auto-feedback（自动反馈）

Feedback Loop (v2)

Overview

The feedback-loop skill is a dual-mode, self-improving intelligence layer. It runs completely on its own with no external dependencies, and automatically unlocks richer analysis when intent-engineering or dark-factory are present.

Use this skill when:

• You want to analyze the performance of any process, script, or agent execution.
• You need prioritized improvement suggestions without running a full pipeline.
• You want to auto-generate regression tests from observed failures or edge cases.
• You need to track whether a system's behavior is drifting from its stated goals.
• You want a self-contained improvement report you can act on immediately.
• You are running the full intent-engineering → dark-factory → feedback-loop triad.

Dual-Mode Operation

The skill detects what inputs are available and automatically selects the richest mode:

There is no configuration switch — the skill detects what is available and adapts automatically. You never need to change anything to switch modes.

Architecture

┌──────────────────────────────────────────────────────────────────┐
│                     FEEDBACK LOOP (v2)                           │
│                                                                  │
│  ┌─────────────────────────────────────────────────────────────┐ │
│  │  INPUT LAYER (auto-detects mode)                            │ │
│  │                                                             │ │
│  │  Standalone:  any JSON log / plain text / prior obs        │ │
│  │  Enhanced:    + outcome_report.json (dark-factory)         │ │
│  │  Full Triad:  + specification.json (intent-engineering)    │ │
│  └───────────────────────┬─────────────────────────────────────┘ │
│                          │                                       │
│                          ▼                                       │
│  ┌─────────────────────────────────────────────────────────────┐ │
│  │  OBSERVER  (observer.py)                                    │ │
│  │  Normalizes all inputs → observation.json                  │ │
│  └───────────────────────┬─────────────────────────────────────┘ │
│                          │                                       │
│                          ▼                                       │
│  ┌─────────────────────────────────────────────────────────────┐ │
│  │  ANALYZER  (analyzer.py)                                    │ │
│  │  Scores performance · detects regressions                  │ │
│  │  Generates suggestions · checks alignment                  │ │
│  │  Auto-creates regression tests                             │ │
│  └───────────────────────┬─────────────────────────────────────┘ │
│                          │                                       │
│                          ▼                                       │
│  ┌─────────────────────────────────────────────────────────────┐ │
│  │  ORCHESTRATOR  (orchestrator.py)                            │ │
│  │  Assembles signed improvement_report.json                  │ │
│  │  Produces updated_observation.json for next cycle          │ │
│  │  Optionally produces updated_specification.json            │ │
│  └─────────────────────────────────────────────────────────────┘ │
└──────────────────────────────────────────────────────────────────┘

Quick Start

Standalone — any JSON log

python scripts/orchestrator.py --input my_execution_log.json --goal "Process tickets in under 2 minutes"

Standalone — plain text description

python scripts/orchestrator.py --text "Script ran but missed 3 null input cases and took 4 minutes" --goal "Handle all inputs"

Standalone — continuing a prior cycle

python scripts/orchestrator.py --observation observation.json

Dark Factory Enhanced

python scripts/orchestrator.py --outcome outcome_report.json --goal "Achieve 98% test pass rate"

Full Triad (after intent-engineering + dark-factory)

python scripts/orchestrator.py --outcome outcome_report.json --spec specification.json

Run stages independently

python scripts/observer.py --input log.json --goal "..." --output observation.json
python scripts/analyzer.py --observation observation.json --output analysis.json
python scripts/orchestrator.py --analysis analysis.json --output-dir ./reports/

Outputs

Every run produces files in the current directory (or --output-dir):

The Six-Step Internal Workflow

Step 1 — Normalize Input

observer.py accepts any input format and normalizes it into a standard observation.json. In standalone mode it extracts metrics from JSON or text. In enhanced/triad mode it also ingests the structured fields from outcome_report.json and specification.json.

Step 2 — Score Performance

analyzer.py computes a performance_score (0.0–1.0) from extracted metrics. It compares against the previous cycle's score (if available) to detect regressions and trends.

Step 3 — Generate Improvement Suggestions

The analyzer produces concrete, actionable suggestions ranked critical → high → medium → low. Each suggestion includes a description, rationale, effort estimate, and expected impact. In triad mode, suggestions are also cross-referenced against the original specification's success criteria.

Step 4 — Auto-Generate Regression Tests

Every failure and edge case is automatically converted into a regression test with a concrete input and expected_output. These are appended to updated_observation.json and (in triad mode) to updated_specification.json.

Step 5 — Check Goal Alignment

The analyzer checks the observation against references/alignment_values.json (your organization's principles) and, in triad mode, against the original specification's stated goal. It produces an alignment_score (0.0–1.0) and flags any drift.

Step 6 — Generate Signed Report

orchestrator.py assembles all outputs into a single improvement_report.json with a SHA-256 integrity digest, making every report independently verifiable.

Self-Improving Loop

The skill is designed to be run repeatedly. Each run produces an updated_observation.json that serves as the input for the next run. Over time, the regression test suite grows, the alignment score stabilizes, and the improvement suggestions become more targeted.

Cycle 1:  any input → observation → analysis → improvement_report_1.json + updated_observation.json
Cycle 2:  updated_observation.json → analysis → improvement_report_2.json + updated_observation.json
Cycle N:  ...

In full triad mode, the updated_specification.json feeds back into intent-engineering to close the loop across all three skills.

Configuration

All configuration lives inside the skill — no external files required.

references/alignment_values.json — Edit to define your organization's goals and values. The analyzer checks every observation against these values to produce the alignment score.

references/scoring_weights.json — Edit to change how the performance score is calculated (e.g. weight pass rate more heavily than speed).

references/suggestion_rules.json — Edit to add custom rules for generating improvement suggestions.

Resources

feedback-loop/
├── SKILL.md                              ← this file
├── scripts/
│   ├── observer.py                       ← normalizes any input → observation.json
│   ├── analyzer.py                       ← scores, detects regressions, generates suggestions
│   └── orchestrator.py                   ← runs all stages, produces signed report
├── references/
│   ├── alignment_values.json             ← org goals and values (edit this)
│   ├── scoring_weights.json              ← performance score weights
│   ├── suggestion_rules.json             ← rules for improvement suggestions
│   └── operations_guide.md              ← detailed ops and troubleshooting guide
└── templates/
    ├── improvement_report_template.md    ← human-readable report template
    └── observation_template.json         ← blank observation template