dsh-anchored-standard: a two-phase DeepSeek Harness preset that is already frozen

The problem: a session's first request decides its reasoning style

The README frames the whole project around one observation: the first reasoning chain of a session has a style, and that style is selected by the conditions of request number one. It calls this the trajectory, and it splits it into two visible shapes. The Minimal condition produces first lines that begin "We need...". The Standard condition produces "Let me..." first lines, which the README calls standard-like. The levers that select between them, isolated in issue #11 and listed in the terminology section, are the tool schema, the output budget, and the injected reminders.

That is the problem the presets exist to solve. If you want the lighter trajectory but still need the heavier tools later in the session, you have to change the conditions mid-session without destroying the trajectory you started with. The base mode does this by beginning with only the Minimal tool pair visible and no auto-injected context, then promoting once the session is durable. The audience is narrow: people running DeepSeek Harness who care about which reasoning style a session settles into, and who are willing to read a terminology list before installing anything.

Bootstrap, promotion signal, resident catalog: the two-phase mechanism

The mode table in the README is the clearest description of the mechanism. Anchored Standard starts request number one with two tools, described as the Minimal pair, and a Minimal tool schema as the anchor. Promotion happens on the first durable tool/call or assistant/message, whichever arrives first, which the table records as promoteOn: either. Cost is listed as none, meaning no extra model call is spent on the anchor.

The variants change one variable each. Zero-Anchored Standard starts with zero tools and uses one fixed anchor turn, costing one extra model call. Whoami Standard also starts with zero tools but uses a self-introduction turn, the README writing the prompt as "你是谁", and also costs one extra model call. Prefab Anchored Standard seeds rolled history from a bundled successful trajectory and is already promoted inside the seed, so instantiating it costs no model call at all. Eternal Minimal never promotes: the visible catalog never grows, and heavier tools are reached through something the README calls the dshx bash gateway.

After promotion, the session runs on what the README calls the resident catalog: the bootstrap pair, plus the discovery tools, plus every tool the model explicitly unlocked. The discovery tools are named as dev_tool_search, skill_search and skill_load. Two details matter for anyone reasoning about failure. Phase state is derived from durable events rather than held in memory, so resume and reload preserve it. And the prefab hydrates the blank session in place when its preset is selected, with no per-workspace import required.

Installing a single mode directory and running the sync check

The repository is a private npm package, so there is no registry install. The README states that every mode directory is self-contained and installs alone under whatever id you copy it to. The package.json sets type to module and requires Node 22.19.0 or newer through the engines field, so check that first.

The repository ships three scripts. sync regenerates the materialized copies of the shared plugin inside each mode directory, test runs the Node test runner, and check runs the sync check followed by the tests.

npm run sync

Running sync writes the committed copies of shared/ into the mode directories. If you have edited anything under shared/, this is what propagates it. The check script runs the same generator in verification mode before the test suite, which is the quickest way to see whether a mode directory is internally consistent:

npm run check

The README does not document a rollback path for a bad sync, and it does not give a CLI entry point for selecting a mode. Mode selection happens in the harness, by copying a directory under an id of your choosing. If you want the base behaviour, copy preset/. If you want the variant with no first-request output cap and promotion on the first durable event, the README points to the separately developed both-anchored contribution rather than to a directory in this repository.

Where the design breaks down: frozen status, unmeasured successor, retired model

The largest limitation is not technical. The README states the project is frozen as of 2026-09-10, that new issues and pull requests are no longer accepted or handled, and that the repository stays online with its history intact. The stated reason is that on that date DeepSeek released V4.1 Flash and announced V4 Pro, the model these presets were measured on and tuned for, is retired on 2026-09-14 12:00, with its traffic routed to V4.1 Flash. A frozen repository with an unmaintained harness dependency is a real constraint, not a footnote.

The second limitation is stated by the project itself and is worth repeating exactly. The README says V4.1 Flash's own first-round trajectory has NOT been measured, because there was no API budget in maintenance mode, and that this should be treated as an inference from earlier flash data rather than a result. The earlier data it points to is a first-round dose-response probe in which the flash family anchored under every condition, 32 out of 32. So the claim that the successor does not need patching rests on an inference, and the project says so.

There is also an earlier status note kept for the record, dated 2026-08-17, which says active development had already stopped after price increases on the DeepSeek official API and the opencode go subscription. The evaluation loops the presets depend on, described as Project2-class runs and multi-trial roll and probe experiments, were no longer affordable. Anyone expecting harness-compatibility updates should read that note before assuming any will arrive. The licence field in package.json says MIT, but the repository metadata reports the licence as NOASSERTION, and the README does not resolve the discrepancy.

Alternatives the README itself points to, and how they differ

The README lists three community projects plus a fourth entry from the research repository, and the differences are structural rather than cosmetic. both-anchored, from the DeepseekCotexplorations contributions directory, is the closest comparison: it is also a two-phase anchored preset, but it has no first-request output cap and no promotion gate. Instead it promotes on the first durable tool/call or assistant/message and keeps the full PTC SDK intact through post-promotion wire trimming. If your objection to the base mode is the gate or the cap, that is the variant to read, and the README notes the feedback and experiment data live in the research repository with sample sizes and boundary caveats stated in the original contribution.

dsh-routing-suite takes a different approach entirely: it is a runtime injector with task-aware thinking-mode routing presets, the router-standard family, rather than a two-phase bootstrap. J-Space Cognition Suite is model-agnostic and packaged as a Skill, so it does not depend on the harness preset mechanism at all. dsh-recovery solves a neighbouring problem rather than competing: it is a zero-dependency self-healing CLI plus watchdog plugin offering broken-preset quarantine and rollback, safe-mode whitelists, and boot probes. If you are worried about what happens when a preset breaks mid-session, that is the project addressing it, and this repository does not.

Maintenance cost and what the licence field does not settle

Maintenance cost here is effectively zero because the project has stopped. The last push was on 2026-09-10, and the README states plainly that new issues and pull requests are no longer accepted or handled. There is no upgrade path to plan for, and no release was retrieved. The practical cost is the opposite one: if DeepSeek changes the harness interface, nothing in this repository will be updated to match, and the note from 2026-08-17 already described the mode as maintenance only with bug fixes and harness-compatibility updates only when feasible.

On licensing, package.json declares MIT, while the repository metadata reports NOASSERTION. The README and the file listing do not explain the gap, and the LICENSE and NOTICE files are present but their contents are not reproduced in the repository's README. Treat the discrepancy as something to check in the files themselves before you reuse any code, and note that this article is not legal advice. The README also states that this is a community project, not an official DeepSeek preset, and is not affiliated with or endorsed by DeepSeek.

Editorial conclusion

Adopt this only if you are studying how first-request conditions shape a model trajectory, or if you already run DeepSeek Harness and want the context-gate injection control and the prefab pipeline as reference code. Do not adopt it as a dependency for a production agent: the project is frozen, new issues and pull requests are not accepted, and it was tuned for V4 Pro, which the README says is retired on 2026-09-14 12:00. Before copying anything, verify which mode directory you actually need, confirm your Node version satisfies the engines field, and check whether the model you are running is one the project ever measured.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. xiaobright/dsh-anchored-standard on GitHub