PromptLoop

A Bayesian prompt observer and quality loop for your AI agent.

PromptLoop sits between you and your IDE's agent as an MCP server. Every prompt is scored, routed, refined if needed, and then verified — all inside a single tool call using MCP Sampling. The system learns over time: it runs online gradient descent on scoring dimension weights every time you rate a response, and selects clarifying questions via Thompson Sampling on Beta priors.

The underlying CLI and tool API ship as loopllm (the original name). PromptLoop is the project brand.

---

VS Code Extension

Install the companion extension for a live quality scratchpad and prompt history dashboard directly in the sidebar.

Prompt Lab — live quality scratchpad

Scores on every keystroke (350 ms debounce). Grade badge, 5 dimension bars, issues + suggestions tags, Copy and Send to Chat.

History — learning curve + metrics

Learning curve sparkline, grade distribution, SGD learned weights per dimension. Updates after every loopllm_feedback call.

The extension VSIX is at vscode-loopllm/loopllm-prompt-gauge-0.1.0.vsix — install with code --install-extension.

---

Primary tool: loopllm_run_pipeline

This is the intended entry point for PromptLoop. A single call runs the full observe → elicit → refine → verify loop:

loopllm_run_pipeline("add retry logic to the download function")

What happens inside:

1. Score the prompt across 5 dimensions (< 1 ms, deterministic)
2. Elicit clarifying questions if quality < 0.6 — selected by Thompson Sampling on historical Beta priors
3. Decompose into subtasks if complexity > 0.5
4. Execute each subtask: ctx.sample(prompt) → evaluate → retry if below threshold
5. Verify the assembled output against quality criteria via a second ctx.sample() call
6. Log result + quality score to SQLite; update scoring weights via online SGD

Everything runs inline via MCP Sampling — no extra chat turns, no polling.

---

Quickstart

git clone https://github.com/azank1/loop-llm   # GitHub repo still named loop-llm
cd loop-llm
pip install -e ".[mcp]"
code .

.vscode/mcp.json is committed — VS Code picks up the server automatically.
On first load verify with:

use loopllm_intercept with prompt: add retry logic to the download function

Then:

use loopllm_run_pipeline with prompt: add retry logic to the download function

---

Install as a package

pip install loopllm[mcp]

Add .vscode/mcp.json to your project:

{
  "servers": {
    "loopllm": {
      "type": "stdio",
      "command": "loopllm",
      "args": ["mcp-server", "--provider", "agent"]
    }
  }
}

For Cursor use .cursor/mcp.json with "mcpServers" as the top-level key.

---

How it works

You type a prompt
      ↓
loopllm_intercept          ← scores it across 5 dimensions (~0ms, deterministic)
      ↓
route decision
  < 0.4  → elicitation     ← Thompson Sampling picks the highest-gain question
  0.4–0.6 → elicit + refine
  ≥ 0.6  → refine directly
      ↓
loopllm_refine (if needed)
  → ctx.sample(prompt)     ← MCP Sampling: calls host LLM mid-execution
  → evaluate output        ← deterministic evaluators (length, regex, JSON schema)
  → if score < threshold: ctx.sample(improved prompt)   ← retry inline
  → return best result
      ↓
result logged to ~/.loopllm/store.db
      ↓
loopllm_feedback (optional rating 1–5)
  → one SGD step updates scoring dimension weights

Scoring dimensions (each 0–1, composited by learned weights into grade A–F):

Dimension	What it catches
Specificity	Vague, generic requests
Constraint Clarity	Missing format, length, or rule requirements
Context Completeness	No background or goal stated
Ambiguity	Unclear references, pronouns without antecedents
Format Specification	No output format specified

MCP Sampling — generation tools call ctx.sample() to invoke the host agent's LLM inline. The entire score→generate→evaluate→retry loop is a single tool call. Falls back to agent_execute passthrough if the client doesn't declare sampling capability.

