qbridge-agent 0.4.0

Local, compiler-backed CUDA-Q optimization agent

QBridge Agent

Natural language → QBridge DSL → local CUDA-Q solve.
No cloud compiler. No coding-agent detour.

QBridge-Agent is a lean optimization agent for quantum developers who already know they want QAOA on CUDA-Q, not a general-purpose LLM rewriting Python files until something runs. You describe a bounded binary optimization problem — knapsack, MaxCut, portfolio, scheduling — and the agent:

1. Translates your request into validated QBridge DSL
2. Compiles locally with the installed qbridge package (deterministic, auditable)
3. Executes on your NVIDIA CUDA-Q target and prints the decoded assignment

No Aider loop. No hand-written @cudaq.kernel. No cloud compile endpoint.

qbridge "Knapsack with 5 items, values [6,5,8,9,6], weights [2,3,4,5,3], capacity 8"

---

Why this beats a coding agent for QUBO problems

	QBridge-Agent	General coding agent (e.g. Aider)
Output	Validated QBridge DSL → compiler-lowering	Ad-hoc CUDA-Q scripts
Compile path	Deterministic local qbridge	LLM edits until syntax runs
Verification	Independent constraint/objective check	Usually none
Token cost (T1/T2 suite)	~1,294 mean tokens/trial	~6,602 (5.10×)
Wall time (same suite)	~6.6 s	~42.7 s (~6.49×)
Verified trials (18-run benchmark)	12 / 18	0 / 18

Benchmark: frozen six-task T1/T2 CUDA-Q suite, tencent/hy3:free (default CLI model), NVIDIA cusvsim_fp32, three trials per system. Prior openai/gpt-4o-mini run: 16/18 verified, 2.68× tokens, ~3.36× wall time — see evaluations/README.md.

---

Table of contents

1. Install
2. CLI experience
3. Environment setup (WSL + CUDA-Q)
4. Mode A — DSL file (no API key)
5. Mode B — natural language
6. Workspace layout
7. Programmatic API
8. Worked examples
9. CLI reference
10. Troubleshooting
11. Benchmarks & evaluations
12. Scope & non-goals
13. License & releases

---

Install

Requirements

Component	Version / notes
Python	3.9+
qbridge[quantum]	≥ 0.4.0 — the CUDA-Q compiler (presolve, dual compile, init_strategy)
CUDA-Q	0.15.0, NVIDIA target available
OpenRouter API key	First-run setup for NL mode (stored locally)

PyPI

pip install "qbridge[quantum]>=0.4.0" qbridge-agent

Source checkout

git clone https://github.com/ProfessionalQwerty/qbridge-agent.git
cd qbridge-agent
pip install -e ".[dev]"

Verify:

qbridge --help
qbridge-compile --help   # from qbridge package

---

CLI experience

QBridge-Agent is a single-shot CLI — one quoted problem or one DSL file per invocation. There is no persistent chat shell; output is structured, timed, and color-coded for terminals.

First launch

Running qbridge with no arguments (or your first NL prompt) opens setup:

──────────────────────────────────────────────────────────
  QBridge-Agent  local compiler-backed CUDA-Q optimization
──────────────────────────────────────────────────────────

 Welcome to QBridge.
 Describe an optimization problem in plain English.
 We translate to QBridge DSL, compile locally, and run CUDA-Q.

 OpenRouter setup  keys at https://openrouter.ai/keys
 Paste your key and press Enter (visible input — paste works reliably on Windows/WSL)
 OpenRouter API key: sk-or-v1-...
 Received  sk-or-v1-abc...xyz (73 chars)
 Model [openai/gpt-4o-mini]:
 ✓ Ready  model=tencent/hy3:free

Credentials are saved under ~/.qbridge-agent/ (credentials.json + config.json). Nothing is sent to a QBridge cloud endpoint — only OpenRouter for NL translation.

Everyday commands

Command	What it does
qbridge	Show configured model and usage hints
qbridge "…"	NL → DSL → compile → CUDA-Q solve
qbridge problem.py	DSL file → compile → solve (no API key)
qbridge --setup	Re-enter OpenRouter key and default model
qbridge --doctor	Test env vs saved key; show which is active
qbridge --no-solve problem.py	Compile only, skip GPU solve

Terminal output (NL mode)

Each run prints a banner, timed steps, structured blocks, and token usage:

 ● Request  Knapsack with 5 items, values [6,5,8,9,6], weights …
 ● Model  tencent/hy3:free
 ● Translating DSL and compiling locally
 ✓ Translating DSL and compiling locally  12.69s

 Generated DSL
   path: qbridge-output/problem.py

 Solution
   Backend:    cudaq:nvidia
   Objective:  14.0
   Feasible:   True
   q0: 1  …

 Token usage
   prompt: 1078  completion: 158  total: 1236  wall time: 12.68s

 ✓ Done  workspace: …/qbridge-output

On failure, the CLI prints a red error line and actionable hints (e.g. qbridge --setup after a 401).

