Plugins

QA de fidelidade de design para modelos somente texto: uma ferramenta `deepseek_vision` empresta um olho de qualquer rota de visão compatível com OpenAI, para que o modelo possa julgar se uma implementação corresponde ao seu mock — fornecida com o benchmark por trás desse julgamento (quatro fixtures, 23 defeitos injetados, transcrições brutas) e a disciplina de questionamento da qual ele depende.

Configura o contextWindow para modelos registados no dsh, para que sessões longas deixem de ser incorretamente avaliadas como excesso de contexto.

OMK — Observe. Measure. Know. Mudanças de conhecimento respaldadas por evidências para aplicações de IA.

Mapeia o áudio do sistema analisado localmente para efeitos de conversação reativos à música, julgamentos de jogo de ritmo, feedback de pontuação e combo, e barras de espectro RGB opcionais no DSH Web.

Harness de memória de agente em três níveis: CLAUDE.md L1 sempre ativo, índice L2, esqueleto detalhado L3 com modelos, um gate que bloqueia gravações em arquivos protegidos sem um token de aprovação de uso único, registro de eventos sem julgamento, destilação semanal e uma skill de loop de execução TDD.

Executa 3 ou 5 candidatos de coding agent em worktrees Git desanexados, valida seus patches com comandos do projeto, classifica apenas os candidatos aprovados e exige aprovação separada antes de aplicar o vencedor.

Plugin de modo de aprendizagem Feynman: ciclo ensinar → reexplicar → avaliar → reexplicar, renderizado como cartões de aula em HTML (diagramas mermaid + destaque de código shiki)

Resumes an agent turn that ended after narrating its next action without calling a tool: the guard replays the turn log on agent/turn-stopping, asks a judge model one true/false question, and steers the same turn to run one more step.

Um subagente de aprovação independente julga cada escalonamento de sandbox, com um modelo configurável e um log de auditoria.

Automatic approval and review for tool calls that need approval: keyword buckets match red lines without context, then a judge model classifies what was requested and sends it to allow, reject or a human according to the matching row and risk level; a judgment that never ran always goes to a human, and rejections tell the model why.

Peak / off-peak rate badges for the model selector (every row) and the composer tool row, judged per provider by its own time zone and schedule; the live catalog is served from the host over a fenced route.

Commitment write-gate: operator-authored rules enforced before a tool call runs — a deterministic guard tier plus an LLM-judge tier, fail-closed by default, every block written to a contradictions log.

Self-learning habit engine for the DeepSeek Harness — detects user-correction signals, judges habits with a low-cost model on threshold, settles candidates behind a two-level human gate

Codex-style approval autopilot for DeepSeek Harness: ordered glob rules (allow/ask/deny) plus an AI risk judge (low/medium/high) mapped through a risk tolerance, as an approval answerer.

Autonomous permission classifier for the auto preset: tool-scoped allow/deny rules, an LLM semantic judge, and git checkpointing for unattended sessions.

Destructive-command interception gate for dsh: parses shell semantics, judges risk against 41 built-in rules, and holds irreversible rm -rf, git reset --hard, and git push --force style commands at a confirmation gate.

Roteador de modelos em dois níveis para o DeepSeek Harness — roteamento por LLM-Judge, cadeias de fallback multimodelo, failover com backoff exponencial e orquestração em nível de tarefa (adaptação do pi-shift-router para o DSH)

Cross-project experience recall for DeepSeek Harness: when the running work matches something your memory store already learned elsewhere, that lesson is injected before the next step. Runs fully locally, judged by a local small model.

DeepSeek Harness plugin bundle: grill_tree, a fixed-policy design-space search tool. Expands one decision per node, fans out independent proposers, has a separate judge prune the branches, beam-limits each depth, and ends with complete designs or an infea

DeepSeek Harness plugin for CVE impact retesting: fingerprint a target, judge it against a CVE's affected range, gather passive evidence, and report a verdict with confidence and limitations.

Resume screening plugin for HR in DeepSeek Harness: ingest up to 100k Word/PDF/Excel resumes into one searchable library with skills / coarse categories / ingest time, auto-ingest & de-dupe new resumes, then screen by plain-language criteria — rule pre-filter in the database plus LLM fine-judge on the shortlisted few, CSV export.

Shared toolkit for authoring DeepSeek Harness plugins, published as @perrylink/dsh-plugin-kit: a pluggable provider registry seam, fail-closed approval and session-event gates, shared sanitize/pricing/judge modules and a new-plugin skeleton.

Multi-role parallel argumentation for DeepSeek Harness: runs the real Codex and Claude CLIs in parallel, has the DSH model (Judge) write a third-party verdict back into the conversation, plus Obsidian-style single-agent direct chat and per-role model config.

DeepSeek Harness 插件：生成前注入「人味」风格约束 + AI 味检测报告（纯代码、零 token、结果可复现）。dsh plugin: human-flavor style injection + AI-flavor detection report, judged by code.