dsh-pdf-edit

🚨 Temporary Maintenance Pause

This project will be placed on a maintenance pause for an extended period. No updates will be made during this time. The plugin may continue to function for a while, but please use it at your own risk. A resumption date has not been set. We appreciate your understanding. English | 中文

🧩 dsh-std ecosystem ready (Community v0.15): ships a dsh-plugin.json static manifest and a standard FacetModule entry (dist/std/host.js), loadable by dsh-std hosts such as @dsh-std/adapter-dsh — no @deepseek-ai/* packages required; the legacy cordis entry is kept as-is. Both entries share the exact same tool implementations, so editing behavior is identical.

DeepSeek Harness plugin — AI edits PDF text, automatically preserving the original layout.

What Is This

An AI-powered PDF editing plugin. Tell it what to change in natural language, and it will:

• Edit text only, keep layout intact — Fonts, sizes, colors, and positions are all locked. The result looks identical to the original
• Handle overflow automatically — When replacement text is longer, font size shrinks or text truncates. Layout never breaks
• Support CJK — Auto-detects and embeds system CJK fonts (SimHei / MS Gothic / Noto Sans CJK)

Use Cases

Install

Node >= 22. Compatible with dsh 0.1.1-rc.2 (cordis-direct) and dsh-std Community v0.15 hosts.

# Via Harness plugin CLI (same as official plugins)
dsh plugin --profile web add dsh-pdf-edit@latest

# Or via npm
npm install dsh-pdf-edit

# Inside a dsh-std host (adapter scans profile dependencies for dsh-plugin.json)
dsh plugin --profile web add @dsh-std/adapter-dsh
dsh plugin --profile web add dsh-pdf-edit

dsh-std Ecosystem Adaptation (v0.4.0+)

This plugin is also a dsh-std Community v0.15 standard plugin: hosts read the static dsh-plugin.json at the package root to decide compatibility without executing plugin code, and all runtime traffic goes through an adapter layer — future upstream breaking changes are absorbed by the adapter, so this plugin stays maintenance-free.

Under std hosts (no cordis config injection), configuration comes from environment variables:

Note: the dsh-std tool/model protocols (v1alpha1) do not yet expose host LLM inference to plugins, so AI editing under std hosts uses direct DeepSeek API calls; tool registration, lifecycle, and cleanup semantics are fully standard.

Seeing "Cannot read properties of undefined (reading 'prepare')"?

This is a known dsh-host issue from the rc stage: when @deepseek-ai/dsh-tools is loaded more than once in the same process, the tool-scheduler Symbol misses. Since v0.1.7 this plugin pins it via peerDependencies to prevent that at the source — but a leftover copy can still trigger if you ever ran pnpm install manually inside the profile directory.

Troubleshoot in order:

# 1. Check for a materialized local copy (real dir, not a symlink)
ls -l ~/.dsh/profiles/web/node_modules/@deepseek-ai/

# 2. If present, remove the core-package copies
cd ~/.dsh/profiles/web
pnpm remove @deepseek-ai/dsh-tools @deepseek-ai/cordis

# 3. Restart dsh and verify in a NEW session (old crashed sessions are unrecoverable)

The plugin also probes for this at load time: if detected, it throws an error containing these exact recovery commands instead of crashing silently.

Changelog

v0.4.0

• dsh-std Community v0.15 adaptation: static dsh-plugin.json manifest at the package root + standard FacetModule entry (src/std/host.ts → dist/std/host.js). Under std hosts (e.g. @dsh-std/adapter-dsh), the 5 tools are published as tools.dsh/v1alpha1 Tool extensions with local ToolHandlers; lifecycle/cleanup follow the standard activation scope
• No hard dependency on official packages: @deepseek-ai/dsh-tools peerDependency is now optional — zero @deepseek-ai/* dependencies under std hosts; future upstream breaking changes are absorbed by the adapter layer. The cordis-direct entry is preserved unchanged and reports clear guidance when the package is missing
• std-host configuration switches to environment variables (DSH_PDF_EDIT_* / DEEPSEEK_API_KEY)
• Tool parameter schemas upgraded to standard JSON Schema; tool implementations (pdfEditPreview, etc.) are exported for both entries to share

