Skip to content

πŸ™Œ OpenHands: AI-Driven Development

fix(router): complete Flash template prompt and inject model table (#17860)

Juan Pedro Michelini Jorge443b2cd9c3d5
diff --git a/.github/pr-assets/pr-17860-router-flash-prompt-before-after.png b/.github/pr-assets/pr-17860-router-flash-prompt-before-after.png
new file mode 100644
index 0000000000..57d88eecbc
Binary files /dev/null and b/.github/pr-assets/pr-17860-router-flash-prompt-before-after.png differ
diff --git a/src/components/features/settings/meta-llm-profiles/default-meta-profile.ts b/src/components/features/settings/meta-llm-profiles/default-meta-profile.ts
index efec6f6937..de3a2e27ec 100644
--- a/src/components/features/settings/meta-llm-profiles/default-meta-profile.ts
+++ b/src/components/features/settings/meta-llm-profiles/default-meta-profile.ts
@@ -122,12 +122,16 @@ ANTI-PATTERNS TO AVOID:
- Escalating routine repo fixes to a premium model on a single vague difficulty signal β€” most such tasks are solved far cheaper.
- Routing research/QA tasks to premium coding flagships instead of GPT-5.4/GPT-5.5.
- Using claude-opus-4-8 as the buildout workhorse when GPT-5.4 covers standard scope at half the cost.
-- Downgrading protected-
-Task:
-{{ instance_text }}
+You may choose any model listed in the model table below; the recommendations above should be followed unless the task text gives a strong, specific reason to deviate.
-Return ONLY JSON: {"model": "<exact model name>", "reason": "<short reason>"}`;
+{{ model_table }}
+
+Return ONLY valid JSON in this exact shape, with the model field containing an exact model name from the table:
+{"model": "<exact model name>", "reason": "<short reason: category + difficulty tier + why this model>"}
+
+Task:
+{{ instance_text }}`;
export const DEFAULT_ROUTER_FLASH_META_PROFILE_DEFAULT: MetaProfile = {
classifier_model: "minimax-m3",