Skip to content
PerryLinkPublic

About

HCB data/gate subproject of dsh-standards: compatibility matrix, ENTRY/FLOW/BOOST tracks, risk, hcb-gate/1.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

HCB is the data/gate subproject of dsh-standards. Rules, checker and ENTRY/FLOW/BOOST definitions live in dsh-standards. This repo holds the compatibility matrix, tracks data, the v0.5 evidence layer, risk and hcb-gate/1. Independent community project. Not affiliated with DeepSeek.

Official: GitHub (canonical) · npm · Gitee mirror · GitCode mirror

npx dsh-hcb check dsh-auto-review

HCB — Host Compatibility Baseline for DeepSeek Harness plugins

Declared compatibility + declared adoption tracks, with an explicit declared / checked / tested evidence layer. Not a score, not a certification, and not a runtime guarantee. Independent community project. Not affiliated with DeepSeek.

DSH ships fast and breaks compatibility often. Plugin authors and enterprise buyers have no single place to answer:

  • Which DSH versions does this plugin declare support?
  • Does it declare DSH-ENTRY / DSH-FLOW / DSH-BOOST status — and with what evidence?
  • Has anybody independently checked the published tarball, or is the row still only a claim?
  • What capability surface does it request?
  • Is it ready to pilot, or does it need review first?

HCB answers these from package metadata (engines.dsh, peerDependencies, dsh.capabilities, dsh.standards), DSH GitHub releases, and — for rows that have been audited — a stored checker report.

Three tracks, not three levels

HCB follows the dsh-standards model. ENTRY / FLOW / BOOST are independent tracks; they are never summed into a score.

Track Question Gates Evidence states
DSH-ENTRY can it be recognized, installed, booted? G0–G5 runtime / recognized / static / blocked / failed / not-run / stale
DSH-FLOW can a user install, configure, use, recover, uninstall? J1–J8 runtime / manual / not-run / failed / stale
DSH-BOOST does it use Cordis natively and efficiently? LEV-1–8 measured / not-measured / failed / stale

Rules: a skip is not a pass; static is not runtime; BOOST is not a gate; compatibility is a claim by the package author, not a test result. See docs/TRACKS.md.

What is in v0.14.0 (schema v0.5)

File Meaning
data/hosts.json DSH releases (tag → version → date)
data/matrix.json plugin × host compatibility plus standards tracks, evidence, risk, and gate
data/tracks.json counts per ENTRY/FLOW/BOOST state and gate decision
data/capabilities.json capability / risk aggregation
spec/tracks.json HCB's consumer mapping of dsh-levels/1
schema/hcb.schema.json JSON Schema (v0.5) for the data format
data/roster.json canonical dsh-kit family roster (112) + external
data/coverage.json coverage vs roster
scripts/coverage.mjs reproducibility check
scripts/collect.mjs reproducible collector; preserves checked evidence for the same version
scripts/audit-entry.mjs v0.5: pack + static ENTRY check -> evidence patch (optional --merge)
scripts/migrate-v05.mjs one-shot v0.4 -> v0.5 matrix migration
scripts/validate.mjs validates the matrix, evidence, summary and gate
docs/EVIDENCE.md evidence contract and re-run guide
.github/actions/check drop-in CI check
index.html static viewer (GitHub Pages)

Evidence levels:

Level Produced by Proves
declared npm metadata / package.json#dsh.standards the author's claim only
checked scripts/audit-entry.mjs -> dsh-standards@1.7.0 --scope entry --entry-claim static (default 1.7.0; historical rows are labeled by row.evidence.checker — v0.6–v0.8.2 rounds used 1.6.10, except dsh-permission-rules in v0.8.2) the published tarball was statically examined; no boot
tested a real DSH host boot (ENTRY G0–G3; future harness) runtime behaviour was observed

v0.5 ships the declared and checked machinery. tested is defined but is not produced by audit-entry.mjs; the script is intentionally static-only. See docs/EVIDENCE.md for the field contract, failure / blocked semantics, and the exact re-run commands.

