Tools

The agent acts only through tools, and every tool call passes the permission engine first. /tools lists what is available in a session. The effect column below is what the permission engine and the modes reason about.

Built-in tools

Other tools appear in some situations:

• team starts a team run for large, separable tasks ("team": { "enabled": false } turns it off).
• work_declare, work_check_overlap, work_update and work_ack_overlap appear while you are signed in to Govern. See Organisation context.
• The fw_* and hdl_* tools appear when a developer environment is active.

Editing

Tools per model family (tools.profile, default auto). OpenAI-family models (gpt-*, o-series, codex) are offered apply_patch instead of edit_file and multi_edit; Anthropic, Gemini and other models keep the exact-replacement tools. edit, patch and both force a choice.

The patch format:

*** Begin Patch
*** Update File: src/app.ts
@@ function main() {
   const a = 1;
-  const b = 2;
+  const b = 3;
*** Add File: src/new.ts
+export const created = true;
*** Delete File: src/old.ts
*** Update File: src/name.ts
*** Move to: src/renamed.ts
@@
-old line
+new line
*** End Patch

Every section is computed before anything is written, so a failing hunk changes nothing. A hunk whose context differs from the file only in whitespace is applied when it matches exactly one place, and the result says so. A hunk that does not match is reported with its number, its line in the patch, the lines it expected and the closest region of the file.

Edit fallback (tools.fuzzyEdits, default on). When edit_file or multi_edit cannot find the exact text, they retry with indentation-relative matching and then whitespace-normalised matching. A fallback is applied only when it matches exactly one place, and the result names the strategy and the lines, followed by the diff. When nothing matches, the error lists the three closest regions.

Batched reads (tools.batchedReads, default on). read_file takes up to 10 paths and returns each under a ==> path <== header.

Every agent edit is also checked by the checks gate before it is written.

Shell commands

Commands are parsed into a full syntax tree before the permission check, so pipelines, && chains, substitutions, here-documents, redirections and wrappers such as sudo, env, timeout, xargs, sh -c and find -exec are each checked on their own. Commands that cannot be analysed statically, such as eval "$cmd", always ask when deny rules exist.

Commands the agent starts do not inherit secret-named environment variables; see Permissions and modes.

The repl tool

repl keeps an interpreter running for the conversation, so variables, imports and loaded data carry over between calls. It is offered when python3 (Windows: python) or node is on PATH. "tools": { "repl": false } turns it off.

Standard input is not available. On a timeout or interruption the cell is interrupted and the state kept; an interpreter that does not respond within a couple of seconds is restarted and its state lost. Interpreters start in the project root with the same environment and sandbox as run_shell. For rules, the subject is repl:<language> <first line of code>, for example "allow": ["repl(repl:python *)"]. Calls are never treated as read-only.

Web tools

web_fetch fetches a URL as readable text and blocks private addresses. webFetch.allowedDomains, webFetch.deniedDomains and webFetch.maxBytes limit it, as do rules such as web_fetch(domain:docs.rs).

web_search is available only when a backend is configured: BRAVE_SEARCH_API_KEY, TAVILY_API_KEY or SEARXNG_URL.

Helper agents and worktrees

delegate hands a self-contained job to a helper agent with its own conversation, which keeps the main context small. Helpers can run in the background (agents.maxBackground, default 4). Named profiles give helpers a brief, a restricted tool set or another model; see Instructions, commands and skills. worktree creates checkouts under worktree.directory (default .neuralinverse/worktrees).

Verification after edits

After a turn in which the agent changed files, the CLI can run the project's own fast checks (type check, lint, related tests) and hand failures back to the model, for a bounded number of rounds.

Language servers

code_intel and the diagnostics shown after edits use language servers already installed on your machine; nothing is installed for you. Supported: TypeScript and JavaScript (tsc --lsp from TypeScript 7+, typescript-language-server, vtsls), Python (pyright, basedpyright, pylsp, jedi-language-server), Go (gopls), Rust (rust-analyzer), Java (jdtls), C and C++ (clangd), C# (csharp-ls), Ruby (ruby-lsp, solargraph), PHP (intelephense, phpactor), Swift (sourcekit-lsp), Lua, YAML, JSON and Bash. After an edit, new errors and warnings in that file are shown to the agent.

{ "lsp": { "disabled": ["pyright"], "servers": { "zls": { "command": "zls", "extensions": [".zig"] } } } }

Code index and repository map

On first use the CLI builds a local index of the workspace in ~/.neuralinverse/projects/<workspace>/index/, never inside the repository: symbols per file for many languages, imports, and a keyword index over identifiers, paths and text. It honours .gitignore and skips vendored, build and cache folders, generated or minified files, binaries and files over 1 MB. A file watcher (or a rescan every 60 seconds) keeps it current. /index shows its status; /index rebuild scans again.

A short repository map (about 1500 tokens: the most imported and most recently changed files with their top-level declarations) is added to the system prompt.

{ "index": { "enabled": true, "maxFiles": 50000, "repoMapTokens": 1500 } }

Semantic search is off by default, and the index is purely local. Setting index.embeddings turns on vector search and sends chunks of your indexed source code and your search queries to that provider's embeddings endpoint. Use a provider you trust with your code; a local Ollama model keeps everything on your machine:

{ "index": { "embeddings": { "provider": "ollama", "model": "nomic-embed-text" } } }

Supported: OpenAI-compatible /embeddings (OpenAI, Azure OpenAI, vLLM, LM Studio), Ollama /api/embed, Gemini batchEmbedContents, and Bedrock Titan and Cohere embedding models.

How the agent loop is tuned

Each default is meant to be changed only after measurement with the evals.