cs-source-reliable

English · 简体中文

Source reliability for CS teaching in DeepSeek Harness: a bundled registry of 119 verified, agent-collectable CS sources with authority/evidence tiers, plus two model-facing tools to find sources and audit citations.

A plugin for DeepSeek Harness (dsh). The registry covers the 17 CS2023 knowledge areas and ships as data; both tools are local lookups (no network calls), so they are fast and have no side effects.

What it ships

• 119 sources — 68 teaching-layer (textbooks, courses, official docs, standards) and 51 frontier-layer (preprints, proceedings, release notes, advisories).
• Per source: authority tier (A primary spec/official docs/releases, B peer-reviewed/authoritative, C curated secondary, D lead only), evidence type, content_access (static / portal / js / redirect / frameset / api / binary), license_class, cache_policy, attribution, refresh_days, last_checked, and the CS2023 areas it serves.
• Two tools: cs_source_find and cs_source_audit.
• A short system-prompt section that tells the model when to use them and how to cite.

Tools

cs_source_find

Find a source by URL, keyword, or CS2023 area code.

• url — match a page (host + path prefix).
• query — keywords, e.g. "algorithm textbook" or "RFC".
• ka — one of AL AR CN DM DS FPL GV HCI IS MSF OS PDC SEC SE SEP SF SPD.
• layer — breadth (teaching) or depth (frontier).
• limit — max results (default 5, ceiling 20).

Returns the compact record with tier, access mode and refresh policy. Example result:

- [python-docs] Python Documentation — A/official-doc; access=static; refresh=30d; https://docs.python.org/3/

Keyword ranking uses IDF-like token weighting with a subject-token boost, so "python 3.13 release notes" ranks Python sources above generic "release notes" pages.

cs_source_audit

Audit the URL citations in a draft answer: map each URL to a known source, report its tier and access mode, flag unknown domains and C/D-tier citations, count substantive paragraphs without a source, and suggest fixes.

Citations: 2/2 audited, 1 known, 1 unknown; A=1 B=0 C=0 D=0; uncited paragraphs=0
- https://docs.python.org/3/ -> python-docs (A/official-doc)
- https://example.com/blog/post -> unknown; not in the registry: verify the domain yourself before relying on it

Trigger rules

The plugin does not intercept requests. The model decides, guided by a system-prompt section (order 240, in the plugin-guidance band):

• call cs_source_find before stating CS facts that depend on versions, APIs, standards, or current state;
• prefer A/B sources, treat C as secondary and D as a lead, never as fact;
• cite as [source-id] URL plus the version or date the source states, and say "unsourced" when no registry source fits;
• call cs_source_audit on a draft that cites URLs before sending it;
• skip both tools for non-CS topics.

Measured footprint (2026-11-06)

Item	Measured
cs_source_find schema (name + description + parameters)	893 chars ≈ 224 tokens
cs_source_audit schema	454 chars ≈ 114 tokens
Guidance section	642 chars ≈ 161 tokens
Combined per-request overhead	≈ 498 tokens (schemas + guidance)
Registry load	0.7–2.6 ms
cs_source_find latency	avg 0.081 ms, p95 0.15 ms (5,000 calls)
cs_source_audit latency	avg 0.022 ms (3,064-char draft, 500 calls)

Token counts are character-based estimates at 4 chars/token; the two tool schemas and the guidance are prompt-cached along with the rest of the system prompt. Tool output is bounded (≤20 results, notes truncated to 180 chars).

Install

Requirements: DeepSeek Harness >=0.2.1-rc.1 <0.3.0-0, Node.js 20+.

From the release tarball (recommended — no build step):

dsh plugin --profile web add https://github.com/uhygyuf/cs-source-reliable/releases/latest/download/cs-source-reliable.tgz

From the source repository:

dsh plugin --profile web add github:uhygyuf/cs-source-reliable

From a local directory during development:

dsh plugin --profile web add /absolute/path/to/cs-source-reliable

Restart or reload the profile after installing. The plugin row appears under Settings → Plugins as CS source reliability.

Development

The registry source of truth is data/cs-source-matrix.yaml (119 sources with keywords, tiers and probe metadata); data/sources.json is generated from it.

npm install            # dev dependency: yaml
npm run build:data     # rebuild data/sources.json from the matrix
node test/check.mjs    # 27 checks: registry, tools, lossless JSON, latency budgets

tools/apply-keywords.cjs re-applies tools/source-keywords.json to the matrix; tools/make-icon.py regenerates the icon PNGs.

Data provenance and maintenance

• Every source was probed on 2026-11-06: 118/119 returned HTTP 200; Semantic Scholar returned 429 (it needs an API key).
• refresh_days is a suggestion per source class: advisories 2 d, feeds/releases 7 d, official docs 30 d, peer-reviewed 90 d, specs 180 d, textbooks 365 d. last_checked is the probe date.
• license_class / cache_policy are policy buckets, not legal advice: link to the source and quote briefly; full-text caching is only allowed for public-domain/open sources.

Limitations

• The tools look up a curated registry; they do not fetch pages or verify claims on their own. Source-tier data is only as fresh as the last registry rebuild.
• cs_source_audit is a mechanical check (URL extraction, tier mapping, uncited-paragraph count). It does not judge whether a cited source actually supports a sentence.
• A/B tiers reduce risk; they do not guarantee correctness. Version-sensitive claims still need the version/date the source states.
• Non-registry URLs are flagged as unknown even when legitimate (e.g. a project's own GitHub repository is not a CS knowledge source).

License

MIT. The registry contains only source names, URLs and policy metadata; the referenced content remains under each source's own terms.