opencode-swarm-plugin 0.37.0 → 0.39.1

package/dist/compaction-prompt-scoring.d.ts

@@ -0,0 +1,124 @@
+/**
+ * Compaction Prompt Quality Scoring - Pure Functions
+ *
+ * Evaluates the quality of continuation prompts generated after context compaction.
+ * **Problem**: Post-compaction coordinators often "wake up" confused, forget their role,
+ * and start editing files instead of checking worker status.
+ *
+ * **Solution**: Score prompts on 5 dimensions that predict coordinator success:
+ *
+ * 1. **Epic ID Specificity (0.20)**: Real IDs (`mjkw...`) not placeholders (`<epic-id>`, `bd-xxx`)
+ *    - Placeholders = coordinator can't check actual swarm status
+ *
+ * 2. **Actionability (0.20)**: Tool calls with real values (e.g., `swarm_status(epic_id='mjkw81rkq4c')`)
+ *    - Generic instructions like "check status" don't work
+ *
+ * 3. **Coordinator Identity (0.25)**: ASCII header + strong mandates (NEVER/ALWAYS)
+ *    - Visual + semantic cues reinforce role post-compaction
+ *
+ * 4. **Forbidden Tools Listed (0.15)**: Explicitly lists Edit, Write, swarmmail_reserve, git commit
+ *    - Naming forbidden tools reduces violations
+ *
+ * 5. **Post-Compaction Discipline (0.20)**: First suggested tool is swarm_status or inbox (not Edit)
+ *    - First tool sets the pattern - "check status" vs "dive into code"
+ *
+ * **Pure functions**: These can be tested without evalite. The evalite wrappers are in
+ * `evals/scorers/compaction-prompt-scorers.ts`.
+ *
+ * **Data source**: Captured from `captureCompactionEvent()` with `compaction_type: "prompt_generated"`.
+ * The payload includes the FULL prompt content (not truncated) for scoring.
+ *
+ * **Integration**: `compaction-prompt.eval.ts` uses these scorers to track prompt quality over time.
+ * Progressive gates enforce quality: bootstrap → stabilization → production.
+ *
+ * @module compaction-prompt-scoring
+ */
+/**
+ * Compaction prompt structure (from LLM generation)
+ */
+export interface CompactionPrompt {
+    content: string;
+}
+/**
+ * Scorer result type
+ */
+export interface ScorerResult {
+    score: number;
+    message: string;
+}
+/** Matches real epic/cell IDs (mjkw prefix + 7+ base36 chars) */
+export declare const REAL_EPIC_ID: RegExp;
+/** Matches common placeholder patterns */
+export declare const PLACEHOLDERS: RegExp[];
+/** Matches ASCII box-drawing characters (for headers) */
+export declare const ASCII_BOX: RegExp;
+/** Matches strong mandate language */
+export declare const STRONG_LANGUAGE: RegExp[];
+/**
+ * Score epic ID specificity
+ *
+ * Validates that epic IDs are REAL, not placeholders.
+ * Placeholders like <epic-id>, bd-xxx, <path> indicate
+ * the prompt generator failed to inject actual values.
+ *
+ * @returns 1.0 if real IDs, 0.0 if placeholders found
+ */
+export declare function scoreEpicIdSpecificity(prompt: CompactionPrompt): ScorerResult;
+/**
+ * Score actionability of tool calls
+ *
+ * Validates that the prompt includes SPECIFIC actionable tool calls.
+ * Generic instructions like "check status" are useless.
+ * Good: swarm_status(epic_id='mjkw81rkq4c', project_key='/path')
+ * Bad: "Check the status of workers"
+ *
+ * @returns 1.0 if actionable tool calls with real values, 0.0 otherwise
+ */
+export declare function scoreActionability(prompt: CompactionPrompt): ScorerResult;
+/**
+ * Score coordinator identity reinforcement
+ *
+ * Validates that the prompt has STRONG coordinator identity reinforcement.
+ * Post-compaction coordinators lose their identity without visual+semantic cues.
+ *
+ * Checks:
+ * 1. ASCII box header (visual anchor)
+ * 2. Strong language (NEVER/ALWAYS, not "should"/"consider")
+ *
+ * @returns 1.0 for ASCII header + strong mandates, 0.5 for header only, 0.0 otherwise
+ */
+export declare function scoreCoordinatorIdentity(prompt: CompactionPrompt): ScorerResult;
+/**
+ * Score forbidden tools listing
+ *
+ * Validates that the prompt LISTS forbidden tools by name.
+ * Coordinators must know exactly which tools to avoid.
+ *
+ * Required forbidden tools:
+ * 1. Edit
+ * 2. Write
+ * 3. swarmmail_reserve (only workers reserve)
+ * 4. git commit (workers commit)
+ *
+ * @returns ratio of forbidden tools mentioned (0.0 to 1.0)
+ */
+export declare function scoreForbiddenToolsPresent(prompt: CompactionPrompt): ScorerResult;
+/**
+ * Score post-compaction discipline (first tool correctness)
+ *
+ * Validates that the FIRST suggested tool is correct.
+ * Coordinators should check status FIRST, not edit files.