---

How the system learns

Online Gradient Descent on scoring weights

Default weights are {specificity: 0.25, constraint_clarity: 0.20, context_completeness: 0.20, ambiguity: 0.20, format_spec: 0.15}. Each time you call loopllm_feedback(rating) one SGD step adjusts them:

$$\hat{y} = \sum_i w_i \cdot d_i \quad \text{(predicted quality, ambiguity inverted)}$$

$$\mathcal{L} = (\hat{y} - t)^2 \quad t = \frac{\text{rating} - 1}{4}$$

$$w_i \leftarrow w_i - 0.02 \cdot 2(\hat{y} - t) \cdot d_i$$

After each step weights are clipped to $[0.05, 0.50]$ and renormalised to sum to 1 (simplex projection). After ~50 ratings the weights converge to reflect what actually matters in your prompts. Persisted in learned_weights table (schema v4).

Thompson Sampling for question ordering

Each question type (scope, format, constraints, etc.) maintains a $\text{Beta}(\alpha, \beta)$ prior where $\alpha$ = historical positive-impact count + 1 and $\beta$ = negative-impact count + 1. When the pipeline needs to ask a clarifying question it draws one sample per candidate type:

$$s_i \sim \text{Beta}(\alpha_i, \beta_i)$$

and selects $\arg\max_i s_i$. This is a multi-armed bandit: questions that historically improve output quality get asked more often (exploitation) while new/untested types still get surfaced occasionally (exploration).

Beta-Binomial Bayesian priors

Every call observation updates a conjugate $\text{Beta}(\alpha, \beta)$ prior on per-(task_type, model) convergence probability. The adaptive exit criterion for refinement loops is:

$$P(\text{converge on next iteration}) = \frac{\alpha}{\alpha + \beta}$$

The loop stops early when this probability weighted by expected improvement drop falls below a cost-threshold. Implemented in adaptive_exit.py using BetaPrior.prob_above(threshold) (normal approximation via math.erf).

Welford's online algorithm for variance

NormalPrior tracks the running mean and variance of output quality scores with $O(1)$ memory using Welford's online algorithm with optional exponential decay ($\lambda = 0.95$) to down-weight stale observations:

$$\mu_n = \mu_{n-1} + \frac{x_n - \mu_{n-1}}{n}, \quad \sigma^2_n = \frac{(n-1)\sigma^2_{n-1} + (x_n - \mu_{n-1})(x_n - \mu_n)}{n}$$

Rolling plan confidence

$$C_{\text{rolling}} = \frac{\sum_i c_i \cdot 0.85^{n-i}}{\sum_i 0.85^{n-i}}$$

where $c_i$ is the confidence of the $i$-th completed task. Exponential decay weights recent tasks more heavily. Plans gate on confidence ≥ threshold before advancing to the next task.

---

Tools (24)

Tool	What it does
loopllm_run_pipeline	Primary. Elicit → decompose → execute → verify in one call
loopllm_intercept	Score + route a prompt; logs to history
loopllm_gauge	Instant quality bars, no DB write
loopllm_refine	Score → sample → retry loop via MCP Sampling
loopllm_plan_tasks	Decompose a goal into ordered subtasks via MCP Sampling
loopllm_verify_output	Keyword pre-check + deep sample against quality criteria
loopllm_elicitation_start/answer/finish	Multi-turn clarifying question session
loopllm_plan_register	Create a confidence-gated plan saved to SQLite
loopllm_plan_next	Advance to next task; returns needs_replan if quality dropped
loopllm_plan_update	Record task scores; recalculates rolling confidence
loopllm_plan_list	Dashboard: all plans with gauges and task counts
loopllm_plan_delete	Remove a completed or abandoned plan
loopllm_context_history	Browse prompt history with sparklines
loopllm_context_clear	Wipe prompt history (scoped or all)
loopllm_prompt_stats	Prompting quality trend and learning curve
loopllm_feedback	Rate a response (1–5); triggers SGD weight update
loopllm_suggest_config	Bayesian-optimal loop config for a task type
loopllm_classify_task	Label a prompt's task type
loopllm_analyze_prompt	Generate clarifying questions ranked by Thompson-sampled gain
loopllm_list_tasks	List tasks from the persistent store
loopllm_show_task	Detail view for a single task
loopllm_report	Learned weights, Bayesian priors, question effectiveness stats

Plans and learned weights persist to ~/.loopllm/store.db and survive server restarts (schema v4).