Current coverage (2026-11-15, v0.14.0)

  • 143 plugins = the full 112-package dsh-kit family (34 mainline-default + 53 opt-in + 10 catalog-only + 15 domain) + 31 external
  • 123 declare engines.dsh; 20 are unknown
  • 21 declare dsh.standards; 122 are not-declared
  • ENTRY: 14 runtime + 127 recognized + 2 failed (129 checked + 14 tested), 0 not-run, 0 unknown — runtime rows: guide, budget, defend, mask, memento, observe, test-drive, checkpoint-rewind, claude-move, data-quality, doublecheck, session-sync, auto-review, permission-rules (all G0–G3, host 0.2.2-rc.3)
  • FLOW: 3 manual + 18 not-run + 122 unknown — manual is not runtime; no plugin has a runtime FLOW claim
  • Evidence: 0 declared, 129 checked, 14 tested. The 14 tested rows passed a real G0–G3 boot on host 0.2.2-rc.3; the rest are static checked.
  • Checker: 143/143 rows on dsh-standards@1.7.0 (0 rows still on 1.6.10). v0.10.2 re-audited the 95 rows that still carried 1.6.10; v0.10.3 re-audited the four bumped rows; v0.11.1 re-audited the 12 release refreshes. ENTRY is unchanged (1 runtime + 140 recognized + 2 failed) and every row is labeled by row.evidence.checker.
  • Risk: 12 high, 66 medium, 65 unknown (up from 3 / 63 / 77 — the 12 refreshed rows now declare capabilities)
  • Gate: 2 blocked (dsh-checkers and dsh-session-fork, both failed ENTRY), 30 review, 110 pilot, 1 production-candidate (dsh-plugin-guide: ENTRY runtime + FLOW manual + risk medium). Nine of the refreshed high-capability rows moved pilot -> review.

The near-zero dsh.standards adoption is itself a finding. The gate is a compatibility/risk heuristic, not a quality grade.

Coverage vs canonical roster (2026-11-15)

HCB is measured against the canonical dsh-kit family roster (112 packages), not an ad-hoc list:

Group Total Covered by HCB Checked Missing Declared tracks Compat declared
mainline-default 34 34 34 0 21 34
mainline-opt-in 53 53 53 0 0 53
catalog-only 10 10 10 0 0 10
domain 15 15 15 0 0 15
external 31 31 31 0 0 11

112 / 112 family packages (100%) are covered and checked, and so are all 31 external plugins. Coverage is complete, but evidence depth is not: only 21 plugins declare tracks, and 1 has ENTRY runtime. v0.6.0 added the static checked pipeline for the whole mainline-default set (34/34); v0.7.0 started mainline-opt-in (19/53); v0.8.0 finishes all 53 opt-in rows and adds the first 2 catalog-only rows (96 checked overall); v0.8.2 refreshes three changed package versions, keeps checked at 96, and re-audits dsh-permission-rules with the 16-ID checker 1.7.0 (restoring recognized). v0.9.0 audits the last 8 catalog-only and all 15 domain rows with dsh-standards@1.7.0 (23 new; 96 -> 119 checked) and closes the family at 112/112 checked. v0.10.0 audits the last 24 external declared rows with dsh-standards@1.7.0 (119 -> 143 checked): 23 recognized, 1 failed (dsh-session-fork@0.4.0, published tarball missing its declared lib/index.js). v0.10.2 re-audits the 95 rows still labeled 1.6.10 with dsh-standards@1.7.0 (0 ENTRY/gate moves; all 143 rows now record 1.7.0), and 1.7.0 becomes the default checker. v0.10.3 absorbs four new releases — dsh-lsp-actions@0.5.21, dsh-doublecheck@0.9.32, dsh-memento@0.5.31, dsh-session-sync@0.3.2 — re-audits all four with 1.7.0 (all recognized, 13/14 checks, 0 blocking) and records their new dsh.standards declarations (declared tracks 5 -> 9; gate split unchanged). v0.11.1 absorbs 12 new releases — dsh-research-report@0.4.2, dsh-local-ai@0.2.27, dsh-industry-research@0.3.26, dsh-checkpoint-rewind@0.6.35, dsh-defend@0.3.28, dsh-claude-move@0.4.20, dsh-budget@0.4.26, dsh-data-quality@0.3.27, dsh-mask@0.2.25, dsh-fund-research@0.5.4, dsh-observe@0.2.28, dsh-test-drive@0.3.26 — re-audits all 12 with 1.7.0 (all recognized, 13/14 checks, 0 blocking), records 52 new capability entries and 12 new dsh.standards declarations (declared tracks 9 -> 21). The remaining work is a runtime tested matrix, not more static coverage.

Run npm run coverage to reproduce; the matrix-derived output is authoritative (the committed data/coverage.json is a historical snapshot).