+ *
+ * Good first tools:
+ * - swarm_status
+ * - swarmmail_inbox
+ *
+ * Bad first tools:
+ * - Edit
+ * - Write
+ * - Read (should check status first)
+ *
+ * @returns 1.0 if first tool is swarm_status or inbox, 0.0 otherwise
+ */
+export declare function scorePostCompactionDiscipline(prompt: CompactionPrompt): ScorerResult;
+//# sourceMappingURL=compaction-prompt-scoring.d.ts.map

package/dist/compaction-prompt-scoring.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"compaction-prompt-scoring.d.ts","sourceRoot":"","sources":["../src/compaction-prompt-scoring.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AAEH;;GAEG;AACH,MAAM,WAAW,gBAAgB;IAChC,OAAO,EAAE,MAAM,CAAC;CAChB;AAED;;GAEG;AACH,MAAM,WAAW,YAAY;IAC5B,KAAK,EAAE,MAAM,CAAC;IACd,OAAO,EAAE,MAAM,CAAC;CAChB;AAID,iEAAiE;AACjE,eAAO,MAAM,YAAY,QAAqB,CAAC;AAE/C,0CAA0C;AAC1C,eAAO,MAAM,YAAY,UAKxB,CAAC;AAEF,yDAAyD;AACzD,eAAO,MAAM,SAAS,QAAiB,CAAC;AAExC,sCAAsC;AACtC,eAAO,MAAM,eAAe,UAAoD,CAAC;AAIjF;;;;;;;;GAQG;AACH,wBAAgB,sBAAsB,CAAC,MAAM,EAAE,gBAAgB,GAAG,YAAY,CAuB7E;AAED;;;;;;;;;GASG;AACH,wBAAgB,kBAAkB,CAAC,MAAM,EAAE,gBAAgB,GAAG,YAAY,CA+BzE;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,wBAAwB,CACvC,MAAM,EAAE,gBAAgB,GACtB,YAAY,CA6Bd;AAED;;;;;;;;;;;;;GAaG;AACH,wBAAgB,0BAA0B,CACzC,MAAM,EAAE,gBAAgB,GACtB,YAAY,CAiCd;AAED;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,6BAA6B,CAC5C,MAAM,EAAE,gBAAgB,GACtB,YAAY,CAiCd"}

package/dist/eval-capture.d.ts

@@ -70,7 +70,7 @@ export type PartialEvalRecord = Partial<EvalRecord> & {
     task: string;
 };
 /**
- * Coordinator Event - captures coordinator decisions, violations, and outcomes
+ * Coordinator Event - captures coordinator decisions, violations, outcomes, and compaction
  */
 export declare const CoordinatorEventSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
     session_id: z.ZodString;
@@ -108,6 +108,19 @@ export declare const CoordinatorEventSchema: z.ZodDiscriminatedUnion<[z.ZodObjec
         epic_complete: "epic_complete";
     }>;
     payload: z.ZodAny;
+}, z.core.$strip>, z.ZodObject<{
+    session_id: z.ZodString;
+    epic_id: z.ZodString;
+    timestamp: z.ZodString;
+    event_type: z.ZodLiteral<"COMPACTION">;
+    compaction_type: z.ZodEnum<{
+        detection_complete: "detection_complete";
+        prompt_generated: "prompt_generated";
+        context_injected: "context_injected";
+        resumption_started: "resumption_started";
+        tool_call_tracked: "tool_call_tracked";
+    }>;
+    payload: z.ZodAny;
 }, z.core.$strip>], "event_type">;
 export type CoordinatorEvent = z.infer<typeof CoordinatorEventSchema>;
 /**
@@ -154,6 +167,19 @@ export declare const CoordinatorSessionSchema: z.ZodObject<{
             epic_complete: "epic_complete";
         }>;
         payload: z.ZodAny;
+    }, z.core.$strip>, z.ZodObject<{
+        session_id: z.ZodString;
+        epic_id: z.ZodString;
+        timestamp: z.ZodString;
+        event_type: z.ZodLiteral<"COMPACTION">;
+        compaction_type: z.ZodEnum<{
+            detection_complete: "detection_complete";
+            prompt_generated: "prompt_generated";
+            context_injected: "context_injected";
+            resumption_started: "resumption_started";
+            tool_call_tracked: "tool_call_tracked";
+        }>;
+        payload: z.ZodAny;
     }, z.core.$strip>], "event_type">>;
 }, z.core.$strip>;
 export type CoordinatorSession = z.infer<typeof CoordinatorSessionSchema>;
@@ -294,6 +320,60 @@ export declare function ensureSessionDir(): void;
  * Appends the event as a JSONL line to ~/.config/swarm-tools/sessions/{session_id}.jsonl
  */
 export declare function captureCoordinatorEvent(event: CoordinatorEvent): void;
+/**
+ * Capture a compaction event to the session file
+ *
+ * Helper for capturing COMPACTION events with automatic timestamp generation.
+ * Tracks compaction hook lifecycle: detection → prompt generation → context injection → resumption.
+ *
+ * **Part of eval-driven development pipeline:** Compaction events are used by `compaction-prompt.eval.ts`
+ * to score prompt quality (ID specificity, actionability, coordinator identity).
+ *
+ * **Lifecycle stages:**
+ * - `detection_complete` - Compaction detected (confidence level, context type)
+ * - `prompt_generated` - Continuation prompt created (FULL content stored for eval)
+ * - `context_injected` - Prompt injected into OpenCode context
+ * - `resumption_started` - Coordinator resumed from checkpoint
+ * - `tool_call_tracked` - First tool called post-compaction (measures discipline)
+ *
+ * @param params - Compaction event parameters
+ * @param params.session_id - Coordinator session ID
+ * @param params.epic_id - Epic ID being coordinated
+ * @param params.compaction_type - Stage of compaction lifecycle
+ * @param params.payload - Event-specific data (full prompt content, detection results, etc.)
