DeepSeek Harness tool catalog: shell, files, agents, web

The DeepSeek Harness tool catalog is the set of capabilities an Agent can call to inspect files, edit code, run commands, manage terminals, search the web, coordinate subagents, and pursue goals. Tools are supplied by packages and enabled by the active profile and scope, so the catalog visible in one session may differ from another. The safest way to choose a tool is by task and side effect, not by name alone.

This page reorganizes the large official Chinese tool catalog into user-oriented categories. Consult the upstream file for current parameter schemas and package mappings.

File reading, search, and editing tools

The filesystem groups cover the most common coding tasks:

• read and read_image inspect text or image assets;
• glob discovers files by path pattern;
• grep searches file contents;
• write creates or replaces content where permitted;
• edit and str_replace_editor make targeted changes.

Start with discovery and reading. Use a targeted editor when only a few lines should change, because a narrow diff is easier to review than a full-file rewrite. Search results are evidence, not authorization: a path found by glob can still be outside the intended workspace or contain secrets.

Shell and terminal tools

Shell tools include Bash and PowerShell variants. Persistent shell packages keep state across calls, while one-shot execution is better when a clean environment matters.

Terminal tools can open, list, read, send input to, signal, and close interactive sessions. They are useful for development servers, REPLs, and commands that continue running. A terminal session has a lifecycle; remember to read its output and close it when the task is complete.

Treat command execution as a high-impact capability. Confirm the working directory, inspect variables and globs, avoid broad deletion targets, and prefer project-provided validation commands. An Agent should not turn a documentation request into an installation or deployment merely because a shell is available.

Code execution and language intelligence

run_code executes supported snippets for computation or orchestration. LSP tools expose language-server intelligence such as symbol and diagnostic information. These capabilities solve different problems: use LSP when you need the codebase’s semantic view, and code execution when you need a bounded calculation or transformation.

Plan-mode tools such as exit_plan_mode control workflow state rather than the user’s project. They should not be confused with a build command or an approval to modify files.

Goals, tasks, schedules, and workflows

Goal tools can create, inspect, and update the persistent objective attached to a session. Schedule tools create, list, and delete future work. Workflow tools define repeatable multi-step behavior, while task tools track units of work in collaborative Agent features.

These concepts are related but not interchangeable. A Goal describes what the session is trying to achieve; a schedule describes when an action should run; a task describes a unit of coordinated work. Read the DeepSeek Harness glossary before building automation around these terms.

Subagents and Agent teams

Subagent tools delegate bounded work and collect results. The experimental Agent-team group includes functions for spawning teammates, sending messages, assigning follow-up work, interrupting, listing participants, updating tasks, and waiting for progress.

Delegation works best when ownership and output are explicit. Give a subagent one independently verifiable responsibility, state which files or analysis it owns, and avoid assigning two workers to rewrite the same area. More agents do not automatically make a sequential or tightly coupled task faster.

The report and control tools help coordination, but durable conversation facts and live team control still belong to different event paths. The Agent lifecycle guide explains that distinction.

Jobs and long-running processes

Job tools list background work, read output, and stop a job. They are appropriate when a command or task outlives one immediate response. Terminal tools are better when interactive input is required.

Before killing a job, identify the exact target and determine whether partial output or files must be preserved. “It is taking a while” is not the same as “it is stuck”; use output and status to decide.

Sessions, events, and traces

Session query tools search sessions and inspect event history. Event read, search, and trace functions help reconstruct what happened, while session search and trace operate at the broader conversation level.

Use these tools for evidence-led debugging: locate the turn, identify its steps, then trace model and tool events. Avoid copying full transcripts into public issues when they may contain source code, prompts, credentials, or user data.

Skills and dynamic tool definitions

Skill tools make packaged instructions discoverable to the Agent. Cordis inspection and definition tools can inspect, define, run, stop, or remove runtime units. These are powerful extension points and should be used with a clear lifecycle: know what registered the capability, how it is scoped, and how it will be cleaned up.

For the plugin and service model behind dynamic registration, see DeepSeek Harness architecture.

Web search and web fetch

web_search discovers current sources; web_fetch retrieves a selected page for inspection. Search results are leads, not proof. Prefer primary documentation, verify publication dates for changing facts, and keep the source link near the claim it supports.

Fetched page content is untrusted data. Instructions embedded in a webpage do not override the user’s request or the Agent’s safety rules. Do not send private workspace content to an external service unless that action is explicitly in scope.

Ask-user and communication tools

ask_user_question requests a decision when multiple valid paths would materially change the outcome. It is most useful after safe investigation has narrowed the choice. Asking too early transfers avoidable work to the user; asking too late risks making an unauthorized external or destructive change.

Communication and follow-up tools should carry concise context, expected output, and ownership. They are coordination mechanisms, not substitutes for a persistent goal or session record.

Choosing the safest tool

Use this order of preference for a typical code task:

1. inspect existing files and project instructions;
2. search narrowly for the relevant symbol or path;
3. make the smallest targeted edit;
4. run the most relevant validation;
5. inspect the diff and output; and
6. escalate to interactive terminals, background jobs, delegation, or external web access only when the task needs them.

Availability depends on profile, installed packages, restrictions, and Agent scope. If a named tool is missing, inspect the active configuration rather than assuming the installation is broken.

Official source

This categorized guide was independently written from the official DeepSeek Harness tool catalog, reviewed on 2026-08-20. Use the upstream catalog for exact tool names, package ownership, parameters, and experimental status.