ChatGPTClaudeGrok / X / CursorGemini

DeepSeek Harness Presets Fail: parent Undefined (DSH)

TroubleshootingPublished 2026-11-03Author: DeepSeek Plugin Market
DeepSeek HarnessAgent presetsCannot read properties of undefinedwithinFiberalpha.2preset load failure
DeepSeek Harness Agent presets fail to load with parent undefined: the 0.2.2-alpha.2 registry reads parent off the fiber chain unguarded. Checks and fixes.

In DeepSeek Harness (DSH), all four built-in Agent presets showing "failed to load" with the identical hover text Cannot read properties of undefined (reading 'parent') almost certainly means your preset configuration is fine and the preset registry on the 0.2.2-alpha.2 preview line is crashing on a shared code path. Three checks define the case: all four presets report exactly the same error; a fresh DSH_HOME with only @deepseek-ai/dsh-base and @deepseek-ai/dsh-web-app reproduces it; and switching the same machine to 0.2.1-rc.5 loads all four presets normally. When all three hold, stop asking "what did I break in my preset" and start asking "what did this release break in the registry" (#11994).

Why all four presets failing at once points at the registry, not one preset

Short answer: preset loading is diagnose-then-display, and the registry diagnoses every implementation in one pass; one broken ancestry link fails the whole batch, which is exactly what "all four are broken" looks like. Community members read the 0.2.2-alpha.2 bundle line by line and found this in @deepseek-ai/dsh-agent-preset-registry:

js
function withinFiber(fiber, root) {
  let current = fiber;
  while (true) {
    if (current === root) return true;
    const parent = current.parent.fiber;   // TypeError when parent is empty
    if (parent === current) return false;
    current = parent;
  }
}

Both leakedServices() and serviceForMount() call this for every implementation in the root service store. If any fiber's ancestry chain has a hop with no parent, the thrown TypeError takes down the entire diagnostic batch. That is why the four presets return one identical message instead of four different content errors.

This also explains a misleading observation: sessions and chat work perfectly while preset selection alone is unavailable. Preset diagnosis and the main chat path are separate initialization paths; the first can fail without touching the second, so "I can still chat" is not evidence that presets are healthy.

Three checks that separate a registry failure from a broken preset

The decisive signal: identical errors on all four built-in presets, reproduced in a minimal environment with no third-party plugins. Work in this order.

  1. Start a clean instance with a fresh DSH_HOME. Use a temporary directory so nothing from your old profile leaks in:
bash
DSH_HOME=$(mktemp -d) npx @deepseek-ai/dsh@0.2.2-alpha.2 web --no-open

Expected: after first-run initialization, open Settings → Agent presets and all four built-ins (Standard / PTC / Minimal / Creative) still show "failed to load".

  1. Hover each preset and compare the raw messages. Expected: four identical tooltips, all Cannot read properties of undefined (reading 'parent'). Four different messages (missing model, missing tool, and so on) would point at preset content instead, and this article would not apply.

  2. Confirm the profile has no third-party plugins. It should contain only @deepseek-ai/dsh-base and @deepseek-ai/dsh-web-app. If the failure still reproduces here, you can rule out "some plugin corrupted the service store"; if it only reproduces after installing a plugin, move your investigation to the plugin side.

  3. Run a stable-line control on the same machine and profile. This is the most valuable step in the whole diagnosis; see the next section.

Version comparison: why clearing caches and switching Node changes nothing

Short answer: the defect lives in the 0.2.2 alpha code itself, not in where you installed from, not in the cache, and not in the Node version; switching to 0.2.1-rc.5 on the same machine proves it. The reporter's isolation experiments are complete enough to reproduce as-is.

  1. Delete the old npx install copy. The previous copy was about 756MB; removing it prevents any reuse of a stale unpacked tree.

  2. Reinstall from a brand-new npm cache directory. Expected: installation succeeds, but after startup all four presets still show "failed to load" with the same tooltip. This rules out npm cache pollution.

  3. Run once on Node v24.22.2 (LTS) and once on v26.13.0. Expected: both reproduce. This rules out a Node incompatibility.

  4. Run 0.2.1-rc.5 with the same machine, the same profile, and the same Node. Expected: all four presets load with no broken markers, and settings writes persist into the profile patch in the same instance. This pins the scope to the alpha line.

bash
npx @deepseek-ai/dsh@0.2.1-rc.5 web --no-open

After those four steps only one conclusion survives: 0.2.2-alpha.2 introduced a regression. Whether it is only this registry change or another initialization change in the same release needs the maintainers to diff the alpha line; there is nothing left for users to tune in their environment.

A second independent failure in the same release: settings snap back

Short answer: preset failures and settings rollback are two mechanisms that happen to meet on one version and one machine; merging them into one bug report sends the debugging in the wrong direction. On 0.2.2-alpha.2, changing the permission dropdown in general settings or clicking "Continue" on the first-run preview notice both snap back immediately, cordis.patch.yml does not change, and the UI shows no error. An independent reproduction shows /api/settings/mutate returning HTTP 200 while the RPC layer returns ok: false with rejection code settings/rejected and message next is not a function.

That points at Cordis 4.0.6-alpha.2 changing internal/update to the two-argument form (resolved, restart) while dsh-app-boot@0.2.2-alpha.2 still registers a three-argument handler (_config, _noSave, next) and calls next(), so the third parameter never receives a continuation. Check the actually resolved vendor version and the app-boot calling convention rather than comparing DSH version numbers or clearing the npm cache.