v0.2.1

• Dynamically read DSH default model: use ctx.agentDefaultModel.currentSelection() to get the user's current provider/model, replacing hardcoded agnes
• Priority chain: user config (config.provider/config.model) > DSH default model > agnes fallback
• Whether the user is on deepseek, kimi, glm, minimax, openpangu, mino, claude, grok, gpt, etc., the plugin follows automatically — zero config needed

v0.1.8

• Reuse DSH built-in LLM service (ctx.llm), no manual API Key configuration needed
• inject adds "llm" dependency, plugin calls DSH LLM via ctx.llm.stream()
• DeepSeek API direct connection kept as fallback (when ctx.llm is unavailable)

v0.1.7

• Dependency restructure: @deepseek-ai/dsh-tools moved from dependencies into peerDependencies, pinned exactly to 0.1.1-rc.2, preventing pnpm from materializing a second copy inside profiles (dual copies break the tool-scheduler Symbol and crash every tool)
• Load-time guard: apply() probes for the scheduler before touching ctx.tools; on failure it throws a clear error with recovery commands instead of crashing silently
• README install section now documents the "Cannot read properties of undefined (reading 'prepare')" troubleshooting flow; engines field declares Node >= 22

v0.1.6

• Adapt to the dsh v0.1.1-rc.2 plugin contract: export name / inject / apply(ctx, config), register all four tools via ctx.tools.register(defineTool(...)), configuration passed through the cordis patch row's config: field
• Add path whitelist guard: pdfPath/outputPath validated against allowedRoots with symlink resolution and extension/size checks, preventing injection-driven arbitrary file read/write
• Prompt injection defense: PDF text wrapped in a data container, hardened system prompt, second-stage injection scan on AI output with fallback to original text
• Browser rendering hardening: JavaScript disabled, outbound requests intercepted, CSP and CSS sanitizer, background dataUrl and font name whitelists
• Engineering robustness: API key env-first, request timeout, chunked concurrency limit, 429-aware backoff, AI output length caps
• New test suite (120 cases) and editing benchmark (10 cases, npm run bench)

v0.1.5

• Rename package from @whatsmore-nf/dsh-plugin-pdf-edit to dsh-pdf-edit, display plugin name directly in the marketplace

v0.1.4

• Fix embedCustom passing fontkit object to doc.embedFont, now passes Uint8Array directly
• Fix loadBytes not supporting string paths (e.g. fonts.cjk: '/path/to/font.ttf')
• Fix CFF-format TTC font compatibility, auto-detect and skip unsupported CFF fonts
• Add Android system font paths (MiSansRoundedSC, NotoSansSC, etc.)
• Simplify cordis.patch.yml to community plugin standard format

v0.1.3

• Fix missing output: { schema, render } field in ctx.tools.register() causing registration failure
• Fix execute signature mismatch (should be (args, exec) two-parameter)

v0.1.2

• Add cordis plugin format name/inject/apply exports, fix "invalid plugin" error

v0.1.1

• Fix plugin name mismatch in cordis.patch.yml that caused load failure

v0.1.0

• Initial release
• Style-locked editing: AI modifies text while preserving original layout
• Native render mode: pdf-lib direct draw, zero browser dependency
• CJK font auto-detection and embedding
• Overflow handling: shrink / clip / wrap / reject
• Glossary global substitution
• Three relayout templates: academic / mobile / briefing

How It Works

Original PDF ──▶ Extract text + styles ──▶ AI generates edits ──▶ Overlay draw back to PDF

1. Extract: pdfjs reads text content from each page, recording position, font, size, and color for every word
2. AI edit: The extracted text is sent to DeepSeek. AI returns only the text fragments that need changing
3. Overlay draw: pdf-lib covers old text with a white rectangle at the original position, then draws new text at the same spot with the same font

No browser needed. No Chromium download. Pure JavaScript.

License

MIT