From da7c4cb709f97471d933469a17f1310e7534547a Mon Sep 17 00:00:00 2001 From: Jason Woltje Date: Mon, 12 Oct 2026 16:51:20 -0500 Subject: [PATCH] decisions: 122, Haiku for low-to-mid work, GPT for review Co-Authored-By: Claude Opus 5.5 --- docs/plans/2026-09-26_lead-decisions.md | 32 +++++++++++++++++++++++++ 1 file changed, 32 insertions(+) diff --git a/docs/plans/2026-09-26_lead-decisions.md b/docs/plans/2026-09-26_lead-decisions.md index dd3401f7b..9c7ff75d3 100644 --- a/docs/plans/2026-09-26_lead-decisions.md +++ b/docs/plans/2026-09-26_lead-decisions.md @@ -2743,3 +2743,35 @@ which stay with him. Each item names who decided it and what happened. message the same way. agents/README.md gets each new thread id when it launches. - The order doesn't change: #232, then #228, then the redeploy. +122. **Model routing revised: Haiku for low-to-mid work, GPT for review + (2026-10-12).** This replaces decision 121's seat plan. Jason added + through Mos (sage-0113) that Haiku 5.5 on high effort handles A/B + testing, research and low-to-mid coding, and that GPT 6 and 6.1 are + very good at code review. Q129's terms still hold until 10-18 01:00 + CDT: gpt-6.1-sol first, gpt-6-sol as fallback, gpt-6-luna only for + very low-level tasks, never Astra (R26), no switch mid-task, and + each model named when its thread is created. + - Larger builds split between Opus and Codex Sol. Of rows 85, 90, + 91, 93 and 94, Codex builds 85, 93 and 94, and Opus builds 90 and + 91. + - Low-to-mid rows go to Haiku on high effort in place of Opus: + rows 71, 80, 84 and 88, plus row 96, which is research and an A/B + test. Row 96's gate doesn't change: it still waits on rows 92 and + 93, the redeploy and Jason's re-test (decision 120). + - Rocko moves to Codex gpt-6.1-sol after row 77 lands and builds + rows 94 and 85. Rows 84 and 88 move from Rocko to Researcher. + - Researcher moves to Haiku 5.5 on high effort after row 86 lands. + They build rows 71, 80, 84 and 88 while row 96 waits, then run + 96. Rows 71 and 80 move from Darkwing to Researcher. + - Filbert moves to Codex gpt-6.1-sol as the GPT reviewer. The move + waits until row 95 lands and Filbert's row 92 round 2 verdict is + posted, whichever is later. On Codex they review and build row + 93. Filbert already reviews rows 71, 80, 84, 88, 90 and 91, so + every Opus- or Haiku-built row gets a GPT review. + - Every Codex-built row keeps a Claude reviewer: Dewey on 94, + Darkwing on 85, and Darkwing and Dewey on 93. + - Darkwing and Dewey stay on Opus. Darkwing keeps rows 63, 69, 91 + and 92, and Dewey keeps rows 89 and 90. + - Opus seats drop to three: Sage, Darkwing and Dewey. + - T3's thread-create call takes a model but no effort level, so + Haiku's high effort is set in the pane and confirmed there.