Credential precedence

Source	Priority
OPENROUTER_API_KEY env	Used only if it looks like a real sk-or-… key (≥ 20 chars)
~/.qbridge-agent/credentials.json	Default after setup
Placeholder env (your-key, etc.)	Ignored — saved file key is used instead

Run qbridge --doctor to see which key is active and whether OpenRouter accepts it.

Styling

• Cyan banner and bullets, magenta section headers, blue values, green success marks
• Respects NO_COLOR and FORCE_COLOR
• Falls back to ASCII glyphs (>, OK, X) when Unicode is unavailable

---

Environment setup (WSL + CUDA-Q)

CUDA-Q with NVIDIA targets is Linux-first. On Windows, use WSL2 — not bare python.exe, not an arbitrary WSL distro.

Verified example (Windows host → Ubuntu WSL)

wsl -d Ubuntu -- /home/dilip/qbridge-env/bin/python3 -c "import cudaq; print(cudaq.__version__)"
# CUDA-Q Version 0.15.0

Suggested virtualenv bootstrap (inside WSL Ubuntu)

python3 -m venv ~/qbridge-env
source ~/qbridge-env/bin/activate
pip install "qbridge[quantum]>=0.4.0" qbridge-agent
python -c "import cudaq; cudaq.set_target(\"nvidia\"); print(cudaq.get_target())"

Run the agent

qbridge "Knapsack with 5 items, values [6,5,8,9,6], weights [2,3,4,5,3], capacity 8"

If you previously exported a placeholder like OPENROUTER_API_KEY=your-key in your shell, remove it:

unset OPENROUTER_API_KEY
qbridge --doctor

---

Mode A — DSL file (no API key)

When you already have — or hand-write — QBridge DSL, the agent is a compile + solve CLI with structured output.

qbridge examples/knapsack_n5.py --workspace ./qbridge-output

Compile only (emit CUDA-Q kernel, skip GPU solve):

qbridge examples/knapsack_n5.py --workspace ./qbridge-output --no-solve

Expected terminal flow

──────────────────────────────────────────────────────────
  QBridge-Agent  local compiler-backed CUDA-Q optimization
──────────────────────────────────────────────────────────

 ● Loading QBridge DSL  qbridge-output
 ● Compiling to CUDA-Q
 ✓ Compiling to CUDA-Q  1.24s
 Artifacts
   CUDA-Q kernel: …/compiled_output/kernel_execution.py
   DSL source:    …/problem.py
 ● Running CUDA-Q on NVIDIA target
 ✓ Running CUDA-Q on NVIDIA target  2.81s
 Solution
   Backend:       cudaq:nvidia
   Objective:     …
   Feasible:      True
 ✓ Done  workspace: /…/qbridge-output

The DSL file must use import qbridge as qb and declare variables, objective, and constraints — not arbitrary Python. Full DSL rules: QBridge compiler README.

---

Mode B — natural language

After first-run setup, describe your problem in plain English:

qbridge "Knapsack: 5 items, values [6,5,8,9,6], weights [2,3,4,5,3], capacity 8"

Reconfigure anytime:

qbridge --setup

Model override for one run

qbridge "MaxCut on a 4-node cycle with edge weights 1,2,3,4" --model openai/gpt-4o-mini

When to use	Mode
You have DSL already, or want zero API cost	A — .py file
Exploring problems, demos, NL interface	B — quoted prompt
CI / regression	A (deterministic input)

Token usage is printed at the end of NL runs.

---

Workspace layout

Every run writes artifacts under --workspace (default: ./qbridge-output):

qbridge-output/
├── problem.py                    # DSL copy (Mode A) or generated DSL (Mode B)
├── prompt.txt                    # Mode B only — your NL request
├── compiled_output/
│   ├── kernel_execution.py       # Generated CUDA-Q kernel
│   ├── compile_meta.json         # Local compiler metadata
│   └── solution.json             # Assignment + objective + feasibility
├── _usage.json                   # Mode B — token counts + timings
└── response.txt                  # Mode B — raw LLM response (if pipeline saves it)

Reading solution.json

{
  "assignment": {"x0": 1, "x1": 0, "x2": 1},
  "objective_value": 19.0,
  "constraints_satisfied": true,
  "backend": "cudaq:nvidia",
  "cudaq_target": "Target nvidia ..."
}

If constraints_satisfied is false, the QAOA sample decoded to an infeasible assignment — tighten SOLVE_KWARGS in the generated DSL or re-run.

