harness-evolver 4.0.1 → 4.0.3

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "harness-evolver",
   "description": "LangSmith-native autonomous agent optimization — evolves LLM agent code using multi-agent proposers, LangSmith experiments, and git worktrees",
-  "version": "4.0.1",
+  "version": "4.0.3",
   "author": {
     "name": "Raphael Valdetaro"
   },

package/README.md

@@ -95,8 +95,8 @@ claude
 <td>Three-gate iteration triggers (score plateau, cost budget, convergence detection) replace blind N-iteration loops. State validation ensures config hasn't diverged from LangSmith.</td>
 </tr>
 <tr>
-<td><b>Self-Scheduling</b></td>
-<td>Background and cron-based evolution modes for unattended optimization. Schedule nightly runs and get notified on improvements.</td>
+<td><b>Background Mode</b></td>
+<td>Run all iterations in background while you continue working. Get notified on completion or significant improvements.</td>
 </tr>
 </table>
 
@@ -137,7 +137,7 @@ claude
   +- 1.8  Analyze per-task failures (adaptive briefings)
   +- 1.8a Synthesize strategy document (coordinator synthesis)
   +- 1.9  Prepare shared proposer context (KV cache-optimized prefix)
-  +- 2.   Spawn 5 proposers in parallel (per-strategy tool restrictions)
+  +- 2.   Spawn 5 proposers in parallel (each in a git worktree)
   +- 3.   Run target for each candidate (code-based evaluators)
   +- 3.5  Spawn evaluator agent (LLM-as-judge via langsmith-cli)
   +- 4.   Compare experiments -> select winner + per-task champion
@@ -165,7 +165,7 @@ Skills (markdown)
   └── /evolver:deploy   → tags and pushes
 
 Agents (markdown)
-  ├── Proposer (x5)     → modifies code in worktrees (per-strategy tool restrictions)
+  ├── Proposer (x5)     → modifies code in isolated git worktrees
   ├── Evaluator          → LLM-as-judge via langsmith-cli
   ├── Critic             → detects gaming + implements stricter evaluators
   ├── Architect          → ULTRAPLAN deep analysis (opus model)

package/agents/evolver-proposer.md

@@ -148,14 +148,6 @@ Prioritize changes that fix real production failures over synthetic test failure
 4. **Commit your changes** — uncommitted changes are lost when the worktree is cleaned up
 5. **Write proposal.md** — the evolve skill reads this to understand what you did
 
-## Tool Restrictions
-
-Your available tools may be restricted based on your strategy:
-- **Exploit/Crossover/Failure-targeted**: Edit-only (no Write). Focus on modifying existing files.
-- **Explore**: Full access including Write. You may create new files if your approach requires it.
-
-If you need to create a file but only have Edit, restructure your approach to modify existing files instead.
-
 ## Return Protocol
 
 When done, end your response with:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "harness-evolver",
-  "version": "4.0.1",
+  "version": "4.0.3",
   "description": "LangSmith-native autonomous agent optimization for Claude Code",
   "author": "Raphael Valdetaro",
   "license": "MIT",

package/skills/evolve/SKILL.md

@@ -87,8 +87,7 @@ If iterations > 3, offer execution mode:
       "multiSelect": false,
       "options": [
         {"label": "Interactive", "description": "I'll watch. Show results after each iteration."},
-        {"label": "Background", "description": "Run all iterations in background. Notify on completion or significant improvement."},
-        {"label": "Scheduled", "description": "Schedule iterations to run on a cron (e.g., nightly optimization)."}
+        {"label": "Background", "description": "Run all iterations in background. Notify on completion or significant improvement."}
       ]
     }
   ]
@@ -98,35 +97,6 @@ If iterations > 3, offer execution mode:
 **If "Background" selected:**
 Run the evolution loop as a background task. Use the `run_in_background` parameter on the main loop execution.
 
