ChatGPTClaudeGrok / X / CursorGemini

DeepSeek Harness Settings Snap Back: settings/rejected (DSH)

TroubleshootingPublished 2026-11-03Author: DeepSeek Plugin Market
DeepSeek Harnesssettings/rejectednext is not a functionCordissettings not persistedRPC rejection code
DeepSeek Harness settings snap back while cordis.patch.yml stays unchanged. HTTP 200 hides settings/rejected and next is not a function; read the code.

DeepSeek Harness (DSH) has a failure mode that is very easy to misread as "the UI froze": you switch the permission dropdown or click Continue on the first-run dialog, the interface briefly shows the change applied, then it snaps back, not a single byte of cordis.patch.yml changes, and no error is shown. The cause is not front-end rendering but the RPC layer of the write path: /api/settings/mutate returns HTTP 200, yet the response body has ok: false, rejection code settings/rejected, and message next is not a function. A 200 only proves the server accepted the request; whether the change persists depends on the RPC ok and the rejection code (#11994).

Read the RPC rejection code, do not stop at HTTP 200

Short answer: the first-hand evidence of a failed settings write is ok:false plus a rejection code in the response body, not the HTTP status and not whether the UI turns red. A community member reproduced this independently on Windows with Node 24.13.1 and a fresh isolated DSH_HOME, with only the @deepseek-ai/dsh-base and @deepseek-ai/dsh-web-app bundles, an empty [] patch, and zero third-party plugins. The symptom was identical: clicking Continue on the preview notice left it on screen, and the acknowledgment state could not be saved.

Evidence steps:

  1. Open the network panel in devtools, filter on settings/mutate, then perform an operation that triggers a write, for example switching permissions from "work in workspace" to "full access".

  2. Check the status code of that request. Expected: 200. Do not stop there; 200 only means the HTTP layer succeeded.

  3. Expand the response body and find the RPC result fields. Expected: ok is false and a rejection code field is present. In the reproduction the value was settings/rejected.

  4. Read the message in that same response. Expected: next is not a function. That message is the real locating clue; if the message instead describes a timeout, your investigation belongs on the request path rather than on a business rejection.

  5. Confirm "nothing was persisted" against the file. Copy cordis.patch.yml before and after the operation and diff them. Expected: identical. That pins "the UI reverted" and "the write failed" to the same event.

Summary: HTTP 200 + ok:false + settings/rejected + next is not a function = the server rejected it deliberately, unrelated to network, permissions, or timeouts. A different rejection code or a different message means this article's conclusion does not apply.

From settings/rejected to next is not a function: the callback contract changed

Short answer: the rejection happens because Cordis changed its update callback signature while an old consumer still registers three parameters and calls next(), so the third parameter never receives a continuation. A read-only review of the published npm artifacts produced two key facts.

  1. The new caller passes two arguments. Cordis 4.0.6-alpha.2 calls internal/update with (resolved, restart).

  2. The old consumer still registers three. In dsh-app-boot@0.2.2-alpha.2, lib/index.js still registers (_config, _noSave, next) and calls next() (around line 4586 of the global prepend hook; the bundled config include implementation near line 218 keeps the three-parameter form as well).

  3. Mating the two forms leaves the third slot permanently empty. When next() runs, the function does not exist, so it throws next is not a function, which the settings write path wraps as settings/rejected.

  4. The separately released loader/include alpha.2 already uses two arguments. Upstream completed the migration, but these two old consumers in app-boot were not updated together with it.

Reading that sequence explains why "the same version number behaves differently": what determines behavior is the signature of that one call between each consumer and Cordis, not the DSH version string. Investigating requires checking both the vendor version that actually resolved and the app-boot calling convention; one without the other is not enough (#11994).

A/B the dependency set to pin the vendor version actually in effect

The decisive signal: keep DSH and app-boot fixed, change only Cordis and loader/include, and the write behavior flips. The comparison matrix in the discussion is already clean enough to reproduce directly.

DSH / dsh-app-bootCordisloader / includeWrite confirmed without plugins
0.2.2-alpha.24.0.6-alpha.11.0.7-alpha.1 / 1.0.11-alpha.1ok: true, completes normally
0.2.2-alpha.24.0.6-alpha.21.0.7-alpha.2 / 1.0.11-alpha.2settings/rejected: next is not a function

Steps:

  1. Check what the current profile actually installed. Do not infer it from the DSH version; open the profile directory and read the real Cordis and loader/include versions in the dependency tree.

  2. Record one complete failing baseline. Include the HTTP status, ok, rejection code, message, and a hash of cordis.patch.yml before and after the operation. That is the starting point for the comparison.

  3. Switch the vendor set back to 4.0.6-alpha.1 with 1.0.7-alpha.1 / 1.0.11-alpha.1. Expected: the same operation returns ok: true and cordis.patch.yml gains the expected keys.

  4. Switch back to the alpha.2 set and reproduce once more. Expected: next is not a function returns. One round trip in each direction establishes the causal link.

  5. Write the comparison into your debugging notes. Re-verify with the same method on future upgrades and do not skip it just because "the DSH version did not change".

Throughout the A/B test, no DSH source or artifact was modified, no acknowledgment state was written by hand, and no model request was sent. That is also a quality test for the conclusion: when one variable can explain a difference, do not guess with three.

Three kinds of "the change does not persist" need separate handling

Short answer: at least three different mechanisms can make settings "impossible to change", and mixing them up sends the investigation in a completely wrong direction. They present different signatures; triage by signature.

