@dsalazardev/ai-hub-cli 0.0.0-stage → 0.1.0

package/skills/shared/prompt-engineering/references/output-contracts.md

@@ -0,0 +1,120 @@
+# Output contracts: per-provider strict recipes
+
+Pick the strongest mechanism the provider gives you. Provider APIs and version notes
+below verified 2026-06-02 (access date); each external figure carries its own source
+date inline — do not read the access banner as the date a cited number was published.
+
+## OpenAI — strict `json_schema` via `response_format`
+
+Highest reliability: the provider compiles the schema into a token-masking FSM, so the
+output is guaranteed to parse against the schema (**<0.1% schema-failure rate** — figure
+published by OpenAI at the Aug 2024 Structured Outputs launch, not a 2026 number). Set
+`strict: true` and mark every field required (use unions/nullable for optional).
+
+```python
+from openai import OpenAI
+from pydantic import BaseModel
+
+class Triage(BaseModel):
+    category: str  # "bug" | "billing" | "other"
+
+client = OpenAI()
+resp = client.chat.completions.parse(
+    model="gpt-…",                       # resolve the model id from config, not here
+    messages=[{"role": "user", "content": prompt}],
+    response_format=Triage,              # SDK emits strict json_schema under the hood
+)
+result = resp.choices[0].message.parsed  # already a Triage, or None on refusal
+```
+
+Raw form (no SDK helper):
+
+```json
+{
+  "type": "json_schema",
+  "json_schema": {
+    "name": "triage",
+    "strict": true,
+    "schema": {
+      "type": "object",
+      "properties": { "category": { "type": "string", "enum": ["bug", "billing", "other"] } },
+      "required": ["category"],
+      "additionalProperties": false
+    }
+  }
+}
+```
+
+## Anthropic — strict tool use
+
+Reliable schema adherence, close behind OpenAI strict. Define a single tool whose
+`input_schema` is your contract and force it. **Assistant-prefill structured output is
+dead in the latest models** — Anthropic's migration guide states prefill returns a `400`
+on Sonnet 4.6 / Opus 4.6 / Opus 4.7 and routes you to structured outputs / `output_config.format`
+(platform.claude.com/docs/en/about-claude/models/migration-guide). Do not use prefill to
+coerce a shape.
+
+```python
+import anthropic
+client = anthropic.Anthropic()
+msg = client.messages.create(
+    model="claude-…",                    # from config
+    max_tokens=256,
+    tools=[{
+        "name": "emit_triage",
+        "description": "Return the ticket category.",
+        "input_schema": {
+            "type": "object",
+            "properties": {"category": {"type": "string", "enum": ["bug", "billing", "other"]}},
+            "required": ["category"],
+        },
+    }],
+    tool_choice={"type": "tool", "name": "emit_triage"},
+    messages=[{"role": "user", "content": prompt}],
+)
+block = next(b for b in msg.content if b.type == "tool_use")
+result = block.input  # {"category": "..."}
+```
+
+**Streaming caveat:** Anthropic tool-use output arrives as ONE block at end of stream —
+you cannot read fields progressively. OpenAI/Gemini stream field-by-field. If your UI
+parses-as-it-streams, that path does not exist on Claude tool use.
+
+## Google Gemini — `responseSchema`
+
+Set `response_mime_type: "application/json"` and a `response_schema`. Validate after,
+same as any JSON path.
+
+```json
+{ "generationConfig": { "response_mime_type": "application/json", "response_schema": { "type": "OBJECT", "properties": { "category": { "type": "STRING", "enum": ["bug","billing","other"] } } } } }
+```
+
+## TypeScript schema surface (zod)
+
+```typescript
+import { z } from "zod";
+const Triage = z.object({ category: z.enum(["bug", "billing", "other"]) });
+// OpenAI: zodResponseFormat(Triage, "triage"); Anthropic: zodToJsonSchema(Triage) as input_schema.
