[codex] add local failure mining workflow

AGENTS.md

@@ -71,6 +71,7 @@ uv run pytest --cov=ofw --cov-report=term-missing --cov-fail-under=90 -q
 uvx --from radon radon cc -s -a src tests plugins/openflywheel/scripts/mcp_server.py
 python3 ~/.codex/skills/.system/skill-creator/scripts/quick_validate.py plugins/openflywheel/skills/trace-query-planner
 python3 ~/.codex/skills/.system/skill-creator/scripts/quick_validate.py plugins/openflywheel/skills/outcome-recorder
+python3 ~/.codex/skills/.system/skill-creator/scripts/quick_validate.py plugins/openflywheel/skills/failure-miner
 python3 ~/.codex/skills/.system/plugin-creator/scripts/validate_plugin.py plugins/openflywheel
 ```
 

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

@@ -1,7 +1,7 @@
 {
   "name": "openflywheel",
-  "version": "0.4.0",
-  "description": "Initialize ITSM-bench harness workspaces, query Langfuse trajectories, and record authoritative verifier outcomes.",
+  "version": "0.5.0",
+  "description": "Initialize ITSM-bench harness workspaces, query Langfuse trajectories, and record outcomes and compact failure diagnoses.",
   "author": {
     "name": "OpenFlyWheel"
   },
@@ -10,8 +10,8 @@
   "skills": "./skills/",
   "interface": {
     "displayName": "OpenFlyWheel",
-    "shortDescription": "Initialize ITSM workspaces and inspect traces",
-    "longDescription": "Initialize an ITSM-bench agent-harness optimization workspace, select and inspect bounded Langfuse trajectory evidence, and record authoritative external-verifier outcomes on exact traces.",
+    "shortDescription": "Prepare ITSM workspaces and mine failures",
+    "longDescription": "Initialize an ITSM-bench agent-harness optimization workspace, inspect bounded Langfuse trajectory evidence, record authoritative outcomes, and persist compact failure diagnoses in the prepared local workspace.",
     "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@6319f24",
+        "git+https://github.com/divo12/OpenFlyWheel.git@0d4af6e0499c54765e2f7784b48b53baf96d1ad5",
         "--with",
         "mcp>=1.13,<2",
         "openflywheel-mcp"

plugins/openflywheel/program_templates/itsm.md

@@ -38,6 +38,12 @@ Use the trace tools in the smallest sufficient sequence:
    deterministic text filter.
 4. `get_span_context` retrieves bounded raw context only for a selected span.
 
+For every verifier-backed failed outcome, use `$failure-miner` to locate the earliest
+unrecovered causal observation and call `record_failure`. Retain the returned compact artifact
+under `.workspace/failures/` before forming a harness hypothesis. Record an inconclusive
+diagnosis when the evidence is insufficient. Do not copy Langfuse trace payloads into the
+workspace; Langfuse remains their source of truth.
+
 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-miner/SKILL.md

@@ -0,0 +1,59 @@
+---
+name: failure-miner
+description: Diagnose one authoritative failed agent outcome from bounded Langfuse evidence and record the supported or inconclusive result in the prepared harness workspace. Use after record_outcome returns a fail receipt; do not use for outcome judging, cross-trace pattern mining, or harness edits.
+---
+
+# Failure Miner
+
+Diagnose why one verifier-backed task failed. The verifier establishes the failure; the
+trajectory supplies causal evidence. Require the prepared workspace root, exact failed
+outcome fields, and its `outcome_score_id`. Never infer or replace the outcome.
+
+## Investigation
+
+Use at most 20 trace-tool calls:
+
+1. Ground the symptom from verifier evidence: state the expected and actual outcome.
+2. Call `get_trace_schema` to skim the trace before reading content.
+3. Call `query_spans` with the smallest useful selectors: entity or observation ID, tool,
+   span type, error flag, UTC range, or deterministic text filter.
+4. Call `get_span_context` only for plausible causal observations. Work backward from the
+   terminal mismatch through finalization, state verification, mutations, tool results, and
+   the evidence that drove those actions.
+5. Select the earliest unrecovered observation whose correction could plausibly change the
+   failed outcome. Earlier recovered errors remain evidence but are not the critical point.
+
+Stop searching once the causal claim is supported. Do not fetch every page or load a complete
+trace when bounded filters answer the question.
+
+## Classification
+
+For supported evidence, select exactly one type:
+
+- `intent_plan_failure`: the task or constraints were misunderstood or planned incorrectly.
+- `tool_interaction_failure`: tool choice, arguments, invocation, or recovery was defective.
+- `evidence_grounding_failure`: evidence was invented, stale, omitted, or misinterpreted.
+- `control_flow_failure`: execution looped, retried badly, lost state, or finalized early.
+- `policy_failure`: an instruction, permission, approval, or safety boundary was violated.
+
+Set `evidence_status=supported` only when the critical observation is among one to ten cited
+observation IDs and the trace supports both a root cause and the action that should have
+occurred there. A failed tool call alone is not causal if the agent recovered.
+
+If the evidence cannot distinguish a causal explanation, set `evidence_status=inconclusive`,
+leave `issue_type`, `critical_observation_id`, `root_cause`, and `counterfactual_action` empty,
+and state the missing evidence in `inconclusive_reason`. Never force a category.
+
+## Record
+
+Call `record_failure` exactly once with the prepared worktree as `workspace_root`, the exact
+failed outcome receipt, `evidence_status`, expected and actual outcomes, and cited observation
+IDs. A supported diagnosis must also pass `issue_type`, `critical_observation_id`,
+`root_cause`, and `counterfactual_action`; an inconclusive diagnosis passes only its
+`inconclusive_reason` for those conditional fields. Retain the returned
+`.workspace/failures/<artifact-id>.json` path. The artifact is compact diagnosis state;
+Langfuse remains the source of trace content.
+
+Do not call `record_outcome`, modify Langfuse, copy trace payloads locally, cluster failures,
+generate datasets, recommend a harness component, edit the harness, or propose a repair while
+following this skill. Those are later workflow phases.

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "openflywheel"
-version = "0.4.0"
+version = "0.5.0"
 description = "A governed self-improving agent harness"
 requires-python = ">=3.11"
 dependencies = [