codeply-cli 0.3.1 → 0.4.1

package/lib/native-tools.mjs

@@ -1,189 +1,189 @@
-/**
- * Native function calling as a transport.
- *
- * Inside the engine everything stays in one format: the model's actions are
- * <codeply:name> blocks in its reply, and results come back as
- * "[tool result: name]" user messages. Every guard in agent.mjs (claim audit,
- * stuck detection, verification gates, batching) works on that format.
- *
- * For models with real function calling this module translates at the edge:
- *   - toNativeMessages(): engine history -> OpenAI-shaped history
- *     (assistant.tool_calls + role "tool" results), which ai.js sends as is to
- *     OpenAI-compatible APIs and converts for Anthropic and Ollama.
- *   - toolCallsToTags(): the model's tool calls -> <codeply:...> blocks, so
- *     the rest of the loop sees exactly what it would have seen in text mode.
- * Structured arguments mean no tag-format mistakes, and the model uses the
- * calling convention it was trained on.
- */
-
-// What the model is told about each action. Parameters not listed as
-// required are optional. Descriptions stay short: the system prompt's RULES
-// section carries the detailed guidance.
-export const TOOL_DOCS = {
-  todo: { d: 'Replace your task checklist for a multi-step task. Send the whole list every time.', req: ['items'], p: { items: 'One task per line: "[ ] pending", "[>] in progress" (only one), "[x] done" (only after a tool confirmed it), "[-] dropped".' } },
-  ask_user: { d: 'Ask the user a multiple-choice question when a real decision is theirs to make and cannot be inferred from the project.', req: ['question'], p: { question: 'The question, one sentence.', options: 'Up to 5 short answers, one per line, recommended first.' } },
-  list_dir: { d: 'List a folder.', req: [], p: { path: 'Folder, relative to the project root. Default ".".' } },
-  read_file: { d: 'Read a text file with line numbers.', req: ['path'], p: { path: 'File path.', offset: '0-based line to start at.', limit: 'Max lines (up to 600).' } },
-  write_file: { d: 'Create a file or fully rewrite one. For changes to an existing file prefer edit_file.', req: ['path', 'content'], p: { path: 'File path.', content: 'The complete file content, verbatim.' } },
-  edit_file: { d: 'Replace one exact block of an existing file.', req: ['path', 'search', 'replace'], p: { path: 'File path.', search: 'Existing text copied exactly from a read (no line-number prefixes), just enough to be unique.', replace: 'The replacement text.', all: '"true" to replace every occurrence (renames).' } },
-  search: { d: 'Regex search across project files, or find files by name when only glob is given.', req: [], p: { pattern: 'Case-insensitive regex.', glob: 'Limit to paths like "**/*.ts", or alone to find files by name.', path: 'Folder or single file to search in.', context: 'Lines of surrounding code per hit (0-5).', files_only: '"true" to list matching files with hit counts.' } },
-  run: { d: 'Run a shell command in the project folder on the user\'s machine.', req: ['command'], p: { command: 'The command line.' } },
-  use_skill: { d: 'Load a skill\'s full instructions.', req: ['name'], p: { name: 'Skill name from the SKILLS list.' } },
-  list_skills: { d: 'Search the full skill library.', req: ['query'], p: { query: 'One or two words.' } },
-  fetch_image: { d: 'Download an image into the project.', req: ['url', 'path'], p: { url: 'Image URL.', path: 'Where to save it, e.g. assets/hero.jpg.' } },
-  browser_check: { d: 'Open a page in a real browser and get errors plus a screenshot.', req: ['url'], p: { url: 'file:///absolute/path or http://localhost:port.', wait: 'Extra milliseconds to wait before capturing.', viewport: 'desktop (default), tablet, mobile, or WIDTHxHEIGHT.' } },
-  gmail_send: { d: 'Send a real email now (no drafts). Only when the user asked for it.', req: ['to', 'subject', 'body'], p: { to: 'Recipient address.', subject: 'Subject line.', body: 'Plain-text body.' } },
-  gmail_search: { d: 'Search the connected Gmail account.', req: ['query'], p: { query: 'Gmail search syntax.' } },
-  slack_post_message: { d: 'Post a real Slack message now. Only when the user asked for it.', req: ['channel', 'text'], p: { channel: 'Channel name or id.', text: 'Message text.' } },
-  vercel_deploy: { d: 'Deploy a folder to the connected Vercel account (live, real).', req: [], p: { path: 'Folder to deploy. Default ".".' } },
-  supabase_create_project: { d: 'Create a new Supabase project (billable, real).', req: ['name'], p: { name: 'Project name.' } },
-  supabase_delete_project: { d: 'Permanently delete a Supabase project. Only when the user clearly asked.', req: [], p: { name: 'Project name.', ref: 'Project ref.' } },
-  github_create_repo: { d: 'Create a GitHub repo in the connected account and push this folder to it.', req: ['name'], p: { path: 'Folder to push. Default ".".', name: 'Repository name.' } },
-  design_reference_search: { d: 'Search the local library of real app screenshots for a kind of screen.', req: ['term'], p: { term: 'What the screen is, e.g. "checkout".', category: 'Optional category filter.' } },
