@semalt-ai/code 1.8.5 → 1.19.0

package/lib/tool_specs.js

@@ -50,7 +50,11 @@ const TOOL_SPECS = {
   },
 
   read_file: {
-    description: 'Read a UTF-8 text file and return its full contents.',
+    description: 'Read a UTF-8 text file. Paginated: by default returns the first ~2000 lines; '
+      + 'if the file is longer the result ends with a [PARTIAL] notice giving the total line count '
+      + 'and the start_line for the next page. Use start_line/end_line to read a specific slice '
+      + '(also capped to ~2000 lines). Set show_line_numbers=true when you need line numbers to '
+      + 'target an edit_file call (off by default — keeps content copyable for replace_in_file and cheaper).',
     parameters: {
       type: 'object',
       properties: {
@@ -58,6 +62,20 @@ const TOOL_SPECS = {
           type: 'string',
           description: 'Absolute or relative path to the file to read',
         },
+        start_line: {
+          type: 'integer',
+          description: 'First line to read (1-based, inclusive). Omit to start at line 1.',
+        },
+        end_line: {
+          type: 'integer',
+          description: 'Last line to read (1-based, inclusive). Omit to read to the line cap. '
+            + 'A range wider than the line cap is still capped — page with start_line.',
+        },
+        show_line_numbers: {
+          type: 'boolean',
+          description: 'Prefix each line with its 1-based line number (aligned with edit_file). '
+            + 'Default false. Turn on when you intend to edit by line number.',
+        },
       },
       required: ['path'],
     },
@@ -289,6 +307,70 @@ const TOOL_SPECS = {
     },
   },
 
+  grep: {
+    description: 'Search file contents by regular expression across the working tree and return the actual matches (file:line:text), so you can navigate to the relevant lines and read only a slice instead of whole files. Honors .gitignore, skips binary files and node_modules/.git. Prefers ripgrep when available with an identical pure-Node fallback. output_mode selects what is returned: "content" (default) gives file:line:text per match ("show me the lines"); "files_with_matches" gives just the unique file paths ("which files contain this"); "count" gives per-file and total match counts ("how many") — count/files_with_matches return almost nothing, so prefer them when you do not need the line text. Results are bounded by head_limit (default 100); when more matches exist than are shown, a truncation notice tells you how many were omitted — narrow the pattern, switch to count/files_with_matches, or raise head_limit.',
+    parameters: {
+      type: 'object',
+      properties: {
+        pattern: {
+          type: 'string',
+          description: 'Regular expression to search for (matched per line)',
+        },
+        path: {
+          type: 'string',
+          description: 'Optional glob limiting which files are searched, e.g. "*.js" or "src/**/*.ts"',
+        },
+        ignore_case: {
+          type: 'boolean',
+          description: 'Case-insensitive match when true',
+          default: false,
+        },
+        output_mode: {
+          type: 'string',
+          enum: ['content', 'files_with_matches', 'count'],
+          description: 'What to return: "content" = file:line:text per match (default); "files_with_matches" = unique file paths only; "count" = match counts per file and total. Use count/files_with_matches when you only need how many / which files.',
+          default: 'content',
+        },
+        head_limit: {
+          type: 'integer',
+          description: 'Maximum number of results to serialize into context (default 100): match lines in content mode, files in files_with_matches/count mode. Excess is omitted with a truncation notice.',
+        },
+        offset: {
+          type: 'integer',
+          description: 'Number of results to skip before serializing (for paging through a large match set). Default 0.',
+        },
+      },
+      required: ['pattern'],
+    },
+  },
+
+  glob: {
+    description: 'List files matching a glob pattern (relative paths). Skips node_modules/.git and hidden entries. The file list is bounded by head_limit (default 100); when more files match than are shown, a truncation notice reports how many were omitted — narrow the glob or raise head_limit.',
+    parameters: {
+      type: 'object',
+      properties: {
+        pattern: {
+          type: 'string',
+          description: 'Glob pattern such as "*.ts" or "src/**/*.js"; matches the basename when it has no slash, otherwise the relative path',
+        },
+        path: {
+          type: 'string',
+          description: 'Base directory to search from; defaults to the current working directory',
+          default: '.',
+        },
+        head_limit: {
+          type: 'integer',
+          description: 'Maximum number of file paths to serialize into context (default 100). Excess is omitted with a truncation notice.',
+        },
+        offset: {
+          type: 'integer',
+          description: 'Number of file paths to skip before serializing (for paging). Default 0.',
+        },
+      },
+      required: ['pattern'],
+    },
+  },
+
   replace_in_file: {
     description: 'Regex-replace occurrences of a pattern in a file; returns the number of replacements made.',
     parameters: {
@@ -384,7 +466,7 @@ const TOOL_SPECS = {
   },
 