-**If "Scheduled" selected:**
-Ask for schedule via AskUserQuestion:
-```json
-{
-  "questions": [
-    {
-      "question": "Schedule?",
-      "header": "Cron Schedule",
-      "multiSelect": false,
-      "options": [
-        {"label": "Every 6 hours", "description": "Run 1 iteration every 6 hours"},
-        {"label": "Nightly (2 AM)", "description": "Run iterations overnight"},
-        {"label": "Custom", "description": "Enter a cron expression"}
-      ]
-    }
-  ]
-}
-```
-
-Then create a cron trigger:
-```
-Use CronCreate tool to schedule:
-  - command: "/evolver:evolve --iterations 1 --no-interactive"
-  - schedule: {selected_cron}
-  - description: "Harness Evolver: scheduled optimization iteration"
-```
-
-Report: "Scheduled evolution iterations. Use `/evolver:status` to check progress. Cancel with CronDelete."
-
 ## The Loop
 
 Read config:
@@ -218,10 +188,11 @@ $EVOLVER_PY $TOOLS/synthesize_strategy.py \
     --trace-insights trace_insights.json \
     --best-results best_results.json \
     --evolution-memory evolution_memory.json \
+    --production-seed production_seed.json \
     --output strategy.md 2>/dev/null
 ```
 
-The `strategy.md` file is included in the proposer `<files_to_read>` block via the shared context (Step 1.9). This replaces raw data dumps with a synthesized, actionable document — proposers receive specific targets, not raw traces.
+The `strategy.md` file is included in the proposer `<files_to_read>` block via the shared context (Step 1.9). It synthesizes trace analysis, evolution memory, and production data into an actionable document. Proposers also receive `production_seed.json` directly for access to raw production traces.
 
 ### 1.9. Prepare Shared Proposer Context
 
@@ -233,6 +204,7 @@ SHARED_FILES_BLOCK="<files_to_read>
 - .evolver.json
 - strategy.md (if exists)
 - evolution_memory.md (if exists)
+- production_seed.json (if exists)
 - {entry_point_file}
 </files_to_read>"
 
@@ -311,33 +283,6 @@ APPROACH: {failure_targeted_or_efficiency}
 {adaptive_briefing_e}
 ```
 
-**Tool restrictions per strategy:**
-
-| Strategy | Allowed Tools | Rationale |
-|----------|--------------|-----------|
-| Exploit (A) | Read, Edit, Bash, Glob, Grep | No Write — can't create new files, only edit existing |
-| Explore (B) | Read, Write, Edit, Bash, Glob, Grep | Full access — may need new files for new architecture |
-| Crossover (C) | Read, Edit, Bash, Glob, Grep | No Write — combines existing patterns, doesn't create |
-| Failure-targeted (D, E) | Read, Edit, Bash, Glob, Grep | No Write — focused fixes on specific files |
-
-Apply via the `tools` parameter in each Agent() call. Example for exploit:
-```
-Agent(
-  subagent_type: "evolver-proposer",
-  tools: ["Read", "Edit", "Bash", "Glob", "Grep"],
-  ...
-)
-```
-
-For explore:
-```
-Agent(
-  subagent_type: "evolver-proposer",
-  tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep"],
-  ...
-)
-```
-
 Wait for all 5 to complete.
 
 **Stuck proposer detection**: If any proposer hasn't completed after 10 minutes, it may be stuck in a loop. The Claude Code runtime handles this via the agent's turn limit. If a proposer returns without committing changes, skip it — don't retry.

package/tools/synthesize_strategy.py

@@ -1,9 +1,9 @@
 #!/usr/bin/env python3
 """Synthesize evolution strategy document from trace analysis.
 
-Reads trace_insights.json, best_results.json, and evolution_memory.json
-to produce a targeted strategy document with specific file paths,
-line numbers, and concrete change recommendations for proposers.
+Reads trace_insights.json, best_results.json, evolution_memory.json,
+and production_seed.json to produce a targeted strategy document with
+specific file paths and concrete change recommendations for proposers.
 
