Files
agent-runtime-framework/model-registry/tiers.yaml
Paul O'Reilly d7decc1dec minimax: move to MiniMax-M3.1-Flash-Preview; drop bare minimax/minimax-m2 slugs
- new context minimax-m31-flash/v1 (model pin; no thinking-budget init, Flash
  ignores it); four composites switch from minimax-m3-max/v1 (kept for rollback)
- models + registry for minimax-m31-flash-{max,medium}; removes the dangling
  model-registry/minimax.yaml (endpoint "minimax" never existed)
- tiers, harness-roles, minimax-* task templates and the *-review-loop@1
  workflows use the Flash slugs (MiniMax silently maps unknown names to M3, so
  the provider-id translation lands in agent-runtimes)
2026-10-13 10:55:03 +13:00

60 lines
2.0 KiB
YAML

# Model capability tiers for typed-pipeline routing.
#
# Each tier defines score floors (against model-registry/*.yaml `scores` dimensions).
# The CP picks the cheapest registry entry meeting ALL floors (MR-9/MR-11 — cheapest-wins).
#
# Stages that need a specific tier set `tier:` on their `kind: agent` node (W2).
# Until W2 lands, templates hardcode `model:` directly and these floors serve as
# documentation of the intended constraint.
#
# See claude/stage-routing.md for the authoritative pipeline stage map.
tiers:
planning:
description: >
Judgment-heavy work: plan generation, spec re-draft (replan-spec-opus),
architecture review escalation (review-spec-opus). Requires sustained
reasoning over large context windows.
min_scores:
complexity: 9
creativity: 9
context_utilisation: 9
qualifies_today:
- deepseek-v4flash # benchmark-inflated 9s; hold to allowed_models ceiling
- minimax-m31-flash-max
- claude-opus-5
note: >
WF-SCH-14 allowed_models ceiling remains authoritative until M7 observed-score
recalibration corrects benchmark-inflated registry entries.
spec-test:
description: >
Spec-adherent structured work: spec drafting, spec arch review, scope/decompose,
test writing, impl review, merge verdict. Needs reliable spec adherence and
test execution.
min_scores:
spec_adherence: 9
test_pass_rate: 9
qualifies_today:
- minimax-m31-flash-max
- minimax-m31-flash-medium
- claude-sonnet-5
- claude-opus-5
coding:
description: >
Narrow TDD implementation against pre-validated tests. Lower floors because
tests already constrain the solution space.
min_scores:
spec_adherence: 7
test_pass_rate: 7
qualifies_today:
- airouter-qwen3
- airouter-deepseekv4flash
- minimax-m31-flash-max
- minimax-m31-flash-medium
- claude-sonnet-5
- claude-opus-5
- claude-fable-5