Tools
The mounted tools come from published Harness plugins and from the session’s Agent preset; omdsh adds the terminal presentation layer, not a second tool registry. /tools lists exactly what the active session exposes, and the list follows the preset you chose with /agent. Access still gates what each mutating tool may touch.
Shell
| Tool | Behavior |
|---|---|
bash / pwsh | One-shot commands in the confined shell for the host. The shipped default timeout is 120 seconds. Settings → Plugins → Shell controls default and maximum timeouts; explicit per-call timeouts are capped by that maximum. Background commands appear in /jobs. |
terminal_open, terminal_send, terminal_read, terminal_signal, terminal_close, terminal_list | Persistent terminal sessions that retain cwd, environment, and interactive children across calls. |
Working directory
working_directory reads the current directory, or changes it with cd to an existing directory. Later file operations, shell launches, file completion and the footer follow that directory. The original project still groups the session and determines its existing permissions; changing directory does not widen Access or move an already running terminal. Minimal keeps its two-tool catalog.
Git worktrees
@deepseek-ai/dsh-experimental-tool-worktree ships enabled by default (and can be disabled in /plugins → Bundles) to expose create_worktree; Minimal still restricts its model tools. omdsh defaults Keep worktrees when archiving (retainOnArchive: true in /settings → Plugins → Tools) so archiving a session in /resume preserves its session-created worktrees unless you turn that setting off. create_worktree creates a separate Git checkout from a local revision (HEAD by default) and switches the session’s working directory into it, while working_directory switches to another checkout or returns to the session’s original directory. The conversation, original project, and Access remain unchanged; file operations, new shells, and the footer follow the current directory.
Creation requires Git 2.45 or newer and Node in the same execution environment. It checks out committed files only, skips configured Git filters and does not initialize submodules or install dependencies. Source changes stay in the original checkout. Leaving, disabling the bundle or cancelling creation never automatically deletes branches or files; failures may retain partial artifacts. Return explicitly and review retained artifacts before managing them with Git.
Files and search
read, write, edit, and str_replace_editor cover file content, while grep and glob search the workspace. Mutating tools honor the session Access preset, and a mutation outside the sandbox returns the shared denial described in Permissions and access. Results over the context budget spill to a private file with a bounded preview in the transcript; the original stays readable with read or grep.
Web
web_search and web_fetch use DeepSeek’s Search and Fetch APIs by default: one direct API request per operation, no auxiliary model turn, and configurable auto, account, or api-key billing in Settings → Plugins → Web search and fetch (defaulting to auto with the credential reference stored by /login or the active DeepSeek account). Hosted fetch does not depend on local DNS, which matters behind a proxy’s fake-IP resolver; the target URL is retrieved by DeepSeek’s Fetch API rather than contacted from this machine. The Messages search and local Direct fetch (http) providers remain mounted: switch to Direct fetch when you need this machine to reach local or intranet addresses directly. A deployment that wants no search turns off search on the tool-web row in its own overlay.
Code intelligence
The read-only lsp tool answers definitions, references, implementations, and hover for configured language servers. No lsp tool appears until an lsp.json registers a server; see Language servers.
Task and session
todo_write tracks implementation items, ask_user_question asks you when inspection cannot decide a user-owned choice, present records finished deliverables so they stay findable after the turn (paired with @deepseek-ai/dsh-deliverable-file-references in Standard, PTC, and Creator for deliverable and #L… workspace file-link guidance), and skill loads a matching SKILL.md on demand.
omdsh uses timed questions with a default 120-second wait. Editing an answer or pressing Ctrl+T pauses the question countdown; Ctrl+S explicitly skips one question, and Esc leaves a finite timed question pending. When the wait ends, the model can continue independent work; a timeout grants no approval. A pending-question indicator and /questions let you answer later, including after resuming the session. Late answers appear in the transcript and reach the next model step. Ctrl+C interrupts instead of submitting an empty answer. In line-input mode, a blank answer line explicitly skips a question; EOF cancels the answer UI.
The model can request timeout: -1 when it must wait indefinitely, and a Profile can still select legacy mode. PTC has a separate execution budget: its default is 180 seconds, so Take time does not keep a run_code program alive indefinitely. If a PTC program ends before its question result is recorded, that question cannot be restored for a late answer.
Five read-only tools let the agent reach work it did in an earlier session: session_search and session_event_search find it, session_trace and session_event_trace follow lineage and relationships, and session_event_read returns exact event data. Without them a model that already solved something can only be told again, because /resume is yours to reach and not the agent’s.
They are workspace-scoped: cross-session reads are authorized only when the target session’s working directory exactly matches the caller’s, and a session with no working directory can inspect only itself. Turning the row off in a home overlay with disabled: true removes all five tools and their guidance.
Runtime inspection
The cordis preset adds read-only runtime inspection on top of the native catalog: cordis_inspect_list discovers providers, and cordis_inspect_query reads exact Service, Event, and Tool APIs. It also adds published Harness authoring Skills and the plugin_manager tool for profile-wide management. That tool requires full host access or approval for the call; dependency scripts require separate explicit approval. /plugins offers the human management surface.
Delegation
subagent, subagent_fork, and subagent_isolated delegate to child agents (with an optional name for sibling or slash-separated path addressing), send_message, wait_agent, and list_agents coordinate continuable children (interrupt_agent stays off by default, matching upstream dsh-base, while manual Ctrl+C remains available in child inspection), and workflow_run fans work out across children from a script. Enabling session Team mode with /team on adds team_task_create, team_task_list, team_task_get, and team_task_update for shared task-board coordination across the root agent and its continuable children. ralph iterates fresh agents toward one objective but ships disabled, matching the upstream base bundle; a home overlay restores it with disabled: false on its row. The code preset denies workflow_run, so a PTC session authors its orchestration through run_code alone. See Subagents and delegation.
Related
- Commands —
/toolsand the rest of the catalog - Permissions and access — what the sandbox gates
- Skills and MCP — extending the catalog with Skills and MCP tools