   http_get: {
-    description: 'Perform an HTTP GET and return the response status code and body, truncated to a configured byte cap.',
+    description: 'Fetch a web page. The `mode` parameter selects how much processing to apply: "summarized" (default) runs the page through Readability + Markdown conversion and a secondary-model summary so only a compact result enters context — use it for find/answer tasks; "extracted" returns the extracted main-content Markdown without the summary — use it to read an article/doc verbatim or grab an exact snippet/quote; "raw" bypasses extraction entirely and returns the ORIGINAL fetched HTML/content (token-capped) — use it to analyze a page\'s HTML/CSS/JavaScript/markup/structure, which extraction destroys. Plain-text/JSON responses pass through unchanged in summarized/extracted modes. Optionally pass intent to focus the summary. The deprecated booleans summarize=false and raw=true are aliases for mode="extracted". DATA EXTRACTION: to pull SPECIFIC VALUES out of a page (hex colors, version strings, URLs, IDs, counts), do NOT use http_get — use `download` (or sandboxed `curl`) to save the page to the working directory, then `grep` it (e.g. `grep -oiE \'#[0-9a-f]{6}\'` for hex colors) so only the matching lines enter context, not the whole page. Reserve mode="raw" for when you actually need to read the markup structure; it loads the entire (token-capped) page and is expensive for simple value extraction. For SPA / asset-heavy sites the values often live in linked assets (e.g. /_nuxt/*.css or *.js, bundled stylesheets) rather than the top-level HTML — download+grep those asset URLs.',
     parameters: {
       type: 'object',
       properties: {
@@ -392,11 +474,46 @@ const TOOL_SPECS = {
           type: 'string',
           description: 'HTTP or HTTPS URL to fetch',
         },
+        mode: {
+          type: 'string',
+          enum: ['summarized', 'extracted', 'raw'],
+          description: 'How to process the page. "summarized" (default): extract main content, convert to Markdown, then summarize with a secondary model — for find/answer tasks. "extracted": extracted Markdown of the main content, no summary — for reading a doc/article verbatim or an exact snippet. "raw": the original fetched HTML/content, token-capped, NO extraction — for analyzing the page\'s HTML/CSS/JS/structure.',
+        },
+        summarize: {
+          type: 'boolean',
+          description: 'DEPRECATED alias (prefer mode). false ≡ mode="extracted" (extracted Markdown, no summary). An explicit mode wins over this.',
+        },
+        raw: {
+          type: 'boolean',
+          description: 'DEPRECATED alias (prefer mode). true ≡ mode="extracted". NOTE: this does NOT return raw HTML — use mode="raw" for the original markup. An explicit mode wins over this.',
+        },
+        intent: {
+          type: 'string',
+          description: 'Why you are fetching this page; focuses the summary on answering it.',
+        },
       },
       required: ['url'],
     },
   },
 
+  web_search: {
+    description: 'Search the web and return a compact list of candidate results — each with a title, url, and snippet. Use this to FIND relevant pages when you do not already know the URL, instead of guessing URLs. IMPORTANT: this returns only short snippets, NOT page content. Read the snippets, pick the single most relevant result (or the few that matter), and then fetch ONLY those with http_get to read them — use http_get mode="summarized" for reading/answering and mode="raw" only for analyzing markup. Do NOT fetch every result; that wastes context. Optional count caps how many results to return (the backend clamps it).',
+    parameters: {
+      type: 'object',
+      properties: {
+        query: {
+          type: 'string',
+          description: 'The search query.',
+        },
+        count: {
+          type: 'integer',
+          description: 'Optional maximum number of results to return (clamped by the backend; defaults to the backend default).',
+        },
+      },
+      required: ['query'],
+    },
+  },
+
   ask_user: {
     description: 'Prompt the interactive user for input and return their answer; auto-answers "y" in non-TTY mode.',
     parameters: {
@@ -462,6 +579,109 @@ const TOOL_SPECS = {
     },
   },
 
