ChatGPTClaudeGrok / X / CursorGemini

DeepSeek Harness Input Disabled: Workspace or Model Gate?

TroubleshootingPublished 2026-11-18Author: DeepSeek Plugin Market
DeepSeek HarnessDSH plugininput disabledworkspacemodel setup
DeepSeek Harness input may be blocked by a missing workspace or an invalid default model, or only look blocked when the request fails. Learn the difference.

A DeepSeek Harness input that will not send is usually one of two real blocks: no workspace selected, or a default model that points to a deleted provider. There is also a false block where typing works but the request fails. The three look similar; the recovery paths are different.

DeepSeek Harness: Separate real blocks from false blocks

The first test is whether you can type. If the box refuses focus or text, it is a real block. If typing works but sending fails, it is a false block, and the failure is in the request path. (source)

Real blocks are controlled by UI state: a grey, unfocusable box, or a Choose Model prompt that refuses input. False blocks leave the box normal and fail after Send, usually with an error, a timeout, or no return.

The evidence lives in different places. Real blocks live in the UI state; false blocks live in the request result. Searching logs for a real block, or staring at the UI for a false block, both waste time.

A ten-second check: put the cursor in the box and type one character. If it appears, investigate the request path; if not, continue with the two real blocks.

There is also a middle state: the box focuses and accepts text, but Send is grey or does nothing. That is usually a missing precondition, such as a blocked model selection or a session that has not finished opening. Look for a nearby Choose Model or Select Workspace prompt before classifying it.

Do not refresh first. Refreshing throws away the state that would tell you which gate is closed, and it can make “no workspace” look like a transient glitch.

DeepSeek Harness: Real block one: no workspace selected

A fresh Web UI selects no workspace, and the message box is unavailable until you select one. The official docs say a fresh Web UI has no workspace selected; you must add one, and the input is unavailable before that (source).

The startup directory is a common source of confusion. The dsh process uses the directory it started in as the default filesystem location, but that does not mean the UI has selected a workspace. Process default and UI selection are different states.

The check is one step: open Select Workspace and look for a selected directory. A directory that appears in the list but is not selected still leaves the input locked. See How to choose a DeepSeek Harness workspace.

A directory in the list proves it was added, not that it is active. Many first runs stop here because the project is visible; the selection marker is elsewhere. Look for the marker, not the list length.

If selecting a directory still leaves the box locked, check two smaller things: whether the directory is accessible, and whether the page is still attached to the original session. A deleted, moved, or permission-changed directory can invalidate the selection.

DeepSeek Harness: Real block two: invalid default model

The second real block comes from the model: the saved default points to a provider that no longer exists. The official docs say the input shows Choose Model and blocks typing until you select another model (source).

It looks different from the workspace block. The box is not entirely grey; it is gated by a model choice. The message is “the request has no valid destination,” not “the agent has no directory.”

Typical triggers: you deleted a custom provider without switching the default first; or you moved to a profile where the old default provider is not configured. Both leave a dangling default.

The check: open the model selector and see whether the selected model still belongs to an existing provider. If the selector asks you to choose first, or the displayed selection does not match the configured list, this is the block.

Pick a configured model and the input usually unlocks immediately. The full provider setup is in DeepSeek Harness model providers. You do not need to rebuild the provider just to unlock input; select a working model first.

DeepSeek Harness: False block: input works, request fails

If the box works and the message leaves, but nothing returns, the request path failed. The most common official errors are MISSING_CREDENTIAL and UNKNOWN_MODEL: no usable credential, or a model that is not in the configured list (source).

Other false blocks include a 401 from a bad key, an unreachable endpoint, and a custom gateway whose protocol or base URL does not match. None of them stop typing; all of them stop the request.

Check in this order:

  1. Read the error — if there is a code, follow the code rather than guessing from symptoms.
  2. Send the shortest task — remove task length and complexity from the equation.
  3. Switch models — if another configured model succeeds, the problem is the original model's key or endpoint.

A failed request does not always show a loud banner. Sometimes it just never completes. Judge by whether the request finished, not by whether a red message appeared.

Distinguish credential from network: if every model fails, suspect network, proxy, or endpoint; if only one model fails, suspect its key or model ID. See DeepSeek Harness model errors and, for page-level failures, Web UI blank or unreachable.

DeepSeek Harness: A table for the three states

Separate “can I type?” from “did it return?” and the three states fall into place. Use this table (source):

SymptomClassSmallest checkRecovery
Grey box, no typingReal block: workspaceIs a directory selected?Add and select one
Choose Model, typing refusedReal block: default modelDoes the default provider exist?Select a configured model
Typing works, no resultFalse block: request pathRead the error, send a short taskFix credential, endpoint, or model
Page blank or unreachableAccess problemCheck terminal address and processFollow startup troubleshooting

The table is exclusive: each row has a different recovery. Restarting the browser cannot fix an unselected workspace, a dangling default, or a missing credential. It only helps with cache-level problems.

If the input works, the model returns, and the task still sits still, you have left the input problem. Move to approval and running-signal triage.

DeepSeek Harness: Recovery order and verification

Recover in this order: workspace, model, credential. The order follows the dependency: the workspace unlocks input, the model sends the request, the credential makes it succeed. (source)

Three checks:

  1. Workspace — after selecting a directory, the box focuses and accepts text.
  2. Model — a usable model is selected, and the UI no longer asks you to choose.
  3. Credential — a short task returns instead of reporting a credential or model error.

When all three pass, the input problem is closed. If it still fails, narrow the scope: only one old session, or only one browser tab. Session-level failures point at session state; tab-level failures point at page connection or cache.

A new session is a useful control. If a new session works while the old one does not, stop retrying the old session and inspect its state instead.

Scope first, then repair: global config, session state, or page connection. That order keeps a local glitch from turning into a configuration rewrite.

One working control session is worth ten speculative edits.

ChatGPTClaudeGrok / X / CursorGemini

FAQ

Why is the DeepSeek Harness message box grey in a freshly opened Web UI?

A fresh DeepSeek Harness Web UI selects no workspace. The message box stays unavailable until you add and select a directory, so check Select Workspace before blaming the browser.

Why does the DeepSeek Harness input ask me to choose a model and refuse typing?

In DeepSeek Harness, the saved default model points to a provider that no longer exists. DeepSeek Harness shows Choose Model and blocks input until you select another configured model.

The DeepSeek Harness input accepts text but sends nothing. Which gate is closed?

In DeepSeek Harness, that is a false gate: input works and the failure is in the request path. Check for a missing credential, an unknown model, or an unreachable endpoint before looking at the input box.

What is the right recovery order for a blocked DeepSeek Harness input?

In DeepSeek Harness, select a workspace, confirm a usable model, then verify credentials. DeepSeek Harness unlocks input with the workspace, sends with the model, and succeeds with the credential.

Related Terms

blocked input
Blocked input is a DeepSeek Harness state where the message box refuses text or shows Choose Model, because a required UI precondition is missing.— DeepSeek Harness docs - Using the Web UI
false block
A false block is a DeepSeek Harness input that looks normal and sends, but the request fails later in the model or credential path.— DeepSeek Harness docs - Configuring models
default model
The default model is the DeepSeek Harness model preselected for new sessions; if its provider is deleted, input is blocked until another model is chosen.— DeepSeek Harness docs - Configuring models
workspace unlock
Workspace unlock is the DeepSeek Harness transition where selecting a workspace makes the session input available.— DeepSeek Harness docs - Using the Web UI

Sources