-  view_images: { d: 'Look at images (e.g. reference screenshots).', req: ['urls'], p: { urls: 'Comma-separated image URLs.' } },
-  supabase_api: { d: 'Call the Supabase Management API.', req: ['method', 'path'], p: { method: 'GET, POST, PATCH, PUT or DELETE.', path: 'Path starting with /v1/.', body: 'JSON body.' } },
-  supabase_sql: { d: 'Run SQL on a Supabase project database.', req: ['ref', 'query'], p: { ref: 'Project ref.', query: 'SQL.' } },
-  vercel_api: { d: 'Call the Vercel REST API.', req: ['method', 'path'], p: { method: 'HTTP method.', path: 'Path like /v9/projects.', body: 'JSON body.' } },
-  web_fetch: { d: 'Fetch a web page or API URL and get its content as readable text.', req: ['url'], p: { url: 'http:// or https:// URL.', format: '"text" (default, HTML converted to readable text), "html" or "raw".' } },
-  web_search: { d: 'Search the web for current information (docs, errors, versions, news).', req: ['query'], p: { query: 'The search query.', num: 'Number of results (1-20, default 8).' } },
-  apply_patch: { d: 'Change several files at once (add, update, delete, move) with one patch. Use it for multi-file changes; use edit_file for a single block.', req: ['patch'], p: { patch: 'The patch text: "*** Begin Patch", then "*** Add File: path" (lines start with +), "*** Update File: path" (optional "*** Move to: path", "@@ anchor line", then lines starting with " ", "-" or "+"), "*** Delete File: path", and "*** End Patch".' } },
-  lsp: { d: 'Code navigation for TypeScript/JavaScript: find the real definition or every reference of a symbol, hover for its type, outline a file, or search symbols.', req: ['operation'], p: { operation: 'definition, references, implementation, hover, documentSymbol or workspaceSymbol.', path: 'File the symbol is in (not needed for workspaceSymbol).', line: '1-based line of the symbol.', character: '1-based column of the symbol (or give symbol instead).', symbol: 'The name on that line, e.g. fetchUser.', query: 'For workspaceSymbol: the name to look for.' } },
-  plan_exit: { d: 'Plan mode only. Call after the plan file is written: asks the user whether to switch to Build mode and start implementing.', req: [], p: { path: 'The plan file, e.g. .codeply/plans/add-login.md. Default: the newest plan file.' } },
-  plan_enter: { d: 'Build mode only. Ask the user to switch to Plan mode first, for a large or ambiguous task. Use sparingly.', req: [], p: { reason: 'Why planning first would help, one sentence.' } },
-  mcp: { d: 'Call a tool on a connected MCP server.', req: ['server', 'tool'], p: { server: 'Server name from MCP SERVERS.', tool: 'Tool name on that server.', args: 'JSON object of the tool\'s arguments.' } },
-};
-
-/** OpenAI-style tool definitions for every action in `params` (name -> param list). */
-export function buildToolSchemas(params) {
-  return Object.entries(params).map(([name, list]) => {
-    const doc = TOOL_DOCS[name] || { d: name.replace(/_/g, ' '), req: [], p: {} };
-    const properties = {};
-    for (const p of list) properties[p] = { type: 'string', description: doc.p[p] || p };
-    return {
-      type: 'function',
-      function: {
-        name,
-        description: doc.d,
-        parameters: { type: 'object', properties, required: doc.req.filter((r) => list.includes(r)) },
-      },
-    };
-  });
-}
-
-/**
- * Engine history -> OpenAI-shaped native history.
- *
- * Each assistant reply with action blocks becomes one assistant message with
- * tool_calls, followed by exactly one "tool" message per call (APIs reject a
- * tool call with no answer). Results are matched to calls in order by name;
- * a call that got no result (skipped by a guard) gets a short stand-in, and
- * any notes that arrived in between go after the tool messages.
- *
- * @param {Array} messages      engine transcript
- * @param {Function} parse      agent.mjs parseReply
- */
-export function toNativeMessages(messages, parse, mcpToNative = null) {
-  // An engine <codeply:mcp> call goes out as the MCP tool's own function.
-  const nativeCall = (c) => {
-    if (c.name === 'mcp' && mcpToNative) {
-      const fn = mcpToNative[`${c.args?.server}/${c.args?.tool}`];
-      if (fn) return { name: fn, arguments: String(c.args?.args || '').trim() || '{}' };
-    }
-    return { name: c.name, arguments: JSON.stringify(c.args || {}) };
-  };
-  const out = [];
-  let seq = 0;
-  let i = 0;
-  while (i < messages.length) {
-    const m = messages[i];
-    if (m.role !== 'assistant' || typeof m.content !== 'string' || !m.content.includes('<codeply:')) {
-      out.push(m);
-      i++;
-      continue;
-    }
-    const { prose, calls } = parse(m.content);
-    if (!calls.length) { out.push(m); i++; continue; }
-    const withIds = calls.map((c) => ({ ...c, id: `call_${++seq}` }));
-    out.push({
-      role: 'assistant',
-      content: prose || '',
-      tool_calls: withIds.map((c) => ({ id: c.id, type: 'function', function: nativeCall(c) })),
-    });
-    // Collect everything up to the next assistant message.
