Trident

Prompt, plan, build. A Claude Code plugin that routes each step of a task to the right Claude model, grounds the build in open, permissively-licensed work, and ends every run with a one-page report your team can read.

/trident:ship build a CLI that converts CSV to Parquet with a --compression flag

Stage	Model	What it does
Refine	Haiku	Turns a rough request into a spec with testable acceptance criteria
Plan	Fable	Researches existing libraries and papers, checks licences, writes a cited plan, picks the builder
Build	Sonnet, or Opus for complex work	Implements the plan and checks every API against the installed version before using it
Verify	Haiku	Independently tests the acceptance criteria and dependency licences, and sends failures back once
Report	—	Writes .trident/reports/<date>-<task>.md: summary, decisions, sources, verification, changes, risks

Why Trident

• Grounded, not guessed. The planner cites only sources it has opened. The builder confirms function signatures in the installed package. A separate verifier, which did not write the code, checks the result.
• Licence-safe by default. Dependencies must be MIT, Apache-2.0, BSD, ISC, Unlicense, CC0, PSF or Zlib. GPL-family, SSPL, BUSL, non-commercial and unknown licences are blocked.
• People stay in charge. You approve a short brief before anything is built. Reports read as plain business updates and double as pull-request descriptions.
• Spend goes where it matters. Fast models handle the routine stages. The frontier model plans, and Opus is used only when the planner judges the work complex.

Install

Requires Claude Code.

claude plugin marketplace add AquarkGR/Trident
claude plugin install trident@aquark

Update later with claude plugin update trident@aquark.

Use

In any Claude Code session:

/trident:ship <what you want built>

Option (add to the request)	Effect
--yes	Skip the approval brief
--pr	Open a pull request with the report as its description (requires an authenticated gh)

Example: /trident:ship add input validation to the upload endpoint --pr

Example run

A real run of /trident:ship build a small Python CLI csv2parquet … plus a test using sample.csv took about 4 minutes:

• Plan: the planner reviewed an existing Apache-2.0 csv2parquet project, used none of its code, and chose pyarrow alone to keep dependencies minimal.
• Build: Sonnet built a packaged CLI with 10 passing tests.
• Verify: the verifier confirmed each criterion with evidence, including round-tripped values and the codec reported by Parquet metadata.

Read the full report: examples/csv2parquet-report.md

Customize

Each stage is one Markdown file in plugins/trident/agents/. Edit its prompt, tools or model. The orchestration and report template live in plugins/trident/skills/ship/SKILL.md. Bump version in plugin.json when you change behaviour.

Other editions

The same pipeline, with the same neutral prompts, licence policy and report format, is available outside Claude Code:

Edition	Runs on	Folder
Claude API	Claude API and Agent SDK, for CI or your own tooling	editions/claude-api
OpenAI	OpenAI Agents SDK	editions/openai
Local	One local Qwen model on Ollama, sized to your RAM. Stages spend reasoning and context as budget instead of switching to larger models	editions/local

Repository layout

.claude-plugin/marketplace.json   Marketplace catalogue (lists the plugin)
plugins/trident/                  The Claude Code plugin
  agents/                         refiner, planner, builder, verifier
  skills/ship/SKILL.md            /trident:ship orchestration and report template
editions/                         API, OpenAI and local editions
examples/                         Reports from real runs

Contributing

Issues and pull requests are welcome. Before opening a PR:

claude plugin validate --strict ./plugins/trident

CI validates the plugin and dry-runs every edition on each push.

License

MIT © Aquark. Trident is an independent project and is not affiliated with Anthropic, OpenAI or Alibaba Cloud.