| name | kg-feedback |
| skill_type | skill |
| description | Records a human/agent correction so the system learns durably — reward adjustments, governance rules, eval cases, reads-avoided and action-outcome loops, and pinned gotchas. Use to teach or correct the KG — "record this correction", "pin a gotcha", "log the outcome of this action". |
| license | MIT |
| tags | ["graph-os","feedback","learning"] |
| tier | core |
| metadata | {"author":"Genius","version":"0.1.0"} |
kg-feedback
Condensed intent-surface note (Seam 8). Under the default intent surface (MCP_TOOL_MODE=intent), graph_feedback is held back from the default tool list (nothing removed — REST + _execute_tool still reach it exactly as documented below). Two ways to use this skill unchanged: (1) load_tools(tools=["graph_feedback"]) once per session (as below), then proceed exactly as documented; or (2) call the why intent verb with the same natural-language request — the resolver routes to graph_feedback for you and returns the result plus a routing justification. Set MCP_TOOL_MODE=condensed/verbose/both to expose the granular tools eagerly instead.
graph_feedback persists corrections keyed by correction_type: outcome (adjust an entity's reward), rule (durable governance/voice/source rule consulted at retrieval), eval (add a regression case), reads_avoided (close the code_context reads-avoided loop, AU-AHE.evaluation.reads-avoided-feedback), action_outcome (close the loop on any autonomous action so routing prefers what works, AU-AHE.evaluation.action-outcome-feedback), gotcha (pin a hard-won trap to a file/module so kg-code surfaces it).
Invoke
- MCP:
load_tools(tools=["graph_feedback"]), then graph_feedback(correction_type="gotcha", target_id="path/to/file.py", corrected_value="...").
- REST twin:
POST /graph/feedback with {"correction_type": "action_outcome", "target_id": "...", "corrected_value": "{...}"}.
Example
graph_feedback(correction_type="action_outcome", target_id="model_route:chat",
corrected_value='{"success":true,"reward":1.0}')
