@cognipeer/agent-sdk 0.5.4 → 0.6.1

package/README.md

@@ -137,12 +137,8 @@ const agent = createSmartAgent({
   summarization: { summaryTriggerTokens: 8000, summaryMode: "incremental" },
   context: { policy: "hybrid", lastTurnsToKeep: 8 },
   toolResponses: {
-    smallResponseChars: 1200,
-    smallResponsePolicy: "keep_full",
-    defaultPolicy: "keep_structured",
-    largeResponsePolicy: "summarize_archive",
-    fallbackPolicy: "keep_full",
-    keepRecentFullCount: 2,
+    defaultPolicy: "summarize_archive",
+    toolResponseRetentionByTool: { read_skills: "keep_full" },
     maxToolResponseChars: 4000,
     maxToolResponseTokens: 1200,
   },
@@ -158,15 +154,13 @@ const result = await agent.invoke({
 console.log(result.content);
 ```
 
-Tool-response retention is size-aware and composable:
+Tool-response retention is lazy and summarizer-driven:
 
-- `smallResponseChars` + `smallResponsePolicy` define how very small payloads are kept.
-- `defaultPolicy` handles normal informative outputs.
-- `largeResponsePolicy` handles verbose outputs that cross the max char/token thresholds.
-- `fallbackPolicy` applies to critical tools and now defaults to `keep_full` so safety-critical payloads are not collapsed accidentally.
-- `keepRecentFullCount` pins the last N successful tool responses at full content during context assembly and excludes them from summarization passes.
-
-Resolution order is: per-tool override, then critical fallback, then large/small classification, then the default policy. Redundant payloads like `ok`, `true`, and `null` are dropped automatically.
+- Tool outputs are stored at full fidelity in `state.toolHistory` and are never reduced at tool-call time.
+- When the summarizer runs (context limits reached), old tool messages are rewritten in place according to `defaultPolicy` (default: `summarize_archive`). The full payload is still recoverable via `get_tool_response` using the execution id embedded in the placeholder.
+- `toolResponseRetentionByTool` lets you opt specific tools out of reduction (e.g. `read_skills: "keep_full"`).
+- `criticalTools` are never reduced. The built-in list covers `response`, `manage_todo_list`, and `get_tool_response`.
+- `maxToolResponseChars` / `maxToolResponseTokens` only drive an eager hard-cap truncation when a single tool output is big enough to blow up the very next model call. The truncated head always points at `get_tool_response` for recovery.
 
 The smart wrapper now supports runtime presets (`fast`, `balanced`, `deep`, `research`), custom profiles layered on top of a base preset, structured summarization, hybrid context compaction, configurable tool-response retention, in-memory fact storage, delegation limits, and an eval harness via `runSmartAgentEvalHarness(...)`.