---

Local models (no MCP)

pip install loopllm[serve]
loopllm serve --port 8765          # REST scoring middleware

from loopllm.local_loop import LocalModelLoop

loop = LocalModelLoop(
    base_url="http://localhost:11434",
    model="llama3.2",
    score_url="http://localhost:8765/score",
    quality_threshold=0.80,
    max_retries=3,
)
result = loop.run("Write a Python function to parse JSON safely.")
print(result.output, result.best_score, result.converged)

---

Contributing

git clone https://github.com/azank1/loop-llm   # GitHub repo still named loop-llm
cd loop-llm
pip install -e ".[dev]"
python -m pytest tests/ -q          # 186 tests, ~2s

Branch strategy:

• main — stable, released code
• testing-production — active development; PRs target this branch first

Key files:

• src/loopllm/mcp_server.py — all 24 MCP tools + SGD weight update + MCP Sampling helpers
• src/loopllm/store.py — SQLite persistence (schema v4: priors, history, plans, learned_weights)
• src/loopllm/priors.py — Beta-Binomial and Normal priors, Welford's algorithm
• src/loopllm/adaptive_exit.py — Bayesian early stopping
• src/loopllm/evaluators.py — deterministic output evaluators

PRs welcome. Add tests for new tools in tests/.

---

Environment variables

Variable	Default	Description
LOOPLLM_PROVIDER	agent	agent, ollama, or openrouter
LOOPLLM_MODEL	agent	Model identifier (ignored in agent mode)
LOOPLLM_DB	~/.loopllm/store.db	SQLite store path
OLLAMA_HOST	http://localhost:11434	Ollama base URL
OPENROUTER_API_KEY	—	OpenRouter API key

---

License

MIT

A Bayesian prompt observer and quality loop for your AI agent.

PromptLoop sits between you and your IDE's agent as an MCP server. Every prompt you write is intercepted, scored across five quality dimensions, and routed to the right tool — clarification, refinement, or task decomposition. It tracks your prompting history over time, lets you manage multi-task plans with confidence gating, and runs iterative improvement loops using the host agent's own LLM via MCP Sampling, with no separate model needed.

Think of it as a prompt observer: it watches what you ask, measures how well-formed it is, tells you where it's weak, and automatically improves it before the agent acts.

Stack: Python 3.11+, FastMCP (stdio), SQLite (Bayesian priors + prompt history + plans), deterministic heuristic scoring, MCP Sampling for mid-execution LLM calls.

---

Quickstart

git clone https://github.com/azank1/loop-llm   # GitHub repo still named loop-llm
cd loop-llm
pip install -e ".[mcp]"
code .

.vscode/mcp.json is committed — VS Code picks up the server automatically. If tools don't appear: MCP: Restart Server → loopllm in the Command Palette.

Verify in Copilot chat:

use loopllm_gauge to score this prompt: write me some code

---

Install as a package

pip install loopllm[mcp]

Add .vscode/mcp.json to your project:

{
  "servers": {
    "loopllm": {
      "type": "stdio",
      "command": "loopllm",
      "args": ["mcp-server", "--provider", "agent"]
    }
  }
}

For Cursor use .cursor/mcp.json with "mcpServers" as the top-level key.

---

How it works

You type a prompt
      ↓
loopllm_intercept          ← scores it across 5 dimensions (~0ms, deterministic)
      ↓
route decision
  < 0.4  → elicitation     ← agent asks the 1-2 highest-gain clarifying questions
  0.4–0.6 → elicit + refine
  ≥ 0.6  → refine directly
      ↓
loopllm_refine (if needed)
  → ctx.sample(prompt)     ← MCP Sampling: calls host LLM mid-execution
  → evaluate output        ← deterministic evaluators (length, regex, JSON schema)
  → if score < threshold: ctx.sample(improved prompt)   ← retry inline
  → return best result
      ↓
result logged to ~/.loopllm/store.db

Scoring dimensions (each 0–1, composited into a grade A–F):

Dimension	What it catches
Specificity	Vague, generic requests
Constraint Clarity	Missing format, length, or rule requirements
Context Completeness	No background or goal stated
Ambiguity	Unclear references, pronouns without antecedents
Format Specification	No output format specified

MCP Sampling means generation tools (loopllm_refine, loopllm_run_pipeline, loopllm_plan_tasks, loopllm_verify_output) call ctx.sample() to invoke the host agent's LLM inline — the entire score→generate→evaluate→retry loop happens inside a single tool call, not across multiple chat turns. Falls back to agent_execute passthrough if the client doesn't declare sampling capability.

