From 2b48bf591a01e08883532f0d9d94b2ba597b5440 Mon Sep 17 00:00:00 2001 From: Suleiman Shahbari Date: Sat, 22 Aug 2026 03:05:19 +0300 Subject: [PATCH] Haiku is out of scope, and the reason is written down (#1334) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Rom's call on #1334's last open question: neither a model floor nor a refusal to arm the hand-off. Haiku simply isn't supported, and the dashboard's existing warn-never-block (#1439) is the whole treatment. Recorded in MEMORY.md because it is a decision with a rationale that no future session can infer from the code — the code only shows a warning, not that the two workarounds were considered and rejected. --- MEMORY.md | 1 + 1 file changed, 1 insertion(+) diff --git a/MEMORY.md b/MEMORY.md index 0d82c1a4e..74b965473 100644 --- a/MEMORY.md +++ b/MEMORY.md @@ -6,6 +6,7 @@ - **Never interrupt a running session because of low quota.** Quota gates whether a session may *start*; a session already running is never paused, degraded or cut short. - **The unit of work is an *agent*** — not a run or a session. The CLI that drives it (`claude`, `codex`) is the *driver*. - **Zero migration code.** Manual migration is good; code that migrates is not. A renamed or deleted thing leaves nothing behind — no fallback to the old name, no key alias, no old-format branch, no upgrade step. Nor a notice about what was dropped: a warning about data nobody has is still code nobody needs. +- **Haiku is out of scope.** It skips the session-finish protocol, so unattended work started on it always ends as an unmerged draft PR that a human has to finish. We deliberately don't work around this: no model floor that overrides the model a user picked, and no refusal to arm the hand-off. The dashboard's warning is the whole treatment — it teaches and never blocks. ## Before modifying this file