---

Programmatic API

Full NL pipeline

from pathlib import Path
from agent.cudaq_entry import run_cudaq_case

report = run_cudaq_case(
    prompt="Select at most 3 projects maximizing value under budget 100 ...",
    workspace=Path("./qbridge-output"),
    model="openai/gpt-4o-mini",
    api_key="...",  # or ~/.qbridge-agent/credentials.json
)

print(report["ok"], report["solution"], report["usage"])

DSL-only pipeline (no LLM)

from pathlib import Path
from agent.compiler_interface.pipeline import run_quantum_pipeline

dsl = Path("examples/knapsack_n5.py").read_text()
report = run_quantum_pipeline(
    "ignored",
    complete_fn=lambda _: dsl,
    workspace_root=Path("./qbridge-output"),
    include_solve=True,
)
print(report.ok, report.solution)

---

Worked examples

Example	File	Command
5-item knapsack	examples/knapsack_n5.py	qbridge examples/knapsack_n5.py --workspace ./out
MaxCut (NL)	examples/maxcut_agent_run.md	See file for quoted prompt
Portfolio (NL)	examples/optimize_portfolio.md	Budget + group cover constraints
Scheduling (NL)	examples/scheduling_agent_run.md	Resource + time windows

MaxCut prompt example

qbridge "MaxCut on triangle graph nodes 0-1-2 with edge weights 1,2,3. Binary vars x0,x1,x2." \
  --workspace ./maxcut-out

---

CLI reference

qbridge [request] [--workspace WORKSPACE] [--no-solve] [--setup] [--doctor] [--model MODEL]
qbridge-agent   # alias

Argument	Description
request	Quoted NL problem or path to .py DSL file
--setup	Re-enter OpenRouter key and model
--doctor	Diagnose credential sources and OpenRouter connectivity
--workspace PATH	Output directory (default: ./qbridge-output)
--no-solve	Compile only (DSL mode)
--model ID	Override saved model for one run

Credentials: ~/.qbridge-agent/credentials.json (key) and config.json (model). A valid OPENROUTER_API_KEY env var overrides the saved file; placeholder values are ignored.

Exit codes

Code	Meaning
0	Success
1	Compile/solve/agent failure
2	Setup cancelled / missing credentials

Environment variables

Variable	Used by
OPENROUTER_API_KEY	NL mode (optional if setup completed)
QBRIDGE_BENCH_MODEL / QBRIDGE_MODEL	Default model override
NO_COLOR / FORCE_COLOR	Terminal styling

---

Troubleshooting

Symptom	Cause	Fix
401 Unauthorized from OpenRouter	Bad/stale key or placeholder env	unset OPENROUTER_API_KEY; qbridge --doctor; qbridge --setup
--doctor shows env your-key	Doc placeholder left in shell	unset OPENROUTER_API_KEY
Setup rejects pasted key	Hidden input mangled paste (older versions)	Upgrade to ≥ 0.3.1; use visible paste prompt
No module named 'cudaq'	Wrong Python / no CUDA-Q	Use WSL Ubuntu + qbridge[quantum] env
No module named 'qbridge.dsl_build'	Wrong qbridge package installed	pip install "qbridge[quantum]>=0.3.0" from PyPI
constraints_satisfied: false	Weak QAOA / penalty settings	Edit generated DSL SOLVE_KWARGS (↑ alpha, p, n_seeds)
Garbled Unicode on Windows terminal	Legacy code page	Use Windows Terminal; agent falls back to ASCII glyphs
Agent generates invalid DSL	LLM drift	Retry; or write DSL manually (Mode A)

---

Benchmarks & evaluations

Frozen six-task T1/T2 CUDA-Q suite (evaluations/benchmark_tasks.json). Each trial records OpenRouter token usage, wall time, and independent constraint verification.

Latest run (tencent/hy3:free, 20 July 2026)

System	Total tokens (mean)	Wall time (mean)	Verified
QBridge-Agent	1,294	6.58 s	12 / 18
Aider	6,602	42.70 s	0 / 18

Artifacts: evaluations/results/20260720_hy3_token_tracking/results.jsonl

Reproduce (WSL + CUDA-Q):

wsl -d Ubuntu -- bash -c '
  cd /mnt/c/Users/you/qbridge-agent &&
  export PYTHONPATH=$(pwd) &&
  unset OPENROUTER_API_KEY &&
  export OPENROUTER_API_KEY=$(python3 -c "import json; from pathlib import Path; print(json.loads(Path.home().joinpath(\".qbridge-agent/credentials.json\").read_text())[\"openrouter_api_key\"])") &&
  /home/you/qbridge-env/bin/python3 evaluations/run_benchmark.py \
    --model tencent/hy3:free --trials 3 \
    --results evaluations/results/my_hy3_run
'

Summarize a completed run:

python evaluations/summarize_results.py evaluations/results/my_hy3_run/results.jsonl

See evaluations/README.md for methodology, the gpt-4o-mini baseline, token capture (AIDER_USAGE_LOG), and timing field definitions.

---

Scope & non-goals

Supported

• Small constrained binary/integer optimization expressible in current QBridge DSL
• Local compile + NVIDIA CUDA-Q execution
• NL → DSL translation via OpenRouter

Not supported (by design)

• Arbitrary Python optimization scripts
• General quantum algorithms (VQE chemistry, Shor, custom ansätze)
• Cloud QBridge compile/solve APIs
• Interactive multi-turn chat UI (one qbridge "..." per problem)

Unsupported inputs are errors, not best-effort guesses.

---

License & releases

• Package: BUSL-1.1 (see LICENSE)
• Release process: RELEASE.md
• Compiler DSL docs: github.com/ProfessionalQwerty/qbridge

Current release: qbridge-agent 0.3.1 with qbridge[quantum]>=0.3.0.