Prompt history is stored in SQLite with per-prompt scores, grades, task types, and session tags. Use loopllm_context_history to browse it, loopllm_prompt_stats for your trend sparkline, or loopllm_context_clear to reset.

---

Tools (24)

Tool	What it does
loopllm_intercept	Score + route a prompt — call this first on every request
loopllm_gauge	Instant quality bars, no DB write
loopllm_refine	Score → sample → retry loop via MCP sampling
loopllm_run_pipeline	Elicit → decompose → execute → verify in one tool call
loopllm_plan_tasks	Decompose a goal into ordered subtasks via MCP sampling
loopllm_verify_output	Keyword pre-check + deep sample against quality criteria
loopllm_elicitation_start/answer/finish	Multi-turn clarifying question session
loopllm_plan_register	Create a confidence-gated plan saved to SQLite
loopllm_plan_next	Advance to next task; returns needs_replan if quality dropped
loopllm_plan_update	Record task scores; recalculates rolling confidence
loopllm_plan_list	Dashboard: all plans with gauges and task counts
loopllm_plan_delete	Remove a completed or abandoned plan
loopllm_context_history	Browse prompt history with sparkline
loopllm_context_clear	Wipe prompt history (scoped or all)
loopllm_prompt_stats	Prompting quality trend and learning curve
loopllm_feedback	Rate a response (1–5) to improve future scoring
loopllm_suggest_config	Bayesian-optimal loop config for a task type
loopllm_classify_task	Label a prompt's task type
loopllm_analyze_prompt	Generate clarifying questions ranked by information gain
loopllm_list_tasks	List tasks from the persistent store
loopllm_show_task	Detail view for a single task
loopllm_report	Learned Bayesian priors and question effectiveness

Plans persist to ~/.loopllm/store.db and survive server restarts.

Rolling confidence:

confidence(task) = prompt_score × 0.35 + output_score × 0.65
rolling            = Σ(confidence_i × 0.85^(n-i)) / Σ(0.85^(n-i))

---

Local models (no MCP)

pip install loopllm[serve]
loopllm serve --port 8765          # REST scoring middleware

from loopllm.local_loop import LocalModelLoop

loop = LocalModelLoop(
    base_url="http://localhost:11434",
    model="llama3.2",
    score_url="http://localhost:8765/score",
    quality_threshold=0.80,
    max_retries=3,
)
result = loop.run("Write a Python function to parse JSON safely.")
print(result.output, result.best_score, result.converged)

---

Contributing

git clone https://github.com/azank1/loop-llm   # GitHub repo still named loop-llm
cd loop-llm
pip install -e ".[dev]"
python -m pytest tests/ -q          # 186 tests, ~0.5s

Branch strategy:

• main — stable, released code
• testing-production — active development; PRs target this branch first

Key files:

• src/loopllm/mcp_server.py — all 24 MCP tools + MCP sampling helpers
• src/loopllm/plan_registry.py — confidence-gated plan tracking
• src/loopllm/store.py — SQLite persistence (schema v3)
• src/loopllm/evaluators.py — deterministic output evaluators
• src/loopllm/serve.py — REST server for local models
• src/loopllm/local_loop.py — Ollama wrapper with score→retry loop

PRs welcome. Add tests for new tools in tests/.

---

Environment variables

Variable	Default	Description
LOOPLLM_PROVIDER	agent	agent, ollama, or openrouter
LOOPLLM_MODEL	agent	Model identifier (ignored in agent mode)
LOOPLLM_DB	~/.loopllm/store.db	SQLite store path
OLLAMA_HOST	http://localhost:11434	Ollama base URL
OPENROUTER_API_KEY	—	OpenRouter API key

---

License

MIT