+ *
+ * @example
+ * // Capture detection complete
+ * captureCompactionEvent({
+ *   session_id: "session-123",
+ *   epic_id: "bd-456",
+ *   compaction_type: "detection_complete",
+ *   payload: {
+ *     confidence: "high",
+ *     context_type: "full",
+ *     epic_id: "bd-456",
+ *   },
+ * });
+ *
+ * @example
+ * // Capture prompt generated (with full content for eval)
+ * captureCompactionEvent({
+ *   session_id: "session-123",
+ *   epic_id: "bd-456",
+ *   compaction_type: "prompt_generated",
+ *   payload: {
+ *     prompt_length: 5000,
+ *     full_prompt: "You are a coordinator...", // Full prompt, not truncated - used for quality scoring
+ *     context_type: "full",
+ *   },
+ * });
+ */
+export declare function captureCompactionEvent(params: {
+    session_id: string;
+    epic_id: string;
+    compaction_type: "detection_complete" | "prompt_generated" | "context_injected" | "resumption_started" | "tool_call_tracked";
+    payload: any;
+}): void;
 /**
  * Read all events from a session file
  */

package/dist/eval-capture.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"eval-capture.d.ts","sourceRoot":"","sources":["../src/eval-capture.ts"],"names":[],"mappings":"AAmBA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAMxB;;GAEG;AACH,eAAO,MAAM,oBAAoB;;;;;;;;;;iBAmB/B,CAAC;AACH,MAAM,MAAM,cAAc,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,oBAAoB,CAAC,CAAC;AAElE;;GAEG;AACH,eAAO,MAAM,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA2D3B,CAAC;AACH,MAAM,MAAM,UAAU,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,gBAAgB,CAAC,CAAC;AAE1D;;GAEG;AACH,MAAM,MAAM,iBAAiB,GAAG,OAAO,CAAC,UAAU,CAAC,GAAG;IACpD,EAAE,EAAE,MAAM,CAAC;IACX,SAAS,EAAE,MAAM,CAAC;IAClB,IAAI,EAAE,MAAM,CAAC;CACd,CAAC;AAEF;;GAEG;AACH,eAAO,MAAM,sBAAsB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iCA2CjC,CAAC;AACH,MAAM,MAAM,gBAAgB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,sBAAsB,CAAC,CAAC;AAEtE;;GAEG;AACH,eAAO,MAAM,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAMnC,CAAC;AACH,MAAM,MAAM,kBAAkB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,wBAAwB,CAAC,CAAC;AAM1E;;GAEG;AACH,eAAO,MAAM,sBAAsB,8BAA8B,CAAC;AAElE;;GAEG;AACH,wBAAgB,eAAe,CAAC,WAAW,EAAE,MAAM,GAAG,MAAM,CAE3D;AAED;;GAEG;AACH,wBAAgB,iBAAiB,CAAC,WAAW,EAAE,MAAM,GAAG,IAAI,CAM3D;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAC9B,WAAW,EAAE,MAAM,EACnB,MAAM,EAAE,UAAU,GAAG,iBAAiB,GACrC,IAAI,CAKN;AAED;;GAEG;AACH,wBAAgB,eAAe,CAAC,WAAW,EAAE,MAAM,GAAG,UAAU,EAAE,CAajE;AAED;;GAEG;AACH,wBAAgB,kBAAkB,CAAC,WAAW,EAAE,MAAM,GAAG,iBAAiB,EAAE,CAU3E;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAC9B,WAAW,EAAE,MAAM,EACnB,EAAE,EAAE,MAAM,EACV,OAAO,EAAE,OAAO,CAAC,UAAU,CAAC,GAC3B,OAAO,CAgBT;AAWD;;;;;GAKG;AACH,wBAAgB,oBAAoB,CAAC,MAAM,EAAE;IAC3C,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,QAAQ,EAAE,YAAY,GAAG,eAAe,GAAG,YAAY,GAAG,MAAM,CAAC;IACjE,SAAS,EAAE,MAAM,CAAC;IAClB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,QAAQ,EAAE,KAAK,CAAC;QACd,KAAK,EAAE,MAAM,CAAC;QACd,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,KAAK,EAAE,MAAM,EAAE,CAAC;QAChB,YAAY,CAAC,EAAE,MAAM,EAAE,CAAC;QACxB,oBAAoB,CAAC,EAAE,MAAM,CAAC;KAC/B,CAAC,CAAC;CACJ,GAAG,iBAAiB,CAsBpB;AAED;;;;;GAKG;AACH,wBAAgB,qBAAqB,CAAC,MAAM,EAAE;IAC5C,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,YAAY,EAAE,MAAM,EAAE,CAAC;IACvB,WAAW,EAAE,MAAM,EAAE,CAAC;IACtB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,OAAO,CAAC;IACjB,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB,GAAG,IAAI,CAwBP;AAED;;;;;GAKG;AACH,wBAAgB,kBAAkB,CAAC,MAAM,EAAE;IACzC,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;CACrB,GAAG,UAAU,GAAG,IAAI,CAoDpB;AAED;;GAEG;AACH,wBAAgB,oBAAoB,CAAC,MAAM,EAAE;IAC3C,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,QAAQ,EAAE,OAAO,CAAC;IAClB,QAAQ,EAAE,OAAO,CAAC;IAClB,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB,GAAG,IAAI,CAMP;AAMD;;;;GAIG;AACH,wBAAgB,gBAAgB,CAAC,WAAW,EAAE,MAAM,GAAG,KAAK,CAAC;IAC3D,KAAK,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IAC1C,QAAQ,EAAE;QACR,WAAW,EAAE,MAAM,CAAC;QACpB,YAAY,EAAE,MAAM,CAAC;QACrB,aAAa,CAAC,EAAE,MAAM,EAAE,CAAC;QACzB,cAAc,CAAC,EAAE,OAAO,CAAC;KAC1B,CAAC;IACF,MAAM,EAAE,UAAU,CAAC;CACpB,CAAC,CAkBD;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAAC,WAAW,EAAE,MAAM,GAAG;IACrD,YAAY,EAAE,MAAM,CAAC;IACrB,eAAe,EAAE,MAAM,CAAC;IACxB,WAAW,EAAE,MAAM,CAAC;IACpB,WAAW,EAAE,MAAM,CAAC;IACpB,aAAa,EAAE,MAAM,CAAC;IACtB,gBAAgB,EAAE,MAAM,CAAC;IACzB,cAAc,EAAE,MAAM,CAAC;CACxB,CAuCA;AAMD;;GAEG;AACH,wBAAgB,aAAa,IAAI,MAAM,CAEtC;AAED;;GAEG;AACH,wBAAgB,cAAc,CAAC,SAAS,EAAE,MAAM,GAAG,MAAM,CAExD;AAED;;GAEG;AACH,wBAAgB,gBAAgB,IAAI,IAAI,CAKvC;AAED;;;;GAIG;AACH,wBAAgB,uBAAuB,CAAC,KAAK,EAAE,gBAAgB,GAAG,IAAI,CAWrE;AAED;;GAEG;AACH,wBAAgB,iBAAiB,CAAC,SAAS,EAAE,MAAM,GAAG,gBAAgB,EAAE,CAavE;AAED;;;;;GAKG;AACH,wBAAgB,WAAW,CAAC,MAAM,EAAE;IAClC,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;CACjB,GAAG,kBAAkB,GAAG,IAAI,CAoB5B"}
+{"version":3,"file":"eval-capture.d.ts","sourceRoot":"","sources":["../src/eval-capture.ts"],"names":[],"mappings":"AAyBA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAMxB;;GAEG;AACH,eAAO,MAAM,oBAAoB;;;;;;;;;;iBAmB/B,CAAC;AACH,MAAM,MAAM,cAAc,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,oBAAoB,CAAC,CAAC;AAElE;;GAEG;AACH,eAAO,MAAM,gBAAgB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA2D3B,CAAC;AACH,MAAM,MAAM,UAAU,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,gBAAgB,CAAC,CAAC;AAE1D;;GAEG;AACH,MAAM,MAAM,iBAAiB,GAAG,OAAO,CAAC,UAAU,CAAC,GAAG;IACpD,EAAE,EAAE,MAAM,CAAC;IACX,SAAS,EAAE,MAAM,CAAC;IAClB,IAAI,EAAE,MAAM,CAAC;CACd,CAAC;AAEF;;GAEG;AACH,eAAO,MAAM,sBAAsB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iCA0DjC,CAAC;AACH,MAAM,MAAM,gBAAgB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,sBAAsB,CAAC,CAAC;AAEtE;;GAEG;AACH,eAAO,MAAM,wBAAwB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAMnC,CAAC;AACH,MAAM,MAAM,kBAAkB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,wBAAwB,CAAC,CAAC;AAM1E;;GAEG;AACH,eAAO,MAAM,sBAAsB,8BAA8B,CAAC;AAElE;;GAEG;AACH,wBAAgB,eAAe,CAAC,WAAW,EAAE,MAAM,GAAG,MAAM,CAE3D;AAED;;GAEG;AACH,wBAAgB,iBAAiB,CAAC,WAAW,EAAE,MAAM,GAAG,IAAI,CAM3D;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAC9B,WAAW,EAAE,MAAM,EACnB,MAAM,EAAE,UAAU,GAAG,iBAAiB,GACrC,IAAI,CAKN;AAED;;GAEG;AACH,wBAAgB,eAAe,CAAC,WAAW,EAAE,MAAM,GAAG,UAAU,EAAE,CAajE;AAED;;GAEG;AACH,wBAAgB,kBAAkB,CAAC,WAAW,EAAE,MAAM,GAAG,iBAAiB,EAAE,CAU3E;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAC9B,WAAW,EAAE,MAAM,EACnB,EAAE,EAAE,MAAM,EACV,OAAO,EAAE,OAAO,CAAC,UAAU,CAAC,GAC3B,OAAO,CAgBT;AAWD;;;;;GAKG;AACH,wBAAgB,oBAAoB,CAAC,MAAM,EAAE;IAC3C,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,QAAQ,EAAE,YAAY,GAAG,eAAe,GAAG,YAAY,GAAG,MAAM,CAAC;IACjE,SAAS,EAAE,MAAM,CAAC;IAClB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,QAAQ,EAAE,KAAK,CAAC;QACd,KAAK,EAAE,MAAM,CAAC;QACd,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB,KAAK,EAAE,MAAM,EAAE,CAAC;QAChB,YAAY,CAAC,EAAE,MAAM,EAAE,CAAC;QACxB,oBAAoB,CAAC,EAAE,MAAM,CAAC;KAC/B,CAAC,CAAC;CACJ,GAAG,iBAAiB,CAsBpB;AAED;;;;;GAKG;AACH,wBAAgB,qBAAqB,CAAC,MAAM,EAAE;IAC5C,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,YAAY,EAAE,MAAM,EAAE,CAAC;IACvB,WAAW,EAAE,MAAM,EAAE,CAAC;IACtB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,OAAO,CAAC;IACjB,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB,GAAG,IAAI,CAwBP;AAED;;;;;GAKG;AACH,wBAAgB,kBAAkB,CAAC,MAAM,EAAE;IACzC,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;CACrB,GAAG,UAAU,GAAG,IAAI,CAoDpB;AAED;;GAEG;AACH,wBAAgB,oBAAoB,CAAC,MAAM,EAAE;IAC3C,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,QAAQ,EAAE,OAAO,CAAC;IAClB,QAAQ,EAAE,OAAO,CAAC;IAClB,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB,GAAG,IAAI,CAMP;AAMD;;;;GAIG;AACH,wBAAgB,gBAAgB,CAAC,WAAW,EAAE,MAAM,GAAG,KAAK,CAAC;IAC3D,KAAK,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IAC1C,QAAQ,EAAE;QACR,WAAW,EAAE,MAAM,CAAC;QACpB,YAAY,EAAE,MAAM,CAAC;QACrB,aAAa,CAAC,EAAE,MAAM,EAAE,CAAC;QACzB,cAAc,CAAC,EAAE,OAAO,CAAC;KAC1B,CAAC;IACF,MAAM,EAAE,UAAU,CAAC;CACpB,CAAC,CAkBD;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAAC,WAAW,EAAE,MAAM,GAAG;IACrD,YAAY,EAAE,MAAM,CAAC;IACrB,eAAe,EAAE,MAAM,CAAC;IACxB,WAAW,EAAE,MAAM,CAAC;IACpB,WAAW,EAAE,MAAM,CAAC;IACpB,aAAa,EAAE,MAAM,CAAC;IACtB,gBAAgB,EAAE,MAAM,CAAC;IACzB,cAAc,EAAE,MAAM,CAAC;CACxB,CAuCA;AAMD;;GAEG;AACH,wBAAgB,aAAa,IAAI,MAAM,CAEtC;AAED;;GAEG;AACH,wBAAgB,cAAc,CAAC,SAAS,EAAE,MAAM,GAAG,MAAM,CAExD;AAED;;GAEG;AACH,wBAAgB,gBAAgB,IAAI,IAAI,CAKvC;AAED;;;;GAIG;AACH,wBAAgB,uBAAuB,CAAC,KAAK,EAAE,gBAAgB,GAAG,IAAI,CAWrE;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+CG;AACH,wBAAgB,sBAAsB,CAAC,MAAM,EAAE;IAC7C,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;IAChB,eAAe,EACX,oBAAoB,GACpB,kBAAkB,GAClB,kBAAkB,GAClB,oBAAoB,GACpB,mBAAmB,CAAC;IACxB,OAAO,EAAE,GAAG,CAAC;CACd,GAAG,IAAI,CAWP;AAED;;GAEG;AACH,wBAAgB,iBAAiB,CAAC,SAAS,EAAE,MAAM,GAAG,gBAAgB,EAAE,CAavE;AAED;;;;;GAKG;AACH,wBAAgB,WAAW,CAAC,MAAM,EAAE;IAClC,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,MAAM,CAAC;CACjB,GAAG,kBAAkB,GAAG,IAAI,CAoB5B"}

package/dist/eval-gates.d.ts

@@ -0,0 +1,84 @@
+/**
+ * Result from a gate check
+ */
+export interface GateResult {
+    /** Whether the gate passed */
+    passed: boolean;
+    /** Current phase */
+    phase: "bootstrap" | "stabilization" | "production";
+    /** Human-readable message */
+    message: string;
+    /** Baseline score (mean of history) */
+    baseline?: number;
+    /** Current score */
+    currentScore: number;
+    /** Regression percentage (negative = improvement) */
+    regressionPercent?: number;
+}
+/**
+ * Configuration for gate thresholds
+ */
+export interface GateConfig {
+    /** Regression threshold for stabilization phase (default: 0.1 = 10%) */
+    stabilizationThreshold?: number;
+    /** Regression threshold for production phase (default: 0.05 = 5%) */
+    productionThreshold?: number;
+}
+/**
+ * Default regression thresholds by phase
+ */
+export declare const DEFAULT_THRESHOLDS: {
+    readonly stabilization: 0.1;
+    readonly production: 0.05;
+};
+/**
+ * Check if the current eval score passes the quality gate
+ *
+ * Progressive gates adapt based on data maturity:
+ * - **Bootstrap (<10 runs)**: Always pass, focus on collecting baseline data
+ * - **Stabilization (10-50 runs)**: Warn on >10% regression (default), but pass
+ * - **Production (>50 runs + variance <0.1)**: Fail on >5% regression (default)
+ *
+ * **Baseline calculation**: Mean of all historical scores for this eval (not just last run).
+ *
+ * **Regression formula**: `(baseline - current) / baseline`
+ * - Positive = regression (score dropped)
+ * - Negative = improvement
+ * - Returns 0 if baseline is 0 (avoids division by zero)
+ *
+ * **Variance threshold (0.1)**: High variance keeps eval in stabilization phase even with >50 runs.
+ * This prevents premature production gates when scores are still unstable.
+ *
+ * **CI Integration**: Production gates can fail PRs. Use `swarm eval status` to check phase before merging.
+ *
+ * @param projectPath - Absolute path to project root (contains `.opencode/eval-history.jsonl`)
+ * @param evalName - Name of the eval (e.g., "swarm-decomposition", "coordinator-behavior")
+ * @param currentScore - Current score to check (typically 0-1 range)
+ * @param config - Optional threshold configuration (defaults: stabilization=0.1, production=0.05)
+ * @returns Gate check result with pass/fail, phase, baseline, regression details
+ *
+ * @example
+ * ```typescript
+ * import { checkGate } from "./eval-gates.js";
+ *
+ * const result = checkGate("/path/to/project", "swarm-decomposition", 0.89);
+ *
+ * if (!result.passed) {
+ *   console.error(`❌ Gate FAILED: ${result.message}`);
+ *   process.exit(1); // Fail CI
+ * }
+ *
+ * console.log(`✅ ${result.phase} phase: ${result.message}`);
+ * ```
+ *
+ * @example
+ * ```typescript
+ * // Custom thresholds for sensitive eval
+ * const result = checkGate("/path", "critical-eval", 0.92, {
+ *   stabilizationThreshold: 0.05,  // 5% threshold in stabilization
+ *   productionThreshold: 0.02,     // 2% threshold in production
+ * });
+ * ```
+ */
+export declare function checkGate(projectPath: string, evalName: string, currentScore: number, config?: GateConfig): GateResult;
+//# sourceMappingURL=eval-gates.d.ts.map

package/dist/eval-gates.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"eval-gates.d.ts","sourceRoot":"","sources":["../src/eval-gates.ts"],"names":[],"mappings":"AAYA;;GAEG;AACH,MAAM,WAAW,UAAU;IAC1B,8BAA8B;IAC9B,MAAM,EAAE,OAAO,CAAC;IAChB,oBAAoB;IACpB,KAAK,EAAE,WAAW,GAAG,eAAe,GAAG,YAAY,CAAC;IACpD,6BAA6B;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,uCAAuC;IACvC,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,oBAAoB;IACpB,YAAY,EAAE,MAAM,CAAC;IACrB,qDAAqD;IACrD,iBAAiB,CAAC,EAAE,MAAM,CAAC;CAC3B;AAED;;GAEG;AACH,MAAM,WAAW,UAAU;IAC1B,wEAAwE;IACxE,sBAAsB,CAAC,EAAE,MAAM,CAAC;IAChC,qEAAqE;IACrE,mBAAmB,CAAC,EAAE,MAAM,CAAC;CAC7B;AAED;;GAEG;AACH,eAAO,MAAM,kBAAkB;;;CAGrB,CAAC;AAoCX;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgDG;AACH,wBAAgB,SAAS,CACxB,WAAW,EAAE,MAAM,EACnB,QAAQ,EAAE,MAAM,EAChB,YAAY,EAAE,MAAM,EACpB,MAAM,CAAC,EAAE,UAAU,GACjB,UAAU,CAiFZ"}

package/dist/eval-history.d.ts

@@ -0,0 +1,117 @@
+/**
+ * Progressive phases based on run count and variance
+ */
+export type Phase = "bootstrap" | "stabilization" | "production";
+/**
+ * Single eval run record
+ */
+export interface EvalRunRecord {
+    /** ISO-8601 timestamp */
+    timestamp: string;
+    /** Name of the eval (e.g., "swarm-decomposition") */
+    eval_name: string;
+    /** Score (0-1 range typically) */
+    score: number;
+    /** Run count (monotonically increasing per eval) */
+    run_count: number;
+}
+/**
+ * Default path for eval history
+ */
+export declare const DEFAULT_EVAL_HISTORY_PATH = ".opencode/eval-history.jsonl";
+/**
+ * Variance threshold for production phase
+ */
+export declare const VARIANCE_THRESHOLD = 0.1;
+/**
+ * Run count thresholds for phase transitions
+ */
+export declare const BOOTSTRAP_THRESHOLD = 10;
+export declare const STABILIZATION_THRESHOLD = 50;
+/**
+ * Get the eval history file path
+ */
+export declare function getEvalHistoryPath(projectPath: string): string;
+/**
+ * Ensure the eval history directory exists
+ */
+export declare function ensureEvalHistoryDir(projectPath: string): void;
+/**
+ * Record an eval run to JSONL history
+ *
+ * Appends atomically to `.opencode/eval-history.jsonl`. Each line is a complete JSON object
+ * representing one eval run (timestamp, eval name, score, run count).
+ *
+ * **Auto-creates directory** if `.opencode/` doesn't exist.
+ *
+ * **Thread-safe**: Uses `appendFileSync` for atomic writes (safe for concurrent eval runs).
+ *
+ * **Integration**: Called automatically by evalite runner after each eval completes.
+ * Also callable manually for custom eval tracking.
+ *
+ * @param projectPath - Absolute path to project root
+ * @param run - Eval run record with timestamp, eval_name, score, run_count
+ *
+ * @example
+ * ```typescript
+ * import { recordEvalRun } from "./eval-history.js";
+ *
+ * recordEvalRun("/path/to/project", {
+ *   timestamp: new Date().toISOString(),
+ *   eval_name: "swarm-decomposition",
+ *   score: 0.92,
+ *   run_count: 15,
+ * });
+ * ```
+ */
+export declare function recordEvalRun(projectPath: string, run: EvalRunRecord): void;
+/**
+ * Get score history for a specific eval
+ *
+ * Returns runs in chronological order (oldest first)
+ */
+export declare function getScoreHistory(projectPath: string, evalName: string): EvalRunRecord[];
+/**
+ * Calculate statistical variance of scores
+ *
+ * Variance = mean of squared deviations from the mean
+ * Formula: Σ((x - μ)²) / n
+ */
+export declare function calculateVariance(scores: number[]): number;
+/**
+ * Get the current phase for an eval based on run count and score variance
+ *
+ * Progressive phase logic ensures quality gates adapt to data maturity:
+ *
+ * - **Bootstrap (<10 runs)**: No gates, just collect baseline data
+ * - **Stabilization (10-50 runs)**: Warn on >10% regression (but pass)
+ * - **Production (>50 runs AND variance <0.1)**: Fail on >5% regression
+ *
+ * **Variance check**: If >50 runs but variance ≥0.1, stays in stabilization.
+ * This prevents premature production gates when scores are still unstable.
+ *
+ * **Why variance matters**: An eval with wildly fluctuating scores isn't ready for
+ * strict gates. Variance threshold (0.1) ensures the eval is consistent before
+ * enforcing production-level quality control.
+ *
+ * @param projectPath - Absolute path to project root (contains `.opencode/eval-history.jsonl`)
+ * @param evalName - Name of the eval (e.g., "swarm-decomposition")
+ * @returns Current phase: "bootstrap" | "stabilization" | "production"
+ *
+ * @example
+ * ```typescript
+ * import { getPhase } from "./eval-history.js";
+ *
+ * const phase = getPhase("/path/to/project", "swarm-decomposition");
+ *
+ * if (phase === "production") {
+ *   console.log("🚀 Production phase - strict gates enabled");
+ * } else if (phase === "stabilization") {
+ *   console.log("⚙️ Stabilization phase - warnings only");
+ * } else {
+ *   console.log("🌱 Bootstrap phase - collecting data");
+ * }
+ * ```
+ */
+export declare function getPhase(projectPath: string, evalName: string): Phase;
+//# sourceMappingURL=eval-history.d.ts.map

package/dist/eval-history.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"eval-history.d.ts","sourceRoot":"","sources":["../src/eval-history.ts"],"names":[],"mappings":"AAaA;;GAEG;AACH,MAAM,MAAM,KAAK,GAAG,WAAW,GAAG,eAAe,GAAG,YAAY,CAAC;AAEjE;;GAEG;AACH,MAAM,WAAW,aAAa;IAC5B,yBAAyB;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,qDAAqD;IACrD,SAAS,EAAE,MAAM,CAAC;IAClB,kCAAkC;IAClC,KAAK,EAAE,MAAM,CAAC;IACd,oDAAoD;IACpD,SAAS,EAAE,MAAM,CAAC;CACnB;AAED;;GAEG;AACH,eAAO,MAAM,yBAAyB,iCAAiC,CAAC;AAExE;;GAEG;AACH,eAAO,MAAM,kBAAkB,MAAM,CAAC;AAEtC;;GAEG;AACH,eAAO,MAAM,mBAAmB,KAAK,CAAC;AACtC,eAAO,MAAM,uBAAuB,KAAK,CAAC;AAE1C;;GAEG;AACH,wBAAgB,kBAAkB,CAAC,WAAW,EAAE,MAAM,GAAG,MAAM,CAE9D;AAED;;GAEG;AACH,wBAAgB,oBAAoB,CAAC,WAAW,EAAE,MAAM,GAAG,IAAI,CAM9D;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AACH,wBAAgB,aAAa,CAC3B,WAAW,EAAE,MAAM,EACnB,GAAG,EAAE,aAAa,GACjB,IAAI,CAKN;AAoBD;;;;GAIG;AACH,wBAAgB,eAAe,CAC7B,WAAW,EAAE,MAAM,EACnB,QAAQ,EAAE,MAAM,GACf,aAAa,EAAE,CAIjB;AAED;;;;;GAKG;AACH,wBAAgB,iBAAiB,CAAC,MAAM,EAAE,MAAM,EAAE,GAAG,MAAM,CAa1D;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,wBAAgB,QAAQ,CAAC,WAAW,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,GAAG,KAAK,CAqBrE"}

package/dist/eval-learning.d.ts

@@ -0,0 +1,216 @@
+/**
+ * Eval-to-Learning Feedback Loop
+ *
+ * Automatically stores eval failures to semantic memory for learning.
+ * When eval scores drop significantly from rolling average (default >15%),
+ * stores context to semantic-memory with tags for future prompt generation.
+ *
+ * ## Usage
+ *
+ * ```typescript
+ * import { learnFromEvalFailure } from "./eval-learning";
+ * import { getMemoryAdapter } from "./memory-tools";
+ * import { getScoreHistory } from "./eval-history";
+ *
+ * const memoryAdapter = await getMemoryAdapter();
+ * const history = getScoreHistory(projectPath, "compaction-test");
+ *
+ * const result = await learnFromEvalFailure(
+ *   "compaction-test",
+ *   currentScore,
+ *   history,
+ *   memoryAdapter
+ * );
+ *
+ * if (result.triggered) {
+ *   console.log(`📉 Regression detected: ${(result.drop_percentage * 100).toFixed(1)}% drop`);
+ *   console.log(`Memory ID: ${result.memory_id}`);
+ * }
+ * ```
+ *
+ * ## Integration Points
+ *
+ * - **After each eval run**: Call to detect regressions automatically
+ * - **Memory tags**: `eval-failure`, `{eval-name}`, `regression`
+ * - **Future prompts**: Query memories with these tags for context
+ * - **Scorer context**: Optional detail about which scorer failed
+ *
+ * ## Customization
+ *
+ * ```typescript
+ * const customConfig = {
+ *   dropThreshold: 0.10,  // 10% threshold (more sensitive)
+ *   windowSize: 10,        // Last 10 runs for baseline
+ * };
+ *
+ * await learnFromEvalFailure(
+ *   "test",
+ *   score,
+ *   history,
+ *   adapter,
+ *   { config: customConfig }
+ * );
+ * ```
+ *
+ * @module eval-learning
+ */
+import type { EvalRunRecord } from "./eval-history";
+import type { MemoryAdapter } from "./memory-tools";
+/**
+ * Configuration for eval-to-learning feedback
+ */
+export interface EvalLearningConfig {
+    /** Threshold for significant drop (0-1, default 0.15 = 15%) */
+    dropThreshold: number;
+    /** Rolling average window size (default 5 runs) */
+    windowSize: number;
+}
+/**
+ * Default configuration
+ */
+export declare const DEFAULT_EVAL_LEARNING_CONFIG: EvalLearningConfig;
+/**
+ * Result from learning check
+ */
+export interface LearningResult {
+    /** Whether the check triggered memory storage */
+    triggered: boolean;
+    /** Baseline score from rolling average */
+    baseline: number;
+    /** Current score */
+    current: number;
+    /** Drop percentage (0-1, e.g., 0.20 = 20% drop) */
+    drop_percentage: number;
+    /** Memory ID if stored */
+    memory_id?: string;
+}
+/**
+ * Calculate rolling average of recent scores
+ *
+ * Uses last N runs (default 5) to establish baseline.
