Bug Description

The agents.defaults.compaction.model configuration is completely ignored when a session is bound to a cliBackends backend (e.g., claude-cli). The CLI backend performs compaction using the session's primary model instead of the configured compaction model.

Expected Behavior

When compaction.model is set (e.g., to a smaller, cheaper model), ALL compaction operations should use that model — including compactions triggered for sessions bound to CLI backends.

Actual Behavior

• compaction.model is only read by compactEmbeddedPiSessionDirect (internal compaction path)
• Sessions bound to cliBackends (e.g., claude-cli) delegate compaction to the CLI backend, which does NOT read compaction.model
• The CLI backend always uses the session's primary model for compaction

Reproduction

1. Configure two local LLM servers: a large model (e.g., 122B) and a small model (e.g., 9B)
2. Set agents.defaults.compaction.model to the small model
3. Configure agents.defaults.cliBackends.claude-cli
4. Use Discord or other channel that creates sessions bound to claude-cli
5. Fill the session until auto-compaction triggers
6. Observe: the large model handles compaction, the small model receives no requests

Impact

• Users cannot offload compaction to a cheaper/lighter model when using CLI backends
• On GPU-constrained setups (e.g., running large local models), this causes unnecessary GPU memory pressure from the main model performing compaction, potentially leading to system instability

Environment

• OpenClaw version: 2026.3.28
• OS: macOS (Apple Silicon)
• CLI backend: Claude Code

Relevant Code Paths

• compactEmbeddedPiSessionDirect in auth-profiles-*.js: Reads compaction.model ✓
• delegateCompactionToRuntime in delegate-*.js: Routes to compactEmbeddedPiSessionDirect ✓
• CLI backend compaction: Does NOT call compactEmbeddedPiSessionDirect, uses own model ✗

Suggested Fix

Either:

1. Pass compaction.model config to the CLI backend process so it can use the specified model for compaction
2. Or intercept CLI-backend compaction and route it through compactEmbeddedPiSessionDirect which already handles the model override