+  // ──────────────────────────────────────────────────────────────────
+  // Native git tools (Task 5.1). Implemented by shelling out to `git`
+  // through the same sandbox + deny-list chokepoint as <shell>, but they
+  // parse output into structured results. Read-only: git_status, git_diff,
+  // git_log (and the LIST ops of git_branch / git_worktree). Mutating:
+  // git_add, git_commit, git_branch (create/delete), git_checkout,
+  // git_worktree (add/remove). NOTE: git operations are NOT reversible via
+  // /rewind — checkpoints (Task 4.3) snapshot file-tool mutations only.
+  // ──────────────────────────────────────────────────────────────────
+
+  git_status: {
+    description: 'Show the working-tree status as structured lists of staged, unstaged, and untracked files plus the current branch. Read-only.',
+    parameters: { type: 'object', properties: {}, required: [] },
+  },
+
+  git_diff: {
+    description: 'Show changes as a structured diff (files, hunks, +/- counts). Defaults to unstaged changes; set staged=true for the staged (index) diff. Read-only.',
+    parameters: {
+      type: 'object',
+      properties: {
+        staged: { type: 'boolean', description: 'Diff the staged (index) changes instead of the working tree', default: false },
+        path: { type: 'string', description: 'Optional path to limit the diff to' },
+      },
+      required: [],
+    },
+  },
+
+  git_log: {
+    description: 'List recent commits as structured records (hash, author, email, date, subject). Read-only.',
+    parameters: {
+      type: 'object',
+      properties: {
+        count: { type: 'integer', description: 'Maximum number of commits to return (default 20)', minimum: 1, default: 20 },
+        path: { type: 'string', description: 'Optional path to limit history to' },
+      },
+      required: [],
+    },
+  },
+
+  git_add: {
+    description: 'Stage changes for commit. Provide `paths` (one or more files) or set all=true to stage everything. Mutating — requires approval.',
+    parameters: {
+      type: 'object',
+      properties: {
+        paths: { type: 'string', description: 'File path(s) to stage; space-separated for multiple' },
+        all: { type: 'boolean', description: 'Stage all changes (git add -A)', default: false },
+      },
+      required: [],
+    },
+  },
+
+  git_commit: {
+    description: 'Create a commit with the given message and return the new commit hash and branch. The message is required and must be non-empty (the tool never invents one). Mutating — requires approval. NOTE: a commit is not reversible via /rewind.',
+    parameters: {
+      type: 'object',
+      properties: {
+        message: { type: 'string', description: 'Commit message (required, non-empty)' },
+        all: { type: 'boolean', description: 'Automatically stage modified/deleted tracked files before committing (git commit -a)', default: false },
+      },
+      required: ['message'],
+    },
+  },
+
+  git_branch: {
+    description: 'List branches (no name — read-only) or create/delete a branch (name given — mutating). Set delete=true to delete (force=true for -D). Branch creation/deletion is not reversible via /rewind.',
+    parameters: {
+      type: 'object',
+      properties: {
+        name: { type: 'string', description: 'Branch name to create or delete; omit to list branches' },
+        delete: { type: 'boolean', description: 'Delete the named branch instead of creating it', default: false },
+        force: { type: 'boolean', description: 'Force the operation (e.g. delete an unmerged branch with -D)', default: false },
+      },
+      required: [],
+    },
+  },
+
+  git_checkout: {
+    description: 'Switch to a branch or ref (set create=true to create and switch with -b). Mutating — requires approval. WARNING: checkout can overwrite or discard uncommitted changes in the working tree, and those changes are NOT recoverable via /rewind (checkpoints snapshot file-tool mutations only, not git operations).',
+    parameters: {
+      type: 'object',
+      properties: {
+        name: { type: 'string', description: 'Branch or ref to switch to' },
+        create: { type: 'boolean', description: 'Create the branch and switch to it (git checkout -b)', default: false },
+        force: { type: 'boolean', description: 'Force checkout, discarding local changes (git checkout -f)', default: false },
+      },
+      required: ['name'],
+    },
+  },
+
+  git_worktree: {
+    description: 'Manage linked worktrees for running parallel agents in isolated working trees. op=list (read-only) lists worktrees; op=add creates one (optionally on a new `branch`); op=remove deletes one (force=true to override). add/remove are mutating and not reversible via /rewind.',
+    parameters: {
+      type: 'object',
+      properties: {
+        op: { type: 'string', description: 'Operation: list | add | remove', enum: ['list', 'add', 'remove'], default: 'list' },
+        path: { type: 'string', description: 'Worktree directory path (required for add/remove)' },
+        branch: { type: 'string', description: 'For add: create and check out this new branch in the worktree' },
+        force: { type: 'boolean', description: 'For remove: remove even with local changes', default: false },
+      },
+      required: [],
+    },
+  },
+
   // ──────────────────────────────────────────────────────────────────
   // Parser-wrapper tags. These are XML envelopes that carry other tool
   // calls; they are not invocable tools themselves. Entries exist only