-    const answers = new Map();
-    const notes = [];
-    let j = i + 1;
-    for (; j < messages.length && messages[j].role !== 'assistant'; j++) {
-      const r = messages[j];
-      const text = typeof r.content === 'string' ? r.content : (r.content.find?.((p) => p.type === 'text')?.text || '');
-      const hit = /^\[tool result: ([a-z_]+)\]/.exec(text);
-      const target = hit && withIds.find((c) => c.name === hit[1] && !answers.has(c.id));
-      if (target) answers.set(target.id, r);
-      else notes.push(r);
-    }
-    const images = [];
-    for (const c of withIds) {
-      const r = answers.get(c.id);
-      if (!r) {
-        out.push({ role: 'tool', tool_call_id: c.id, name: c.name, content: '(not run - see the note that follows)' });
-        continue;
-      }
-      if (typeof r.content === 'string') {
-        out.push({ role: 'tool', tool_call_id: c.id, name: c.name, content: r.content });
-      } else {
-        // Tool messages are text-only on most APIs: screenshots ride in a
-        // user message right after the tool results.
-        out.push({ role: 'tool', tool_call_id: c.id, name: c.name, content: r.content.find((p) => p.type === 'text')?.text || '' });
-        images.push(...r.content.filter((p) => p.type === 'image_url'));
-      }
-    }
-    if (images.length) out.push({ role: 'user', content: [{ type: 'text', text: 'Images from the tool results above:' }, ...images] });
-    out.push(...notes);
-    i = j;
-  }
-  return out;
-}
-
-/**
- * The model's native tool calls -> the <codeply:...> blocks the loop runs.
- * Unknown tools and unparseable arguments are reported back instead of run.
- * @returns {{ text: string, problems: string[] }}
- */
-export function toolCallsToTags(content, toolCalls, params, mcpFromNative = null) {
-  const blocks = [];
-  const problems = [];
-  for (const tc of toolCalls || []) {
-    let name = tc.function?.name;
-    const mcpTarget = mcpFromNative && mcpFromNative[name];
-    if (!name || (!params[name] && !mcpTarget)) { problems.push(`unknown tool "${name}"`); continue; }
-    let args;
-    try { args = JSON.parse(tc.function.arguments || '{}'); }
-    catch { problems.push(`${name}: arguments were not valid JSON`); continue; }
-    if (mcpTarget) {
-      args = { server: mcpTarget.server, tool: mcpTarget.tool, args: JSON.stringify(args || {}) };
-      name = 'mcp';
-    }
-    const inner = params[name]
-      .filter((p) => args[p] !== undefined && args[p] !== null)
-      .map((p) => {
-        const v = typeof args[p] === 'string' ? args[p] : JSON.stringify(args[p]);
-        return `<${p}>\n${v}\n</${p}>`;
-      })
-      .join('\n');
-    blocks.push(`<codeply:${name}>\n${inner}\n</codeply:${name}>`);
-  }
-  return { text: [String(content || '').trim(), ...blocks].filter(Boolean).join('\n'), problems };
-}
-
-// Replaces the tag examples in the system prompt when native tools are on.
-export const NATIVE_FORMAT = `TOOLS
-You act by calling the tools you have been given (native function calling).
-Call a tool, then stop and wait: its result comes back to you. Up to 4 read_file /
-list_dir / search calls may go in one reply and run together; anything that
-changes something (write_file, edit_file, run, ...) goes alone. Before a call,
-write at most one short sentence. When the task is finished, reply with text
-only and no tool call; that ends your turn. In the rules below, "action block"
-means a tool call.`;
-
-/** Errors that mean "this model or endpoint does not do function calling". */
-export const TOOLS_UNSUPPORTED = /does not support tools|tools? (?:are|is) not supported|not support(?:ed)? (?:function|tool)|function calling is not|unsupported parameter:? '?tools|unrecognized request argument.*tools|extra inputs are not permitted.*tools|unknown field .?tools|tool_choice.*not supported|no endpoints found that support tool use/i;
+/**
+ * Native function calling as a transport.
+ *
+ * Inside the engine everything stays in one format: the model's actions are
+ * <codeply:name> blocks in its reply, and results come back as
+ * "[tool result: name]" user messages. Every guard in agent.mjs (claim audit,
+ * stuck detection, verification gates, batching) works on that format.
+ *
+ * For models with real function calling this module translates at the edge:
+ *   - toNativeMessages(): engine history -> OpenAI-shaped history
+ *     (assistant.tool_calls + role "tool" results), which ai.js sends as is to
+ *     OpenAI-compatible APIs and converts for Anthropic and Ollama.
