feat: configurable compaction model

CHANGELOG.md

@@ -6,6 +6,19 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+### Added
+
+- **Configurable compaction model.** `modelSystem.compaction` selects the model
+  used to summarize the conversation during context compaction (both automatic
+  compaction and `/compact`). It defaults to the main conversation model; set it
+  to a `ModelRef` object (`{ "backend": "...", "model": "..." }`) to compact with
+  a different (for example cheaper or longer-context) model. When the configured backend is not ready, compaction
+  inherits the main model and prints a warning in the compaction notice; if the
+  chosen model errors mid-compaction, it retries with the main model before
+  falling back to the deterministic trim, reporting the failure instead of
+  silently degrading. The configured model is shown in `/route status` and
+  `/route template`.
+
 ## [1.2.0] - 2026-07-08
 
 ### Added

README.md

@@ -853,7 +853,10 @@ token counts, and reported OpenRouter costs stay visible.
 `/route` lets you describe a model stack that blends local and frontier models:
 separate orchestrators for low/medium/high planning, coders for
 low/medium/high implementation, play and production review models, a security
-review model, and one selector model that chooses the route for a task.
+review model, a compaction model that summarizes the transcript when context is
+compacted, and one selector model that chooses the route for a task. `/route
+status` shows the configured stack (including the compaction model) and `/route
+template` prints the JSON shape to paste into `agent.config.json`.
 
 ```text
 /route template
@@ -879,6 +882,25 @@ For whole-goal decomposition, `/plan route <goal>` uses `modelSystem.planner`
 or a planner override to create `.small-harness/plan.json`; `/plan execute`
 then runs each ready node with the configured low/medium/high coder model.
 
+Context compaction (summarizing the conversation when the prompt budget fills,
+both automatically and via `/compact`) uses the main conversation model by
+default. Set `modelSystem.compaction` to a `ModelRef` object (a `{ "backend":
+"...", "model": "..." }` pair, the same shape as `planner`/`selector`) to
+summarize with a different model, for example a cheaper or longer-context one:
+
+```json
+{
+  "modelSystem": {
+    "compaction": { "backend": "openrouter", "model": "anthropic/claude-3.5-haiku" }
+  }
+}
+```
+
+Prompt-budget sizing still tracks the main model (that is the context being
+fit); only the summary call uses the compaction model. If the compaction
+backend is not ready (for example a missing API key), compaction falls back to
+the main model and prints a warning.
+
 ---
 
 ## Configuration

src/commands/context_cmds.rs

@@ -308,6 +308,9 @@ pub(super) async fn cmd_compact(args: &str, state: &mut AppState) -> Result<()>
         "  {GREEN}✓{RESET} {DIM}{}{RESET}",
         state.context_guard_notice.as_deref().unwrap_or("")
     );
+    if let Some(warning) = result.warning {
+        println!("  {YELLOW}!{RESET} {DIM}{warning}{RESET}");
+    }
     println!("  {DIM}session → {}{RESET}", state.session_path.display());
     Ok(())
 }

src/commands/route.rs

@@ -171,6 +171,7 @@ fn print_route_status(stack: &ModelSystemConfig) {
     }
     print_model_ref("planner", stack.planner.as_ref());
     print_model_ref("selector", stack.selector.as_ref());
+    print_model_ref("compaction", stack.compaction.as_ref());
     print_tier_set("orchestrator", &stack.orchestrators);
     print_tier_set("coder", &stack.coders);
     print_review_set("review", &stack.reviewers);
@@ -215,6 +216,11 @@ fn print_route_template() {
       "thinkingDepth": "deep",
       "notes": "Chooses the model route for a task."
     }},
+    "compaction": {{
+      "backend": "openrouter",
+      "model": "anthropic/claude-3.5-haiku",
+      "notes": "Summarizes the transcript when context is compacted. Omit to inherit the main model."
+    }},
     "orchestrators": {{
       "low": {{ "backend": "ollama", "model": "qwen2.5-coder:7b" }},
       "medium": {{ "backend": "openrouter", "model": "qwen/qwen-2.5-coder-32b-instruct" }},