Sidecar acts on parent conversation context instead of its own briefing

[ellisjr]

Description

When a sidecar is launched WITH parent conversation context (the default, without --no-context), and the parent conversation involves discussing/launching sidecars, the sidecar model gets confused and acts on the conversation content rather than its briefing.

Steps to Reproduce

1. In a Claude Code session, discuss launching sidecars (e.g., "let's launch sidecars to review the PRD")
2. Launch a sidecar WITHOUT --no-context:

   sidecar start --model google/gemini-2.5-pro --agent Plan --no-ui \
     --prompt "Review _bmad-output/planning-artifacts/prd.md"

3. Read the sidecar's output via sidecar read <id> --summary

Expected Behavior

Sidecar reads the PRD file and produces a review per the briefing.

Actual Behavior

Sidecar reads the parent conversation context (which discusses relaunching sidecars), interprets it as its task, and outputs a "status update" about relaunching sidecars instead of doing the actual briefing:

Relaunching the 3 failed/stuck sidecars using direct APIs.
| # | Model | Status | Notes |
| 1 | Gemini Pro | Relaunching | Was stuck, now using google/gemini-pro |
...

Workaround

Use --no-context for self-contained tasks.

Suggestion

Consider giving the briefing stronger framing in the system prompt (e.g., "Your primary task is the briefing below. The conversation history is background context only.") or filtering out sidecar-related conversation turns from the context.