Wsl Notify

Reusable agent skills for opencode. Each skill is a directory with a SKILL.md entry point (YAML frontmatter: name, description) plus optional assets/, scripts/, and tests/.

At runtime everything a skill writes lives in one project-local workspace root, .pdca/ — helpers it runs (.pdca/scripts/), cycle status files (.pdca/status/) and evidence (.pdca/<task>/rv<k>/). The skill's own scripts/ is the source of those helpers; the project runs its copies from .pdca/scripts/.

Contents

Skills live under opencode/.

Skill	Description
pdca-entry	Single entry point (umbrella) for the family: classifies a task and routes it to exactly one contract — collection, incremental, .NET, coder or the neutral cycle.
pdca	Domain-neutral PDCA cycle: PLAN → DO → CHECK → ACT with transition gates, evidence-based CHECK, escalation accounting, durable status, and recovery.
pdca-coder	Stack-agnostic PDCA executor for software-engineering tasks (code, tests, docs, refactoring, config). Picks the build/test/lint toolchain per language.
pdca-collection	Runs a batch of tasks as parallel PDCA lanes (worktree per group, branch chain + auto-commit, --no-ff merge of completed groups); a batch may contain single-cycle units and sequential incremental-campaign units.
pdca-dotnet	PDCA cycle for .NET/C# tasks: parallel DO/CHECK streams, role routing, ACT handoff.
pdca-incremental	Drives a goal without a committed whole-volume plan as a chain of small PDCA increments: one accepted measurement per step, baseline chain, one campaign status file, explicit stop conditions, plus repair and verify step kinds; a campaign can also be a unit of a collection batch and then stays sequential.

The five contracts share the same cycle: a cheap orchestrator drives roles (planner, coder, check, scout, escalate), every CHECK is evidence-based, and a required approval that cannot be obtained is a STOP rather than an open question. pdca-entry sits above them: it does not run a cycle, it only classifies the task and hands it to exactly one contract.

Install

Install a single skill with the skills CLI — one command per skill:

npx skills add AlexeyShirshov/skills@pdca-entry
npx skills add AlexeyShirshov/skills@pdca
npx skills add AlexeyShirshov/skills@pdca-coder
npx skills add AlexeyShirshov/skills@pdca-dotnet
npx skills add AlexeyShirshov/skills@pdca-collection
npx skills add AlexeyShirshov/skills@pdca-incremental

Add -g for a user-level install (all projects) and -y to skip prompts:

npx skills add AlexeyShirshov/skills@pdca-coder -g -y

Without -g the CLI targets the current project.

Installing a skill is only half of the setup: the cycle dispatches subagents, so the host needs the role agents and the runner pdca-orchestrator (both ship in the skill's assets/agents/), nested delegation (experimental.subagent_depth) and subagent permissions. Every agent id is pdca-prefixed (pdca-planner, pdca-executor, pdca-check, pdca-escalate, pdca-scout, pdca-security-auditor, pdca-orchestrator), so it cannot collide with an existing planner/coder/scout. The full checklist — with a verification step and a troubleshooting table — is in docs/install.md (docs/ru/install.md); an agent can perform the whole checklist from docs/install-prompt.md.

From a checkout

Clone the repository and copy the skill(s) you want into an opencode skills directory.

Global (all projects):

cp -r opencode/pdca opencode/pdca-coder opencode/pdca-collection opencode/pdca-dotnet \
  opencode/pdca-incremental ~/.config/opencode/skills/

Per project:

mkdir -p .opencode/skills
cp -r opencode/pdca* .opencode/skills/

Skills are discovered by the name field in their frontmatter.

Usage

Only pdca-entry is advertised to the model for automatic selection. The five contracts set metadata.opencode/autoinvoke: false: they stay registered but load only when the user names them explicitly (skill pdca-dotnet, @pdca-dotnet) or when the entry hands off.

Invoke a skill by name, e.g. in an opencode session:

• pdca-entry — "by the cycle" / "via PDCA" / "по циклу": the umbrella that routes the task to one contract below (an explicitly named contract bypasses it).
• pdca — only when explicitly named (domain-neutral cycle).
• pdca-coder — explicitly, for code tasks of any stack.
• pdca-dotnet — explicitly, for .NET/C# work.
• pdca-collection — for a list of independent tasks / a milestone; a collection batch may contain cycle units and campaign units, and a campaign unit stays sequential (one lane, one step at a time).
• pdca-incremental — for a goal without a committed whole-volume plan (per-task plans may already exist), moved forward in steps; includes per-step repair and an optional integrated verify. A campaign can also run as a unit of a collection batch.

Tests

Each skill ships a tests/ suite (plain unittest, standard library only). Run one skill:

python3 -m unittest discover -s opencode/pdca/tests -p 'test_*.py' -v

Some checks compare the skill against host-side artifacts that live in a separate config repository (config/agents/, config/profiles/, config/AGENTS.md, the host README.md). Those checks skip unless you point at that checkout:

OPENCODE_CONFIG_REPO=/path/to/opencode-config \
  python3 -m unittest discover -s opencode/pdca-dotnet/tests -p 'test_*.py'

pdca-collection additionally ships a dependency-free bash check of its git model (branch chain, --no-ff merge of completed groups, conflict detection, single-group isolation, and an incomplete group preserved but never merged) on throwaway temp repositories. unittest discover runs it through tests/test_git_model.py; it can also be run directly:

bash opencode/pdca-collection/tests/git-model-sim.sh

Plugins

Plugins live under opencode/plugins/, one directory per package.

Plugin	Description
wsl-notify	Windows toast notifications for OpenCode — permission, question, completion and error events. Zero dependencies, one entrypoint for OpenCode V1 and V2, works from WSL (powershell.exe interop) and native Windows.
opencode-roslyn	C# semantic analysis for OpenCode V2 — the roslyn tool answers structure/types/symbols/members/refs/callers/implementations/rename through the roslynq CLI (Roslyn + MSBuildWorkspace), with permissions, progress and cancellation. V2 has no LSP runtime yet; this is the semantic answer in the meantime.
opencode-sidebar-plugins	A Plugins section in the OpenCode V2 CLI sidebar — every loaded plugin and, by default, only those that actually reach the model context (prompt, tools, agents, skills). Classifies built-ins from their registrations, scans local and package plugin sources, and refuses to guess when a package entry cannot be resolved.
zen-failover	Server failover for subscription providers: on an exhausted quota it moves every agent, the default model and the active session to the fallback (ChatGPT → the same model on Console, Go → direct DeepSeek), probes the plan once a minute, and restores the original bindings as soon as the subscription answers again. Transient per-minute 429s are retried, not degraded.

Install the published package in opencode.json(c) (plugins on V2, plugin on V1), or copy a plugin's index.js into <config>/plugins/ and let opencode discover it locally. See each plugin's README for options and requirements.

License

MIT © 2026 Alexey Shirshov