SAFLA

Self-Aware Feedback Loop Algorithm (python)

🔁 SAFLA: Self-Aware Feedback Loop Algorithm (Agentic Edition)

You are building a state-of-the-art autonomous system that implements SAFLA — a fully recursive, self-aware feedback loop rooted in the latest agentic AI research. This build is not theoretical. It is real, test-driven, self-reflecting, and capable of patching its own policy via continual divergence, reward, and error analysis.

---

⚙️ Environment Setup

• Codespace: Internet access enabled
• MCP configuration: @/.roo/mcp.json
• Secrets: Stored in .env (used by requesty.ai, context7, perplexityai)
• Source control: Commit after each phase to https://github.com/ruvnet/SAFLA
• Mode roles: Defined in .roomodes
• Instruction template: @/mcp-instructions.md

---

🧠 Step 1 — Guided Deep Research (🔍 deep-research)

Use MCPs to perform scoped deep research into:

1. Using perplexityai MCP + Sonar LLM:

   • Topic: “self-aware feedback loop architectures in autonomous agents”
   • Output: /research/05_final_report/03_findings.md

2. Using context7 MCP:

   • Topic: “top libraries for divergence detection, vector memory, delta patching, and loop evaluation”
   • Output: /research/02_data_collection/02_secondary_findings.md

---

🔄 SAFLA Orchestration (via MCP Protocol)

Define and implement a full Model-Context Protocol (MCP) orchestration layer. Each cycle of the loop shares, stores, and updates the following:

{
  "context": {
    "goals": [...],
    "constraints": [...],
    "state_snapshot": {...},
    "memory_refs": [...],
    "prior_rationale": "...",
    "prior_reflection": "...",
    "patches_applied": [...]
  },
  "action_proposal": "...",
  "result_observed": "...",
  "metrics": {
    "reward": 0.92,
    "divergence": 0.14,
    "delta_improvement": 0.07
  },
  "status": "pending | complete | patched",
  "task": "reflection | scoring | implementation | critic"
}

MCP payloads must be passed between agents (reflection, scorer, prompt-generator, code) via .roo message passing or new_task invocations. Store all serialized contexts in /memory-bank/context_logs/.

---

📐 Delta Evaluation and Reflection Reuse Loop

Implement a formal delta evaluation model to quantify improvement across iterations:

📊 Delta Metric (Δ-score)

Δ = (rewardᵢ - rewardᵢ₋₁) / tokens_usedᵢ

• If Δ < ε (e.g. 0.01): trigger mini-reflection and policy patch loop
• If Δ > τ (e.g. 0.2): promote patch to core-policy-guidelines.md

Track all Δ-scores across iterations in:

• /scores/delta_scores.json
• /logs/loop_deltas.md

🔁 Reflection Reuse

Each reflection_LS{n}.md must be scored by the scorer:

• If patch_effectiveness_score ≥ 0.8, reuse it in future prompt-generator calls
• Save reusable patches to: /patches/library/generic-patches.md

Add reflection_id tags in prompt history and memory logs for traceability.

---

🧠 Agent Modes (from .roomodes)

Use the following roles in your recursive loop:

Mode	Function
🧠 memory-manager	Stores embeddings, computes JSD, and prunes
💬 prompt-generator	Produces next-step action prompts with deltas and memory
🔄 reflection	Writes new patch notes, checks past patch effectiveness
🧐 critic	Detects logic flaws, offers commentary
🎯 scorer	Calculates Δ, JSD, reward change, and patch success
🧪 tdd	Red-green-refactor + test scaffolding
🧠 code	Writes modular implementations (<500 lines)
♾️ mcp	Handles integrations with requesty, Sonar, context7
🏁 final-assembly	Merges code, metrics, and results into /final.md

Each role must call spawn new_task and conclude with attempt_completion.

---

📦 Execution Phases

1. /memory-bank/

   • store.py: SQLite + JSD pruning
   • README.md: define memory layers, delta logic, and storage structure

2. /cli/

   • safla run: Full loop executor
   • safla step: Trigger one mode at a time

3. /tests/

   • Test coverage for: JSD, patch success, prompt selection, MCP health

4. /orchestration/

   • run_loop.py: Recursive mode orchestrator
   • workflow.json: DAG or step-graph for mode transitions

---

✅ Evaluation Conditions

At each loop, log:

• Δ-score vs. previous
• JSD vs. previous state
• token usage
• patch applied (y/n)
• reflection reused (y/n)

If no improvement is detected for 3 loops:

• Trigger escalation to meta-reflection
• Write loop_stall_report.md

---

📘 Output Artifacts

• @/plans/research.md: From Perplexity + Context7
• /final.md: Full output with diagrams and scores
• /reflection/: All patches, reused or rejected
• /memory-bank/: Snapshot DBs, embedding deltas
• /scores/: Delta history, patch effectiveness, loop cost

---

🧾 Identity

Created by rUv This is recursive, modular autonomy in motion. A feedback loop that reflects, scores, patches, and evolves — autonomously.