+ *   - toolCallsToTags(): the model's tool calls -> <codeply:...> blocks, so
+ *     the rest of the loop sees exactly what it would have seen in text mode.
+ * Structured arguments mean no tag-format mistakes, and the model uses the
+ * calling convention it was trained on.
+ */
+
+// What the model is told about each action. Parameters not listed as
+// required are optional. Descriptions stay short: the system prompt's RULES
+// section carries the detailed guidance.
+export const TOOL_DOCS = {
+  todo: { d: 'Replace your task checklist for a multi-step task. Send the whole list every time.', req: ['items'], p: { items: 'One task per line: "[ ] pending", "[>] in progress" (only one), "[x] done" (only after a tool confirmed it), "[-] dropped".' } },
+  ask_user: { d: 'Ask the user a multiple-choice question when a real decision is theirs to make and cannot be inferred from the project.', req: ['question'], p: { question: 'The question, one sentence.', options: 'Up to 5 short answers, one per line, recommended first.' } },
+  list_dir: { d: 'List a folder.', req: [], p: { path: 'Folder, relative to the project root. Default ".".' } },
+  read_file: { d: 'Read a text file with line numbers.', req: ['path'], p: { path: 'File path.', offset: '0-based line to start at.', limit: 'Max lines (up to 600).' } },
+  write_file: { d: 'Create a file or fully rewrite one. For changes to an existing file prefer edit_file.', req: ['path', 'content'], p: { path: 'File path.', content: 'The complete file content, verbatim.' } },
+  edit_file: { d: 'Replace one exact block of an existing file.', req: ['path', 'search', 'replace'], p: { path: 'File path.', search: 'Existing text copied exactly from a read (no line-number prefixes), just enough to be unique.', replace: 'The replacement text.', all: '"true" to replace every occurrence (renames).' } },
+  search: { d: 'Regex search across project files, or find files by name when only glob is given.', req: [], p: { pattern: 'Case-insensitive regex.', glob: 'Limit to paths like "**/*.ts", or alone to find files by name.', path: 'Folder or single file to search in.', context: 'Lines of surrounding code per hit (0-5).', files_only: '"true" to list matching files with hit counts.' } },
+  run: { d: 'Run a shell command in the project folder on the user\'s machine.', req: ['command'], p: { command: 'The command line.' } },
+  use_skill: { d: 'Load a skill\'s full instructions.', req: ['name'], p: { name: 'Skill name from the SKILLS list.' } },
+  list_skills: { d: 'Search the full skill library.', req: ['query'], p: { query: 'One or two words.' } },
+  fetch_image: { d: 'Download an image into the project.', req: ['url', 'path'], p: { url: 'Image URL.', path: 'Where to save it, e.g. assets/hero.jpg.' } },
+  browser_check: { d: 'Open a page in a real browser and get errors plus a screenshot.', req: ['url'], p: { url: 'file:///absolute/path or http://localhost:port.', wait: 'Extra milliseconds to wait before capturing.', viewport: 'desktop (default), tablet, mobile, or WIDTHxHEIGHT.' } },
+  gmail_send: { d: 'Send a real email now (no drafts). Only when the user asked for it.', req: ['to', 'subject', 'body'], p: { to: 'Recipient address.', subject: 'Subject line.', body: 'Plain-text body.' } },
+  gmail_search: { d: 'Search the connected Gmail account.', req: ['query'], p: { query: 'Gmail search syntax.' } },
+  slack_post_message: { d: 'Post a real Slack message now. Only when the user asked for it.', req: ['channel', 'text'], p: { channel: 'Channel name or id.', text: 'Message text.' } },
+  vercel_deploy: { d: 'Deploy a folder to the connected Vercel account (live, real).', req: [], p: { path: 'Folder to deploy. Default ".".' } },
+  supabase_create_project: { d: 'Create a new Supabase project (billable, real).', req: ['name'], p: { name: 'Project name.' } },
+  supabase_delete_project: { d: 'Permanently delete a Supabase project. Only when the user clearly asked.', req: [], p: { name: 'Project name.', ref: 'Project ref.' } },
+  github_create_repo: { d: 'Create a GitHub repo in the connected account and push this folder to it.', req: ['name'], p: { path: 'Folder to push. Default ".".', name: 'Repository name.' } },
+  design_reference_search: { d: 'Search the local library of real app screenshots for a kind of screen.', req: ['term'], p: { term: 'What the screen is, e.g. "checkout".', category: 'Optional category filter.' } },
+  view_images: { d: 'Look at images (e.g. reference screenshots).', req: ['urls'], p: { urls: 'Comma-separated image URLs.' } },
+  supabase_api: { d: 'Call the Supabase Management API.', req: ['method', 'path'], p: { method: 'GET, POST, PATCH, PUT or DELETE.', path: 'Path starting with /v1/.', body: 'JSON body.' } },
