jadezane6634-ops/dsh-purge-flash
AI EngineeringA DSH plugin that reshapes outbound LLM requests and retries refusals, aimed at people probing model refusal behavior.
This is a narrow, self-aware plugin for DSH users who want more control over how a model begins its answer and what happens after a refusal. Its creator is explicit that it is a probability tool rather than a switch, and its own limitations section is the most useful part of the documentation.
The problem
A user asks a question the model declines, gets a refusal, and then has to manually rephrase, retry, or abandon the thread. The failure is not just the answer; it is the repetition of composing new prompts and losing the thread's momentum. dsh-purge-flash changes the outbound request instead: it injects framing into the system context, seeds a possible answer opening, and, when a refusal is detected, removes traces of that refusal and resends with one more pressure step. The user keeps working in the same conversation while the plugin handles the retry loop.
What's different
The central mechanism is early-token steering. The README argues that a refusal is effectively decided by the first characters of a reply, so the plugin supplies an opening (a code fence, 1., or 结论:) rather than letting the model choose one. It layers a three-tier skill router: strong keyword hits load a leaf toolchain of roughly 350 tokens, weak hits load a parent skill at up to 134 tokens, and no hit loads nothing. That conservative fallback is deliberate: the author says a wrong leaf is worse than no leaf. A refusal then triggers a retry ladder that scrubs the refusal history and increases pressure, with up to six pairs on retry.
The cost side is concrete: 1,597 resident tokens (1,448 kernel plus 149 tail echo), plus 67–91 tokens for many-shot pairs and 10 tokens for semantic continuation when a category hits, which the README puts under 1.3% of a 128k window. The plugin claims zero host-file patches and uses DSH's public event surface, so DSH upgrades should not require repatching. The trade-off is coupling: stripOfficial regular expressions are written for DSH 0.2.1-rc.2 identity text and must be updated when the host changes wording. The retry judgment also only inspects the first 24 characters, so a model that writes a long preamble before refusing slips past it.
Who should try it
It fits someone already running DSH who wants to experiment with request-body shaping and refusal retries, and who is comfortable reading status output and configuration. The Windows ZIP-and-install.bat path is the lowest-friction route; macOS and Linux users must edit package.json and create a symlink manually, so a non-Windows user needs to be comfortable with that setup before trying it.
It is a poor fit for anyone who needs a guaranteed outcome or who expects the plugin to make prohibited categories negotiable. The README states three object-level categories remain non-negotiable, and refusals are probabilistic: the same sentence may pass today and fail tomorrow. It is also a poor fit for token-constrained contexts where even 1.6k resident tokens are material.
Requirements and limitations
DSH 0.2.1-rc.2 or newer is required for the request-body shaping path; older versions still get other layers but no shaping, and the absence of sessionController.fork means rewind is unavailable. Configuration is read once at assembly, so changes require a full process restart—closing the window is not enough. reasoningScrub is off by default and explicitly unverified on real hardware: DeepSeek with tools requires the full reasoning_content returned, and whether rewriting triggers a 400 is untested.
Sampling control is limited. In thinking mode, temperature, presence_penalty, and frequency_penalty are reported ineffective, leaving only top_p in [0.95, 1.0]. patches: gateway (Tier B host patch) is not implemented, and boN resampling and sampling knobs beyond response_format are reserved but unwired, so changing them has no effect. The router is keyword matching, so purely semantic requests with no category terms fall to the fallback group and load nothing.
HotRepo's take
The project was created on 2026-10-25 and had 150 total stars when it appeared at #9 on HotRepo's NEW-REPOSITORY board for JavaScript on 2026-10-26—that is a board rank, not a daily growth figure. HotRepo has only one recorded observation date, so persistence cannot be judged. The included test suite and the 479 KB dependency-free repository are documented as self-checkable project instructions, and the README's limitations section is unusually specific. Before adopting it, verify that your DSH version exposes the llm service and sessionController.fork, because without them the request-shaping and rewind layers are absent.
Written by AI from the project's README, updated October 26, 2026. The project's own docs are the source of truth.