Before submitting
Area
apps/desktop
Steps to reproduce
- Run a session whose main-loop model is
claude-fable-5[1m], served through a local proxy (ANTHROPIC_BASE_URL=http://localhost:8317).
- Define custom agents in
~/.claude/agents/ with frontmatter model pins to OpenAI models served through the same proxy: Explore → model: gpt-5.6-sol-fast(medium), codex-sol → model: gpt-5.6-sol(high).
- Have the main loop spawn
codex-sol (depth 1).
- Have that agent call the Agent tool itself to spawn several
Explore agents (depth 2).
- Open the Agents panel and compare the model column against the per-response
model field in the agent records.
Expected behavior
Model column reflects the resolved frontmatter model (or the per-response model field) at any spawn depth — gpt-5.6-sol for all five agents.
Actual behavior
- Depth-1 spawns: correct —
gpt-5.6-sol · high.
- Depth-2 spawns: labeled
fable-5 · high (the session default), even though every response in those agents was served by gpt-5.6-sol — verifiable from the per-response model field and the OpenAI resp_… response-id format (genuine Fable responses carry claude-fable-5 with Anthropic msg_… ids, so the two are unambiguous).
- Depth-2 spawn metadata does record the correct
agentType: "Explore", parentAgentId, and spawnDepth: 2; only the model label falls back to the inherited session model at that hop.
Impact
Major degradation or frequent failure — users auditing model usage see false positives (the panel attributed ~1.6M tokens to fable-5 that were actually served by gpt-5.6-sol). If the same spawn-record label feeds usage/quota attribution, nested-agent traffic may be counted against the wrong provider/model bucket (here: Anthropic instead of OpenAI). Please confirm whether attribution uses the panel label or the per-response model field.
Environment
macOS 26 (Darwin 25.5.0), t3code desktop app, local CLIProxyAPI/VibeProxy on port 8317
Before submitting
Area
apps/desktop
Steps to reproduce
claude-fable-5[1m], served through a local proxy (ANTHROPIC_BASE_URL=http://localhost:8317).~/.claude/agents/with frontmatter model pins to OpenAI models served through the same proxy:Explore→model: gpt-5.6-sol-fast(medium),codex-sol→model: gpt-5.6-sol(high).codex-sol(depth 1).Exploreagents (depth 2).modelfield in the agent records.Expected behavior
Model column reflects the resolved frontmatter model (or the per-response
modelfield) at any spawn depth —gpt-5.6-solfor all five agents.Actual behavior
gpt-5.6-sol · high.fable-5 · high(the session default), even though every response in those agents was served bygpt-5.6-sol— verifiable from the per-responsemodelfield and the OpenAIresp_…response-id format (genuine Fable responses carryclaude-fable-5with Anthropicmsg_…ids, so the two are unambiguous).agentType: "Explore",parentAgentId, andspawnDepth: 2; only the model label falls back to the inherited session model at that hop.Impact
Major degradation or frequent failure — users auditing model usage see false positives (the panel attributed ~1.6M tokens to
fable-5that were actually served bygpt-5.6-sol). If the same spawn-record label feeds usage/quota attribution, nested-agent traffic may be counted against the wrong provider/model bucket (here: Anthropic instead of OpenAI). Please confirm whether attribution uses the panel label or the per-response model field.Environment
macOS 26 (Darwin 25.5.0), t3code desktop app, local CLIProxyAPI/VibeProxy on port 8317