Complete your first task

By the end of this walkthrough you will have omdsh installed, signed in, and one finished agent task.

Install and launch

Install the command globally, or use npx @agi-fans/oh-my-dsh to try the current package without a global installation:

npm install --global @agi-fans/oh-my-dsh

Then change into the project you want the agent to understand and start omdsh there:

cd /path/to/your/project
omdsh

The startup directory becomes the session workspace. If it sits inside a Git repository, the status line picks that repository up.

Sign in securely

Run /login. For DeepSeek, omdsh opens the API Key page, asks for the key in a masked prompt, validates it, and saves it to the credential store.

A mounted provider that registered an authorization flow — for example a browser sign-in — appears in the /login list with that flow’s methods, and omdsh renders only the notices and prompts the flow asks for.

Never append a key to the command: /login <key> is deliberately rejected so the secret does not enter command history or the transcript.

An externally managed DEEPSEEK_API_KEY remains a supported fallback. A key saved through /login takes priority on later requests and across restarts; /logout removes only the omdsh-managed choice and falls back to the environment when available.

Configure the session before the first prompt

omdsh keeps three session controls separate instead of collapsing them into one mode. If you are unsure, keep the defaults for the first task.

ConceptCommandChoices
Agent/agentStandard is the full coding agent; PTC defaults to programmatic tool calling; Minimal keeps persistent Bash and str_replace_editor; Cordis adds live runtime inspection and plugin experimentation.
Workflow/workflowDefault works directly; Plan investigates and presents a reviewable plan before implementation.
Access/permissionRead only, Workspace write, or Full access.

Agent changes what the model can see, so select it before the first prompt; it is locked once model history exists. Each preset owns its tool exposure: PTC uses run_code with the generated TypeScript SDK, while Standard, Minimal, and Cordis expose their tools as native functions. A PTC program calls tools from inside its own code, so the calls it makes are recorded as sub-calls: each one has its own content, indented under run_code, and is included when Ctrl+O opens the completed turn. Workflow and Access are durable session state and can change later.

Choose safe Access

Run /permission before a task that may change files. The interactive selector offers three policies:

ModeUse it when
Read onlyYou want inspection without workspace writes; an escalation still requires approval.
Workspace writeThe task may edit the current workspace, while wider access still requires approval.
Full accessYou trust the workspace and intentionally want unrestricted filesystem access without approval prompts.

Full access requires a second confirmation. Access is the enforcement boundary; the Plan workflow is guidance and does not replace sandbox or approval policy. The sandbox and approval behavior behind each preset is documented in Permissions and access.

Send a concrete request

Start with an outcome, scope, and verification target. For example:

Find why the user settings are not persisted, fix the smallest responsible module, and run the focused tests. Do not change files under refs/.

While the agent works, Deep Driving marks the active turn. Thinking, intermediate replies, and tool previews stay visible until the turn ends, then fold to a duration summary such as Worked for 16s and the complete final answer. Failures remain visible. Press Ctrl+O to revisit the completed turn’s process, and press it again to return to the folded view.

Use the mouse wheel or PgUp to read earlier messages in either view. The prompt for the turn you are reading stays at the top once it scrolls out of view. Press End or click Jump to latest message to return to the live tail. The two-line footer keeps model, workspace, Git state, and token telemetry visible. See Keyboard and keys for preview limits, tool-result inspection, search, and mouse behavior.