experiment-team — Stage 6 multi-talent repo

A multi-talent Talent Market repo that ships the two specialists Stage 6 (Auto Experiment) needs in an adversarial research pipeline. Each talent is a separate directory with its own profile.yaml, registered under its own talent_id on Talent Market.

Talents

Directory	talent_id	Owns	Skill
experiment-code-writer/	experiment-code-writer	Stage 6a (Implementation): translate Stage 5 prose plan into runnable Python	code_implementer
experiment-runner/	experiment-runner	Stage 6b (Execution): drive remote experiment HTTP API end to end	experiment_runner

Why a multi-talent repo

Stage 6 splits into two sub-phases (impl + exec) with very different skill sets:

• Code Writer needs Python + research-spec translation chops, low temperature for deterministic implementation, and knowledge of smoke-mode + chat-template conventions.
• Runner needs infra/HTTP API discipline, secret-handling, and polling/cancel semantics — closer to DevOps than data science.

Shipping them as two separate talents from one repo means:

1. They version together — when the code-writer's smoke contract changes, the runner's exec contract changes too; one repo, one PR.
2. The market exposes both with their own pricing, descriptions, and avatars.
3. Pipeline engines that route impl/exec to different employees hire each talent independently.

clone_talent_repo handles the multi-talent layout: it iterates the repo's subdirs, reads each profile.yaml's id, and caches each talent at _TALENTS_CLONE_DIR / <id>/.

Install

Register the repo URL on Talent Market once — the market discovers both talents automatically:

Add Talent → paste https://github.com/YihangChen9/experiment-team

Both experiment-code-writer and experiment-runner then become hireable by their respective talent_id.

Repo layout

experiment-team/
├── README.md
├── LICENSE
├── experiment-code-writer/
│   ├── profile.yaml             id: experiment-code-writer
│   ├── DESCRIPTION.md           verbatim copy of the code-implementation-runbook
│   ├── avatar.jpg
│   ├── skills/
│   │   ├── core.md
│   │   └── code-implementation-runbook/
│   │       └── SKILL.md         Stage 6a implementation runbook
│   └── tools/
│       └── manifest.yaml
└── experiment-runner/
    ├── profile.yaml             id: experiment-runner
    ├── DESCRIPTION.md           verbatim copy of the experiment-execution-runbook
    ├── avatar.jpg
    ├── skills/
    │   ├── core.md
    │   ├── experiment-execution-runbook/
    │   │   └── SKILL.md         Stage 6b execution orchestrator runbook
    │   └── experiment-infra/    HTTP API runbook + fast_*.sh scripts + assets
    │       ├── SKILL.md
    │       ├── scripts/         fast_query_budget / fast_submit / fast_query_exp_status / fast_cancel / ...
    │       ├── assets/          base.conf.json + run YAMLs + sample Python scripts
    │       ├── references/      exp-configuration docs + runtime_images.json
    │       ├── receipt/
    │       └── experiment_infra_credentials.example.json   (real credentials NEVER committed)
    └── tools/
        └── manifest.yaml

Upstream source

Extracted from the OneManCompany AutoResearch pipeline at https://github.com/Memento-Teams/Memento-Research. The runbooks are byte-identical to that repo's feat/stage6-implementation-subphase branch:

• experiment-code-writer/skills/code-implementation-runbook/SKILL.md ← src/onemancompany/default_skills/code-implementation-runbook/SKILL.md
• experiment-runner/skills/experiment-execution-runbook/SKILL.md ← src/onemancompany/default_skills/experiment-execution-runbook/SKILL.md
• experiment-runner/skills/experiment-infra/ ← src/onemancompany/default_skills/experiment-infra/ (with the real credentials file stripped)

License

See LICENSE.