+  supabase_sql: { d: 'Run SQL on a Supabase project database.', req: ['ref', 'query'], p: { ref: 'Project ref.', query: 'SQL.' } },
+  vercel_api: { d: 'Call the Vercel REST API.', req: ['method', 'path'], p: { method: 'HTTP method.', path: 'Path like /v9/projects.', body: 'JSON body.' } },
+  web_fetch: { d: 'Fetch a web page or API URL and get its content as readable text.', req: ['url'], p: { url: 'http:// or https:// URL.', format: '"text" (default, HTML converted to readable text), "html" or "raw".' } },
+  web_search: { d: 'Search the web for current information (docs, errors, versions, news).', req: ['query'], p: { query: 'The search query.', num: 'Number of results (1-20, default 8).' } },
+  apply_patch: { d: 'Change several files at once (add, update, delete, move) with one patch. Use it for multi-file changes; use edit_file for a single block.', req: ['patch'], p: { patch: 'The patch text: "*** Begin Patch", then "*** Add File: path" (lines start with +), "*** Update File: path" (optional "*** Move to: path", "@@ anchor line", then lines starting with " ", "-" or "+"), "*** Delete File: path", and "*** End Patch".' } },
+  lsp: { d: 'Code navigation for TypeScript/JavaScript: find the real definition or every reference of a symbol, hover for its type, outline a file, or search symbols.', req: ['operation'], p: { operation: 'definition, references, implementation, hover, documentSymbol or workspaceSymbol.', path: 'File the symbol is in (not needed for workspaceSymbol).', line: '1-based line of the symbol.', character: '1-based column of the symbol (or give symbol instead).', symbol: 'The name on that line, e.g. fetchUser.', query: 'For workspaceSymbol: the name to look for.' } },
+  plan_exit: { d: 'Plan mode only. Call after the plan file is written: asks the user whether to switch to Build mode and start implementing.', req: [], p: { path: 'The plan file, e.g. .codeply/plans/add-login.md. Default: the newest plan file.' } },
+  plan_enter: { d: 'Build mode only. Ask the user to switch to Plan mode first, for a large or ambiguous task. Use sparingly.', req: [], p: { reason: 'Why planning first would help, one sentence.' } },
+  mcp: { d: 'Call a tool on a connected MCP server.', req: ['server', 'tool'], p: { server: 'Server name from MCP SERVERS.', tool: 'Tool name on that server.', args: 'JSON object of the tool\'s arguments.' } },
+};
+
+/** OpenAI-style tool definitions for every action in `params` (name -> param list). */
+export function buildToolSchemas(params) {
+  return Object.entries(params).map(([name, list]) => {
+    const doc = TOOL_DOCS[name] || { d: name.replace(/_/g, ' '), req: [], p: {} };
+    const properties = {};
+    for (const p of list) properties[p] = { type: 'string', description: doc.p[p] || p };
+    return {
+      type: 'function',
+      function: {
+        name,
+        description: doc.d,
+        parameters: { type: 'object', properties, required: doc.req.filter((r) => list.includes(r)) },
+      },
+    };
+  });
+}
+
+/**
+ * Engine history -> OpenAI-shaped native history.
+ *
+ * Each assistant reply with action blocks becomes one assistant message with
+ * tool_calls, followed by exactly one "tool" message per call (APIs reject a
+ * tool call with no answer). Results are matched to calls in order by name;
+ * a call that got no result (skipped by a guard) gets a short stand-in, and
+ * any notes that arrived in between go after the tool messages.
+ *
+ * @param {Array} messages      engine transcript
+ * @param {Function} parse      agent.mjs parseReply
+ */
+export function toNativeMessages(messages, parse, mcpToNative = null) {
+  // An engine <codeply:mcp> call goes out as the MCP tool's own function.
+  const nativeCall = (c) => {
+    if (c.name === 'mcp' && mcpToNative) {
+      const fn = mcpToNative[`${c.args?.server}/${c.args?.tool}`];
+      if (fn) return { name: fn, arguments: String(c.args?.args || '').trim() || '{}' };
+    }
+    return { name: c.name, arguments: JSON.stringify(c.args || {}) };
+  };
+  const out = [];
+  let seq = 0;
+  let i = 0;
+  while (i < messages.length) {
+    const m = messages[i];
+    if (m.role !== 'assistant' || typeof m.content !== 'string' || !m.content.includes('<codeply:')) {
+      out.push(m);
+      i++;
+      continue;
+    }
+    const { prose, calls } = parse(m.content);
+    if (!calls.length) { out.push(m); i++; continue; }
+    const withIds = calls.map((c) => ({ ...c, id: `call_${++seq}` }));
+    out.push({
+      role: 'assistant',
+      content: prose || '',
+      tool_calls: withIds.map((c) => ({ id: c.id, type: 'function', function: nativeCall(c) })),
+    });
+    // Collect everything up to the next assistant message.