+const parsed = Triage.parse(JSON.parse(raw));   // throws -> trigger retry
+```
+
+## Retry-on-parse-fail (for JSON mode / freeform only)
+
+Strict mechanisms above rarely need this. When you only have JSON mode or freeform,
+bound the retries and feed the parser error back:
+
+```python
+def call_with_retry(prompt, schema, parse, max_tries=2):
+    last_err = None
+    for _ in range(max_tries + 1):
+        raw = call_model(prompt if last_err is None
+                         else f"{prompt}\n\nYour previous output failed: {last_err}. Return only valid output.")
+        try:
+            return parse(raw, schema)          # raises on invalid
+        except Exception as e:
+            last_err = str(e)
+    raise ValueError(f"contract not met after retries: {last_err}")
+```
+
+Keep `max_tries` small (1-2). If a strict mechanism is available, use it instead of
+spending tokens on retries.

package/skills/shared/prompt-engineering/scripts/verify.sh

@@ -0,0 +1,84 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Lints prompt artifacts in a target dir (default: cwd) against this skill.
+# Read-only: greps only, never writes, never installs, no network.
+# Scope: *.prompt, *.prompt.md, *.prompt.txt, prompts/**, and *.md/*.txt that
+#   contain a prompt marker (a "Role:"/"Task:" block or a fenced ```prompt).
+# Exit 0 = clean OR no prompt artifacts found (empty target is not a failure).
+# Exit 1 = a banned vague hedge was found in prompt text.
+# WARN (still exit 0) = JSON requested with no schema/contract nearby.
+
+TARGET="${1:-.}"
+if [ -t 1 ] && command -v tput >/dev/null 2>&1 && [ "$(tput colors 2>/dev/null || echo 0)" -ge 8 ]; then
+  YELLOW="$(tput setaf 3)"; RED="$(tput setaf 1)"; GREEN="$(tput setaf 2)"; RESET="$(tput sgr0)"
+else
+  YELLOW=""; RED=""; GREEN=""; RESET=""
+fi
+rc=0
+warn() { printf '%s[WARN]%s %s\n' "$YELLOW" "$RESET" "$1" >&2; }
+info() { printf '[INFO] %s\n' "$1"; }
+fail() { printf '%s[FAIL]%s %s\n' "$RED" "$RESET" "$1" >&2; rc=1; }
+
+if [ ! -e "$TARGET" ]; then
+  warn "target '$TARGET' does not exist; nothing to lint"
+  exit 0
+fi
+
+PRUNE='-path */node_modules/* -o -path */.venv/* -o -path */venv/* -o -path */.git/* -o -path */dist/* -o -path */build/*'
+
+# Collect candidate prompt files (newline-delimited; bash 3.2 safe, no mapfile).
+candidates=""
+add() { [ -n "$1" ] && candidates="${candidates}$1
+"; return 0; }
+while IFS= read -r f; do add "$f"; done <<EOF
+$(find "$TARGET" \( $PRUNE \) -prune -o -type f \
+   \( -name '*.prompt' -o -name '*.prompt.md' -o -name '*.prompt.txt' \) -print 2>/dev/null || true)
+EOF
+# Promote .md/.txt that carry an explicit ```prompt fence. Skill docs (SKILL.md and
+# files under references/) describe prompts and quote banned phrases on purpose, so they
+# are NOT prompt artifacts and must not be linted as such.
+while IFS= read -r f; do
+  [ -z "$f" ] && continue
+  case "$f" in */SKILL.md|*/references/*) continue ;; esac
+  if grep -qE '```[[:space:]]*prompt' "$f" 2>/dev/null; then add "$f"; fi
+done <<EOF
+$(find "$TARGET" \( $PRUNE \) -prune -o -type f \( -name '*.md' -o -name '*.txt' \) -print 2>/dev/null || true)
+EOF
+
+# De-dup.
