agentic-loop-triad（动能环三联征）

Unified Orchestrator v2

A meta-skill that unifies intent-engineering, dark-factory, and feedback-loop into a single self-evolving system — and unlocks five paradigm shifts that none of the three skills can achieve alone.

When to Use This Skill

Use this skill whenever you need to take a goal from raw language all the way to a verified, continuously improving implementation — in one command. It replaces the manual handoff between intent-engineering, dark-factory, and feedback-loop with a single pipeline that:

• Translates goals into machine-executable specifications
• Executes those specifications autonomously
• Measures outcomes against the original intent
• Generates regression tests from every failure
• Feeds improvements back into the next cycle automatically

Use individual skills when you need surgical control over a single stage. Use this skill when you want the full loop.

The Five Paradigm Shifts

These are capabilities that emerge only when the three skills operate as a unified system. They are not features — they are new modes of operation.

Paradigm Shift 1 — Specification Drift Detection

What it is: The system continuously compares what was *intended* (the specification produced by intent-engineering) against what was *executed* (the outcome report from dark-factory) and what was *observed* (the feedback loop analysis). Over multiple cycles, it builds a drift map: a quantified record of how far execution has wandered from original intent.

Why it matters: In isolated skill usage, drift is invisible. A specification can be perfectly written, execution can pass all tests, and the feedback loop can report healthy scores — while the system is quietly drifting away from the original goal. Only by holding all three artifacts simultaneously can the orchestrator detect semantic drift: the gap between what the goal *said* and what the system *does*.

How it works: signal_router.py computes a drift_score (0.0–1.0) by comparing the specification's success_criteria against the feedback loop's alignment_score and the dark-factory's pass_rate. A drift score above 0.15 triggers a re-specification cycle automatically.

Paradigm Shift 2 — Capability Expansion via Meta-Learning

What it is: After each full pipeline cycle, meta_learner.py analyzes the complete history of specifications, outcomes, and improvement reports to identify *patterns* — recurring failure modes, consistently high-performing specification patterns, and improvement suggestions that were applied and worked. It then updates the skill registry and the shared intent framework with these learnings, making every future run smarter.

Why it matters: Individual skills improve within their own domain. The feedback loop makes the feedback loop better. Intent-engineering makes specifications better. But neither knows what the other learned. The meta-learner operates across all three domains simultaneously, compounding learnings: a failure pattern discovered in dark-factory execution informs how intent-engineering writes specifications, which changes what dark-factory executes, which changes what the feedback loop measures.

How it works: meta_learner.py maintains a learnings_log.json that accumulates cross-cycle patterns. After 3+ cycles on the same goal, it generates specification_patches.json — suggested improvements to the specification template itself — and rule_patches.json — new suggestion rules for the feedback loop analyzer.

Paradigm Shift 3 — Autonomous Re-specification

What it is: When the system detects that a specification is no longer achievable given observed constraints (e.g. a success criterion of duration_ms < 500 that has never been met in 5 cycles), it autonomously generates a revised specification with adjusted criteria, documents the revision rationale, and continues the pipeline — without human intervention.

Why it matters: Current AI systems fail silently or loop forever on impossible goals. This system detects impossibility, documents it, adjusts, and continues. The human gets a clear audit trail of what was attempted, what was impossible, and what was substituted — rather than a system that either crashes or silently lowers its own bar.

How it works: pipeline.py tracks consecutive_failures per success criterion. After a configurable threshold (default: 3), it calls capability_expander.py which generates a revised_specification.json with adjusted criteria and a revision_rationale.md explaining the change. The pipeline continues with the revised spec and flags the cycle as auto_revised.

Paradigm Shift 4 — Cross-Goal Skill Transfer

What it is: When you run the unified orchestrator on a new goal, it first searches the learnings_log.json for similar past goals (by semantic similarity of the goal description). If a match is found above a configurable threshold, it bootstraps the new specification from the prior goal's best-performing specification version — skipping the cold-start problem entirely.

Why it matters: Every AI system starts from zero on every new task. This system starts from the closest thing it has already learned. A goal like "process customer support tickets with 98% accuracy" will bootstrap from any prior goal involving classification, accuracy thresholds, or text processing — not just identical goals. The system accumulates institutional knowledge across goals, not just within them.

How it works: meta_learner.py maintains a goal_similarity_index.json. On each new run, pipeline.py queries this index using keyword overlap and structural similarity. If a match scores above 0.6, the matched goal's best_specification.json is used as the starting template, with the new goal's specific criteria overlaid.

Paradigm Shift 5 — Verifiable Improvement Chains

What it is: Every cycle produces a cryptographically signed improvement report. The unified orchestrator chains these reports together — each report contains the SHA-256 hash of the previous report — creating an immutable, independently verifiable improvement chain. Any external party can verify that improvement claim X was produced from observation Y, which was produced from specification Z, in cycle N.

Why it matters: AI systems routinely claim improvement without proof. This system produces a tamper-evident chain of evidence: you can prove to any stakeholder exactly what changed, when, why, and what effect it had. This is the foundation for auditable AI governance — not as a compliance afterthought, but as a first-class output of every run.

How it works: pipeline.py reads the chain_tip from improvement_chain.json (the hash of the last report) and includes it in the current report before signing. signal_router.py provides a verify_chain command that re-computes all hashes and confirms chain integrity.

Architecture

unified-orchestrator-v2/
├── SKILL.md                                   ← this file
├── scripts/
│   ├── pipeline.py                            ← main entry point — runs the full loop
│   ├── meta_learner.py                        ← cross-cycle pattern learning and transfer
│   ├── capability_expander.py                 ← autonomous re-specification on impossible goals
│   └── signal_router.py                       ← drift detection, chain verification, routing
├── references/
│   ├── paradigm_shifts.md                     ← detailed design notes for all five shifts
│   ├── pipeline_config.json                   ← configurable thresholds and behavior
│   ├── learnings_log.json                     ← accumulated cross-cycle learnings (auto-updated)
│   ├── goal_similarity_index.json             ← cross-goal transfer index (auto-updated)
│   └── improvement_chain.json                 ← cryptographic improvement chain (auto-updated)
├── templates/
│   ├── pipeline_run_report_template.md        ← human-readable full pipeline report
│   └── revision_rationale_template.md         ← auto-revised specification rationale
└── examples/
    ├── example_goal_simple.json               ← simple standalone example
    ├── example_goal_triad.json                ← full triad example
    └── example_learnings_log.json             ← sample accumulated learnings

Usage

Quickstart — Any Goal, Any Input

# From a plain goal description (standalone — no other skills required)
python scripts/pipeline.py --goal "Process customer tickets with 98% accuracy in under 2 seconds"

# From an existing specification (skip intent-engineering)
python scripts/pipeline.py --spec specification.json

# Full triad — all three skills
python scripts/pipeline.py --spec specification.json --outcome outcome_report.json

# Continue a prior cycle (self-improving loop)
python scripts/pipeline.py --state pipeline_state.json

# Verify the improvement chain
python scripts/signal_router.py verify --chain references/improvement_chain.json

Options

Running Multiple Cycles Automatically

# Run 5 cycles automatically, feeding each output back as input
python scripts/pipeline.py --goal "Achieve 98% pass rate on ticket classification" --cycles 5 --output-dir ./run_001/

Verifying an Improvement Chain

python scripts/signal_router.py verify --chain ./run_001/improvement_chain.json
# Output: Chain verified: 5 links, all hashes valid. No tampering detected.

Outputs

Every pipeline run produces the following in --output-dir:

Integration with Individual Skills

The unified orchestrator is designed to work with the individual skills, not replace them. You can:

• Run intent-engineering independently to produce a specification, then pass it to the unified orchestrator with --spec.
• Run dark-factory independently to produce an outcome report, then pass it with --outcome.
• Run feedback-loop independently on any observation, then pass the analysis to the unified orchestrator with --analysis.
• Run the unified orchestrator end-to-end and then use the individual skills to drill into specific stages.

Configuration

Edit references/pipeline_config.json to tune behavior:

{
  "drift_threshold": 0.15,
  "auto_revise_after_n_failures": 3,
  "transfer_similarity_threshold": 0.60,
  "chain_enabled": true,
  "meta_learning_enabled": true,
  "min_cycles_for_meta_learning": 3
}

The Self-Improving Loop

Cycle 1:
  goal → specification → execution → observation → analysis → improvement_report_1
                                                                      ↓
Cycle 2:                                                     pipeline_state.json
  pipeline_state → (transfer check) → specification_v2 → execution → observation
                                                                      ↓
                                                             improvement_report_2
                                                                      ↓
Cycle N:                                                     learnings_log grows
  meta_learner fires (after 3+ cycles) → specification_patches → rule_patches
  → all future specifications start smarter
  → all future feedback loop analyses use better rules
  → drift score stabilizes toward 0.0
  → improvement chain grows as verifiable evidence

Dependency Map

This skill coordinates the following skills. Each is optional — the pipeline degrades gracefully:

When all three are absent, the unified orchestrator runs in meta-only mode: it accepts any JSON log or text description, runs the feedback-loop analysis internally, and produces a signed improvement report with cross-goal transfer and drift detection — using only its own embedded logic.