+    const answers = new Map();
+    const notes = [];
+    let j = i + 1;
+    for (; j < messages.length && messages[j].role !== 'assistant'; j++) {
+      const r = messages[j];
+      const text = typeof r.content === 'string' ? r.content : (r.content.find?.((p) => p.type === 'text')?.text || '');
+      const hit = /^\[tool result: ([a-z_]+)\]/.exec(text);
+      const target = hit && withIds.find((c) => c.name === hit[1] && !answers.has(c.id));
+      if (target) answers.set(target.id, r);
+      else notes.push(r);
+    }
+    const images = [];
+    for (const c of withIds) {
+      const r = answers.get(c.id);
+      if (!r) {
+        out.push({ role: 'tool', tool_call_id: c.id, name: c.name, content: '(not run - see the note that follows)' });
+        continue;
+      }
+      if (typeof r.content === 'string') {
+        out.push({ role: 'tool', tool_call_id: c.id, name: c.name, content: r.content });
+      } else {
+        // Tool messages are text-only on most APIs: screenshots ride in a
+        // user message right after the tool results.
+        out.push({ role: 'tool', tool_call_id: c.id, name: c.name, content: r.content.find((p) => p.type === 'text')?.text || '' });
+        images.push(...r.content.filter((p) => p.type === 'image_url'));
+      }
+    }
+    if (images.length) out.push({ role: 'user', content: [{ type: 'text', text: 'Images from the tool results above:' }, ...images] });
+    out.push(...notes);
+    i = j;
+  }
+  return out;
+}
+
+/**
+ * The model's native tool calls -> the <codeply:...> blocks the loop runs.
+ * Unknown tools and unparseable arguments are reported back instead of run.
+ * @returns {{ text: string, problems: string[] }}
+ */
+export function toolCallsToTags(content, toolCalls, params, mcpFromNative = null) {
+  const blocks = [];
+  const problems = [];
+  for (const tc of toolCalls || []) {
+    let name = tc.function?.name;
+    const mcpTarget = mcpFromNative && mcpFromNative[name];
+    if (!name || (!params[name] && !mcpTarget)) { problems.push(`unknown tool "${name}"`); continue; }
+    let args;
+    try { args = JSON.parse(tc.function.arguments || '{}'); }
+    catch { problems.push(`${name}: arguments were not valid JSON`); continue; }
+    if (mcpTarget) {
+      args = { server: mcpTarget.server, tool: mcpTarget.tool, args: JSON.stringify(args || {}) };
+      name = 'mcp';
+    }
+    const inner = params[name]
+      .filter((p) => args[p] !== undefined && args[p] !== null)
+      .map((p) => {
+        const v = typeof args[p] === 'string' ? args[p] : JSON.stringify(args[p]);
+        return `<${p}>\n${v}\n</${p}>`;
+      })
+      .join('\n');
+    blocks.push(`<codeply:${name}>\n${inner}\n</codeply:${name}>`);
+  }
+  return { text: [String(content || '').trim(), ...blocks].filter(Boolean).join('\n'), problems };
+}
+
+// Replaces the tag examples in the system prompt when native tools are on.
+export const NATIVE_FORMAT = `TOOLS
+You act by calling the tools you have been given (native function calling).
+Call a tool, then stop and wait: its result comes back to you. Up to 4 read_file /
+list_dir / search calls may go in one reply and run together; anything that
+changes something (write_file, edit_file, run, ...) goes alone. Before a call,
+write at most one short sentence. When the task is finished, reply with text
+only and no tool call; that ends your turn. In the rules below, "action block"
+means a tool call.`;
+
+/** Errors that mean "this model or endpoint does not do function calling". */
+export const TOOLS_UNSUPPORTED = /does not support tools|tools? (?:are|is) not supported|not support(?:ed)? (?:function|tool)|function calling is not|unsupported parameter:? '?tools|unrecognized request argument.*tools|extra inputs are not permitted.*tools|unknown field .?tools|tool_choice.*not supported|no endpoints found that support tool use/i;

package/lib/oauth-connectors.js

@@ -423,7 +423,8 @@ function vercelApi(accessToken, teamId, method, apiPath, body) {
 // simpler registration than Vercel/Supabase: no scopes picker or store
 // listing, just a name + callback URL. `repo` scope is what lets
 // githubCreateRepo actually create repositories under the connected account.
-const GITHUB_SCOPES = 'repo gist';
+// `workflow` lets Craft Cloud write craft-cloud.yml into the user's private mirror repo.
+const GITHUB_SCOPES = 'repo gist workflow';
 
 function buildGithubAuthUrl(clientId, redirectUri, state) {
   const params = new URLSearchParams({ client_id: clientId, redirect_uri: redirectUri, scope: GITHUB_SCOPES });

package/lib/permissions.js

@@ -1,109 +1,109 @@
-/**
- * Standing permission rules, set once in a file instead of clicked per chat.
