after', '\n' +
+    'Use `request` to learn the input format for `hl ask`; `response` to learn\n' +
+    'the envelope `hl ask` prints (schema id "humanloop.response/v2").\n' +
+    '\n' +
+    'Examples:\n' +
+    '  hl schema > deck.schema.json       # save the request schema\n' +
+    '  hl schema response | jq            # pretty-print the response schema\n')
+    .action((kind) => {
+    const schema = kind === 'response' ? RESPONSE_SCHEMA : REQUEST_SCHEMA;
     process.stdout.write(JSON.stringify(schema, null, 2) + '\n');
 });
 program.parse();

package/dist/editor/review.d.ts

@@ -0,0 +1,24 @@
+import type { FeedbackResult } from '../types.js';
+export interface ReviewOptions {
+    /** Where the answers JSON is written (live autosave + finalized on exit). */
+    output: string;
+    /** Editor binary override. Default: first of nvim, vim on PATH. */
+    editor?: string;
+    /** Force running in the current terminal even when $TMUX is set. */
+    noTmux?: boolean;
+}
+/**
+ * Compact stdout rendering for the agent: per comment just the line:col
+ * range, the original text in that span, and the note — plus the source
+ * path, a pointer to the full JSON on disk, and a one-line schema hint.
+ * The verbose fields are deliberately not printed so they don't clog context.
+ */
+export declare function formatFeedbackSummary(result: FeedbackResult, feedbackJsonPath: string): string;
+/**
+ * Open a markdown file in a clean, read-only Neovim/Vim review session. The
+ * human anchors comments to source lines/selections with native vim motions
+ * and quits to submit. Blocks until the editor exits, then finalizes and
+ * returns the feedback. Autosaved continuously so a kill is recoverable and
+ * the next run resumes.
+ */
+export declare function launchReview(file: string, opts: ReviewOptions): Promise<FeedbackResult>;