dsh-anchored-standard: Two-Phase DeepSeek Harness Presets for Trajectory Control
Two-phase DeepSeek Harness preset: Minimal-aligned bootstrap, then full Standard tools (Project2 98/99)
At a glance
- What is it?
- dsh-anchored-standard is a community collection of DeepSeek Harness agent preset configurations that anchor a session's first-request reasoning style to the Minimal condition, then promote to a fuller tool catalog once the session is durable. The project was frozen on 2026-09-10 when the target model, DeepSeek V4 Pro, was retired.
- Who is it for?
- dsh-anchored-standard is worth examining for researchers who want to replicate trajectory-anchoring experiments conducted on DeepSeek V4 Pro before its 2026-09-14 retirement. The measurement base is explicit in the README: 32 first-round probes on the flash family, with formal V4 Pro evaluation loops no longer running.
- 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 21 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 October 1, 2026, and from our analysis. They are not legal advice.
Editorial analysis
Anchoring a DeepSeek Session Trajectory on the First Request
dsh-anchored-standard is a set of JavaScript agent preset configurations for the DeepSeek Harness that address a specific behavior: the reasoning style of a DeepSeek V4 Pro session differs depending on what tools and context are present on the very first request. With the Minimal tool pair (two tools, no auto-injected context), the model produces reasoning chains that open with "We need..." phrasing. With the full Standard tool catalog, it opens with "Let me..." phrasing. The README refers to these as the Minimal and Standard trajectories.
The project exists to reliably produce the Minimal trajectory on request 1 even when the eventual session goal requires the full Standard catalog. The target audience is engineers and researchers who work through the DeepSeek Harness API and want predictable first-request reasoning behavior. The README is clear that this is a community project: it is not an official DeepSeek preset and is not affiliated with or endorsed by DeepSeek.
Bootstrap Phase, Promotion, and the Resident Catalog
Each session using the base Anchored Standard preset passes through two phases. The bootstrap phase covers request 1: the model sees exactly two tools (the Minimal pair), receives no auto-injected context, and may have an optional output cap applied. These three conditions, which the README traces to issue #11 findings (the tool schema, the output budget, and injected reminders), reliably select the Minimal trajectory.
Promotion ends the bootstrap phase when the session records its first durable event. In the base mode, promotion fires on either the first `tool/call` or the first `assistant/message`, whichever arrives first. This is configured as `promoteOn: either` in the preset. After promotion, the resident catalog expands: the two bootstrap tools remain, the discovery tools (`dev_tool_search`, `skill_search`, `skill_load`) are added, and any tool the model explicitly unlocked during the session joins the catalog. Phase state is derived from the durable event log, so a session that is resumed or reloaded after interruption preserves its promotion status without re-running the bootstrap.
Repository Layout and Running the Validation Suite
The repository contains seven mode directories, each self-contained and installable independently. They are: preset/ (Anchored Standard), zero-anchored-standard/, whoami-standard/, prefab/, eternal-minimal/, wire-think-standard/, and combo-anchored/. A shared/ directory holds plugin files that are synchronized into each mode directory by a script. The test/, verify/, and scripts/ directories support the test and sync infrastructure.
The package.json file declares MIT as the license and requires Node.js 22.19.0 or later. Two commands cover the main maintenance tasks. To generate the materialized copy of each shared plugin inside every mode directory:
npm run syncTo run both the sync consistency check and the full test suite together:
npm run checkThe README states that each mode directory installs alone under whatever identifier you copy it to, and that the prefab mode hydrates the blank session in place when its preset is selected, requiring no per-workspace import step.
Seven Modes and Their Cost Profiles
The README presents a table of seven modes covering different anchoring strategies and cost trade-offs.
Anchored Standard (preset/) starts the session with the Minimal tool pair and promotes on the first durable event. It costs no extra model calls. Zero-Anchored Standard (zero-anchored-standard/) starts with zero tools, inserts one fixed anchor turn, and promotes on the anchor reply, at a cost of one extra model call. Whoami Standard (whoami-standard/) starts with zero tools and anchors through a self-introduction turn using the Chinese phrase "你是谁", promoting on the reply, also at one extra model call cost.
Prefab Anchored Standard (prefab/) arrives with a seeded rolled history representing a successful trajectory; the session is already promoted at instantiation and costs no model call to set up. Eternal Minimal (eternal-minimal/) uses the Minimal tool pair but never promotes; heavier tools run through a `dshx` bash gateway instead of joining the resident catalog. Wire Think-Execute Standard (wire-think-standard/) has tools present but sets `tool_choice: none` on the wire, routing each think step through a sibling provider, at one extra model call per turn plus prefix-cache churn. Combo Anchored (combo-anchored/) starts with zero tools on every user turn and layers three independent control mechanisms (think/execute split, depth gate, and deliberation drip), also at one extra model call per turn.
