v0.2.12 — zero-step auto-bootstrap

pip install simplicio-cli → just works.

The first time simplicio is invoked after install, if ~/.claude/ is present and the hook is missing, the skill + UserPromptSubmit hook are installed automatically. PEP 517 wheels can't run code on pip install, so the bootstrap happens on first CLI use — the closest equivalent that works on every machine.

```bash
pip install simplicio-cli
simplicio smoke # ← also installs skill + hook (idempotent)
```

• Idempotent — never double-writes
• Opt-out: `export SIMPLICIO_SKIP_AUTO_INIT=1`
• Subcommands `init` / `detect` excluded (no loops)
• Errors are caught + logged; the CLI never breaks because of auto-activation

New README sections explain the runtime flow (skill = semantic layer, hook = deterministic layer), why UserPromptSubmit is the right pre-hook (not PreToolUse), and a disable/re-enable matrix.

Available on PyPI: https://pypi.org/project/simplicio-cli/0.2.12/

See CHANGELOG.md for the full entry.

v0.2.11 — simplicio init + detect

One-shot installer that drops the skill and a UserPromptSubmit hook into ~/.claude/.

• simplicio init — installs ~/.claude/skills/simplicio-cli/SKILL.md + ~/.claude/hooks/simplicio-userpromptsubmit.sh + merges the hook entry into ~/.claude/settings.json (with .bak backup). Idempotent. --dry-run shows the plan.
• simplicio detect — pure-Python heuristic (no LLM) that scores a prompt for code-edit intent and emits a [SIMPLICIO_PROMPT_HINT] block on stderr above the threshold.
• Heavy imports (numpy, provider SDKs) made lazy: init / detect / --help start instantly.

See CHANGELOG.md for the full entry.

v0.2.10 — Claude Code skill

Adds the simplicio-cli skill so Claude Code auto-considers simplicio task for code-edit prompts on a known file.

• .skills/simplicio-cli/SKILL.md + global mirror at ~/.claude/skills/simplicio-cli/SKILL.md
• Pushy trigger description covering explicit invocations ($simplicio, "use simplicio") plus implicit cues (small/local model, verify-loop / pass-rate / 6-layer keywords)
• Registered in .skills/README.md

See CHANGELOG.md for the full entry.

v0.2.9 - 4-quadrant bench + wider multi-model run (Q4 76%)

What's new

4-quadrant benchmark matrix landed: every combination of simplicio prompt yes/no x agent loop yes/no measured side-by-side, so the contribution of each axis is now isolable.

Headline (wider run, 3 models x 10 cases, 25 observed tuples)

Quadrant	Prompt	Execution	Pass rate	Tokens / pass	ms / pass
Q1	raw goal	1-shot	0/25 (0%)	22,387	817,437
Q2	simplicio 6-layer	1-shot	16/25 (64%)	1,093	14,797
Q3	raw goal	loop w/ feedback	11/25 (44%)	7,154	106,382
Q4	simplicio 6-layer	loop w/ feedback	19/25 (76%)	1,914	24,170

Q4 (composition) wins on every axis: pass rate, tokens per passing case, and wall-clock per passing case.

Hypothesis verdicts (threshold |delta| >= 5 pts)

• Loop alone closes the gap. Q4 - Q3 = +32 pts -> REJECTED.
• Simplicio alone is enough. Q4 - Q2 = +12 pts -> REJECTED.
• Gains stack linearly. Q4 - linear = -32 pts -> REJECTED.

Both axes contribute, neither is redundant, and they compose sub-linearly (there is interaction, not double-counting).

Added

• New 4-quadrant harness: bench/run_4quadrant.py.
• Focused run report: bench/results_4quadrant.md (gemma-3-4b, 5 cases, max_iters=3).
• Wider run report: bench/results_4quadrant_wide.md (3 models x 10 cases, max_iters=5, qwen partial).
• Raw artefacts: bench/results_4quadrant.json + bench/results_4quadrant_wide.json.
• PDFs + SVG charts under bench/charts/4q_*.svg and bench/results_4quadrant{,_wide}.pdf.
• Methodology doc: docs/benchmark-4quadrant.md.

Changed

• pyproject.toml: 0.2.8 -> 0.2.9.
• README: new "Run 1 - focused" / "Run 2 - wider multi-model" structure.

Notes

• The wider run was killed mid-execution; qwen-2.5-7b covers only the first 5 of 10 cases and claude-3.5-haiku was never reached. Aggregate treats every observed (model, case, quadrant) tuple as one observation. Reproduce instructions are inside bench/results_4quadrant_wide.md.

Install

pip install simplicio-cli==0.2.9
# or for benchmark extras (fpdf2):
pip install "simplicio-cli[bench]==0.2.9"

v0.2.3 — English translation (breaking)

Breaking change — everything is English now

Mixed-language internals were hurting onboarding. v0.2.3 translates the
entire codebase: docstrings, comments, identifiers, CLI flags, prompt
template, and bench case data.

CLI flag renames

Old	New
--alvo	--target
--criterios	--criteria
--restricoes	--constraints
objetivo (positional)	goal

Internal API renames (for anyone importing simplicio.*)

• prompt.montar → prompt.build_prompt
• providers.gerar → providers.generate
• precedent.montar_bloco_precedente → precedent.build_precedent_block
• precedent.grep_candidatos → precedent.grep_candidates
• skill_router.montar_bloco_skill → skill_router.build_skill_block

Bench JSON key renames

objetivo/alvo/criterios/restricoes → goal/target/criteria/constraints
in both bench/cases.json and bench/cases_offline.json.

Prompt template

Emits [GOAL] / [TARGET] / [CONTRACT] / [OUTPUT] blocks instead of the
Portuguese [OBJETIVO] / [ALVO] / [CONTRATO] / [SAIDA].

Install

pip install simplicio-cli==0.2.3

Benchmark numbers and pipeline behavior unchanged from v0.2.2.

v0.2.2 — Cost measurement: tokens + wall-clock

Highlights — measured, not estimated

Metric	Without simplicio	With simplicio	Δ
Pass rate (156 checks)	35%	90%	+55 pts (+156%)
Wall-clock / run	12.4s	9.9s	−21%
Tokens / run	759	770	+1%
DIFF block present	0%	100%	—
Target file mentioned	0%	96%	—

Same model, same task, same provider. Only the prompt structure changes.

Per model

• Gemma 3 12B IT: 34% → 92% (+58)
• Llama 3.1 8B Instruct: 36% → 90% (+54)
• Qwen 2.5 7B Instruct: 34% → 88% (+54)

What changed

• bench/run_offline.py now captures usage.{prompt,completion,total}_tokens and time.perf_counter() elapsed per call.
• bench/results.md gained "Cost — tokens & wall-clock" section.
• README updated with honest cost numbers.

Install

pip install simplicio-cli==0.2.2

Full diff and notes in CHANGELOG.md.

v0.2.0 — bench v2 + PyPI publish

Highlights

• Published on PyPI: `pip install simplicio-cli` → https://pypi.org/project/simplicio-cli/0.2.0/
• Bench v2: 3 models · 10 cases · 156 checks · SVG charts.

Numbers (same model, different prompt)

Model	Without	With simplicio	Gain
Llama 3.1 8B Instruct	34%	98%	+64 pts
Gemma 3 12B IT	38%	94%	+56 pts
Qwen 2.5 7B Instruct	38%	80%	+42 pts
Average	37%	91%	+54 pts (+145%)

Output-quality signals

• DIFF block present: 0% → 100%
• Target file mentioned: 3% → 96%

Install

```bash
pip install simplicio-cli
```

Full report: bench/results.md · bench/results.pdf