private-agent-skills

Readme

# Private Agent Skills

This directory is the source-of-truth workspace for privately maintained, harness-agnostic agent skills. The skills follow the standard `SKILL.md` format and work with any compatible harness, including Codex, Claude Code, and DeepSeek Harness (DSH).

## Structure

- `project-lifecycle-workflow/`: route software work across the full lifecycle and choose the right branch workflow.
- `integrated-development-workflow/`: orchestrate the complete project workflow with lifecycle routing, method selection, proportional gates, evidence, release, and handoff.
- `construction-plan-system/`: build a gated construction-planning system for large multi-phase software work.
- `long-task-planner/`: create or normalize a durable task book for long-running work.
- `execution-launch/`: convert a finalized plan into a fresh-session full-execution launch brief.
- `session-handoff/`: compress current progress into a restart-ready handoff packet.
- `company-lab-2-server/`: connect to the internal DGX Spark server through the `company-lab` jump host.
- `scripts/sync_skills.ps1`: copy selected skills into a target skills root (default `$HOME\.codex\skills`, overridable with `-DestinationRoot` or the `SKILLS_ROOT` environment variable).
- `scripts/validate_skills.ps1`: run the repository-bundled `quick_validate.py` against selected skills. No agent installation is required.

## Current workflow

1. Route and execute incoming project work with `$integrated-development-workflow`.
2. For lifecycle branch semantics, preserve and invoke `$project-lifecycle-workflow` as the specialist source workflow.
3. For method-specific practice, invoke `$agentic-project-development` when selected by the integrated workflow.
4. For large multi-phase work, build or refresh the gated planning system with `$construction-plan-system`.
5. Create or normalize the canonical task book with `$long-task-planner`.
6. Once the plan is fully approved, generate an execution-first launch brief with `$execution-launch`.
7. Execute work and keep the progress ledger current in the canonical progress file.
8. Before context pressure or session switching on in-progress work, update the canonical progress file first.
9. Generate a restart-ready continuity packet with `$session-handoff`.

For a copy-paste Chinese workflow script that chains all three skills, see `workflow-playbook-zh.md`.

## Current skills

### `integrated-development-workflow`

- Purpose: provide the default end-to-end project development entry point, combining lifecycle routing, development-method selection, proportional planning, implementation control, evidence-based verification, review, release, maintenance, and resumable handoff.
- Design: an orchestrator that preserves the capabilities of `project-lifecycle-workflow` and `agentic-project-development`, escalating to `construction-plan-system`, `long-task-planner`, `execution-launch`, or `session-handoff` as needed.
- Main references:
  - `references/stages/`
  - `references/governance/`
  - `assets/templates/`

### `project-lifecycle-workflow`

- Purpose: route project work across lifecycle phases and branch into the correct execution workflow.
- Best for: project intake, task classification, lifecycle governance, branch selection, and deciding when to escalate into deeper planning.
- Main references:
  - `governance/`
  - `templates/`

### `construction-plan-system`

- Purpose: build a full gated construction-planning system for large software efforts before implementation begins.
- Best for: multi-phase delivery plans, preflight review, long-horizon execution governance, evidence management, and readiness gates.
- Main references:
  - `governance/`
  - `templates/`

### `long-task-planner`

- Purpose: maintain a stable task book for long-running execution.
- Best for: initial planning, normalizing existing governance docs, progress updates, replanning.
- Main references:
  - `references/task-book-template-zh.md`
  - `references/usage-playbook-zh.md`

### `execution-launch`

- Purpose: turn a finalized plan into a one-shot execution brief for a fresh session.
- Best for: approved designs, execution kickoff, reduced mid-task check-ins, Agent Team-first implementation prompts.
- Main references:
  - `references/launch-template-zh.md`
  - `references/usage-playbook-zh.md`

### `session-handoff`

- Purpose: generate a compact, continuity-focused handoff for the next session.
- Best for: context-pressure checkpoints, in-progress session switches, restart prompt generation for unfinished work that already has real execution state.
- Main references:
  - `references/handoff-template-zh.md`
  - `references/usage-playbook-zh.md`

### `company-lab-2-server`

- Purpose: connect through `company-lab` to the internal DGX Spark server for controlled inspection and development commands.
- Best for: environment checks, GPU/CPU/memory inspection, Docker/package discovery, model-training readiness checks, and task-scoped commands under `/home/user02`.
- Security note: the publishable repository copy does not store the server password; provide it through an approved local-only mechanism such as `DGX_SERVER_PASSWORD` or an interactive prompt.

## Maintenance

- Edit skill content here first.
- Run `scripts/validate_skills.ps1` after changes.
- Run `scripts/sync_skills.ps1` to refresh the installed copy (default target `$HOME\.codex\skills`, or pass `-DestinationRoot <skills-root>` / set `SKILLS_ROOT` for another harness).
- Keep skill folders self-contained and avoid adding non-essential files inside each skill directory.
- Save markdown references as UTF-8 to avoid template or prompt corruption in Chinese workflows.

## Harness compatibility

The skill bundle uses the standard `SKILL.md` layout and is harness-agnostic:

- **Codex**: install into `$CODEX_HOME\skills` (default) — `agents/openai.yaml` provides invocation metadata.
- **Claude Code**: copy the skill directories into `~/.claude/skills` (or a project's `.claude/skills`).
- **DeepSeek Harness (DSH)**: this repository is itself a DSH bundle plugin — `package.json` declares `dsh.bundle.patch` (→ `cordis.patch.yml`) and `index.js` registers the seven-skill suite on `ctx.skills`. Once the `private-agent-skills` package is published to npm, install it from the DSH plugin market (or the [1024 Store](https://deepseek1024.com/)); until then, copy the seven skill directories into `~/.dsh/skills` or `~/.agents/skills` (or a project's `.dsh/skills`).

> Note: `index.js` registers exactly the seven-suite skills by whitelist. Personal/environment-specific skills in this repository (e.g. `company-lab-2-server`) are intentionally not part of the published bundle.

The `agents/openai.yaml` files are optional Codex metadata; harnesses without that convention ignore them and the bundle still validates.

## Install the integrated workflow bundle

The integrated workflow depends on six companion skills. Install the supported bundle instead of copying only `integrated-development-workflow`.

```powershell
git clone https://github.com/MikumikuDAIFans/PrivateAgentSkills.git
Set-Location PrivateAgentSkills
# Optional: target a specific harness skills root (default: $CODEX_HOME\skills or $HOME\.codex\skills)
powershell -ExecutionPolicy Bypass -File .\scripts\install_integrated_workflow.ps1 -DestinationRoot "$HOME\.claude\skills"
```

If an older copy is already installed, add `-Force`. The installer copies exactly the seven required skill directories into the destination root (resolution order: `-DestinationRoot`, `SKILLS_ROOT` environment variable, `$CODEX_HOME\skills`, `$HOME\.codex\skills`). It then validates skill names, dependencies, and representative routing cases. Restart the agent harness (or start a new session) after installation.