opencode/picoclaw
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

feat: wire image_model config to agent loop for vision-capable models

Browse source 
- Add media parameter to selectCandidates function
- Check for media attachments and route to image_model when present

Fixes issue #1578
This commit is contained in:
AmChocolate 2026-03-15 21:57:56 +08:00
parent 32c7637a5e
commit 15772c4e6a
1 changed files with 16 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

17
pkg/agent/loop.go
View file

@ -1004,7 +1004,7 @@ func (al *AgentLoop) runLLMIteration(
// selectCandidates evaluates routing once and the decision is sticky for // selectCandidates evaluates routing once and the decision is sticky for
// all tool-follow-up iterations within the same turn so that a multi-step // all tool-follow-up iterations within the same turn so that a multi-step
// tool chain doesn't switch models mid-way through. // tool chain doesn't switch models mid-way through.
activeCandidates, activeModel := al.selectCandidates(agent, opts.UserMessage, messages) activeCandidates, activeModel := al.selectCandidates(agent, opts.UserMessage, messages, opts.Media)
for iteration < agent.MaxIterations { for iteration < agent.MaxIterations {
iteration++ iteration++
@ -1406,6 +1406,9 @@ func (al *AgentLoop) runLLMIteration(
// message scores below the complexity threshold, it returns the light model // message scores below the complexity threshold, it returns the light model
// candidates instead of the primary ones. // candidates instead of the primary ones.
// //
// If the message contains media attachments (images) and an image_model is
// configured, the image model candidates are returned instead.
//
// The returned (candidates, model) pair is used for all LLM calls within one // The returned (candidates, model) pair is used for all LLM calls within one
// turn — tool follow-up iterations use the same tier as the initial call so // turn — tool follow-up iterations use the same tier as the initial call so
// that a multi-step tool chain doesn't switch models mid-way. // that a multi-step tool chain doesn't switch models mid-way.
@ -1413,7 +1416,19 @@ func (al *AgentLoop) selectCandidates(
agent *AgentInstance, agent *AgentInstance,
userMsg string, userMsg string,
history []providers.Message, history []providers.Message,
media []string,
) (candidates []providers.FallbackCandidate, model string) { ) (candidates []providers.FallbackCandidate, model string) {
// If image_model is configured and there are media attachments, route to image model
if len(media) > 0 && agent.ImageModel != "" && len(agent.ImageModelCandidates) > 0 {
logger.InfoCF("agent", "Image model routing: media attachments detected",
map[string]any{
"agent_id": agent.ID,
"image_model": agent.ImageModel,
"media_count": len(media),
})
return agent.ImageModelCandidates, agent.ImageModel
}
if agent.Router == nil || len(agent.LightCandidates) == 0 { if agent.Router == nil || len(agent.LightCandidates) == 0 {
return agent.Candidates, agent.Model return agent.Candidates, agent.Model
} }