  1. settings/rejected + next is not a function: the vendor callback contract problem in this article. Signature: a clear RPC rejection code, HTTP 200, and every settings item fails.

  2. Starts fine but saving reports BAD_INDENT: the loader and the editor use two different YAML parsers that disagree about flow indentation. Signature: only saving fails for certain profiles, while loading works. See DeepSeek Harness settings saving fails with BAD_INDENT.

  3. The profile manifest itself cannot be parsed: bad data at the file layer. Signature: Unexpected token at startup, before the UI even opens. See DeepSeek Harness startup failure caused by a profile manifest BOM.

Use a fixed triage order: first check whether the RPC returned a rejection code, then check whether the file changed across the operation, and only then suspect YAML and file encoding.

Temporary workarounds and fix status

Short answer: this mismatch is something upstream has to align; on the user side you can return to a working vendor set or edit the file and restart. Details:

  1. Return to the working vendor set. That is the first row of the table above: Cordis 4.0.6-alpha.1 with loader/include alpha.1. The acceptance test is ok: true for the same operation and file content that actually changed.

  2. Edit the profile file directly and restart. Suitable for one-off changes to a few keys: back up cordis.patch.yml, edit it, restart, and the setting should take effect. The downside is that you lose interactive verification through the settings page.

  3. Do not keep clicking the save button. With the rejection code unchanged, retrying only repeats the same failing path and produces no new information.

  4. Do not delete and rebuild the whole profile. That throws away working configuration, and the problem is not in the configuration content.

  5. Record fix status accurately. As of this discussion the calling-convention mismatch is unfixed, and 0.2.2-alpha.3 cannot be installed because it pins an unpublished leaf package, so it cannot be used to verify a fix. Details are in DeepSeek Harness alpha channel install fails with ETARGET.

Things to keep in mind while debugging

The core principle for settings-write failures: first prove the request was rejected, then prove the file did not change, and only then discuss why it was rejected. Six points.

  1. The HTTP status is not a conclusion. A 200 only means the request arrived; the business outcome lives in the RPC ok and rejection code.
  2. The message is the second key. settings/rejected only names the category; next is not a function points at the specific mechanism.
  3. File comparison is hard evidence. Save a copy of cordis.patch.yml before and after the operation; identical hashes mean nothing was persisted.
  4. Do not substitute version numbers for dependency checks. The Cordis and loader/include versions that actually resolved are the ones that decide behavior.
  5. Change one variable at a time. Do not move the vendor set, the Node version, and the plugin list simultaneously, or the comparison loses its meaning.
  6. Do not mix in the preset failure. In the same release, four Agent presets failing to load at once is a separate mechanism; see DeepSeek Harness Agent presets all fail to load.

Source: Discussion #11994.

ChatGPTClaudeGrok / X / CursorGemini

FAQ

DeepSeek Harness settings revert — rejected or unsaved?

In DeepSeek Harness, read the RPC layer, not just the HTTP status. Settings writes go through /api/settings/mutate, and that request can return HTTP 200 while the response body carries ok: false with rejection code settings/rejected. That combination means the request reached the server and business logic rejected it on purpose, so it is not a network or timeout problem.

DeepSeek Harness settings/rejected + next is not a function?

It means DeepSeek Harness's hot-reload callback contract does not line up. Cordis 4.0.6-alpha.2 changed internal/update to the two-argument form (resolved, restart), while dsh-app-boot@0.2.2-alpha.2 still registers the three-argument form (_config, _noSave, next) and calls next(). The third parameter never receives a continuation, so calling it throws next is not a function.

Why do version checks miss DeepSeek Harness rejections?

Because DeepSeek Harness behavior depends on which vendor dependencies actually resolve, not on the DSH version string. With the same DSH and app-boot, Cordis 4.0.6-alpha.1 plus loader/include alpha.1 writes normally; swap in 4.0.6-alpha.2 plus loader/include alpha.2 and you get next is not a function. You have to check the Cordis and loader/include versions that really landed in the profile.

Is there a workaround for DeepSeek Harness rejections?

The approach in the DeepSeek Harness discussion is to downgrade the vendor set to a combination that works, namely Cordis 4.0.6-alpha.1 with loader/include alpha.1, after which writes complete on the same DSH version. Another option is editing the profile file directly and restarting, but back the file up first. As of that discussion the calling-convention mismatch is unfixed and belongs to upstream alignment work.

Related Terms

settings/rejected
settings/rejected is the RPC rejection code returned by the DeepSeek Harness settings write endpoint. It means the server accepted the request and entered business logic but rejected it because a precondition was not met. It differs from a network failure or a timeout; you need the message in the same response body to locate the specific cause.— https://github.com/deepseek-ai/deepseek-harness/discussions/11994
Cordis internal/update
Cordis internal/update is the internal entry point in the plugin runtime that resolves a new configuration and applies it to already-loaded plugins. Its callback signature determines how callers obtain the continuation. 4.0.6-alpha.2 changing it to the two-argument (resolved, restart) form is the direct cause of this settings write failure.— https://github.com/deepseek-ai/deepseek-harness/discussions/11994
cordis.patch.yml
cordis.patch.yml is the file inside a DeepSeek Harness profile that records configuration patches; changes made in the settings page are supposed to end up here. When a setting looks applied and then reverts, comparing this file before and after the operation is the most direct way to see whether anything was actually persisted.— https://github.com/deepseek-ai/deepseek-harness/discussions/11994

Sources