The fix for the preset branch does not apply to the settings branch, and vice versa. The full evidence trail for settings writes is in DeepSeek Harness settings snap back immediately after saving.

What you can do now: roll back, bypass the dialog, leave the registry alone

Short answer: until the alpha line is fixed, the only complete workaround is rolling back to the stable line; writing welcomeNoticeVersion by hand only skips one dialog and does not restore presets. In priority order:

  1. Roll back to 0.2.1-rc.5. This is the option verified in the discussion: all four presets load and settings writes persist on the same machine and profile.
bash
npx @deepseek-ai/dsh@0.2.1-rc.5 web
  1. If you only want to skip the first-run "preview notes" dialog, you can add ui-settings-general.welcomeNoticeVersion to the profile's cordis.patch.yml by hand. Expected: the dialog stops appearing. Be explicit about what this is — it bypasses the welcome dialog, presets remain completely unavailable, and it is not a fix.

  2. Do not hand-patch the preset registry bundle. withinFiber is compiled into lib/index.js; editing it breaks integrity checks and gets overwritten on the next install, so it cannot be a real workaround.

  3. Do not count on 0.2.2-alpha.3. That release cannot even be installed: npm i @deepseek-ai/dsh@0.2.2-alpha.3 fails with ETARGET because it pins a leaf package that was never published. Details are in DeepSeek Harness alpha channel install fails with ETARGET.

  4. Wait for the official fix before upgrading. The acceptance test is simple: after upgrading, start a fresh DSH_HOME, confirm all four presets load, and confirm a settings change persists. Both must hold before you call this release usable.

Things to keep in mind while debugging

When an entire feature class fails at once, the most important rule is to find the shared path before hunting individual differences. Six points.

  1. Count the failures and compare the messages first. Four presets down with word-for-word identical errors means a shared path; one preset down, or four different messages, means individual content.
  2. A minimal environment beats reading your config. A fresh DSH_HOME with only built-in bundles is faster than reading your profile line by line, and it excludes interference better.
  3. A control experiment beats any guess. Running a different version on the same machine with every other variable unchanged is the cleanest causal evidence you can get.
  4. Do not use "chat still works" to prove presets are fine. Preset diagnosis and the main chat path are independent, so chat keeps working while presets are down.
  5. Do not merge two bugs from the same release. Presets and settings writes fail at different rejection points; combining them costs you both clues.
  6. Be accurate about fix status. As of this discussion, preset loading on 0.2.2-alpha.2 is not fixed, and alpha.3 cannot be installed because a package is missing. The one verified exit today is the stable line.

Source: Discussion #11994.

ChatGPTClaudeGrok / X / CursorGemini

FAQ

DeepSeek Harness presets all fail to load — my fault?

No — in DeepSeek Harness, four built-in presets failing with one identical message means the problem sits on a shared code path in the preset registry, not in the content of any single preset. After reading the 0.2.2-alpha.2 bundle, community members traced it to withinFiber walking the fiber ancestry chain and dereferencing parent without a null check; a single broken link in the service store makes diagnosis fail for every preset.

Why do DeepSeek Harness presets fail after cache clears?

Because this DeepSeek Harness failure is not about where you installed from or which Node you run; it is in the code of the 0.2.2 alpha line. The reporter reinstalled with a fresh npm cache, deleted a 756MB npx copy, and reproduced on both Node v24.22.2 and v26.13.0 with identical results, while switching the same machine and profile to 0.2.1-rc.5 loaded all four presets normally. The control experiment settles it.

Is DeepSeek Harness preset failure the same as snapback?

No, in DeepSeek Harness they are different mechanisms that happen to appear on the same version and machine. Preset failures come from the empty-parent traversal in dsh-agent-preset-registry's withinFiber. The settings rollback comes from a separate reproduction where /api/settings/mutate returns HTTP 200 but the RPC layer rejects with settings/rejected and message 'next is not a function'. Investigate them separately and never fix presets with a settings fix.

Can DeepSeek Harness Agent presets be made selectable again?

As of that discussion DeepSeek Harness has no fix on the 0.2.2 alpha line, and the practical answer is to go back to the 0.2.1-rc.5 stable line, where all four presets load. A local trick mentioned in the thread is writing ui-settings-general.welcomeNoticeVersion into the profile's cordis.patch.yml to skip the first-run dialog, but it only skips the dialog — presets stay unavailable, so it is not a fix.

Related Terms

Agent preset
An Agent preset is a prebuilt assembly of model, tools, and behavior policy in DeepSeek Harness. Users pick built-ins such as Standard, PTC, Minimal, or Creative from the settings page to create a session. A preset is data plus an assembly description, not a config file you write by hand.— https://github.com/deepseek-ai/deepseek-harness/discussions/11994
withinFiber
withinFiber is a helper inside @deepseek-ai/dsh-agent-preset-registry that decides whether a given fiber sits below a specified root fiber by walking up through current.parent.fiber. In the 0.2.2-alpha.2 bundle it does not guard against an empty parent, which is the root cause of the four simultaneous preset failures.— https://github.com/deepseek-ai/deepseek-harness/discussions/11994
fiber ancestry chain
The fiber ancestry chain is the service mount hierarchy in the DeepSeek Harness plugin runtime, linked by parent references, expressing which service mounted which and which root it belongs to. Any code that uses this chain for ownership checks must handle an empty parent explicitly, or it throws a TypeError.— https://github.com/deepseek-ai/deepseek-harness/discussions/11994

Sources