feat: add evidence-bound failure curation

plugins/openflywheel/.codex-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "openflywheel",
-  "version": "0.6.0",
-  "description": "Initialize ITSM-bench harness workspaces, query Langfuse trajectories, and mine compact failure diagnoses and exact patterns.",
+  "version": "0.7.0",
+  "description": "Initialize ITSM-bench workspaces, mine exact failure patterns, and record evidence-bound curations.",
   "author": {
     "name": "OpenFlyWheel"
   },
@@ -10,8 +10,8 @@
   "skills": "./skills/",
   "interface": {
     "displayName": "OpenFlyWheel",
-    "shortDescription": "Diagnose and mine ITSM failure patterns",
-    "longDescription": "Initialize an ITSM-bench agent-harness optimization workspace, inspect bounded Langfuse evidence, record authoritative outcomes and compact diagnoses, and mine exact recurring patterns without copying trace payloads.",
+    "shortDescription": "Mine and curate ITSM failure patterns",
+    "longDescription": "Initialize an ITSM-bench agent-harness optimization workspace, inspect bounded Langfuse evidence, record authoritative outcomes and compact diagnoses, mine exact recurring patterns, and persist evidence-bound cross-task curations without copying trace payloads.",
     "developerName": "OpenFlyWheel",
     "category": "Productivity",
     "capabilities": ["Read", "Write"],

plugins/openflywheel/.mcp.json

@@ -4,7 +4,7 @@
       "command": "uvx",
       "args": [
         "--from",
-        "git+https://github.com/divo12/OpenFlyWheel.git@ab0ef62cbe1e6cddf0bfd8ec61374d10120c61aa",
+        "git+https://github.com/divo12/OpenFlyWheel.git@726594a56c6683d1d32207a462fe524b202f4b8b",
         "--with",
         "mcp>=1.13,<2",
         "openflywheel-mcp"

plugins/openflywheel/program_templates/itsm.md

@@ -49,6 +49,14 @@ After recording the diagnoses for one bounded run, use `$failure-pattern-miner`
 type and exact normalized root cause; results are not semantic clusters. Keep inconclusive
 diagnoses separate, and reread each supporting diagnosis before forming a shared hypothesis.
 
+After every failed outcome has a retained artifact, use `$failure-curator` for one bounded
+cross-task debugger pass. Use only the current run's retained `record_failure` receipt IDs.
+Do not glob `.workspace/failures/`, which may contain earlier runs. Submit the complete set of at
+most 50 IDs to `record_failure_curation`, assigning each exactly once to a repeated
+evidence-bound group or a deferred entry. Retain the `.workspace/failure-curations/` artifact
+before forming a harness hypothesis. If the run has more than 50 failure artifacts, report the
+curation bound and stop before forming a harness hypothesis.
+
 An intermediate tool error is evidence, not an outcome failure, when the agent recovered
 and the verifier passed. A technically clean trajectory is still a failure when the ITSM
 verifier shows that the required environment state was not achieved.

plugins/openflywheel/skills/failure-curator/SKILL.md

@@ -0,0 +1,33 @@
+---
+name: failure-curator
+description: Group a completed bounded set of recorded failure diagnoses into evidence-bound cross-task patterns before forming a harness hypothesis. Use after failure-miner has recorded every verifier-backed failure; do not query traces, edit the harness, or evaluate candidates.
+---
+
+# Failure Curator
+
+Turn only the retained `record_failure` receipt IDs produced for the current completed baseline
+or candidate run into a compact cross-task debugger overview. Do not glob or enumerate
+`.workspace/failures/`; it may also contain artifacts from earlier runs. Langfuse remains the
+source of trace content; read only the current run's recorded diagnosis artifacts in this phase.
+
+## Curate
+
+1. Read every retained failure artifact for the run, up to the tool's 50-artifact bound. If
+   the run exceeds that bound, stop and report that complete curation is unavailable.
+2. Defer every inconclusive diagnosis with the specific evidence or recurrence gap. A
+   supported diagnosis may also be deferred when it has no matching failure from another task.
+3. Group supported diagnoses only when at least two distinct task IDs share the same causal
+   mechanism and top-level `issue_type`. Shared entities, words, tools, or symptoms alone do
+   not establish a pattern.
+4. Give each group a stable lowercase `pattern_key`, concise title, causal mechanism, general
+   prevention mechanism, and exactly one most relevant harness `target_component`. Keep the
+   prevention mechanism structural rather than task-specific; do not describe a file edit yet.
+5. Assign every source artifact exactly once, either to one group or one deferred entry.
+
+Call `record_failure_curation` once with the prepared worktree as `workspace_root`, the full
+source artifact ID set, the proposed groups, and deferred entries. Retain the returned
+`.workspace/failure-curations/<curation-id>.json` path for the hypothesis phase.
+
+Do not call trace tools, copy trace content, invent missing causal evidence, combine different
+failure types, form a repair hypothesis, edit the harness, run a benchmark, or promote a change
+while following this skill.

