Model or dataset
xiaobright/dsh-anchored-standard avatar
xiaobright/dsh-anchored-standard

dsh-anchored-standard: frozen four days after its last commit, and its key claim is an inference

Two-phase DeepSeek Harness preset: Minimal-aligned bootstrap, then full Standard tools (Project2 98/99)

3,715 stars121 forksJavaScriptNOASSERTION

At a glance

What is it?
dsh-anchored-standard is a community collection of DeepSeek Harness presets that start a session on a minimal tool set and promote to a resident catalog once the session is durable. It froze itself on 10 September 2026 because the model it was tuned for was being retired, and the file is unusually candid: it states that its own central claim has not been measured and should be read as an inference.
Who is it for?
This repository is worth reading as a documented negative result rather than as something to install. Its value is that it shows a hypothesis about model behaviour, the experiments claimed behind it, and an explicit note that the part which mattered most was never tested.
Can I use it commercially?
Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
Is it still maintained?
Yes. The repository last received commits 25 days ago.
What is it written in?
Mainly JavaScript, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • xiaobright/dsh-anchored-standard3,715 ★ · -89
Star history
Repository2026-09-212026-11-15
xiaobright/dsh-anchored-standard3.8k3.7k

-89 Stars in the last 55 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The project froze itself the same day it last committed

The status section is dated, and the date is the same as the last commit. On 2026-09-10 the file announces that the project is frozen and that new issues and pull requests are no longer accepted or handled. The reason is specific: DeepSeek released a new model that day and announced that the model these presets were measured on and tuned for would be retired on 2026-09-14 at 12:00, with its traffic rerouted. So the retirement deadline was four days after the last commit, and the repository was closed to contributions before it happened rather than after. What remains is the point: the repository stays online with its history, its data and its acknowledgement record intact, and the research material was moved to a separate repository.

The central claim is labelled an inference by the project itself

The most valuable sentence in the file is an admission. The presets exist to re-anchor a trajectory that the successor model does not appear to need patched, on the grounds that the earlier flash family anchored under every condition in the project's first round of dose-response probes, 32 out of 32, which is why the project never tuned for it. Then the sentence that matters: the successor model's own first-round trajectory has not been measured, because there was no API budget in maintenance mode, and the reader is told to treat it as an inference from the earlier data rather than a result. The earlier status note, kept for the record, says work on the project had already stopped a month earlier, after price increases on both the vendor's official API and a subscription service made the evaluation loops unaffordable.

The introduction names four modes and the table lists seven

The opening paragraph describes the collection as a base mode, two live-anchor variants and one seeded prefab mode, which is four. The table below it has seven rows, and the tree has seven mode directories, so the summary sentence is out of date relative to the rest of the file. The three modes it does not mention are an eternal minimal mode whose visible catalog never grows, a wire mode that routes each think step through a sibling provider with tool choice suppressed on the wire, and a combo mode that combines a think and execute split with a depth gate and a deliberation drip as three independent rows. The table is the more informative artefact, because its cost column prices each mode in model calls.

Phase state is derived from the session event log

The terminology section is the mechanism. A session's phase is not stored as a flag; it is derived from durable events, meaning events recorded in the session log, which is why resuming or reloading preserves it. The bootstrap phase is the first request of a session, carrying the bootstrap tool pair, no auto-injected context, and an optional cap on output. Promotion is the durable event that ends that phase, and in the base mode it fires on the first tool call or the first assistant message, whichever arrives first, while the variants wait for their anchor reply. What gets promoted is a resident catalog: the bootstrap pair, the discovery tools, and every tool the model explicitly unlocked afterwards. The three discovery tools are named individually.

The same plugin is committed once per mode directory

Every mode directory is self-contained and installs alone under whatever identifier you copy it to, which is what makes the collection forkable at the directory level rather than the package level. The cost of that design is visible in the tree and in the manifest. There is a shared plugin directory at the root, and each mode directory holds a materialized copy of it, a committed duplicate generated by a sync script. Two commands keep that honest: one regenerates the copies, and the check command runs the generator in verification mode before running the tests, so a stale copy fails the check rather than the suite. The prefab mode hydrates a blank session in place when selected, with no per-workspace import step.

A private package with a patch-level Node floor and no dependencies

The package manifest describes something that is never published. It is marked private, carries version 0.1.0, declares MIT, is an ES module, and requires Node 22.19.0 or newer, which is a floor set to a patch release rather than a major one. Most telling is what is absent: there is no dependency block and no development dependency block at all, and the test script invokes Node's own test runner directly. So the collection ships with nothing installed around it, which is consistent with a project whose whole purpose is the contents of a few configuration directories. Note that the repository metadata names no licence while the manifest declares MIT and the tree carries both a licence file and a notice file.

A farewell file with a closing note from the successor model

The project's last document is a personal note in Chinese, and the file explains what it was asked to say: that the next model should encounter more varied harness environments during training, so that patch engineering of this kind stops being necessary. According to the status section, that file now ends with a closing note written by the successor model itself. The parts the project says remain valid are the model-agnostic ones: a unified injection control, the prefab pipeline, the dose-response data for the discovery tools, and the general observation that the API-visible surface conditions the trajectory. Alongside the English file there is a Chinese translation, an acknowledgements record, and a list of four community projects that grew out of it, by two different authors.

Editorial conclusion

This repository is worth reading as a documented negative result rather than as something to install. Its value is that it shows a hypothesis about model behaviour, the experiments claimed behind it, and an explicit note that the part which mattered most was never tested. Three things to understand before reading anything else in it. The project is frozen and accepts nothing, so there is no upstream to wait for. Its measurements were taken against a model that no longer serves traffic, and the successor was never probed. And its strongest surviving claim is a model-agnostic observation about how the visible API surface conditions behaviour, which the file itself says is an inference from older data rather than a result.

Frequently asked questions

What is dsh-anchored-standard?

A frozen community collection of DeepSeek Harness agent presets: a base mode, two live-anchor variants, a seeded prefab mode, and three more, each a directory that installs alone. A session starts on a minimal tool set with no auto-injected context and promotes to a small resident catalog once the session log records a durable event. The file states it is not an official DeepSeek preset.

Why is dsh-anchored-standard frozen?

Because the model the presets were measured and tuned for was retired on 2026-09-14 at 12:00, with traffic rerouted to a newer model released on 2026-09-10. The freeze was declared on the release date, and new issues and pull requests are no longer accepted or handled. The repository stays online with its history and data.

What modes does dsh-anchored-standard ship?

Seven mode directories, though the introduction names only four. The table lists a base anchored mode with two bootstrap tools and no extra cost, a zero-anchored variant, a self-introduction variant, a seeded prefab that needs no model call to instantiate, an eternal minimal mode with no phases at all, a wire think-execute mode costing one extra call per turn, and a combo mode costing one extra call per turn as well.

Official sources

  1. Issues
  2. Project website
  3. README
  4. xiaobright/dsh-anchored-standard on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/xiaobright-dsh-anchored-standard.svg)](https://hysenlabs.com/projects/xiaobright-dsh-anchored-standard)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/xiaobright-dsh-anchored-standard.svg)](https://hysenlabs.com/projects/xiaobright-dsh-anchored-standard)

Was this useful?