天机阁 · Algorithm Codex

English · 中文 · Español · Português · हिन्दी

The plugin is the interface; the codex itself lives in the algo-hub repository one level up (data/algorithms.json + data/taxonomy.json). A standalone install carries a seed snapshot (seed/) and seeds ~/.dsh/tianjige on first run, so the same package works without the repository.

Features

• Search — tianji_search finds algorithms by natural language or keywords, with filters for domain, function key, maturity and “has a ready implementation”. Strict match first, partial match as a fallback, clickable repo/package links in the result.
• Dossier — tianji_get returns one full entry: how it works, tradeoffs, complexity, papers, implementations, sibling variants of the same function, caveats.
• Ask — tianji_ask lets the model answer a design question from the codex first, and opt in to a web search for algorithms outside it (findNew: true).
• Sweep — tianji_sweep searches the web for new algorithms, turns them into candidate drafts under data/pending/, and can run on a daily schedule (action=status|start|config).
• Curation — tianji_stage / tianji_update / tianji_review / tianji_review_mark / tianji_stats / tianji_usage cover adding, editing, reviewing, auditing and token accounting.
• UI — a left-sidebar row opens a full-page view with four tabs (Codex / Arena / Sweep / Ask), and the same page is also available as a right-sidebar tab.

Install

The package is the tianjige-plugin/ directory. Install it into a DSH profile and enable it:

# from the GitHub release (no npm account needed)
dsh plugin --profile desktop add https://github.com/MengXingXiaoMing/dsh-tianjige/releases/download/v1.3.0/dsh-tianjige-1.3.0.tgz

# from npm, once the package is published
dsh plugin --profile desktop add dsh-tianjige

# from a local checkout (development): add the directory itself
dsh plugin --profile desktop add C:\path\to\algo-hub\tianjige-plugin

Use your own profile name instead of desktop if you run a different one. Restart DSH afterwards: the host half is imported once at startup and a page refresh does not reload it. Before restarting you can smoke-check the sources with node --check host.js and node --check lib/*.mjs.

Configuration

The plugin inserts one row (id: tianjige) through cordis.patch.yml. Both keys are optional:

Key	Default	Meaning
dataDir	parent directory of this package (= repository root)	codex root containing data/ and server/ai.mjs
defaultLimit	8	default number of rows returned by tianji_search

- insert:
    - id: tianjige
      name: dsh-tianjige
      config:
        dataDir: D:\wherever\algo-hub   # optional, derived by default
        defaultLimit: 8

Data layout

algo-hub/
  data/
    algorithms.json   # merged codex (219 entries)
    taxonomy.json     # the 13 domains
    review.json       # review ledger
    coverage.json     # sweep coverage matrix (written once a sweep ran)
    staging/          # per-agent material waiting for merge
    pending/          # AI drafts waiting for human review
    runs/             # one ledger per sweep run
    logs/             # page error reports (client-YYYY-MM-DD.log)
  server/ai.mjs       # AI logic for the old web edition (the plugin uses lib/ai.mjs)
  tools/algohub.py    # CLI: validate / merge / export / stats / coverage / record-run / snapshot …

When the plugin is installed from npm there is no repository around it, so the first run copies the bundled starter library (seed/algorithms.json, seed/taxonomy.json) into ~/.dsh/tianjige and keeps every write there. Set dataDir (or DSH_TIANJIGE_DATA_DIR) to override; a link / development install keeps using the repository root.

Compatibility

• DSH >=0.2.1-rc.2 (dsh.engines.dsh), Node ^22.19.0 || >=24.0.0
• No @deepseek-ai/* dependency at all: the host entry has zero imports, and both the data layer and the AI layer are loaded by dynamic import, so a copy inside a profile can never shadow the runtime version.
• AI tools (ask / sweep / AI-assisted review) need DEEPSEEK_API_KEY: an environment variable, or a DEEPSEEK_API_KEY: line in ~/.dsh/.credentials.yaml. Without a key the other tools keep working and the AI tools return AI_NO_KEY.
• The only peer dependency is react (client half, provided by the harness).

Tools

Tool	Purpose
tianji_search	Search entries (query, optional domain / functionKey / maturity / hasImpl / limit)
tianji_get	Full dossier by id
tianji_ask	Natural-language design question (task, optional findNew web search)
tianji_sweep	Sweep: action=status / start / config (daily auto-sweep switch and interval)
tianji_stage	Stage a new candidate (writes data/staging/, waits for human merge)
tianji_update	Edit an existing entry in place (automatic backup)
tianji_review / tianji_review_mark	Review queue: list items / record a verdict
tianji_stats	Codex audit: entry count, domain spread, implementation coverage, review progress
tianji_usage	AI token ledger (read / delete / clear / retention)

UI surfaces

The plugin registers three seats:

• sidebar.panellist — one row “天机阁” in the left sidebar (order: 120)
• main — a full-page view opened from that row (tabs: Codex / Arena / Sweep / Ask)
• sidebar.right.pane.tab — the same page as a right-sidebar tab

Page data goes through host routes /api/tianjige/{library,meta,sweep,ask,review,usage,write,log}, which accept loopback requests only. Without a webServer service the plugin degrades to “tools only, no page”.

Troubleshooting

• Nothing shows up after install — restart DSH first; then check the application field returned by plugin_manager.
• Page says “sweep status unavailable” — usually a dependency module failed to load (for example a missing lib/ai.mjs); the route writes the error into the /api/tianjige/meta 500 response. Check with node --check lib/sweep.mjs.
• AI_NO_KEY from tianji_ask — configure the key as described under Compatibility.
• Data validation — python tools/algohub.py validate (expect 0 errors, 0 warnings) and doctor for reference checks.

Development

• Host: host.js (entry, tools + routes), lib/sweep.mjs (sweep / review / ask), lib/ai.mjs (AI calls and prompts), lib/usage.mjs (token ledger), lib/entries.mjs (entry read/write helpers).
• Client: lib/client.js (React, single-file bundle registered into a slot).
• Tests: npm test runs node --test test/smoke.test.mjs — offline, zero dependencies, against test/fixtures/; it never touches real data and does not need the plugin to be installed.
• Static checks: dsh-plugin-dev check --strict (from the dsh-plugin-guide skill package) and node --check on every entry.

Publishing

There are two distribution channels; both are optional and GitHub stays the source of truth.

Cut a release (maintainers). Bump version, run npm test and dsh-plugin-dev check --strict, then pack the tarball and attach it to a GitHub release:

npm pack
gh release create v<version> dsh-tianjige-<version>.tgz --title "v<version>" --notes "..."

npm (optional). After npm login, run npm publish --access public. Users then install with dsh plugin --profile desktop add dsh-tianjige. Publishing to npm is not required for the marketplace.

DSH marketplace. Registration lives in awesome-dsh-plugin/awesome-dsh-plugin: one file per plugin at data/plugins/<owner>__<repo>.yml, holding url, name, category, an optional tarball, and description.en (plus description.zh if you can). The requirements are a public repo, dsh.bundle.patch in package.json with cordis.patch.yml at the root, the dsh-plugin topic, a description that matches the code, and a repository that is at least one day old — that last one is checked automatically, to filter out repos created minutes before a PR. Open the PR from your fork; once merged, the registry entry is what the marketplace lists.

Curated lists (optional). DshMarketPlace/awesome-dsh-plugin keeps data/curated.yml. A PR there is a recommendation rather than a registration, so send it after the registry entry has been merged.

License

MIT — see LICENSE.