Alchemist: Pipeline Permutations

Reference for the pipeline shapes a ticket can run through.

---

The Alchemist agent pipeline is content-shape, not time-shape. Which planner is chosen, and which model fills each slot, depends on the ticket's tier and the phase being run.

This page is a reference companion to Agent Pipeline Configuration, which is where you actually change these settings.

• Plan. The default planner runs at the high rung. The max tier, or an explicit request, upgrades to the strongest available planning model. The fast tier drops to a quicker model, and economy (explicit only) drops further.
• Implement. The default is tuned for fix quality per dollar, which is where most of a ticket's cost lands. The max tier upgrades to a frontier model. Trivial changes fall through to a mid-tier model.
• Review, security, reconcile, and merge resolution. Frontier models throughout. These phases are the safety net, so they are not cost-optimized.
• Docs. A large-context mid-tier model, since the docs pass has to hold the documentation corpus and the diff at once.
• Investigate. A fast, cheap model unless the max tier is set, which upgrades to frontier.

The cheap tier forces the fastest model across the board and skips the docs pass entirely.

Note: Specific model assignments change as the platform re-benchmarks. Treat the tiering logic above as stable and the individual model names as current-state. Check the in-product Model preferences panel for the live per-tier caps.

• Client-facing production apps. Default or max. The review and security passes are the reason you can hand the result to a client.
• Internal tools and prototypes. fast is usually fine.
• Bulk refactors and cleanup. cheap works, but read the diff before you ship it, since the docs pass is skipped.

*AVCodex · Your AV expertise. Amplified by AI.*