Tools

The agent has access to the following tools:

Tool	Description
read_file	Read file contents with optional offset/limit
read_image	Send a local PNG, JPEG, GIF, or WebP image to a multimodal model (not registered for text-only models)
write_file	Create or overwrite files, creating parent directories
edit_file	Surgical text replacement in files (whitespace-tolerant match)
multi_edit	Apply an ordered batch of edits to one file atomically (all-or-nothing)
ls	List a directory's immediate contents (gitignore-aware)
exec_command	Run foreground shell commands with real-time streaming output
wait	Pause before subsequent sequential tool calls, with an interruptible countdown
glob_files	Find files matching a glob pattern (gitignore-aware)
grep_files	Search file contents using ripgrep regex (gitignore-aware)
web_search	Search the web via DuckDuckGo
web_fetch	Fetch and parse content from a URL with an interruptible 30-second overall timeout; honors HTTP_PROXY, HTTPS_PROXY, ALL_PROXY, SOCKS_PROXY, and NO_PROXY from the environment
remove_file	Remove a file or directory
load_skill	Load the full prompt instructions for a skill by name
start_process	Start a managed background process with file-backed output and live status
inspect_processes	Inspect one or all managed background processes, with an optional bounded log tail for one process
wait_processes	Block until a selected managed process exits, optionally writes output, or an optional timeout expires; omit the timeout for ordinary finite jobs
stop_process	Stop a managed background process and its child process tree

Images loaded with read_image remain in conversation history for subsequent tool rounds and follow-up questions. Saved sessions retain the image bytes, so resuming does not require the original file. Images remain available until their messages are removed by context compaction, rewind, or clearing the conversation. Repeated requests to image-capable models include retained images and their image-token costs, subject to a separate byte budget. Inline image data is capped at 24 MiB per request to leave headroom below common 32 MB HTTP body limits. Older images are replaced with text placeholders in outgoing requests when necessary; their saved bytes are not deleted. The newest image is kept (an image larger than 24 MiB must be resized or compressed). A gateway body-size rejection (content_length_limit or HTTP 413) triggers one recovery attempt: reduce older image payloads, or try context compaction if images cannot shrink. The reduced image budget applies to subsequent requests in the live context, while still preserving the newest image. If recovery fails, reduce attachment sizes or use /clear to start a new session; raising context_length does not raise the gateway's byte limit. Switching to a text-only model sends a text placeholder instead; the saved image bytes remain available when switching back. Image tool payloads are not treated as user prompts in session previews, replay, or rewind. Context sizing uses reported prompt usage when available and a 2,048-token planning allowance per added or unmeasured image; actual costs vary by model, resolution, and detail. Image costs are kept separate from text calibration and included when deciding which older messages to compact.

Managed background process tools are built into ene so permitted model calls, the /ps command, and the live terminal/web status use the same process registry. Like other built-in model tools, their advertisement is subject to the active persona's tool policy; /ps and live status remain available to the UI.

While jobs are active, the terminal shows running and finished counts followed by one line per running process. The Web UI presents the same activity as individual rows in the composer dock. Both include the managed ID, label, automatically formatted runtime, and latest log line. Managed IDs are short session-local numbers (1, 2, ...). Use /ps to list all jobs with their latest output, /ps <label|process-id> [tail-chars] for details and recent output, or /ps stop <label|process-id> to stop one manually. Combined output is stored in .ene/processes/<process-id>-<unique-suffix>.log.

Processes are live-session-scoped. They survive terminal detach and session switching, but are terminated on explicit exit or ene kill. The bundled monitor skill adds an active-monitoring workflow. For periodic monitoring, call the core wait tool first and put the inspection or status calls after it in the same sequential tool-call batch; do not group the wait and checks in parallel.

Skill-provided tools

A trusted skill may ship executable Python in a root tools.py (a module-level TOOLS list of {schema, run, describe, describe_output} entries; both descriptors are optional). describe(arguments) returns a ToolCallDescription for the call label. describe_output(result) returns a concise string for the successful result; failures use the standard error formatter. This keeps each skill's call and result semantics beside its tools while the shared UI owns rendering. The full result still goes to the model, while the concise output is persisted for consistent live and replay display. A skill's tools are registered on its first load and remain available for that session. /skills reload removes tools only when their skill is no longer discovered; use /clear or restart Ene to reload an edited tools.py. A schema collision or broken tools.py fails the skill load instead of partially registering its tools.

The bundled batch skill uses the agent's already configured provider to run direct, tool-free model requests without exposing credentials to the skill. It adds bounded parallelism, structured output, durable JSONL results, and resume support. See its batch instructions for the workflow.

Tool	Description
run_batch	Map one direct LLM transformation over independent text or image items, appending results to a resumable JSONL file