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "openflywheel"
-version = "0.6.0"
+version = "0.7.0"
 description = "A governed self-improving agent harness"
 requires-python = ">=3.11"
 dependencies = [

src/ofw/__init__.py

@@ -27,16 +27,23 @@
     WorkspaceFile,
 )
 from ofw.evaluation import (
+    DeferredFailure,
+    FailureCuration,
+    FailureCurationErrorCode,
+    FailureCurationFailure,
     FailureDiagnosis,
     FailureDiagnosisError,
     FailureErrorCode,
     FailureEvidenceStatus,
+    FailureGroup,
+    FailureGroupMember,
     FailurePatternMiningError,
     FailurePatternMiningErrorCode,
     FailurePatternMiningObservation,
     FailurePatternMiningStatus,
     FailurePatternOrdering,
     FailurePatternSummary,
+    FailureSource,
     FailureType,
     LangfuseOutcomeStore,
     MineFailurePatternsInput,
@@ -109,19 +116,26 @@ def editable(self, path: Path) -> EditableFile:
     "CollectionErrorCode",
     "CommandLoop",
     "CommandVerifier",
+    "DeferredFailure",
     "E2BSandbox",
     "EditableFile",
     "EvidenceReference",
+    "FailureCuration",
+    "FailureCurationErrorCode",
+    "FailureCurationFailure",
     "FailureDiagnosis",
     "FailureDiagnosisError",
     "FailureErrorCode",
     "FailureEvidenceStatus",
+    "FailureGroup",
+    "FailureGroupMember",
     "FailurePatternMiningError",
     "FailurePatternMiningErrorCode",
     "FailurePatternMiningObservation",
     "FailurePatternMiningStatus",
     "FailurePatternOrdering",
     "FailurePatternSummary",
+    "FailureSource",
     "FailureType",
     "GitCommit",
     "Harness",

src/ofw/evaluation/__init__.py

@@ -7,6 +7,15 @@
     FailureEvidenceStatus,
     FailureType,
 )
+from ofw.evaluation.failure_curation import (
+    DeferredFailure,
+    FailureCuration,
+    FailureCurationErrorCode,
+    FailureCurationFailure,
+    FailureGroup,
+    FailureGroupMember,
+    FailureSource,
+)
 from ofw.evaluation.failure_patterns import (
     FailurePatternMiningError,
     FailurePatternMiningErrorCode,
@@ -31,16 +40,23 @@
 )
 
 __all__ = [
+    "DeferredFailure",
+    "FailureCuration",
+    "FailureCurationErrorCode",
+    "FailureCurationFailure",
     "FailureDiagnosis",
     "FailureDiagnosisError",
     "FailureErrorCode",
     "FailureEvidenceStatus",
+    "FailureGroup",
+    "FailureGroupMember",
     "FailurePatternMiningError",
     "FailurePatternMiningErrorCode",
     "FailurePatternMiningObservation",
     "FailurePatternMiningStatus",
     "FailurePatternOrdering",
     "FailurePatternSummary",
+    "FailureSource",
     "FailureType",
     "LangfuseOutcomeStore",
     "MineFailurePatternsInput",