+files="$(printf '%s' "$candidates" | awk 'NF && !seen[$0]++')"
+if [ -z "$files" ]; then
+  info "no prompt artifacts found under '$TARGET'; nothing to lint (ok)"
+  exit 0
+fi
+
+# Eval-set presence: a written prompt should ship with cases somewhere in the target.
+if find "$TARGET" \( $PRUNE \) -prune -o -type f -name 'cases.y*ml' -print 2>/dev/null | grep -q . \
+   || grep -rqsE '^[[:space:]]*should_trigger:|^[[:space:]]*cases:' "$TARGET" 2>/dev/null; then
+  info "eval cases present near the prompt(s)"
+else
+  warn "prompt artifacts found but no eval cases (cases.yaml or 'cases:'/'should_trigger:') — write a small set before tuning"
+fi
+
+# Banned vague hedges in prompt text -> hard fail.
+BANNED='try to|please attempt|if possible|as best you can|do your best'
+while IFS= read -r f; do
+  [ -z "$f" ] && continue
+  if grep -niE "$BANNED" "$f" >/dev/null 2>&1; then
+    fail "vague hedge in prompt text: $f"
+    grep -niE "$BANNED" "$f" | sed 's/^/    /' >&2
+  fi
+  # JSON requested but no schema/contract signal nearby -> warn only.
+  if grep -qiE 'json' "$f" 2>/dev/null \
+     && ! grep -qiE 'schema|strict|response_format|json_schema|tool[_ ]?use|responseSchema' "$f" 2>/dev/null; then
+    warn "JSON requested with no schema/contract signal in $f (JSON mode guarantees valid JSON, not your schema)"
+  fi
+done <<EOF
+$files
+EOF
+
+if [ "$rc" -eq 0 ]; then
+  printf '%s[OK]%s prompt lint passed\n' "$GREEN" "$RESET"
+fi
+exit "$rc"

package/skills/shared/python/SKILL.md

@@ -0,0 +1,320 @@
+---
+name: python
+description: "Use when the task is Python itself, in any framework or none: PEP 695 generics, mypy --strict typing, dataclass/Protocol/TypedDict/Enum choices, asyncio.TaskGroup, stdlib idioms, src/ layout + pyproject.toml with uv, ruff+mypy+pytest gate. NOT a FastAPI/ASGI service (that is fastapi), NOT a deep pytest suite (that is testing-py)."
+tags: [python, typing, async, packaging, uv]
+recommends: [fastapi, secure-coding, deployment]
+origin: ai-hub
+---
+
+# Modern Python at the language level
+
+Write, review, modernize, type, and package Python that reads like a typed,
+flat-control-flow, stdlib-first program whose dependencies and tooling all live in one
+`pyproject.toml`. Types are part of the design, not decoration; the stdlib is large and you
+reach for it before a dependency; correctness is `ruff` + a type checker + `pytest` in one gate.
+
+Targets **Python 3.12+ (floor) / 3.14 (current**, released 7 Oct 2025): PEP 695 inline
+type parameters (`class Box[T]:`, `type Alias = ...`), `asyncio.TaskGroup`, and — in 3.14 —
+deferred annotation evaluation by default (PEP 649/749, no more `from __future__ import
+annotations`), PEP 750 template strings (`t"..."`), and `compression.zstd`. Tooling pins:
+**uv 0.11** (project + package manager), **ruff 0.15** (lint + format), **mypy 1.20
+`--strict`** (or Astral's `ty`, still preview — default to mypy), **pytest 8**.
+
+## Scope — what this skill delegates
+
+- Building a FastAPI / ASGI service (routes, Pydantic models, SQLAlchemy, uvicorn) ->
+  [`fastapi`](../fastapi/SKILL.md). That skill owns the service shape; this one owns the
+  language it is written in.
+- A pytest suite as the primary task — fixtures, parametrize matrices, mocking, coverage,
+  property-based testing -> `testing-py` (this skill keeps only the *baseline*: a few tests
+  so `verify.sh` has something to run, then hands off).
+- Language-agnostic threat modeling / authz / OWASP review ->
+  [`secure-coding`](../secure-coding/SKILL.md) (this skill keeps Python-specific safety: no
+  `eval`/`pickle` of untrusted data, `subprocess` without `shell=True`, `secrets` over `random`).
+- Containerfile / CI pipeline / deploy mechanics -> [`deployment`](../deployment/SKILL.md)
+  (this skill ships only a uv-based CI note).
+- Another language -> `go`, `typescript`, `rust`, etc. Django ORM/models/migrations -> `django`.
+
+Python typing, async language semantics, and uv packaging live **here**, not in a separate
+skill — this skill is the canonical authority for the language substrate under any Python program.
+
+## Decision rules
+
+Apply on every Python edit:
+
+1. **Type the boundary.** Every public function, method, and module-level name gets an
+   annotation; let inference handle locals. Untyped public API defeats `--strict`.
+2. **Flat control flow.** Guard-clause and `return`/`raise` early; keep the happy path
+   unindented — arrow code hides the logic.
+3. **Stdlib before a dependency.** `pathlib`, `itertools`, `functools`, `dataclasses`,
+   `collections` cover most needs; a new dep is a maintenance liability you must justify.
+4. **Prefer immutability.** `@dataclass(frozen=True, slots=True)` for value objects; mutate
+   only where you must — shared mutable state is the bug you debug at 2am.
+5. **Everything passes the gate.** `ruff check` + `ruff format --check` + `mypy --strict` +
+   `pytest` — green locally via `scripts/verify.sh` before you call it done.
+
+## Typing
+
+Type the boundary; run `mypy --strict` so untyped code and implicit `Any` are errors, not
+silent gaps. Use **PEP 695** inline syntax for all new generic code — no explicit `TypeVar`
+objects:
+
+```python
+# Good (3.12+): inline type parameter and the `type` alias statement.
+def first[T](xs: list[T]) -> T:
+    return xs[0]
+
+class Box[T]: ...
+type UserId = int                 # `type` statement: a real alias, lazily evaluated
+# Bad (legacy): `T = TypeVar("T")` then a Generic[T] — fine in old code, don't write it new.
+```
+
+Core choices: `Protocol` (structural, no inheritance) over an ABC when you only need "has these
+methods"; `X | None` (not `Optional[X]`); `Literal`/`Enum` for closed value sets; `Final` for
+constants; `Self` for fluent returns; `@overload` for signature families. Narrow with
+`isinstance`, `assert`, or an `is None` guard — mypy follows the flow. In **3.14** annotations
+are lazy by default (PEP 649/749), so forward references resolve without `from __future__ import
+annotations`.
+
+Full PEP 695 bounds/constraints/variance, `Protocol` vs ABC, `TypedDict`
+`Required`/`NotRequired`, `ParamSpec`/`TypeVarTuple`, `TypeGuard`/`TypeIs`, `cast`, and common
+`--strict` errors with fixes -> [`references/typing.md`](references/typing.md).
+
+## Data modeling
+
+Pick the carrier by what the data *is*, not by habit:
+
+| Need | Use | Why |
+| --- | --- | --- |
+| Immutable value object, typed fields, methods | `@dataclass(frozen=True, slots=True)` | Hashable, no `__dict__` overhead, real types |
+| Small fixed tuple, positional + named, immutable | `NamedTuple` | Tuple semantics + field names; cheap |
+| Shape of an external/JSON dict, no runtime class | `TypedDict` | Types a plain `dict` without wrapping it |
+| Closed set of named constants | `Enum` / `StrEnum` / `IntEnum` | One source of truth; `Literal`-friendly |
+| Mutable bag of related state with behavior | plain class / `@dataclass` | When you genuinely need mutation |
+| Runtime-validated I/O model (parse untrusted data) | Pydantic -> [`fastapi`](../fastapi/SKILL.md) | Validation is a service concern, not language |
+
+```python
+from dataclasses import dataclass
+
+@dataclass(frozen=True, slots=True)
+class Point:
+    x: float
+    y: float
+    def translated(self, dx: float, dy: float) -> "Point":
+        return Point(self.x + dx, self.y + dy)   # returns a new value, never mutates
+```
+
+Frozen-slots dataclass when you want methods + hashability + clear types; `NamedTuple` when
+the thing genuinely *is* a small tuple you also unpack positionally.
+
+## Stdlib idioms
+
+Reach into the stdlib before adding a dependency.
+
+- **`pathlib`** for all filesystem paths: `Path("data") / name`, `p.read_text()`,
+  `p.glob("*.json")`, `p.with_suffix(".bak")` — typed and composable, never `os.path.join`.
+- **`collections`**: `defaultdict(list)`, `Counter(words)`, `deque(maxlen=100)` for ring buffers.
+- **`itertools`**: `chain`, `groupby`, `islice`, `batched` (3.12+) instead of hand-rolled loops.
+- **`functools`**: `@cache` / `@lru_cache` for pure memoization, `@cached_property`, `partial`.
+- **`contextlib`**: `@contextmanager`, `ExitStack` for dynamic resource sets, `suppress(FileNotFoundError)`.
+- **`logging`**, never `print` in a library: `print` writes to a caller's stdout you do not own;
+  `logging.getLogger(__name__)` lets them configure it.
+
+Prefer comprehensions over `map`/`filter`+`lambda`; prefer a generator (`(... for ...)`) when
+you only iterate once. Use `match` for structural dispatch over a chain of `isinstance`:
+
+```python
+from functools import cache
+from pathlib import Path
+
+@cache
+def config_dir() -> Path:                 # computed once, memoized
+    return Path.home() / ".config" / "myapp"
+
+def area(shape: object) -> float:
+    match shape:                          # structural dispatch, captures by attribute/key
+        case {"kind": "circle", "r": float(r)}:
+            return 3.14159 * r * r
+        case _:
+            raise TypeError(f"unknown shape: {shape!r}")
+```
+
+f-strings for formatting; note 3.14's PEP 750 `t"..."` template strings yield a `Template`
+(not a `str`) for *safe custom interpolation* (e.g. escaping) — use them when an f-string would
+inject untrusted text. Full cookbook (itertools/functools/collections recipes, `match`
+patterns, dataclass `field`/`default_factory`/`__post_init__`, `Enum`/`StrEnum`/`IntFlag`) ->
+[`references/stdlib.md`](references/stdlib.md).
+
+## Errors & resources
+
+Define a small exception hierarchy rooted in one base so callers can catch broadly or
+narrowly; chain causes with `raise ... from`; never write a bare `except:`.
+
+```python
+class AppError(Exception): ...
+class NotFoundError(AppError): ...
+
+def load(path: Path) -> str:
+    try:
+        return path.read_text()
+    except FileNotFoundError as e:
+        raise NotFoundError(f"missing {path}") from e   # preserves the cause chain
+```
+
+Use `except*` to handle an `ExceptionGroup` (what a `TaskGroup` raises) by member type.
+Prefer **EAFP** (try the operation, handle the failure) over LBYL race-prone pre-checks.
+Always release resources with `with` (a context manager), not manual try/finally close.
+
+## Async (language level)
+
+Use `asyncio.run(main())` as the single entry point. **`asyncio.TaskGroup` (3.11+) over bare
+`gather`** — it is the structured-concurrency primitive: a child failure cancels its siblings
+and surfaces as an `ExceptionGroup`, and no task outlives the block.
+
+```python
+import asyncio
+
+# Bad: gather leaks the other tasks on first failure and loses structure.
+async def fetch_all_bad(ids: list[int]) -> list[bytes]:
+    return await asyncio.gather(*(fetch(i) for i in ids))
+
+# Good: TaskGroup — sibling cancellation on error, bounded lifetime, real grouping.
+async def fetch_all(ids: list[int]) -> list[bytes]:
+    async with asyncio.TaskGroup() as tg:
+        tasks = [tg.create_task(fetch(i)) for i in ids]
+    return [t.result() for t in tasks]                  # block exited => all done or raised
+```
+
+Bound every wait with `async with asyncio.timeout(5.0):`. On `CancelledError`, clean up and
+**re-raise** — swallowing it breaks cancellation for the whole tree. Async is for **IO-bound**
+concurrency only; CPU-bound work blocks the loop — push it to `asyncio.to_thread` / a
+`ProcessPoolExecutor` (or 3.14's free-threaded build). **HTTP servers belong to
+[`fastapi`](../fastapi/SKILL.md), not here.** Runtime model, `ExceptionGroup`/`except*`, queues
+with backpressure, cancellation discipline, and sync<->async bridging ->
+[`references/async.md`](references/async.md).
+
+## Project layout & packaging (uv)
+
+Use a `src/` layout so tests import the installed package, not the source tree by accident:
+
+```text
+myapp/
+  src/myapp/__init__.py
+  src/myapp/core.py
+  tests/test_core.py
+  pyproject.toml
+  uv.lock              # committed
+  scripts/verify.sh
+```
+
+`pyproject.toml` is the single config — PEP 621 metadata, dependency groups, and tool config.
+Never hand-edit a `requirements.txt`; `uv add` writes the dep and updates `uv.lock`, which you commit.
+
+```toml
+[project]
+name = "myapp"
+version = "0.1.0"
+requires-python = ">=3.12"
+dependencies = ["httpx>=0.27"]
+
+[project.scripts]
+myapp = "myapp.core:main"          # console entry point
+
+[dependency-groups]
+dev = ["ruff>=0.15", "mypy>=1.13", "pytest>=8"]
+
+[build-system]
+requires = ["uv_build>=0.11"]
+build-backend = "uv_build"         # uv's own backend, stable since July 2025
+
+[tool.ruff]
+line-length = 100
+
+[tool.mypy]
+strict = true
+```
+
+Core uv verbs (each updates `uv.lock`, which you commit):
+
+```bash
+uv init --package myapp        # scaffold pyproject.toml + src/ + .venv
+uv add httpx                   # add a runtime dep
+uv add --dev ruff mypy pytest  # add to the dev group
+uv sync --frozen               # install exactly from the lockfile (CI + fresh clones)
+uv run pytest -q               # run inside the managed venv
+uv python install 3.14         # pin/install an interpreter
+```
+
+## Quality gate + verify.sh
+
+One local gate, mirroring CI: `ruff check --fix .`, `ruff format .`, `mypy --strict src`
+(or `ty check`), `pytest -q`. `scripts/verify.sh` runs all of them, skips a missing tool/dir
+with a warning, and exits non-zero on any real failure — run `./scripts/verify.sh` from the
+project root before declaring done. CI is `astral-sh/setup-uv` + `uv sync --frozen` + the same
+four commands; full pipeline -> [`deployment`](../deployment/SKILL.md).
+
+## pytest baseline
+
+Carry just enough to make `verify.sh` meaningful — plain `test_*` functions, `assert`,
+`pytest.raises`, one `parametrize`, `tmp_path` for files:
+
+```python
+import pytest
+from myapp.core import head
+
+@pytest.mark.parametrize("xs, expected", [(["a", "b"], "a"), (["x"], "x")])
+def test_head_returns_first(xs: list[str], expected: str) -> None:
+    assert head(xs) == expected
+
+def test_head_rejects_empty() -> None:
+    with pytest.raises(ValueError):
+        head([])
+
+def test_writes_file(tmp_path) -> None:
+    (tmp_path / "f.txt").write_text("hi")
+    assert (tmp_path / "f.txt").read_text() == "hi"
+```
+
+Deep fixtures, mocking, coverage gates, and property-based testing belong to `testing-py` —
+stop at the baseline and hand off.
+
+## Security (embedded, Python-specific)
+
+Generic threat modeling and authz live in [`secure-coding`](../secure-coding/SKILL.md); these Python-specific controls stay here:
+
+```python
+# Bad                                          # Good
+eval(user_input)                                ast.literal_eval(user_input)   # never eval/exec input
+pickle.loads(network_bytes)                     json.loads(network_bytes)      # never unpickle untrusted data
+subprocess.run(cmd, shell=True)                 subprocess.run(["ls", path])   # list args, no shell=True
+random.random()  # tokens                       secrets.token_urlsafe(32)      # secrets, not random, for secrets
+```
+
+Keep deps locked (`uv.lock`) and audited (`pip-audit` / `uv` resolution); read secrets from
+env or a secret manager, never hardcode or log them.
+
+## Anti-patterns -> STOP
+
+| Tempting move | Reality / do instead |
+| --- | --- |
+| "`def f(xs=[])` is fine, it's empty" | One list shared across all calls; use `= None` then `xs = xs or []`. |
+| "bare `except:` to be safe" | Swallows `KeyboardInterrupt`/bugs; catch a specific type. |
+| "`from module import *`" | Pollutes the namespace, breaks tooling; import names explicitly. |
+| "`print()` to debug this library" | Writes to a stdout you don't own; use `logging.getLogger(__name__)`. |
+| "`os.path.join` is what I know" | `pathlib.Path` is typed and composable; use `/`. |
+| "I'll add types later" | Untyped public API defeats `--strict`; type the boundary now. |
+| "edit `requirements.txt` by hand" | Drifts from the lock; `uv add` / `uv remove` and commit `uv.lock`. |
+| "`asyncio.gather` is simpler" | Leaks siblings on failure; `TaskGroup` for structured concurrency. |
+| "swallow `CancelledError`, it's noise" | Breaks cancellation for the whole tree; clean up and re-raise. |
+| "`time.sleep` inside this coroutine" | Blocks the event loop; `await asyncio.sleep(...)`. |
+| "`pickle.loads` the cache, it's ours" | Any untrusted byte = code execution; use `json`. |
+| "explicit `TypeVar` everywhere" | New code uses PEP 695 `def f[T]` / `class C[T]` / `type X`. |
+
+## Project grounding (02-DOCS)
+
+In a project that has the [`harness`](../harness/SKILL.md) wiki, record this project's Python
+conventions in `02-DOCS/wiki/stack/python.md` and index it in `02-DOCS/wiki/index.md` — the
+interpreter floor, `src/` layout, uv workflow, ruff/mypy config, async-vs-sync stance, and
+data-modeling defaults. Read it first on every use and bump its `Updated` date when a convention
+changes. This is *recorded, not gated*: never block the task on it, and skip silently when there
+is no `02-DOCS/` layer.

package/skills/shared/python/evals/README.md

@@ -0,0 +1,19 @@
+# Eval harness — `python` skill
+
+These cases are read by an **agent harness** (a Claude agent with the skill catalog loaded),
+not a plain shell script. `cases.yaml` is the fixture; this file is the procedure. Two things
+are measured: **triggering** (does `python` fire on language-level prompts and stay silent on
+near-misses that belong to `fastapi`, `testing-py`, `django`, `secure-coding`, `deployment`?)
+and **capability** (does loading the skill make a scaffold answer measurably better?).
+
+**Triggering:** load the full sibling descriptions, start a fresh session per prompt, feed each
+`should_trigger.prompt` and confirm `python` fires; feed each `should_not_trigger.prompt` and
+confirm it routes to the stated `route_to` and does *not* fire `python`. LLM routing is
+non-deterministic — run 3–5 trials each and pass on the majority; target ≥90% accuracy.
+
+**Capability:** run the `capability.scenario` twice — once with the skill body unavailable, once
+with `python/SKILL.md` (and references) loaded — and grade each output against the `must_include`
+rubric (fraction of points present and correct, averaged over 3 trials). Pass bar: ≥80% with the
+skill, and a clear margin over the without-skill baseline. No network is required; grading is
+rubric-based and partly judgment (e.g. "no mutable default argument"). Keep `cases.yaml` faithful
+to SKILL.md's "When to use / When NOT to use" — change them together.

package/skills/shared/python/evals/cases.yaml

@@ -0,0 +1,46 @@
+skill: python
+
+should_trigger:
+  - prompt: "Set up a new Python project with uv — package layout, pyproject, lint and types."
+    why: "Core packaging/layout/uv duty of this skill: src/ layout, pyproject.toml, ruff+mypy wiring."
+  - prompt: "Is this idiomatic Python? Modernize it to 3.12+ typing."
+    why: "Language-level idiom review plus PEP 695 modernization is the spine of the skill."
+  - prompt: "Add type hints to this module and make mypy --strict pass."
+    why: "Typing the boundary and satisfying a strict checker is owned here, even without naming the tool."
+  - prompt: "Convertir este asyncio.gather a algo más seguro con cancelación correcta."
+    why: "Non-obvious + Spanish: structured concurrency / TaskGroup at the language level; user describes the problem without naming TaskGroup."
+  - prompt: "Should I use a frozen slots dataclass or a NamedTuple for this value object?"
+    why: "Non-obvious data-modeling decision the skill's table resolves; no framework or tool named."
+  - prompt: "This function takes xs=[] as a default and the list keeps growing across calls — what's wrong?"
+    why: "Classic mutable-default-argument bug; a language-level anti-pattern this skill diagnoses without the user naming it."
+
+should_not_trigger:
+  - prompt: "Add a FastAPI route that returns a Pydantic response model with async SQLAlchemy."
+    route_to: "fastapi"
+    why: "Service/route/Pydantic/ORM wiring belongs to fastapi; this skill owns only the language it is written in."
+  - prompt: "Write a thorough pytest suite with shared fixtures, mocking, and a coverage gate."
+    route_to: "testing-py"
+    why: "Deep test technique is testing-py; python carries only the baseline and hands off."
+  - prompt: "Model this table as a Django ORM model with migrations and an admin view."
+    route_to: "django"
+    why: "Django ORM/models/migrations/admin is the django skill, not the language substrate."
+  - prompt: "Threat-model this login flow and review it for OWASP-class authorization issues."
+    route_to: "secure-coding"
+    why: "Language-agnostic threat modeling and authz/abuse review is secure-coding; only Python-specific controls stay here."
+  - prompt: "Write a multi-stage Containerfile and a GitHub Actions pipeline to deploy this Python app."
+    route_to: "deployment"
+    why: "Containerfile / CI / shipping mechanics are deployment; this skill ships only a uv-based CI note."
+
+capability:
+  - scenario: "Scaffold a typed src/-layout Python package with uv: one stdlib-idiomatic module
+      using a PEP 695 generic, one asyncio TaskGroup function, a pyproject.toml with ruff+mypy
+      config and a console entry point, a baseline pytest test, and scripts/verify.sh."
+    must_include:
+      - "uv commands: uv init --package, uv add / uv add --dev, and uv sync --frozen for reproducible installs"
+      - "pyproject.toml with a [project] table (PEP 621, requires-python >=3.12) and a build-system/backend"
+      - "a PEP 695 inline generic (e.g. def first[T](xs: list[T]) -> T or class Box[T]), not an explicit TypeVar object"
+      - "an asyncio.TaskGroup function (not bare asyncio.gather) with cancellation-safe structure"
+      - "mypy --strict configured (e.g. [tool.mypy] strict = true) and ruff check/format wired in"
+      - "at least one pytest test_* function (assert and/or pytest.raises)"
+      - "no mutable default argument (uses = None then x = x or [], or field(default_factory=...)) and no bare except"
+      - "an explicit note that HTTP routes/services belong to fastapi, not this skill"