 Usage:
     python3 synthesize_strategy.py \
@@ -11,6 +11,7 @@ Usage:
         --trace-insights trace_insights.json \
         --best-results best_results.json \
         --evolution-memory evolution_memory.json \
+        --production-seed production_seed.json \
         --output strategy.md
 """
 
@@ -42,7 +43,7 @@ def identify_target_files(config):
     return target_files
 
 
-def synthesize(config, insights, results, memory):
+def synthesize(config, insights, results, memory, production=None):
     """Produce strategy recommendations."""
     strategy = {
         "primary_targets": [],
@@ -94,6 +95,28 @@ def synthesize(config, insights, results, memory):
             for eid, data in failing[:10]
         ]
 
+    # Production trace data
+    if production:
+        prod_data = {}
+        stats = production.get("stats", {})
+        if stats:
+            prod_data["total_traces"] = stats.get("total_traces", 0)
+            prod_data["error_rate"] = stats.get("error_rate", 0)
+        categories = production.get("categories", [])
+        if categories:
+            prod_data["traffic_distribution"] = categories[:10]
+        neg = production.get("negative_feedback_inputs", [])
+        if neg:
+            prod_data["negative_feedback"] = neg[:5]
+        errors = production.get("error_patterns", production.get("errors", []))
+        if errors:
+            prod_data["production_errors"] = errors[:5] if isinstance(errors, list) else []
+        slow = production.get("slow_queries", [])
+        if slow:
+            prod_data["slow_queries"] = slow[:5]
+        if prod_data:
+            strategy["production"] = prod_data
+
     return strategy
 
 
@@ -142,6 +165,27 @@ def format_strategy_md(strategy, config):
             lines.append(f"- `{ex['example_id']}` (score: {score:.2f}): {preview}{error}")
         lines.append("")
 
+    prod = strategy.get("production", {})
+    if prod:
+        lines.append("## Production Insights")
+        if prod.get("total_traces"):
+            lines.append(f"- **Traces**: {prod['total_traces']} total, {prod.get('error_rate', 0):.1%} error rate")
+        if prod.get("traffic_distribution"):
+            lines.append(f"- **Traffic**: {', '.join(str(c) for c in prod['traffic_distribution'][:5])}")
+        if prod.get("negative_feedback"):
+            lines.append("- **Negative feedback inputs**:")
+            for nf in prod["negative_feedback"]:
+                lines.append(f"  - {str(nf)[:120]}")
+        if prod.get("production_errors"):
+            lines.append("- **Production errors**:")
+            for pe in prod["production_errors"]:
+                lines.append(f"  - {str(pe)[:120]}")
+        if prod.get("slow_queries"):
+            lines.append("- **Slow queries**:")
+            for sq in prod["slow_queries"]:
+                lines.append(f"  - {str(sq)[:120]}")
+        lines.append("")
+
     return "\n".join(lines)
 
 
@@ -151,6 +195,7 @@ def main():
     parser.add_argument("--trace-insights", default="trace_insights.json")
     parser.add_argument("--best-results", default="best_results.json")
     parser.add_argument("--evolution-memory", default="evolution_memory.json")
+    parser.add_argument("--production-seed", default="production_seed.json")
     parser.add_argument("--output", default="strategy.md")
     args = parser.parse_args()
 
@@ -160,8 +205,9 @@ def main():
     insights = load_json_safe(args.trace_insights)
     results = load_json_safe(args.best_results)
     memory = load_json_safe(args.evolution_memory)
+    production = load_json_safe(args.production_seed)
 
-    strategy = synthesize(config, insights, results, memory)
+    strategy = synthesize(config, insights, results, memory, production)
 
     md = format_strategy_md(strategy, config)
     with open(args.output, "w") as f: