v3.8.0-alpha.7 — Model routing: prefer the session model
Pre-release
Pre-release
·
136 commits
to main
since this release
The plugin no longer hardcodes a stronger model (Opus) for "complex" work, nor pins substantive agents to Sonnet. It now inherits the session model by default — so a newer/stronger model the user launches flows through automatically, with no plugin update needed to "keep up." The only deliberate override is down-shifting to
haikufor trivial mechanical work.
Changed
rules/core/small-to-large-routing.mdrewritten — from aHaiku→Sonnet→Opusladder with "start Opus for hard problems" to "prefer the session model; down-shift tohaikuonly for trivial work; never force-upgrade to a named model." The session model is the ceiling; there is no escalation to a hardcoded model. Defaults table: substantive agents → inherit; only the trivial detectors keephaiku.skills/agent-detector/SKILL.mdModel Selection — wasQuick→haiku, Standard→sonnet (opus for architecture/design), Deep→sonnet (opus for planning); now Quick/classification →haiku, everything else inherits the session model.- Stripped per-agent
model:overrides so they inherit the session model:strategist,tester,security,devops(weremodel: sonnet), plus reasoning skillsproblem-solving,sequential-thinking. Deliberatehaikufloors kept onscanner,agent-detector,git-workflow,session-continuation. - Reframed model-named guidance to be capability-based (future-proof, no stale model names):
rules/core/context-management.mdcompaction strategy now keys off context-window size (small/mid vs large) instead of Haiku/Sonnet/Opus;rules/workflow/dual-llm-review.mdreviewer selection now matches the session model tier instead of naming Opus/Sonnet/Haiku. - Doc sync:
agents/README.mdfrontmatter schema + rules (omitmodel:to inherit;haikuonly as a cheap floor),rules/README.mdone-line description.
Why
New model releases no longer require a plugin update to be used for complex work — the user's session model choice is the single source of truth. Removes the anti-pattern the routing rule itself warned against ("set a fixed model per agent regardless — breaks Opus-session users' intent").