Frozen as of 2026-09-10: The V4 Pro Retirement
The README documents the project's end state in detail. On 2026-09-10, DeepSeek released V4.1 Flash and announced that V4 Pro, the model these presets were built around and measured on, would retire on 2026-09-14 with its traffic routed to V4.1 Flash. The presets exist to patch a trajectory issue that, according to the README, the flash family does not appear to need patched: in first-round dose-response probes, the flash family anchored under every condition across 32 sessions without these presets.
The README is explicit that this observation is an inference from earlier flash data, not a formal result: no API budget exists for measurement in maintenance mode. The last push was on 2026-09-10. No new issues or pull requests are accepted. The repository stays online with its history, data, and acknowledgement record intact. The model-agnostic portions of the work (the context-gate unified injection control, the prefab pipeline, and the discovery-tool dose-response data) remain accessible, and further research continues in the DeepseekCotexplorations repository at github.com/0liveiraaa/DeepseekCotexplorations.
What dsh-anchored-standard Cannot Do
The preset was built for and measured on DeepSeek V4 Pro specifically. V4.1 Flash trajectory behavior under these presets has not been formally measured. The evaluation loops the project depended on (Project2-class runs and multi-trial roll/probe experiments) were no longer affordable after price increases on both the DeepSeek official API and the opencode go subscription, as the README records.
The README does not document rollback behavior if a preset produces unexpected session state. The license field in the repository metadata reads NOASSERTION, though the package.json file declares MIT; users in environments with strict license requirements should confirm the applicable license by reading both files. There are no GitHub releases, so there is no versioned artifact to pin. Engineers who need a stable tagged release for reproducible deployments would need to pin to a specific commit hash instead.
dsh-anchored-standard and the dsh-routing-suite
The README lists dsh-routing-suite as a community project that users have reported performs better in some scenarios. It is described as a runtime injector plus task-aware thinking-mode routing presets (the router-standard family). The functional difference in scope: dsh-routing-suite controls which thinking mode applies to each task at runtime, while dsh-anchored-standard conditions the session trajectory at the very first request. These two concerns are independent: the bootstrap phase is about what the model's reasoning looks like on request 1, while thinking-mode routing applies per subsequent task.
The README also lists dsh-recovery as ecosystem tooling built around these presets: a zero-dependency self-healing CLI plus watchdog plugin that provides broken-preset quarantine and rollback, safe-mode whitelists, and boot probes. Teams that want trajectory anchoring alongside runtime mode routing would need both dsh-anchored-standard and dsh-routing-suite. Teams who need operational resilience tooling would pair dsh-anchored-standard with dsh-recovery.
Editorial conclusion
dsh-anchored-standard is worth examining for researchers who want to replicate trajectory-anchoring experiments conducted on DeepSeek V4 Pro before its 2026-09-14 retirement. The measurement base is explicit in the README: 32 first-round probes on the flash family, with formal V4 Pro evaluation loops no longer running. New deployments on V4.1 Flash should treat the anchoring effect as unconfirmed. To verify harness compatibility before activating any mode, run `npm run check` against the target harness version.
Frequently asked questions
What is the difference between Anchored Standard mode and Zero-Anchored Standard mode in dsh-anchored-standard?
Anchored Standard (preset/) starts the session with two tools, the Minimal pair, and promotes on the first durable tool/call or assistant/message with no extra model call cost. Zero-Anchored Standard (zero-anchored-standard/) starts with zero tools, inserts one fixed anchor turn as the anchoring mechanism, and promotes on the anchor reply, which costs one additional model call.
Does dsh-anchored-standard work with DeepSeek V4.1 Flash?
The presets were built and measured on DeepSeek V4 Pro. The README notes that the flash family anchored under every condition in 32 first-round dose-response probes without these presets, but states explicitly that V4.1 Flash has not been formally measured with the presets applied, and that no API budget is available for further evaluation.
What does trajectory anchoring mean in dsh-anchored-standard?
The README defines trajectory as the style of the model's first reasoning chain. The Minimal condition produces first lines starting with "We need..."; the Standard condition produces first lines starting with "Let me..." Anchoring means configuring request 1 with the Minimal tool pair and no auto-injected context so the model reliably enters the Minimal trajectory, then promoting to the full resident catalog once the session records a durable event.
Official sources
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.
[](https://hysenlabs.com/projects/xiaobright-dsh-anchored-standard)