- *
- *   ~/.codeply/permissions.json            (every project)
- *   <project>/.codeply/permissions.json    (this project, checked first)
- *
- *   {
- *     "deny":  ["run:git push*", "run:rm*", "write_file:*.env", "supabase_delete_project"],
- *     "allow": ["run:npm test", "run:npm run *", "run:git status", "edit_file:src/**"]
- *   }
- *
- * A rule is "<tool>" or "<tool>:<pattern>". For run the pattern matches the
- * command's name (npm run dev, git checkout) and the full command line; for
- * file tools it matches the path relative to the project. * matches within a
- * path segment, ** across them, and a trailing " *" also matches the bare
- * command. Deny beats allow; project rules beat global ones; anything no rule
- * mentions is asked as usual. Evaluation is the same idea as opencode's
- * permission/ rulesets.
- */
-const fs = require('fs');
-const os = require('os');
-const path = require('path');
-
-function readRules(file) {
-  try {
-    const j = JSON.parse(fs.readFileSync(file, 'utf8'));
-    const list = (k) => (Array.isArray(j[k]) ? j[k].map(String).filter(Boolean) : []);
-    return { allow: list('allow'), deny: list('deny') };
-  } catch { return { allow: [], deny: [] }; }
-}
-
-function wildcard(pattern) {
-  let re = '';
-  for (let i = 0; i < pattern.length; i++) {
-    const c = pattern[i];
-    if (c === '*' && pattern[i + 1] === '*') { re += '.*'; i++; }
-    else if (c === '*') re += '[^/]*';
-    else if (c === '?') re += '.';
-    else re += c.replace(/[.+^${}()|[\]\\]/g, '\\$&');
-  }
-  // "npm run *" should also cover plain "npm run".
-  re = re.replace(/ \[\^\/\]\*$/, '(?: .*)?').replace(/ \.\*$/, '(?: .*)?');
-  return new RegExp(`^${re}$`, process.platform === 'win32' ? 'i' : '');
-}
-
-/** Strings a request can be matched against for its tool. */
-function subjects(req, cwd) {
-  if (req.tool === 'run') {
-    const cmd = String(req.detail || '').trim();
-    // Commands match by name AND as full text, so "run:npm run *" and
-    // "run:git push*" both work.
-    return { cmd, names: Array.isArray(req.patterns) ? req.patterns : [] };
-  }
-  if (req.path) {
-    const abs = path.isAbsolute(req.path) ? req.path : path.resolve(cwd || '.', req.path);
-    const rel = path.relative(cwd || '.', abs).replace(/\\/g, '/');
-    return { path: rel.startsWith('..') ? abs.replace(/\\/g, '/') : rel };
-  }
-  return {};
-}
-
-function ruleMatches(rule, req, subj) {
-  const i = rule.indexOf(':');
-  const tool = (i === -1 ? rule : rule.slice(0, i)).trim();
-  const pattern = i === -1 ? null : rule.slice(i + 1).trim();
-  if (tool !== req.tool && tool !== '*') return false;
-  if (!pattern || pattern === '*') return true;
-  const re = wildcard(pattern);
-  if (req.tool === 'run') {
-    if (re.test(subj.cmd)) return true;
-    // For allow on a compound command every part must be covered (checked by
-    // the caller); here a single name matching is enough to count.
-    return subj.names.some((n) => re.test(n));
-  }
-  return subj.path ? re.test(subj.path) : false;
-}
-
-/**
- * @param {object} req  the approval request ({tool, detail, patterns, path, danger})
- * @returns {{decision: 'allow'|'deny'|null, rule?: string, file?: string}}
- */
-function decide(req, cwd) {
-  const files = [cwd && path.join(cwd, '.codeply', 'permissions.json'), path.join(os.homedir(), '.codeply', 'permissions.json')].filter(Boolean);
-  const sets = files.map((f) => ({ file: f, ...readRules(f) }));
-  const subj = subjects(req, cwd);
-  for (const s of sets) {
-    const hit = s.deny.find((r) => ruleMatches(r, req, subj));
-    if (hit) return { decision: 'deny', rule: hit, file: s.file };
-  }
-  // Risky-looking commands (rm -rf, format, curl | sh...) always ask.
-  if (req.danger) return { decision: null };
-  for (const s of sets) {
-    if (req.tool === 'run') {
-      // Allow is judged per command, never on the whole line: "npm run *"
-      // must not let "npm run build && npm install" through. A line that
-      // can't be split into named commands (subshells, redirects) is never
-      // auto-allowed.
-      if (!subj.names.length) continue;
-      const parts = subj.names.length === 1 ? [{ cmd: subj.cmd, names: subj.names }] : subj.names.map((n) => ({ cmd: n, names: [n] }));
-      if (parts.every((p) => s.allow.some((r) => ruleMatches(r, req, p)))) return { decision: 'allow', rule: 'allow rules', file: s.file };
-      continue;
-    }
-    const hit = s.allow.find((r) => ruleMatches(r, req, subj));
-    if (hit) return { decision: 'allow', rule: hit, file: s.file };
-  }
-  return { decision: null };
-}
-
-module.exports = { decide, wildcard };
+/**
+ * Standing permission rules, set once in a file instead of clicked per chat.
