Rubber Duck #4380

Activity

1. tanselmi-appliedsurety

   added

   triage

   on Aug 5, 2026
2. tanselmi-appliedsurety

   added the

   Bug

   issue type on Aug 5, 2026
3. copilot-cli-release-app

   added

   area:agentsSub-agents, fleet, autopilot, plan mode, background agents, and custom agentsSub-agents, fleet, autopilot, plan mode, background agents, and custom agents

   area:modelsModel selection, availability, switching, rate limits, and model-specific behaviorModel selection, availability, switching, rate limits, and model-specific behavior

   and removed

   triage

   on Aug 6, 2026
4. nick-thinks commented on Aug 6, 2026

   

   nick-thinks

   on Aug 6, 2026

   Last edited by nick-thinks

   More actions

   Hi @tanselmi-appliedsurety I agree, typically you should have received a complementary model.

   Are you able to provide a session process log of when this happened to you? That will help to make sure we have identified the exact problem 😄 You can find the session's log file by using /session info and the log file will look like /Users/nick/.copilot/logs/process-1786017957609-61781.log

   Just to double check, can you also confirm if you have overridden the rubber duck subagent model in your subagent settings? If you go to /subagents you are able to change the model mapping behavior for rubber duck from the default of "complementary (different family)". If it is set to "inherit" it will use your session model, and if you have set a specific model it will use that.

   It is also possible that the session agent itself decided to override the rubber duck model choice, which we can identify from the process log. We discourage this option in instructions but leave it open for some flexibility in-session.
5. tanselmi-appliedsurety commented on Aug 6, 2026

   

   tanselmi-appliedsurety

   on Aug 6, 2026

   Author

   More actions

   Thanks @nick-thinks — checked both.

   1. Rubber-duck model mapping

   •  /subagents :  rubber-duck  →  complementary — unavailable, disabled 
   • Overridden: No
   • Status: On
   I did not set this.  settings.json  has no  rubber-duck  model override (only explore/task/research/etc.). So complementary is the built-in default, but the UI treats it as unavailable for the current session model/catalog.

   2. Process log
      Attached a redacted extract from yesterday’s session ( v1.0.78 ) with multiple  rubber-duck  execute/complete events.
      Note: INFO logs show the duck runs, but do not emit complementary selection debug ( parent_model  /  available_models  /  selected_model ). Happy to capture a fresher log if you want a specific debug flag/level.

   That matches the original report: complementary path fails open to same-family/session model (e.g. terra-on-terra) instead of a different family.

   

   copilot-cli-4380-rubber-duck-evidence.REDACTED.txt

   👀React with 👀1Reacted by Nick McKenna
6. eggboy commented on Aug 11, 2026

   

   eggboy

   on Aug 11, 2026

   More actions

   Adding evidence from a case with the same symptom but what looks like a different mechanism — filed separately as #4432 to keep the two apart.

   In my case the parent session was claude-opus-5 and a complementary model was available (gpt-5.6-sol, same high price category). Rubber-duck still ran on claude-opus-4.8.

   Cause: the parent agent passed model explicitly in the task tool call. From events.jsonl:

   {"agent_type":"rubber-duck","name":"reval-review","model":"claude-opus-4.8","prompt":"..."}

   subagent.completed   agentName: "rubber-duck"   model: "claude-opus-4.8"
   

   Driving subagentStartPlan in runtime.node directly (parent claude-opus-5; claude-opus-5 / claude-opus-4.8 / gpt-5.6-sol all at high):

   /subagents setting	explicit model arg	resolved
   (none)	(none)	gpt-5.6-sol ✅
   (none)	claude-opus-4.8	claude-opus-4.8 ❌
   complementary	claude-opus-4.8	claude-opus-4.8 ❌
   gpt-5.6-sol	claude-opus-4.8	claude-opus-4.8 ❌

   So an explicit model argument outranks even a /subagents value the user deliberately set. Since that argument is emitted by the calling model rather than typed by the user, model discretion silently overrides user configuration — with no warning surfaced.

   @tanselmi-appliedsurety — worth checking your events.jsonl for a "model" key in the task tool arguments (tool.execution_start, toolName: "task"). If it is absent, yours is the separate complementary unavailable path your /subagents output points to, and the two need different fixes. If it is present, these are the same bug.

   Repro script and suggested fixes are in #4432.

   👀React with 👀1Reacted by Nick McKenna