Introduction

DeepSeek Harness (DSH) uses a plugin-based architecture intended to solve complex domain-specific workflows through a single plugin. In academic writing, turning fragmented materials into a thesis that can withstand plagiarism review and thesis defense often involves many tedious steps and cross-session information management. dsh-thesis is a full-process thesis plugin maintained by Paloma966. It does not directly generate a thesis; instead, through an “ask first, then act” workflow, it converts vague thesis requirements into a falsifiable Intent Specification and proceeds through nine stages: topic selection, proposal, literature, design, implementation, testing, writing, compliance, and defense.

The plugin’s core value lies in “traceability” and “control”. It requires the user to approve every decision and ensures the traceability of the thesis generation process through a nine-stage directory, a progress ledger, a decision log, search records, and check reports. It does not rely on cloud-based plagiarism systems; all checks are performed locally.

Core Capabilities

The plugin provides a set of tools under namespaces that cover the full thesis lifecycle:

  • Materials and Intent: thesis_ingest converts materials such as docx/xlsx/pptx/pdf/md/csv files and code into text and structural summaries; thesis_intake asks questions step by step and converts materials into an Intent Specification.
  • Workflow Management: thesis_init creates a workspace; thesis_progress generates the nine-stage directory and ledger; thesis_decide records decision logs and gates, and prevents advancement if G1-G4 gates are not passed.
  • Literature Search: thesis_lit_search performs searches; thesis_lit_save saves results; thesis_lit_note records notes. The plugin strictly limits citation sources and recognizes only real search caches.
  • Writing and Compliance: thesis_review reviews chapter by chapter; thesis_check performs five full-document checks; thesis_stylecheck performs a self-check for AI-like phrasing; thesis_build generates a Word document; thesis_originality performs a local originality self-check and similarity reduction.
  • Defense Preparation: thesis_defense conducts defense rehearsals; thesis_slides generates Marp-compatible slides; defense_code_* provides code rehearsal capabilities.
  • Cross-Session Memory: fact_search, fact_remember, and fact_context use a SQLite database to remember school requirements, advisor requirements, and writing preferences, and automatically restore relevant context when needed.

Installation and Activation

When installing from source, you need to clone the repository and build it first, because DSH does not support loading TypeScript source code directly:

git clone https://github.com/Paloma966/dsh-thesis.git && cd dsh-thesis && npm install && npm run build

After building, you can install it through the DSH plugin manager, or reference the built directory in the configuration file:

dsh plugin --profile web add "$PWD"

Note: After upgrading DSH, you may encounter a situation where the plugin “appears installed but has no response”. This happens because the host validates peerDependencies in package.json and silently disables the plugin if the requirement is not met. This plugin declares a dependency on >=0.2.0-rc.2 <0.3.0.

Usage Examples

After installing and restarting DSH, you can start using it.

  1. Initialize the conversation: Enter a command in the conversation to start the workflow.
    这是我的论文材料，帮我做毕业设计

The plugin will automatically run `thesis_ingest` to read materials, `thesis_intake` to ask questions and form an Intent Specification, and `thesis_init` to create the workspace.
  1. Workspace location: For convenience, it is recommended to set the DSH session directory (cwd) directly to the root directory of the thesis project. The plugin automatically searches upward for 00-管理/进度台账.md to locate the workspace. You can also specify it explicitly in the configuration:
    { "workspace": "/path/to/你的论文" }
  1. Stage-by-stage progression: Across the nine stages (topic selection → proposal → literature → design → implementation → testing → writing → compliance → defense), advance progress by invoking the corresponding tool or slash command, such as /thesis-decide.

Limitations and Boundaries

Several guardrails are built into the plugin to prevent misuse and the generation of false content:

  • No fake literature or data: Citations can only come from real search caches; experimental data must come from actual runs; thesis_originality explicitly forbids altering data or conclusions in order to reduce similarity.
  • Standards prevail: All standards should follow school documents. If a required template is missing, the plugin uses a built-in interim template and explicitly marks it as “not a school template”, never pretending to be compliant.
  • Data safety: If a read fails, the tool throws an error and leaves the original file untouched. It never overwrites data with default values and never silently destroys data.
  • Technical limitations:
    • PDF: Extraction is best-effort; scanned or encrypted PDFs may not yield extractable text.
    • Old formats: .doc/.xls/.ppt are not supported; save them in newer formats first.
    • Slide generation: There is no built-in generator; only Marp-compatible Markdown is produced.
    • Subprocesses: pandoc/marp/git subprocesses cannot be cancelled; they have timeouts but no signal propagation.
    • Code rehearsal: The built-in language matrix is Go/TS/Python/Rust; other languages must have gates completed in the configuration.

Configuration

All configuration items have default values and can be overridden in cordis.patch.yml. The following is a key configuration example:

- id: paper
  name: dsh-thesis
  config:
    workspace: /path/to/thesis
    memoryPath: /path/to/paper-memory.db
    similarity: { shingle: 4, threshold: 0.3 }
    ppt: { engine: auto }
    writingGate: true
    destructiveGate: true

Summary

dsh-thesis is a tool that provides structured support for the full thesis lifecycle. It uses strict gate control and local checks to help developers manage the closed loop from materials to the finished thesis. For thesis workflows that require handling large volumes of literature, code, and normative documents, it provides a reproducible workflow path.