+ *
+ *   ~/.codeply/permissions.json            (every project)
+ *   <project>/.codeply/permissions.json    (this project, checked first)
+ *
+ *   {
+ *     "deny":  ["run:git push*", "run:rm*", "write_file:*.env", "supabase_delete_project"],
+ *     "allow": ["run:npm test", "run:npm run *", "run:git status", "edit_file:src/**"]
+ *   }
+ *
+ * A rule is "<tool>" or "<tool>:<pattern>". For run the pattern matches the
+ * command's name (npm run dev, git checkout) and the full command line; for
+ * file tools it matches the path relative to the project. * matches within a
+ * path segment, ** across them, and a trailing " *" also matches the bare
+ * command. Deny beats allow; project rules beat global ones; anything no rule
+ * mentions is asked as usual. Evaluation is the same idea as opencode's
+ * permission/ rulesets.
+ */
+const fs = require('fs');
+const os = require('os');
+const path = require('path');
+
+function readRules(file) {
+  try {
+    const j = JSON.parse(fs.readFileSync(file, 'utf8'));
+    const list = (k) => (Array.isArray(j[k]) ? j[k].map(String).filter(Boolean) : []);
+    return { allow: list('allow'), deny: list('deny') };
+  } catch { return { allow: [], deny: [] }; }
+}
+
+function wildcard(pattern) {
+  let re = '';
+  for (let i = 0; i < pattern.length; i++) {
+    const c = pattern[i];
+    if (c === '*' && pattern[i + 1] === '*') { re += '.*'; i++; }
+    else if (c === '*') re += '[^/]*';
+    else if (c === '?') re += '.';
+    else re += c.replace(/[.+^${}()|[\]\\]/g, '\\$&');
+  }
+  // "npm run *" should also cover plain "npm run".
+  re = re.replace(/ \[\^\/\]\*$/, '(?: .*)?').replace(/ \.\*$/, '(?: .*)?');
+  return new RegExp(`^${re}$`, process.platform === 'win32' ? 'i' : '');
+}
+
+/** Strings a request can be matched against for its tool. */
+function subjects(req, cwd) {
+  if (req.tool === 'run') {
+    const cmd = String(req.detail || '').trim();
+    // Commands match by name AND as full text, so "run:npm run *" and
+    // "run:git push*" both work.
+    return { cmd, names: Array.isArray(req.patterns) ? req.patterns : [] };
+  }
+  if (req.path) {
+    const abs = path.isAbsolute(req.path) ? req.path : path.resolve(cwd || '.', req.path);
+    const rel = path.relative(cwd || '.', abs).replace(/\\/g, '/');
+    return { path: rel.startsWith('..') ? abs.replace(/\\/g, '/') : rel };
+  }
+  return {};
+}
+
+function ruleMatches(rule, req, subj) {
+  const i = rule.indexOf(':');
+  const tool = (i === -1 ? rule : rule.slice(0, i)).trim();
+  const pattern = i === -1 ? null : rule.slice(i + 1).trim();
+  if (tool !== req.tool && tool !== '*') return false;
+  if (!pattern || pattern === '*') return true;
+  const re = wildcard(pattern);
+  if (req.tool === 'run') {
+    if (re.test(subj.cmd)) return true;
+    // For allow on a compound command every part must be covered (checked by
+    // the caller); here a single name matching is enough to count.
+    return subj.names.some((n) => re.test(n));
+  }
+  return subj.path ? re.test(subj.path) : false;
+}
+
+/**
+ * @param {object} req  the approval request ({tool, detail, patterns, path, danger})
+ * @returns {{decision: 'allow'|'deny'|null, rule?: string, file?: string}}
+ */
+function decide(req, cwd) {
+  const files = [cwd && path.join(cwd, '.codeply', 'permissions.json'), path.join(os.homedir(), '.codeply', 'permissions.json')].filter(Boolean);
+  const sets = files.map((f) => ({ file: f, ...readRules(f) }));
+  const subj = subjects(req, cwd);
+  for (const s of sets) {
+    const hit = s.deny.find((r) => ruleMatches(r, req, subj));
+    if (hit) return { decision: 'deny', rule: hit, file: s.file };
+  }
+  // Risky-looking commands (rm -rf, format, curl | sh...) always ask.
+  if (req.danger) return { decision: null };
+  for (const s of sets) {
+    if (req.tool === 'run') {
+      // Allow is judged per command, never on the whole line: "npm run *"
+      // must not let "npm run build && npm install" through. A line that
+      // can't be split into named commands (subshells, redirects) is never
+      // auto-allowed.
+      if (!subj.names.length) continue;
+      const parts = subj.names.length === 1 ? [{ cmd: subj.cmd, names: subj.names }] : subj.names.map((n) => ({ cmd: n, names: [n] }));
+      if (parts.every((p) => s.allow.some((r) => ruleMatches(r, req, p)))) return { decision: 'allow', rule: 'allow rules', file: s.file };
+      continue;
+    }
+    const hit = s.allow.find((r) => ruleMatches(r, req, subj));
+    if (hit) return { decision: 'allow', rule: hit, file: s.file };
+  }
+  return { decision: null };
+}
+
+module.exports = { decide, wildcard };