+ * If history shorter than window, uses all available.
+ *
+ * @param history - Score history (chronological order)
+ * @param windowSize - Number of recent runs to average (default 5)
+ * @returns Average score (0 if empty)
+ */
+export declare function calculateRollingAverage(history: EvalRunRecord[], windowSize?: number): number;
+/**
+ * Check if current score is a significant drop from baseline
+ *
+ * Significant = drop exceeds threshold (default 15%).
+ * Formula: (baseline - current) / baseline >= threshold
+ *
+ * @param currentScore - Current eval score
+ * @param baseline - Baseline score (rolling average)
+ * @param threshold - Drop threshold (default 0.15 = 15%)
+ * @returns True if drop is significant
+ */
+export declare function isSignificantDrop(currentScore: number, baseline: number, threshold?: number): boolean;
+/**
+ * Format failure context for semantic memory storage
+ *
+ * Creates human-readable description of the failure with
+ * quantified metrics and optional scorer context.
+ *
+ * @param evalName - Name of eval that failed
+ * @param currentScore - Current score
+ * @param baseline - Baseline score
+ * @param scorerContext - Optional context about which scorer failed
+ * @returns Formatted context string
+ */
+export declare function formatFailureContext(evalName: string, currentScore: number, baseline: number, scorerContext?: string): string;
+/**
+ * Main learning function - automatically stores eval failures to semantic memory
+ *
+ * **Closed-loop learning**: When eval scores drop significantly from baseline,
+ * this function stores failure context to semantic memory. Future prompt generation
+ * queries these memories for context, preventing repeated mistakes.
+ *
+ * **Trigger condition**: Score drops >15% (default) from rolling average baseline.
+ * Uses last 5 runs (default) to establish baseline, not just previous run.
+ *
+ * **What gets stored**:
+ * - Eval name, baseline score, current score, drop percentage
+ * - Scorer-specific context (which scorer failed, why)
+ * - Timestamp and metadata for querying
+ * - Tags: `eval-failure`, `{eval-name}`, `regression`
+ *
+ * **Future use**: Before generating prompts for the same eval, query semantic memory
+ * with tags to inject learnings from past failures.
+ *
+ * **Integration points**:
+ * - After each eval run (in evalite runner or CI)
+ * - In `checkGate()` when regression detected
+ * - Manual calls for custom eval tracking
+ *
+ * @param evalName - Name of eval (e.g., "compaction-test", "coordinator-behavior")
+ * @param currentScore - Current eval score (typically 0-1 range)
+ * @param history - Score history in chronological order (oldest first)
+ * @param memoryAdapter - Semantic memory adapter (from `getMemoryAdapter()`)
+ * @param options - Optional config (thresholds, window size) and scorer context
+ * @param options.config - Custom thresholds (dropThreshold, windowSize)
+ * @param options.scorerContext - Details about which scorer failed (for context)
+ * @returns Learning result with trigger status, baseline, drop percentage, memory ID
+ *
+ * @example
+ * ```typescript
+ * import { learnFromEvalFailure } from "./eval-learning.js";
+ * import { getMemoryAdapter } from "./memory-tools.js";
+ * import { getScoreHistory } from "./eval-history.js";
+ *
+ * const memoryAdapter = await getMemoryAdapter();
+ * const history = getScoreHistory("/path/to/project", "coordinator-behavior");
+ *
+ * const result = await learnFromEvalFailure(
+ *   "coordinator-behavior",
+ *   0.68,  // Current score
+ *   history,
+ *   memoryAdapter,
+ *   { scorerContext: "violationCount: 5 violations (coordinator edited files)" }
+ * );
+ *
+ * if (result.triggered) {
+ *   console.log(`📉 Regression detected: ${(result.drop_percentage * 100).toFixed(1)}% drop`);
+ *   console.log(`Stored to memory: ${result.memory_id}`);
+ * }
+ * ```
+ *
+ * @example
+ * ```typescript
+ * // Custom threshold (more sensitive)
+ * const result = await learnFromEvalFailure(
+ *   "critical-eval",
+ *   0.85,
+ *   history,
+ *   memoryAdapter,
+ *   {
+ *     config: {
+ *       dropThreshold: 0.10,  // 10% threshold (default is 15%)
+ *       windowSize: 10,       // Last 10 runs for baseline (default is 5)
+ *     },
+ *   }
+ * );
+ * ```
+ */
+export declare function learnFromEvalFailure(evalName: string, currentScore: number, history: EvalRunRecord[], memoryAdapter: MemoryAdapter, options?: {
+    config?: EvalLearningConfig;
+    scorerContext?: string;
+}): Promise<LearningResult>;
+/**
+ * Create custom learning config with specific threshold
+ *
+ * Helper for common use case: custom drop threshold.
+ *
+ * @param dropThreshold - Drop threshold (0-1)
+ * @param windowSize - Optional window size (default 5)
+ * @returns Custom config
+ *
+ * @example
+ * ```typescript
+ * const config = createLearningConfig(0.10); // 10% threshold
+ * await learnFromEvalFailure("test", score, history, adapter, { config });
+ * ```
+ */
+export declare function createLearningConfig(dropThreshold: number, windowSize?: number): EvalLearningConfig;
+//# sourceMappingURL=eval-learning.d.ts.map