dsh-apply-model-all
Apply the model you are using right now to every session — one click, instead of opening
each session and picking it again.
DeepSeek Harness stores model choice per session. The official default-model setting
only covers sessions that have not chosen one yet, and /model only changes the session
you are currently in. So there is no built-in way to move a whole session history onto one
model at once. This plugin adds that control.
What it adds
A compact control next to the composer's model picker. The shipped picker is left
completely untouched:
[⇉ Apply to all sessions] [ deepseek-v4.1-flash ▾ ]
^ the shipped picker, unmodified
Click it, confirm the target model, and it writes that choice into every session that can
hold its own. A live count is shown while it runs.
How it works
It does not patch an official bundle, rewrite session logs, or edit projection caches. It
calls the same public host API the built-in /model dialog calls:
sessionController.list({}) → every session, with its current model
sessionController.selectModel({ sessionId, … }) → set one session's model
Because the write path is the official one, session state stays in the official format,
uninstalling changes nothing, and a future Harness that renames these methods costs you
this plugin only — never the rest of the app.
The batch runs on the host, not in the browser, so closing the tab does not interrupt
it. The client half only starts the run and polls its progress.
Read before writing (deduplication)
selectModel resolves and resumes each target session, which is by far the most expensive
part of the operation, and it has no "already correct, skip" short-circuit. So the plugin
reads every session's current model from list() first and skips the ones that already
match — a repeat click on an already-unified history finishes immediately instead of
redoing all the work.
Comparisons use provider and model only, never reasoning effort. That is deliberate:
reasoning effort belongs to the settings layer and is not part of a session's stored
model selection, so sessions report it as empty. Comparing it made the dedupe never match,
which silently turned every repeat click back into a full rewrite.
Scope and limits
| Case | Behaviour |
|---|
| Sessions that can hold their own model | Updated — this is what the control acts on |
| Subagent sessions | Not touched. They follow their parent session's routing, and the host rejects per-session selection for them (owned by subagent routing). Change the parent and they follow. |
| A session already on the target model | Skipped, no write |
| A session locked by another process | Reported as failed, with the reason |
Sessions held by a live write handle in a different process cannot be written from the
outside; the usual error is already owned by an active write handle. Clicking the control
inside the running app does not hit this, because the lock holder is the same host.
Worth knowing: the underlying call also saves the chosen model as the default in the
background, so new sessions will start on it too.
Install
dsh plugin --profile <your-profile> add dsh-apply-model-all
Restart the app afterwards — bundle patch layers and client bundles are both loaded at
startup, so neither takes effect while it is already running.
Configuration
None. There are two host routes and no settings surface:
POST /apply-model-all/start {provider, model, reasoningEffort?} → {runId}
GET /apply-model-all/status → {run: {…}}
Only one run may be active at a time; a second request returns 409 with the in-flight
progress. Failures are grouped by reason in the response rather than repeated per session,
and failure detail is capped at 50 entries.
Requirements
- DeepSeek Harness
0.1.0-rc.6 or newer.
- The web UI (the control lives in a composer slot).
License
MIT
