simplicio-cli

Your tasks with 99% accuracy using any LLM (Claude, DeepSeek, Codex, Gemini, Hermes, OpenClaw, Cursor).

"hide the Delete button for non-admins" → diff + test + applied + verified.
Works with OpenRouter, OpenAI, Anthropic, GLM, DeepSeek, Ollama — one env var.

pip install simplicio-cli

Why it works — the numbers

Same model. Same task. Only the prompt changes. Measured, reproducible, deterministic.
Fourteen models tested across three runs — five sub-4B tiny models, six
frontier 2026 models, and three mid-tier 7B–12B open models. Every one gained
at least +14 points when wrapped in simplicio's 6-layer contract.

Tiny models — sub-4B, run on 2026-05-26 (50 runs/side, 260 checks)

Not hosted on OpenRouter (requested but skipped): Gemma 3 270M, Gemma 3 1B,
Gemma 2 2B, Qwen3 0.6B, Qwen3 1.7B, Qwen2.5 0.5B, Qwen2.5 1.5B, Qwen 3B,
Nemotron Nano 4B (OR's smallest Nemotron is 9B). Sub-4B substitutes used above.
simplicio still gains +14 to +58 points even on a 1B-param model.

Frontier 2026 models — run on 2026-05-26 (60 runs/side, 312 checks)

Mid-tier 7B–12B open models — earlier run (v0.2.2, 30 runs/side, 156 checks)

Across all 14 models tested across three runs, the average gain is +51
points. Smallest: +14 pts (Llama 3.2 1B — the contract still moves a
1B-param model). Largest: +62 pts (GPT-5.5). The contract helps tiny
sub-4B models, frontier reasoning models, and mid-tier 7B–12B alike — five
of the six frontier models hit 100% pass-rate.

Output-quality signals (rate across all 60 frontier runs)

Cost — tokens & wall-clock (measured, not estimated)

Same provider, same models, same cases. Token counts pulled from the API
usage field; latency from time.perf_counter() around each call.

simplicio wraps the objective in a 6-layer contract — more input tokens up
front, longer completions because the model produces the full DIFF + TEST +
EVIDENCE the contract demands instead of a one-line guess. The bill goes up,
but so does the pass-rate (41% → 99%) and the DIFF-block rate (36% → 98%) —
useful tokens, not chat.

Six frontier models — GPT-5.5, Kimi K2.6, Gemini 3.5 Flash, Qwen 3.7 Max,
Claude Opus 4.7, DeepSeek V4 Pro — gained +52 to +62 points when wrapped
in simplicio's 6-layer contract. Without changing the model. Without
fine-tuning. Five of six landed at 100% pass-rate with simplicio.

Full report: bench/results.md · bench/results.pdf · raw outputs under .simplicio/bench_runs/.

How it works

mapper        WHERE   project structure + latest state
precedent     HOW-1   the real snippet in THIS repo that already does it
skill-router  HOW-2   the ONE mapper skill that matches (ranked, not all)
simplicio     BUILD   stacks the 6 layers into one prompt (cache-friendly)
test          JUDGE   contract written as testable states
verify        PROOF   ran it — did it actually pass? loop-fix up to 3x

The idea in one line: don't ask the model to guess — hand it the path.
Each layer terminates one decision the model would otherwise hallucinate.
Relevant > complete — inject the right context, never all of it.

Install

pip install simplicio-cli           # from PyPI
# or
pip install -e .                    # from this repo

Configure — any LLM, nothing hardcoded

If SIMPLICIO_BASE_URL is unset and the key is ANTHROPIC_API_KEY, it uses the
native Anthropic SDK. Otherwise it uses an OpenAI-compatible client pointed at
your base_url — so any OpenAI-like provider works without code changes.

simplicio smoke      # prints provider config + one test call

Use

# index once (caches embeddings; re-run after big changes)
simplicio index --stack angular

# run a task
simplicio task "hide Delete button for non-admins" \
  --stack angular \
  --target src/app/screen/screen.component.html \
  --criteria "- no admin perm: button absent from DOM
- with admin perm: button present" \
  --constraints "- don't touch save flow
- build passes"

Each task: precedent (from cache) → skill match → 6 layers → LLM generates
(diff + test + Playwright) → apply → run SIMPLICIO_TEST_CMD → pass? done :
send the error back → fix → retry (up to 3x).

Cache — why it doesn't re-map every time

Embeddings are keyed by content hash, stored in .simplicio/. Unchanged
code block → vector reused. Change one file → only that block re-embeds.

Benchmark — reproduce in 30 seconds

OPENROUTER_API_KEY=… \
  BENCH_MODELS="deepseek/deepseek-v4-pro,qwen/qwen3.7-max,moonshotai/kimi-k2.6,openai/gpt-5.5,anthropic/claude-opus-4.7,google/gemini-3.5-flash" \
  python3 bench/run_offline.py

No project required, stdlib only, deterministic regex scoring — no LLM judges
the LLM. Each case runs twice on the same model: raw one-line objective vs
simplicio's 6-layer contract. Outputs scored on target-file mention, DIFF
block, TEST block, contract-state words. Full numbers in bench/results.md.

Full harness (your real project, your real tests)

simplicio bench --cases bench/cases.json --stack angular

Runs each case two ways and runs your real test command (e.g. ng test --watch=false) on each output. Writes the true pass-rate to
bench/results.md.

4-quadrant bench — agent × simplicio matrix

Adds the second axis: not just "does the 6-layer wrap help one call?" but
"does it still help inside a retry loop?". Same model, same cases — only
the cell logic changes.

pip install -e ".[bench]"          # adds fpdf2 for PDF report
OPENROUTER_API_KEY=… \
  BENCH_MODELS="google/gemma-3-4b-it" \
  BENCH_MAX_ITERS=3 \
  python3 bench/run_4quadrant.py

Outputs bench/results_4quadrant.{md,pdf,json} + SVG charts under
bench/charts/4q_*.svg + per-iteration raw outputs under
.simplicio/bench_4q/<model>/case_NN/q*_iter*.txt. Methodology and
hypothesis decomposition: docs/benchmark-4quadrant.md.

The matrix decomposes:

• Prompt effect alone: Q2 − Q1
• Loop effect alone: Q3 − Q1
• Prompt effect inside loop: Q4 − Q3 (does simplicio still matter once you loop?)
• Composition gain over best single axis: Q4 − max(Q2, Q3)
• Synergy vs linear stacking: Q4 − (Q1 + (Q2−Q1) + (Q3−Q1))

Run 1 — focused single-model, google/gemma-3-4b-it, 5 cases, max_iters=3 (2026-05-26)

Decomposition (rejection threshold |Δ| ≥ 5 pts):

Cost per passing case: Q1 = 4,683 tok / 236s — Q2 = 800 tok / 21s — Q3 = 3,135 tok / 109s — Q4 = 1,018 tok / 20s. Full table + charts in bench/results_4quadrant.md.

Run 2 — wider multi-model, 3 models × 10 cases (partial), max_iters=5 (2026-05-26)

Replicated the matrix across more models and more cases. qwen-2.5-7b covers only the first 5 of 10 cases (wide run was killed mid-execution); claude-3.5-haiku not reached. Aggregate counts every observed (model × case × quadrant) tuple as one observation:

Per-model breakdown:

Decomposition (rejection threshold |Δ| ≥ 5 pts):

Same picture at every scale: Q4 (composition) wins on pass-rate, and Q4 stays close to Q2 on cost (1.9k tok / 24s per pass vs. Q2's 1.1k / 15s) while Q3 burns 7.2k tok / 106s per pass for fewer passes. Full table + per-case breakdown in bench/results_4quadrant_wide.md.

Plug points (stubs marked in code)

Layout

simplicio/
  cli.py          # index | task | bench | smoke
  cache.py        # content-hash embedding cache
  precedent.py    # grep + semantic rank (uses cache)
  skill_router.py # picks the ONE matching skill
  prompt.py       # stacks the 6 layers
  providers.py    # any OpenAI-compatible endpoint + Anthropic native
  pipeline.py     # generate → test → fix loop
  bench.py        # with-vs-without harness
  templates/simplicio_prompt.md
bench/
  run_offline.py  # stdlib-only multi-model benchmark
  cases.json      # your benchmark tasks
  cases_offline.json
  results.md      # filled by `simplicio bench` / `run_offline.py`
  charts/         # SVG: overall, delta, by_case, by_stack

License

MIT