deepseek-better-harness
更好的 Harness 证据通道——会话、项目就绪度和代理资产——作为一个包含三项可路由技能的 DeepSeek Harness 捆绑包。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:guyuefangyuanl/deepseek-better-harness#9947e3bca97e6c0bbae8d5155917a371dda6c71dDeepseek Better Harness 以 deepseek-better-harness 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Better Harness reviews an agent project by separating evidence into three independent concerns. This repository carries those three concerns to the DeepSeek Harness as one DSH bundle that registers three independently routable skills.
Install once. Enable the lanes you want — the ones you leave out are never registered, so they cost no routing surface and no context.
Unofficial. Not affiliated with, endorsed by, or released by Qoder or DeepSeek. This is a community port of a boundary, not a redistribution of the product — see NOTICE.
| Lane | Reviews | Answers |
|---|---|---|
session-evidence | Authorized session facts | Where did the agent repeat work, hit friction, or lose a feedback loop? |
project-readiness | The repository itself | Does this repo give an agent a clear change path, bounded risk, and executable verification? |
agent-assets | Skills, presets, hooks, tools, MCP config | Are the agent assets discoverable, correctly routed, intact, and safely owned? |
Each lane contributes exactly one runtime skill through ctx.skills. A project
or user filesystem skill with the same name still overrides the bundled runtime
skill according to DSH precedence.
Each lane is also its own npm package under packages/, so a profile that wants
exactly one can install exactly one — see Install.
Better Harness answers three different questions that fail in three different
ways. Collapsing them into a single "review my project" skill is what makes such
a skill vague enough to misfire. Here each lane is a separate skill with its own
boundary, and config.lanes decides which ones a profile registers at all:
- id: deepseek-better-harness
name: deepseek-better-harness
config:
lanes: [project-readiness]
Wanting a readiness review does not drag in a session analyzer, a scoring engine, and a report renderer. An unknown lane name is a startup error, not a silently smaller skill set.
Every package ships "dependencies": {}. The only declared dep is the DSH peer
package you already have. A bundle is 7 files and about 7 KB — it adds a
capability to your profile without adding a tree to your lockfile.
All three descriptions carry an explicit negative clause:
Use when reviewing authorized DSH session facts for repeated work, agent friction, missing feedback loops, or outcome evidence. Do NOT use to read unapproved raw transcripts, infer hidden reasoning, score a whole repository, or inspect project assets.
Misrouting is the expensive failure in a skill layer — the model loads the wrong skill at the wrong moment, burns context, and drifts. Because the three lanes here are deliberately adjacent, the negative clause is what keeps a session question from pulling in the asset auditor. Most community skills only write the positive half.
harness-session-evidence forces observed behavior, missing evidence, and
inference into separate buckets, and requires at least two comparable
occurrences before something may be called a pattern — one occurrence is a
candidate.
All three skills return evidence candidates rather than final severities or repository-wide scores. That is not modesty: they ship no collectors, so they are not entitled to the score. A skill that cannot see the evidence must report that boundary instead of inventing a finding.
scripts/validate-packages.mjs does not lint manifests. It import()s every
real index.js, captures the skill each one registers, then asserts bundle
identity, frontmatter, resource paths, description limits, and teardown. A
package that is structurally perfect but fails to register is caught here rather
than in your profile. CI runs it on Linux and Windows.
The repository root is the bundle: its manifest declares
dsh.bundle.patch, so DSH treats it as a profile layer and DSH Hub accepts it
as a submission. Tested baseline is DSH 0.1.0-rc.6.
dsh plugin --profile web add "github:guyuefangyuanl/deepseek-better-harness#<commit-sha>"
Pin a commit rather than a branch: the skill text is the product here, and you want a review you can reproduce later.
git clone https://github.com/guyuefangyuanl/deepseek-better-harness.git
dsh plugin --profile web add "link:C:/absolute/path/to/deepseek-better-harness"
On 0.1.0-rc.6 the profile reconciles itself — a dependency whose manifest
declares dsh.bundle.patch is appended to dsh.profile.bundles automatically,
so there is nothing to edit by hand. Confirm the layer landed:
dsh --profile web --dump-config
If the layer is missing, DSH will have said why on stderr:
dsh: warning: <name> declares no dsh.bundle — installed as a plain dependency,
not a profile layer
All three lanes are enabled by default. To register only some of them, give the
bundle a config.lanes list in your profile patch:
- id: deepseek-better-harness
name: deepseek-better-harness
config:
lanes: [project-readiness, agent-assets]
Valid names are session-evidence, project-readiness, and agent-assets. An
unknown name, an empty list, or a non-array fails at startup instead of quietly
registering fewer skills than you asked for.
Each lane is also a standalone package, for a profile that should never be able to route to the other two:
dsh plugin --profile web add "link:C:/absolute/path/to/deepseek-better-harness/packages/project-readiness"
Do not install both the root bundle and a lane package into one profile — the lane would be registered twice.
The selected profile must already provide the DSH skills service and a
user/model skill surface. This bundle installs no lifecycle scripts: there is no
preinstall, install, postinstall, or prepare, and nothing is compiled on
your machine. What is in the tarball is what runs, and CI asserts that.
No dependencies. Node.js 22:
node scripts/validate-packages.mjs
node scripts/pack-dry-run.mjs
parseCanonicalSkill is duplicated verbatim in all three packages. A shared
helper would couple three independently publishable bundles to a common
versioned runtime dependency for about thirty lines of parsing — and break the
zero-dependency property that makes a single-package install cheap. The
validator asserts every copy still produces an identical skill shape. See
the design note.
Publishing is not automated. Registry credentials, package ownership, release tags, and community submission are external actions.
CONTRIBUTING.md covers the checks and the rules that are not negotiable. SECURITY.md covers the trust boundary and how to report a vulnerability.
Created and maintained by @guyuefangyuanl.
MIT. See NOTICE for attribution and the unofficial-status statement.