You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
crew-afk can now hold the reviewer and triage judge to a different model tier than the
coder, explicitly — previously --model applied identically to all three, with no way to
say "review stricter than the coder wrote it." An optional .coding-crew/afk-models.json
({coder, reviewer, triage}, all keys optional) fills this in: an omitted reviewer/triage
still defaults to the coder's own resolved model (the existing guarantee — a role can never
end up weaker by accident, only by an explicit, on-record choice), and --model on the CLI
still overrides the file's coder for that run. No numeric ranking across arbitrary model
strings (most of this repo's own platforms pass an opaque string through to their own CLI, so
a universal rank table would be a lie); the one case this repo does know a real order for —
Claude Code's own opus/sonnet/haiku aliases — gets an advisory (non-blocking) warning
when a role explicitly diverges to a weaker one.
Dispatch logs now name the model and issue actually used. The orchestrator's own trace log
(orchestrator.log) gains model=<value> on every dispatch-coder/dispatch-review/ dispatch-triage step marker — previously only the pi/codex platforms' own dispatch-script
logs recorded a model at all, and none of the four platforms' logs named which issue a [DISPATCH] line belonged to. dispatch-agent.sh/dispatch-codex-agent.sh's [DISPATCH]
line now also carries slug=<issue>, matching the convention their own [TOOL]/[TOOL-ERROR]
lines already used.