DCS

Drunken Coding Skills: a deliberately small, personal library of the Agent Skills I actually use.

The repository root is a skill-only Agent Plugins 1.0 package. Its root plugin.json declares the portable format, and clients discover skills from skills/ by convention. There is intentionally no legacy client-specific plugin manifest.

The package contains nine maintainer-owned skills:

file-pr opens a focused pull request for the current branch.
babysit-pr monitors an existing pull request through reviews and CI.
ask-opus gets a second opinion from Opus through claude -p.
ask-antigravity gets a Gemini 3.8 Flash High review through Antigravity CLI.
ask-grok gets a Grok 4.6 xhigh review through Grok CLI, with a 30-minute default deadline.
ask-others delegates to one GPT-5.6 Luna low coordinator, which invokes those three skills and returns their full reviews unchanged.
chatgpt-pr-review runs an explicit ChatGPT Pro review, waits for its final response, and applies locally verified findings.
wait pauses for work that may need more time and checks again when useful.
review-to-problems investigates review findings with one subagent per issue and writes checked repository-native problem reports.

It also tracks selected public skills:

anti-ui-slop helps agents build product-specific interfaces and check them for generic UI defaults.
architecture-map builds an interactive isometric map from measured repository structure and authored call paths.
codebase-design provides the shared vocabulary and principles for deep-module design.
deslop removes AI-generated code slop. DCS tracks the skill from Cursor's public cursor/plugins repository.
grill-with-docs stress-tests a plan or design while sharpening its domain language and recording durable decisions. DCS also packages the grilling and domain-modeling skills it composes.
improve-codebase-architecture finds architectural deepening opportunities, presents them visually, and grills the selected direction.
pstack provides one explicit, evidence-first engineering workflow distilled from Lauren Tan's pstack principles, quality rubric, and core playbooks.
thermo-nuclear-review performs a strict correctness and security branch audit.
thermo-nuclear-code-quality-review performs a strict maintainability and structure audit.
thermos runs both Thermo-Nuclear reviews concurrently and synthesizes the results.
unslop mirrors P-Stack's writing cleanup skill, removing common AI tells while preserving meaning and intended tone.

The CLI review skills require the corresponding claude, agy, or grok CLI installed and signed in. Antigravity and Grok use their own Python runners; Ask Others only delegates to the individual skills. Review inputs stay in a fixed snapshot, and raw responses remain available alongside the unchanged full reviews.

Future skills belong in skills/ and should be added only when they earn their place.

Upstream skills

upstreams.json is the complete administration file for third-party skills. Each entry names a public Git repository, a ref, the source directory, and its destination under skills/.

Most skill instructions are exact mirrors. Packaging adds a DCS: prefix to every Codex display name so its origin stays visible in skill lists. Adapted skills can also name a small adapter, patch, or file overlay under ports/. The anti-ui-slop adapter packages the last self-contained skill while upstream references runtime scripts that its source directory does not include. It fails when that upstream assumption changes so the port can be reviewed and removed. The pstack adapter tracks the complete upstream plugin but packages one Codex-native pstack skill instead of its Cursor-specific skill stack, agents, model router, and automations. It verifies pstack's identity and main router before packaging the port. The grill-with-docs port preserves manual invocation while replacing its generic Skill-tool handoff with Codex instructions. The Thermos ports remove Cursor-only invocation metadata, replace Cursor's subagent syntax with Codex instructions, and add Codex UI metadata. The Matt Pocock skills bundle their upstream MIT notice, while the architecture orchestrator drops a redundant legacy manual-invocation field and loads its bundled dependencies through Codex-native skill instructions. upstreams.lock.json records each adapted upstream Git tree, so a change remains visible even when an adapter produces identical packaged output. Its managedSkills list tracks every generated skill so removing a registry entry also removes its packaged directory.

Install the synchronizer's YAML dependency once:

python3 -m pip install -r tools/requirements.txt

Check for upstream changes without modifying the repository:

python3 tools/sync_upstreams.py --check

Synchronize registered skills locally:

python3 tools/sync_upstreams.py --sync

The Upstream skills GitHub Actions workflow runs the check every night. A changed upstream makes the run fail so normal GitHub Actions notifications can alert you. Its manual sync operation imports exact upstream directory contents, bumps the DCS patch version, pushes a dedicated update branch, and puts a one-click pull request link in the run summary. It never changes main or merges anything.

The workflow uses only the repository's built-in GITHUB_TOKEN. It needs no personal access token, AI API key, or other secret.

Install

The portable specification leaves installation and distribution to each client. DCS keeps .agents/plugins/marketplace.json only as its Codex distribution index; it is not part of the plugin package format.

Portable Agent Plugins installation requires Codex 0.147.0 or later. Add this repository as a marketplace, then install and enable DCS:

codex plugin marketplace add the-Drunken-coder/DCS --ref main
codex plugin add dcs@dcs

Verify the installation:

codex plugin list --marketplace dcs

Update

Refresh the marketplace snapshot and reinstall DCS:

codex plugin marketplace upgrade dcs
codex plugin add dcs@dcs

Disable or remove

To disable DCS without uninstalling it, open DCS in the Codex Plugins Directory and turn it off. The current CLI has no plugin-disable subcommand.

To uninstall DCS while keeping its marketplace configured:

codex plugin remove dcs@dcs

To also remove the marketplace:

codex plugin marketplace remove dcs

Local development

Clone and validate the plugin:

git clone https://github.com/the-Drunken-coder/DCS.git
cd DCS
python3 -m unittest discover -s tests -v
python3 tools/sync_upstreams.py --validate
python3 tools/sync_upstreams.py --check

Run the wait skill against an actual OpenCode agent without naming the skill in the prompt:

python3 tools/run_wait_skill_trial.py foreground-explicit --duration 20
python3 tools/run_wait_skill_trial.py detached-implicit --duration 20
python3 tools/run_wait_skill_trial.py opaque-underestimate --duration 20
python3 tools/run_wait_skill_trial.py opaque-implicit --duration 20
python3 tools/run_wait_skill_trial.py opaque-buried --duration 20
python3 tools/run_wait_skill_trial.py resumable-terminal --duration 20

Run three reliability trials sequentially across the cheap Go model set:

python3 tools/run_wait_skill_matrix.py --repeat 3 --duration 12

Compare isolated discovery descriptions and names without changing the packaged skill:

python3 tools/run_wait_skill_variants.py --repeat 2 --duration 12

The variant runner also accepts the static-timeout, completed-build-log, and async-code-review near-miss scenarios for checking unwanted activation.

Pass --model opencode-go/deepseek-v4-flash to use the Go subscription model after enabling its China-hosted models. The default uses OpenCode's free DeepSeek V4 Flash endpoint.

The repository validator also enforces DCS release policy: a plain Semantic Versioning release, a non-empty description, and a named author. Those requirements are intentionally stricter than the minimum portable manifest schema.

To test the checkout directly, replace only the DCS marketplace with the local repository and reinstall:

codex plugin remove dcs@dcs
codex plugin marketplace remove dcs
codex plugin marketplace add "$PWD"
codex plugin add dcs@dcs