DeepSeek Harness Model Switch: New Sessions vs Old Logs
A DeepSeek Harness model switch is not one global toggle. The settings page changes the default for new sessions; a session that has already sent a request keeps the model recorded in its own log (source). Old sessions staying on the old model is the rule working, not a cache failing to refresh.
DeepSeek Harness: Two scopes: default and session log
One selection lands in two different places, depending on whether the session has sent a request. A model chosen in the selector first updates the new-session default. A session that has not sent a request follows it; once a session sends a request, the choice is written into that session’s log. (source)
Think in two layers:
| Layer | Written when | Read by | Changes with the settings page |
|---|---|---|---|
| New-session default | Choosing a model | Sessions that have not sent yet | Yes |
| Session model record | First request in that session | Later requests in that session | No |
The purpose is replayability. A session log must reconstruct the context the model saw, including which model saw it. If a later settings change could rewrite history, the log would stop being a trustworthy record.
So “switch model” means “change the default for new sessions,” not “rewrite every session.” Old sessions staying put is the consistency rule.
Choosing a model in the selector already updates the default; you do not need to save it again in Settings. Doing both changes the same default and still does not touch old sessions.
DeepSeek Harness: Why old sessions do not follow
An old session does not follow because it already has a model record; later settings only affect new sessions. The official docs say sessions that have sent requests keep the model recorded in their log (source).
This creates a common misread. You switch from A to B, return to an old session, and the replies feel identical. The switch did work; the old session simply does not read the default.
If old sessions did follow, a single history would contain output from two models while the log recorded one choice. Replay and billing would both drift. Keeping the original model keeps each session’s model attribution single.
To test whether the switch worked, look at a new session, not an old one. An unchanged old session is evidence that the rule is working.
DeepSeek Harness: Use a new model: start a new session
The correct action for a new model is a new session, not repeated changes in an old one. A new session has not sent a request, so it reads the current default. (source)
Three steps:
- Select the target model — this updates the new-session default.
- Create a new session — do not reuse the old one.
- Send the first task — confirm the reply comes from the new model.
If you need old context, copy conclusions into the new session or let it read the relevant files. Do not expect a single old session to keep its history and switch models; those are separate behaviors.
Three common mistakes:
- Changing the selector inside an old session — updates the default, not the old record.
- Deleting the default provider and reselecting — leaves a dangling default and blocks input first.
- Assuming a new session inherits the previous session’s model — new sessions read the current default, not the last session’s record.
The correct order is “select, create, send.” One verification then proves the switch.
Multiple sessions in one workspace can keep different models. To compare two models, create two new sessions and select a different model in each; that is clearer than switching back and forth in one history.
DeepSeek Harness: A dangling default blocks input
If the default model points to a deleted provider, input is blocked; this is a dangling default, not a failed switch. The docs say the input shows Choose Model and blocks typing until another model is selected (source).
Distinguish the two cases:
- Old session does not switch — it has a model record and ignores the new default; input works.
- Default is dangling — a new session cannot find a valid default; input is blocked.
Recovery is to choose a configured model. The new-session default is restored and input unlocks. Old sessions keep their original records.
If you delete a provider to clean up, switch the default to a surviving provider first. The reverse order leaves a dangling default and looks like “messages stopped working right after I deleted something.”
DeepSeek Harness: Multiple sessions, multiple models
Model records follow sessions, not workspaces, so one workspace can host several models. This is useful: a fast model for retrieval and a strong model for refactoring can coexist in one project. (source)
Three boundaries:
- Switching workspace does not unify models — existing sessions keep their records.
- Only new sessions read the default — the default applies before the first request.
- Deleting a provider affects every session that records it — multiple sessions can lose their model at once.
The third is easy to miss. Before deleting a provider, list the sessions that record its models. If they still matter, migrate them to another provider first.
DeepSeek Harness: Verification: three steps for a new session
Verify the switch on a new session, never on an old one. Three steps (source):
- Default — the selector shows the target model as the current choice.
- New session — the first request in a new session uses that model.
- Difference — compare a task that exposes model differences, such as explanation style or answer length.
If the new session still behaves like the old model, confirm it is actually new. Some interfaces can reuse a session and only look new; check the session identity or history, not just the title.
Remember three lines: old sessions keep their model, new sessions read the default, migrate sessions before deleting a provider. Model switching then stops feeling like “the setting did not apply.”
Hidden saved keys and catalog refresh are covered in the neighboring articles.
FAQ
DeepSeek Harness treats model choice in two scopes: a selection becomes the default for new sessions, while a session that has sent a request keeps the model recorded in its log.
Start a new DeepSeek Harness session with the new default. Old sessions keep their recorded model by design; copy conclusions or let the new session read the relevant files.
In DeepSeek Harness, the saved default points to a provider that no longer exists. DeepSeek Harness shows Choose Model and blocks input until you select another configured model.
DeepSeek Harness supports that. DeepSeek Harness records the model in each session, so multiple sessions in one workspace can keep different models; only new sessions read the current default.
Related Terms
- default model
- The default model is the DeepSeek Harness model preselected for new sessions; choosing a model in the selector updates this default.— DeepSeek Harness docs - Configuring models
- session model record
- A session model record is the model choice written into a DeepSeek Harness session log; a session that has sent a request keeps using it.— DeepSeek Harness docs - Configuring models
- dangling default
- A dangling default is a DeepSeek Harness default model that points to a deleted provider, causing the input to ask for a new model.— DeepSeek Harness docs - Configuring models
- new session
- A new session is a DeepSeek Harness session that has not sent a request yet; it reads the current default model.— DeepSeek Harness docs - Configuring models
Sources
- DeepSeek Harness docs - Configuring models· deepseek-harness
- DeepSeek Harness docs - Using the Web UI· deepseek-harness