Alchemist: Agent Pipeline Configuration
Tune the agent pipeline per project: models, effort, and per-phase overrides.
Every build ticket runs through a pipeline of phases — plan, implement, review, security, and docs — plus a read-only investigate phase for research-only tickets. By default each phase uses a platform-tuned model and reasoning setting. The Agent pipeline configuration panel (on a project’s settings page) lets you override any of that per project, per phase.
For each phase you can set:
- Model — pin a specific model for that phase. Choose from the built-in
catalog (Anthropic, OpenAI, xAI, Cerebras) or any OpenRouter model (pick
from the fetched list or paste a custom
openrouter/…id). - Reasoning — the effort and thinking level. Options are gated to what the chosen model supports, so you can’t pick an invalid combination.
- Additional instructions — extra guidance appended to that phase’s prompt. This runs after the built-in framing, so it’s safe — the tool contract and required terminal actions are still enforced.
- Prompt override (advanced) — replace the editable body of the phase’s prompt entirely. Powerful, but you own the result; a bad override can break that phase’s runs.
How a phase resolves its settings
Settings resolve in priority order, highest first. The Effective configuration table at the top of the panel shows what each phase will actually use and which layer won (the source badge):
- project — your per-project override (this panel).
- env — a platform operator environment override.
- tier — the project’s pipeline tier (e.g. a faster/cheaper tier).
- default — the platform’s built-in default for that phase.
Leaving a field blank means that phase falls through to the next layer. Leaving everything blank is identical to the platform default — the panel only changes behavior for the fields you set.
Resetting
Each phase has a Reset button that clears all of its overrides and returns it to the resolved default. Resetting one phase never touches the others.