Plugins

Réécrit chaque requête selon un profil d'invite appris et met à jour automatiquement les habitudes, les règles anti-hallucination et les enregistrements d'apprentissage.

La couche d'autorité pour l'IA agentique, en plugin DeepSeek Harness. Régit chaque appel d'outil selon vos règles métier via le point de décision de politique TapPass /v1/govern.

Claude Code compatibility for DeepSeek Harness: skills, memory loading and writing, scoped rules, and settings switches

Measure whether a dsh setup change actually helped: register repeatable cases, run them, and diff the results before and after you change rules, skills, or models.

A DeepSeek Harness plugin that gives the model a tool to set the session title using rules you define. DeepSeek Harness 插件：给模型一个按你自己定义的规则设置会话标题的工具。

Shared infrastructure for data assetization plugins: business rules loader, file format adapter, report generator, path validator, audit logger Also known as @deepseek-ai/dsh-data-asset-shared.

Maintained fork of dsh-approval-gate. A neutral operation whose confirmation count has reached the threshold auto-approves when the judge is unavailable, instead of prompting a human again; approver-facing explanations are generated in Chinese from the real sandbox mode, command and paths. Also carries deterministic hard-deny for credential exfiltration and system-path destruction, judge-input redaction, a judge model candidate chain, fingerprint-scoped allow rules, and an approval view with unified diff and one-click revert.

Discovers workspace Markdown commands, skills and path-scoped rules from .dsh and .agents directories.

Long-term memory for DeepSeek Harness, read-only as of 0.1.0: a scribe_recall tool that reads a plain-markdown memory room, path rules that refuse directory traversal, Unicode smuggling and NTFS alternate data streams, and a bounded index whose truncation is always reported rather than silent.

Pre-execution safety valve for DSH that judges with TypeSafe Jev: it hooks tools/pre-execute and decides every bash/pwsh call by first applying offline static rules, then by asking the Jev model — the System One model from TypeSafe, which returns a structured decision instead of prose — one yes-no question ("will this command irreversibly delete or overwrite real data?"), turning the answer into allow, revise, block or escalate; revise hands the model a safer rewrite, escalate offers a one-shot human token; exhausted credit or a missing API key degrades loudly instead of failing silent, and a missing key is requested in the conversation itself (`guard key set`, read from stdin); every verdict is appended to a shared audit log and the text people read is bilingual zh-CN/en.

Conway's Game of Life for DeepSeek Harness: a sidebar panel row opens a drawable centre-column board with four rules, seven stampable patterns and a population chart.

Ships a skill about writing DSH skills - SKILL.md frontmatter and root precedence, how to phrase a description so it actually triggers, length and splitting rules, and an A/B check that the finished skill changes agent behaviour. The package also carries the checker the skill uses to validate a skill directory.

Preventive content policy for DeepSeek Harness tool results: rewrite matched text with regex rules, or drop/truncate unpredictable payloads by field path before they are normalised, rendered, or persisted. The upstream counterpart to dsh-plugin-redact.

Replacement session-title provider that compresses oversized or code-heavy first prompts before calling the model (in the tested DSH version the built-in provider rejects inputs above 4096 bytes without a model call) and waits for a real task instead of titling a greeting. Titles use the session’s own provider and model by default and manual titles are protected; title rules cover a character cap, an optional date affix, style, language and exclusion words, plus a per-session lock. Also re-titles stored sessions in bulk, with a progress overlay that keeps running after Settings is closed.

Ponytail (lazy-senior-dev) as a mode-switchable, always-on DSH plugin: injects the COMPACT ruleset every step (off empties it) and mirrors the full ruleset + five companions into ~/.dsh/skills. Attribution: github.com/DietrichGebert/ponytail (MIT).

Make the model believe it is inside the Z.ai Code (Zhipu ZCode) agent environment: replaces the harness identity opener and injects a ZCode rules skeleton with a dynamic project section. 让模型认为自己处于智谱 ZCode Agent 环境。

Persistent, searchable, per-project memory for the DeepSeek Harness: decisions, rules, and sprint context in a queryable DuckDB file, rules enforced on every model request, with a full management UI in the Web Client

Peer communication between conversations in one DeepSeek Harness process: list the conversations the workspace shows, link one under a nickname, message it without waiting for its turn to end, read its progress without waking it, and declare rules for your own tool calls.

Pre-boot health gate and repair for dsh plugin trees: seven static scan rules, fault attribution after failed boots, whitelisted reversible fixes, and a panel with plugin-row management plus a RAG knowledge base of past failures.

Records scoped workflow rules and correction reminders with bounded evidence, local settings, and optional Brain Hub integration.

dsh plugin: cron scheduling — bounded tasks with calendar & interval rules, delivered to live agents

Multi-agent frugality defense plugin: read-ledger dedup metrics, compaction-immune rules, completion gate, cheap-model review lane.

Memory triage for DeepSeek Harness: at the end of each turn it decides which entries to keep and which to archive, routes project-specific ones into a project bucket while global ones stay in the main store, merges near-duplicate entries, and falls back to an LLM when the rules are unsure — automatic actions are recoverable archives only, never deletions.

Time-traveling stream rules for DeepSeek Harness — project rules stay dormant until a regex matches the live token stream, then the request is aborted, the rule injected as a system reminder, and the request retried from the same point. Zero per-turn cont