CLI

node scripts/hcb-cli.mjs check dsh-auto-review
node scripts/badge.mjs dsh-auto-review > badge.svg

# v0.5 evidence audit (dry run by default)
node scripts/audit-entry.mjs dsh-plugin-guide
node scripts/audit-entry.mjs dsh-plugin-guide --merge   # review the patch first
npm run validate
npm test

Quick start

npm install
npm run collect   # needs GITHUB_TOKEN for higher rate limits
npm run validate
python3 -m http.server 8080   # open index.html

Use in CI

- uses: PerryLink/dsh-hcb/.github/actions/check@main
  with:
    package: your-npm-package

It prints compatibility + tracks + gate; fails only on declared-incompatible by default. See docs/CI.md.

How to add a plugin

  1. Add the npm package name to data/plugins.seed.json.
  2. Run npm run collect.
  3. Optional but encouraged: run node scripts/audit-entry.mjs <package> and, after reviewing the patch, --merge; commit the raw report under evidence/.
  4. Open a PR with the regenerated data/*.json, the evidence report/patch, and the reason. A checked row must include its reportPath.

What this is not

  • Not a security audit. Declared capabilities ≠ actual behavior.
  • Not a runtime test. checked is a static tarball check; only tested means a host booted it.
  • Not a ranking. No total score; ENTRY/FLOW/BOOST are separate.
  • Not a certification. production-candidate only means HCB found no blocker in the checks it knows.
  • Not official. Do not use the DeepSeek name/logo to imply endorsement.

Roadmap

  • v0.4 full 112-package family coverage + three tracks + risk/gate.
  • v0.5 evidence layer: declared / checked / tested, static audit pipeline, patches.
  • v0.6 full mainline-default static checked coverage (34/34; 41 checked overall); runtime tested matrix, capability diff and stale alerts remain future work.
  • v0.7 first 19 of 53 mainline-opt-in rows statically checked (60 checked overall).
  • v0.8 finish mainline-opt-in (53/53) and add the first 2 of 10 catalog-only rows statically checked (96 checked overall); the remaining 8 catalog-only, 15 domain and 24 external rows plus runtime tested remain future work.
  • v0.8.1 refresh dsh-auto-review@0.15.2, dsh-permission-rules@0.8.1 and dsh-plugin-guide@0.5.9 from npm; re-audit all three (checked stays 96); hold dsh-permission-rules ENTRY blocked until a 16-ID checker can validate sessions:write; record FLOW manual for auto-review and permission-rules (never runtime).
  • v0.9 finish the family: 8 remaining catalog-only + 15 domain rows statically checked with dsh-standards@1.7.0 (23 new; 119 checked overall; 112/112 family checked); 24 external rows stay declared, runtime tested remains future work.
  • v0.10 finish the matrix: all 24 remaining external rows statically checked with dsh-standards@1.7.0 (119 -> 143 checked; 0 declared), 23 recognized + 1 failed (dsh-session-fork@0.4.0, missing built entry). No tested evidence; runtime tested, capability diff and stale alerts remain future work.
  • v0.10.2 unify the checker: re-audit the 95 rows still on dsh-standards@1.6.10 with 1.7.0 (143/143 now 1.7.0; ENTRY stays 141 recognized + 2 failed), make 1.7.0 the default, and label historical rows by row.evidence.checker.
  • v0.10.3 absorb dsh-lsp-actions@0.5.21, dsh-doublecheck@0.9.32, dsh-memento@0.5.31 and dsh-session-sync@0.3.2: four 1.7.0 re-audits, all recognized (13/14, 0 blocking); declared tracks 5 -> 9; ENTRY/gate unchanged (141 recognized + 2 failed; 2 blocked / 21 review / 120 pilot).
  • v0.11.1 absorb 12 new releases (dsh-research-report@0.4.2 through dsh-test-drive@0.3.26): 12 1.7.0 re-audits, all recognized (13/14, 0 blocking), 52 capability entries, declared tracks 9 -> 21; nine high-capability rows move pilot -> review, so gate is 2 blocked / 30 review / 110 pilot / 1 production-candidate. dsh-mcp-panel (0.6.32 not published) and dsh-click (0.3.29; 0.3.30 was staged at scope time) stay at their old versions.
  • v1.0 stable schema + CI action + weekly refresh + DOI.

License

MIT (code and data). See LICENSE. The track model is Apache-2.0 from dsh-standards; see NOTICE.md.

About

HCB data/gate subproject of dsh-standards: compatibility matrix, ENTRY/FLOW/BOOST tracks, risk, hcb-gate/1.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages