jsmithpkp-llm-client-kit 0.1.3__tar.gz

jsmithpkp_llm_client_kit-0.1.3/PKG-INFO

@@ -0,0 +1,116 @@
+Metadata-Version: 2.4
+Name: jsmithpkp-llm-client-kit
+Version: 0.1.3
+Summary: Cached LLM client with Ollama + Anthropic provider dispatch, fixture mode, and JSON-schema-strict completions.
+Author: Jonathan Smith
+License: MIT
+Project-URL: Homepage, https://github.com/jsmithpkp21/llm-client-kit
+Project-URL: Issues, https://github.com/jsmithpkp21/llm-client-kit/issues
+Requires-Python: >=3.11
+Description-Content-Type: text/markdown
+Provides-Extra: anthropic
+Requires-Dist: anthropic>=0.40; extra == "anthropic"
+
+# llm-client-kit
+
+Cached, fixture-aware LLM client with **Ollama** (default) and **Anthropic** provider dispatch. Carved out of `resume-builder/src/resume_builder/llm_client.py` so it can be shared across internal Python apps without each app rolling its own dispatch + caching layer.
+
+PyPI: [`jsmithpkp-llm-client-kit`](https://pypi.org/project/jsmithpkp-llm-client-kit/). Import path stays `llm_client_kit` (PyPI distribution name and Python import name are allowed to differ).
+
+## Install
+
+```bash
+pip install jsmithpkp-llm-client-kit
+```
+
+For the Anthropic provider path, install the extra:
+
+```bash
+pip install 'jsmithpkp-llm-client-kit[anthropic]'
+```
+
+## Public API
+
+```python
+from llm_client_kit import (
+    LLMClient,
+    LLMResponse,
+    LLMEndpointUnreachableError,
+    AnthropicProviderError,
+    PROVIDER_OLLAMA,
+    PROVIDER_ANTHROPIC,
+)
+```
+
+### Construct from env (resume-builder-compatible)
+
+```python
+client = LLMClient.from_env()  # reads RESUME_BUILDER_LLM_* env vars
+result = client.complete_json(
+    namespace="my_stage",
+    system_prompt="You are a JSON-only classifier...",
+    user_payload={"subject": "...", "body": "..."},
+)
+```
+
+Env vars honored:
+
+| Env var | Default | Notes |
+|---|---|---|
+| `RESUME_BUILDER_LLM_PROVIDER` | `ollama` | `ollama` or `anthropic` |
+| `RESUME_BUILDER_LLM_API_URL` | provider-default | OpenAI-compat for Ollama; ignored by Anthropic adapter |
+| `RESUME_BUILDER_LLM_MODEL` | `llama3.1:8b` / `claude-sonnet-4-6` | per-provider default |
+| `RESUME_BUILDER_LLM_API_KEY` / `ANTHROPIC_API_KEY` | unset | required for Anthropic |
+| `RESUME_BUILDER_LLM_TIMEOUT_SECONDS` | `90` | per-request timeout |
+| `RESUME_BUILDER_LLM_TIMEOUT_<NAMESPACE>` | unset | per-stage timeout override |
+| `RESUME_BUILDER_LLM_MAX_TOKENS` | `2048` | Anthropic only |
+| `RESUME_BUILDER_LLM_FIXTURE` | `0` | fixture mode (no network calls; cache-only) |
+| `RESUME_BUILDER_LLM_CACHE_DIR` | `.llm_cache` / `tests/fixtures/llm_cache` (fixture mode) | response cache location |
+
+> The env var prefix `RESUME_BUILDER_LLM_*` is preserved from the carve-out source for v0.1.0 backward compat. A future release will parameterize the prefix so consumers can use their own namespace.
+
+### Construct directly (no env)
+
+```python
+from pathlib import Path
+from llm_client_kit import LLMClient, PROVIDER_OLLAMA
+
+client = LLMClient(
+    endpoint="http://localhost:11434/v1/chat/completions",
+    model="llama3.2:3b",
+    cache_dir=Path("./.cache"),
+    fixture_mode=False,
+    timeout_seconds=30.0,
+    provider=PROVIDER_OLLAMA,
+)
+```
+
+## Error contract
+
+`complete_json` documents three failure modes:
+
+- `ValueError` — invalid / non-object JSON response, malformed cache payload.
+- `RuntimeError` — provider-agnostic failures (cache I/O, fixture-mode miss) AND Ollama-path transport/parse failures.
+- `LLMEndpointUnreachableError` (`RuntimeError` subclass) — connection refused / DNS failure / host unreachable, on the Ollama path.
+- `AnthropicProviderError` (NOT a `RuntimeError` subclass) — any failure on the Anthropic path. **Intentionally** outside the `(RuntimeError, ValueError)` fallback chain: a paid API call halting loudly is better than silently degrading.
+
+See class docstrings for the full rationale.
+
+## Releasing
+
+Releases are automated. Tag-triggered workflow at `.github/workflows/release.yml`:
+
+1. Bump `version` in `pyproject.toml` and `__version__` in `src/llm_client_kit/__init__.py` on a PR.
+2. After the PR merges, tag `vX.Y.Z` on `main` and push the tag.
+3. The `release` workflow runs `gitleaks` against the working tree AND full git history — any finding aborts the workflow before the build. Then `python -m build` produces sdist + wheel, and `pypa/gh-action-pypi-publish` uploads via PyPI Trusted Publisher OIDC (no API token stored in repo secrets).
+
+```bash
+# example bump-and-release flow
+git switch main && git pull
+# (PR merged that bumps version to 0.1.4)
+git tag v0.1.4
+git push origin v0.1.4
+# watch: gh run watch
+```
+
+The secret-scan step is **mandatory** and is the only thing standing between an accidentally-committed credential and a published wheel. Do not edit the workflow to skip it without first understanding what it catches.

jsmithpkp_llm_client_kit-0.1.3/README.md

@@ -0,0 +1,103 @@
+# llm-client-kit
+
+Cached, fixture-aware LLM client with **Ollama** (default) and **Anthropic** provider dispatch. Carved out of `resume-builder/src/resume_builder/llm_client.py` so it can be shared across internal Python apps without each app rolling its own dispatch + caching layer.
+
+PyPI: [`jsmithpkp-llm-client-kit`](https://pypi.org/project/jsmithpkp-llm-client-kit/). Import path stays `llm_client_kit` (PyPI distribution name and Python import name are allowed to differ).
+
+## Install
+
+```bash
+pip install jsmithpkp-llm-client-kit
+```
+
+For the Anthropic provider path, install the extra:
+
+```bash
+pip install 'jsmithpkp-llm-client-kit[anthropic]'
+```
+
+## Public API
+
+```python
+from llm_client_kit import (
+    LLMClient,
+    LLMResponse,
+    LLMEndpointUnreachableError,
+    AnthropicProviderError,
+    PROVIDER_OLLAMA,
+    PROVIDER_ANTHROPIC,
+)
+```
+
+### Construct from env (resume-builder-compatible)
+
+```python
+client = LLMClient.from_env()  # reads RESUME_BUILDER_LLM_* env vars
+result = client.complete_json(
+    namespace="my_stage",
+    system_prompt="You are a JSON-only classifier...",
+    user_payload={"subject": "...", "body": "..."},
+)
+```
+
+Env vars honored:
+
+| Env var | Default | Notes |
+|---|---|---|
+| `RESUME_BUILDER_LLM_PROVIDER` | `ollama` | `ollama` or `anthropic` |
+| `RESUME_BUILDER_LLM_API_URL` | provider-default | OpenAI-compat for Ollama; ignored by Anthropic adapter |
+| `RESUME_BUILDER_LLM_MODEL` | `llama3.1:8b` / `claude-sonnet-4-6` | per-provider default |
+| `RESUME_BUILDER_LLM_API_KEY` / `ANTHROPIC_API_KEY` | unset | required for Anthropic |
+| `RESUME_BUILDER_LLM_TIMEOUT_SECONDS` | `90` | per-request timeout |
+| `RESUME_BUILDER_LLM_TIMEOUT_<NAMESPACE>` | unset | per-stage timeout override |
+| `RESUME_BUILDER_LLM_MAX_TOKENS` | `2048` | Anthropic only |
+| `RESUME_BUILDER_LLM_FIXTURE` | `0` | fixture mode (no network calls; cache-only) |
+| `RESUME_BUILDER_LLM_CACHE_DIR` | `.llm_cache` / `tests/fixtures/llm_cache` (fixture mode) | response cache location |
+
+> The env var prefix `RESUME_BUILDER_LLM_*` is preserved from the carve-out source for v0.1.0 backward compat. A future release will parameterize the prefix so consumers can use their own namespace.
+
+### Construct directly (no env)
+
+```python
+from pathlib import Path
+from llm_client_kit import LLMClient, PROVIDER_OLLAMA
+
+client = LLMClient(
+    endpoint="http://localhost:11434/v1/chat/completions",
+    model="llama3.2:3b",
+    cache_dir=Path("./.cache"),
+    fixture_mode=False,
+    timeout_seconds=30.0,
+    provider=PROVIDER_OLLAMA,
+)
+```
+
+## Error contract
+
+`complete_json` documents three failure modes:
+
+- `ValueError` — invalid / non-object JSON response, malformed cache payload.
+- `RuntimeError` — provider-agnostic failures (cache I/O, fixture-mode miss) AND Ollama-path transport/parse failures.
+- `LLMEndpointUnreachableError` (`RuntimeError` subclass) — connection refused / DNS failure / host unreachable, on the Ollama path.
+- `AnthropicProviderError` (NOT a `RuntimeError` subclass) — any failure on the Anthropic path. **Intentionally** outside the `(RuntimeError, ValueError)` fallback chain: a paid API call halting loudly is better than silently degrading.
+
+See class docstrings for the full rationale.
+
+## Releasing
+
+Releases are automated. Tag-triggered workflow at `.github/workflows/release.yml`:
+
+1. Bump `version` in `pyproject.toml` and `__version__` in `src/llm_client_kit/__init__.py` on a PR.
+2. After the PR merges, tag `vX.Y.Z` on `main` and push the tag.
+3. The `release` workflow runs `gitleaks` against the working tree AND full git history — any finding aborts the workflow before the build. Then `python -m build` produces sdist + wheel, and `pypa/gh-action-pypi-publish` uploads via PyPI Trusted Publisher OIDC (no API token stored in repo secrets).
+
+```bash
+# example bump-and-release flow
+git switch main && git pull
+# (PR merged that bumps version to 0.1.4)
+git tag v0.1.4
+git push origin v0.1.4
+# watch: gh run watch
+```
+
+The secret-scan step is **mandatory** and is the only thing standing between an accidentally-committed credential and a published wheel. Do not edit the workflow to skip it without first understanding what it catches.

jsmithpkp_llm_client_kit-0.1.3/pyproject.toml

@@ -0,0 +1,43 @@
+[build-system]
+requires = ["setuptools>=68", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+# PyPI distribution name is namespaced with `jsmithpkp-` so the
+# top-level slot isn't claimed under a generic name. The Python
+# import path stays `llm_client_kit` (PyPI name and import name are
+# allowed to differ, e.g. `python-dateutil` → `dateutil`), so
+# consumers don't have to rewrite imports when bumping from the
+# internal-pypiserver build to the public-PyPI build.
+name = "jsmithpkp-llm-client-kit"
+version = "0.1.3"
+description = "Cached LLM client with Ollama + Anthropic provider dispatch, fixture mode, and JSON-schema-strict completions."
+readme = "README.md"
+requires-python = ">=3.11"
+authors = [{ name = "Jonathan Smith" }]
+license = { text = "MIT" }
+# No runtime deps for the Ollama path (stdlib urllib). Anthropic is
+# optional; consumers that use PROVIDER_ANTHROPIC must install it via
+# the `anthropic` extra. Kept optional so a kit consumer that only
+# uses Ollama doesn't carry the SDK install cost.
+dependencies = []
+
+[project.optional-dependencies]
+anthropic = ["anthropic>=0.40"]
+
+[project.urls]
+Homepage = "https://github.com/jsmithpkp21/llm-client-kit"
+Issues = "https://github.com/jsmithpkp21/llm-client-kit/issues"
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.setuptools.package-data]
+# PEP 561 marker: tells mypy/pyright the package ships its own type
+# information so consumers don't have to install a separate stubs
+# package or set ignore_missing_imports.
+llm_client_kit = ["py.typed"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+pythonpath = ["src"]

jsmithpkp_llm_client_kit-0.1.3/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

jsmithpkp_llm_client_kit-0.1.3/src/jsmithpkp_llm_client_kit.egg-info/PKG-INFO

@@ -0,0 +1,116 @@
+Metadata-Version: 2.4
+Name: jsmithpkp-llm-client-kit
+Version: 0.1.3
+Summary: Cached LLM client with Ollama + Anthropic provider dispatch, fixture mode, and JSON-schema-strict completions.
+Author: Jonathan Smith
+License: MIT
+Project-URL: Homepage, https://github.com/jsmithpkp21/llm-client-kit
+Project-URL: Issues, https://github.com/jsmithpkp21/llm-client-kit/issues
+Requires-Python: >=3.11
+Description-Content-Type: text/markdown
+Provides-Extra: anthropic
+Requires-Dist: anthropic>=0.40; extra == "anthropic"
+
+# llm-client-kit
+
+Cached, fixture-aware LLM client with **Ollama** (default) and **Anthropic** provider dispatch. Carved out of `resume-builder/src/resume_builder/llm_client.py` so it can be shared across internal Python apps without each app rolling its own dispatch + caching layer.
+
+PyPI: [`jsmithpkp-llm-client-kit`](https://pypi.org/project/jsmithpkp-llm-client-kit/). Import path stays `llm_client_kit` (PyPI distribution name and Python import name are allowed to differ).
+
+## Install
+
+```bash
+pip install jsmithpkp-llm-client-kit
+```
+
+For the Anthropic provider path, install the extra:
+
+```bash
+pip install 'jsmithpkp-llm-client-kit[anthropic]'
+```
+
+## Public API
+
+```python
+from llm_client_kit import (
+    LLMClient,
+    LLMResponse,
+    LLMEndpointUnreachableError,
+    AnthropicProviderError,
+    PROVIDER_OLLAMA,
+    PROVIDER_ANTHROPIC,
+)
+```
+
+### Construct from env (resume-builder-compatible)
+
+```python
+client = LLMClient.from_env()  # reads RESUME_BUILDER_LLM_* env vars
+result = client.complete_json(
+    namespace="my_stage",
+    system_prompt="You are a JSON-only classifier...",
+    user_payload={"subject": "...", "body": "..."},
+)
+```
+
+Env vars honored:
+
+| Env var | Default | Notes |
+|---|---|---|
+| `RESUME_BUILDER_LLM_PROVIDER` | `ollama` | `ollama` or `anthropic` |
+| `RESUME_BUILDER_LLM_API_URL` | provider-default | OpenAI-compat for Ollama; ignored by Anthropic adapter |
+| `RESUME_BUILDER_LLM_MODEL` | `llama3.1:8b` / `claude-sonnet-4-6` | per-provider default |
+| `RESUME_BUILDER_LLM_API_KEY` / `ANTHROPIC_API_KEY` | unset | required for Anthropic |
+| `RESUME_BUILDER_LLM_TIMEOUT_SECONDS` | `90` | per-request timeout |
+| `RESUME_BUILDER_LLM_TIMEOUT_<NAMESPACE>` | unset | per-stage timeout override |
+| `RESUME_BUILDER_LLM_MAX_TOKENS` | `2048` | Anthropic only |
+| `RESUME_BUILDER_LLM_FIXTURE` | `0` | fixture mode (no network calls; cache-only) |
+| `RESUME_BUILDER_LLM_CACHE_DIR` | `.llm_cache` / `tests/fixtures/llm_cache` (fixture mode) | response cache location |
+
+> The env var prefix `RESUME_BUILDER_LLM_*` is preserved from the carve-out source for v0.1.0 backward compat. A future release will parameterize the prefix so consumers can use their own namespace.
+
+### Construct directly (no env)
+
+```python
+from pathlib import Path
+from llm_client_kit import LLMClient, PROVIDER_OLLAMA
+
+client = LLMClient(
+    endpoint="http://localhost:11434/v1/chat/completions",
+    model="llama3.2:3b",
+    cache_dir=Path("./.cache"),
+    fixture_mode=False,
+    timeout_seconds=30.0,
+    provider=PROVIDER_OLLAMA,
+)
+```
+
+## Error contract
+
+`complete_json` documents three failure modes:
+
+- `ValueError` — invalid / non-object JSON response, malformed cache payload.
+- `RuntimeError` — provider-agnostic failures (cache I/O, fixture-mode miss) AND Ollama-path transport/parse failures.
+- `LLMEndpointUnreachableError` (`RuntimeError` subclass) — connection refused / DNS failure / host unreachable, on the Ollama path.
+- `AnthropicProviderError` (NOT a `RuntimeError` subclass) — any failure on the Anthropic path. **Intentionally** outside the `(RuntimeError, ValueError)` fallback chain: a paid API call halting loudly is better than silently degrading.
+
+See class docstrings for the full rationale.
+
+## Releasing
+
+Releases are automated. Tag-triggered workflow at `.github/workflows/release.yml`:
+
+1. Bump `version` in `pyproject.toml` and `__version__` in `src/llm_client_kit/__init__.py` on a PR.
+2. After the PR merges, tag `vX.Y.Z` on `main` and push the tag.
+3. The `release` workflow runs `gitleaks` against the working tree AND full git history — any finding aborts the workflow before the build. Then `python -m build` produces sdist + wheel, and `pypa/gh-action-pypi-publish` uploads via PyPI Trusted Publisher OIDC (no API token stored in repo secrets).
+
+```bash
+# example bump-and-release flow
+git switch main && git pull
+# (PR merged that bumps version to 0.1.4)
+git tag v0.1.4
+git push origin v0.1.4
+# watch: gh run watch
+```
+
+The secret-scan step is **mandatory** and is the only thing standing between an accidentally-committed credential and a published wheel. Do not edit the workflow to skip it without first understanding what it catches.

jsmithpkp_llm_client_kit-0.1.3/src/jsmithpkp_llm_client_kit.egg-info/SOURCES.txt

@@ -0,0 +1,11 @@
+README.md
+pyproject.toml
+src/jsmithpkp_llm_client_kit.egg-info/PKG-INFO
+src/jsmithpkp_llm_client_kit.egg-info/SOURCES.txt
+src/jsmithpkp_llm_client_kit.egg-info/dependency_links.txt
+src/jsmithpkp_llm_client_kit.egg-info/requires.txt
+src/jsmithpkp_llm_client_kit.egg-info/top_level.txt
+src/llm_client_kit/__init__.py
+src/llm_client_kit/client.py
+src/llm_client_kit/py.typed
+tests/test_smoke.py

jsmithpkp_llm_client_kit-0.1.3/src/jsmithpkp_llm_client_kit.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

jsmithpkp_llm_client_kit-0.1.3/src/jsmithpkp_llm_client_kit.egg-info/requires.txt

@@ -0,0 +1,3 @@
+
+[anthropic]
+anthropic>=0.40

jsmithpkp_llm_client_kit-0.1.3/src/jsmithpkp_llm_client_kit.egg-info/top_level.txt

@@ -0,0 +1 @@
+llm_client_kit

jsmithpkp_llm_client_kit-0.1.3/src/llm_client_kit/__init__.py

@@ -0,0 +1,21 @@
+"""Shared LLM client with Ollama + Anthropic provider dispatch."""
+
+from llm_client_kit.client import (
+    PROVIDER_ANTHROPIC,
+    PROVIDER_OLLAMA,
+    AnthropicProviderError,
+    LLMClient,
+    LLMEndpointUnreachableError,
+    LLMResponse,
+)
+
+__all__ = [
+    "AnthropicProviderError",
+    "LLMClient",
+    "LLMEndpointUnreachableError",
+    "LLMResponse",
+    "PROVIDER_ANTHROPIC",
+    "PROVIDER_OLLAMA",
+]
+
+__version__ = "0.1.3"

jsmithpkp_llm_client_kit-0.1.3/src/llm_client_kit/client.py

@@ -0,0 +1,733 @@
+#!/usr/bin/env python3
+"""Cached LLM client with Ollama + Anthropic provider dispatch.
+
+Lifted from resume-builder's internal ``llm_client.py`` to be a shared
+dependency consumed via the internal pypiserver. Behavior is identical
+to the resume-builder original at v0.1.0; the only changes are dropping
+the host-app-specific path-safety guard (``assert_not_blocked_runtime_input``)
+and broadening the docstrings. The env var convention (``RESUME_BUILDER_LLM_*``)
+is preserved verbatim for v0.1.0 so resume-builder's adoption is a no-op
+import swap — a follow-up release will parameterize the prefix.
+
+Consumers that don't want to use the env var convention can construct
+``LLMClient`` directly via ``__init__`` instead of ``from_env``.
+"""
+
+from __future__ import annotations
+
+import errno
+import hashlib
+import json
+import os
+import socket
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, TypeVar
+
+_LLMClientT = TypeVar("_LLMClientT", bound="LLMClient")
+from urllib.error import URLError
+from urllib.request import Request, urlopen
+
+
+_DEFAULT_CHAT_ENDPOINT = "http://localhost:11434/v1/chat/completions"
+_DEFAULT_MODEL = "llama3.1:8b"
+# Per-provider endpoint default — picked at `from_env` time when no
+# `RESUME_BUILDER_LLM_API_URL` override is set. The Anthropic adapter
+# doesn't actually use `endpoint` at request time (the SDK manages the
+# URL internally), but the value still gets baked into cache keys and
+# diagnostic output, so defaulting to the Anthropic API URL when the
+# provider is Anthropic keeps both honest.
+_DEFAULT_ANTHROPIC_ENDPOINT = "https://api.anthropic.com/v1/messages"
+# Bumped from 30s to 90s in #414: on local llama3.1:8b the longer
+# transform_for_role / enrich_data prompts routinely exceeded 30s on a
+# 12GB consumer GPU. Connection-refused / DNS failures still return
+# instantly (they bypass timeout), so the higher ceiling does not slow
+# down the "endpoint unreachable" diagnostic path. Per-stage env vars
+# below let callers tighten or loosen individual stages.
+_DEFAULT_TIMEOUT_SECONDS = 90.0
+_FIXTURE_ENV = "RESUME_BUILDER_LLM_FIXTURE"
+_CACHE_DIR_ENV = "RESUME_BUILDER_LLM_CACHE_DIR"
+_ENDPOINT_ENV = "RESUME_BUILDER_LLM_API_URL"
+_MODEL_ENV = "RESUME_BUILDER_LLM_MODEL"
+_TIMEOUT_ENV = "RESUME_BUILDER_LLM_TIMEOUT_SECONDS"
+_PROVIDER_ENV = "RESUME_BUILDER_LLM_PROVIDER"
+_API_KEY_ENV = "RESUME_BUILDER_LLM_API_KEY"
+_MAX_TOKENS_ENV = "RESUME_BUILDER_LLM_MAX_TOKENS"
+
+# Supported provider identifiers.
+PROVIDER_OLLAMA = "ollama"
+PROVIDER_ANTHROPIC = "anthropic"
+_DEFAULT_PROVIDER = PROVIDER_OLLAMA
+
+# When the user opts into the Anthropic provider but doesn't override
+# the model env var, default to Sonnet 4.6 — cost-efficient frontier
+# model with strong instruction following on long, rule-dense prompts
+# (`_BODY_SYSTEM_PROMPT_BASE` is ~4K tokens of HARD RULES + addendum).
+# Override via RESUME_BUILDER_LLM_MODEL (e.g. claude-opus-4-7).
+_DEFAULT_ANTHROPIC_MODEL = "claude-sonnet-4-6"
+
+# Per-request output cap for the Anthropic adapter. Cover-letter body
+# drafts return 600-1200 tokens in practice; 2048 gives headroom without
+# inviting runaway output. Override via RESUME_BUILDER_LLM_MAX_TOKENS.
+_DEFAULT_MAX_TOKENS = 2048
+# Per-stage timeout env var prefix (#414). Resolution order at the
+# request site: per-stage override (RESUME_BUILDER_LLM_TIMEOUT_<NS>) ->
+# global override (RESUME_BUILDER_LLM_TIMEOUT_SECONDS) -> default.
+_PER_STAGE_TIMEOUT_ENV_PREFIX = "RESUME_BUILDER_LLM_TIMEOUT_"
+
+
+def _per_stage_timeout_env(namespace: str) -> str:
+    """Return the env var name that overrides timeout for ``namespace``.
+
+    Namespace strings are lowercase snake_case in callers
+    (``transform_for_role``, ``enrich_data``, ...). The matching env var
+    upper-cases the suffix only — the rest of the prefix is fixed so the
+    var name is predictable and greppable.
+    """
+    return f"{_PER_STAGE_TIMEOUT_ENV_PREFIX}{namespace.upper()}"
+
+
+def _parse_timeout_env(raw: str | None) -> float:
+    """Parse RESUME_BUILDER_LLM_TIMEOUT_SECONDS, falling back to the default.
+
+    Empty / unset / non-numeric values fall back to the default; values
+    <= 0 also fall back since urlopen requires a positive timeout (or
+    None, but we don't expose 'no timeout' as a knob).
+    """
+    if raw is None:
+        return _DEFAULT_TIMEOUT_SECONDS
+    try:
+        value = float(raw.strip())
+    except ValueError:
+        return _DEFAULT_TIMEOUT_SECONDS
+    if value <= 0:
+        return _DEFAULT_TIMEOUT_SECONDS
+    return value
+
+
+@dataclass(frozen=True)
+class LLMResponse:
+    content: str
+    cache_key: str
+    from_cache: bool
+
+
+class LLMEndpointUnreachableError(RuntimeError):
+    """Raised when the LLM endpoint refuses connection or DNS resolution fails.
+
+    A subclass of ``RuntimeError`` so existing callers that catch
+    ``(RuntimeError, ValueError)`` to degrade to baseline keep working.
+    Build-pipeline code that wants to emit the "endpoint unreachable"
+    operator warning and fast-fail subsequent stages can narrow on this
+    subclass without disturbing the broader transport-error path
+    (timeouts, mid-stream resets, etc.) which keeps using bare
+    ``RuntimeError`` since those are recoverable per-stage failures.
+    """
+
+
+class AnthropicProviderError(Exception):
+    """Raised on any Anthropic-provider failure — **explicitly NOT a
+    subclass of ``RuntimeError``**.
+
+    The pipeline's existing fallback handlers catch
+    ``(RuntimeError, ValueError)`` to degrade gracefully to deterministic
+    logic when the local LLM is flaky. That's the right behavior when
+    a user is running on Ollama and the local model is intermittently
+    slow or broken — fall back to the deterministic baseline and keep
+    moving.
+
+    When the user has explicitly opted into the Anthropic provider via
+    ``RESUME_BUILDER_LLM_PROVIDER=anthropic``, that fallback is wrong:
+    they paid for the API call expecting frontier-model quality, and
+    silently degrading to non-LLM-tailored output produces something
+    far worse than what they asked for (and takes time + tokens to
+    fail at). Raising a separate exception base means Anthropic
+    failures propagate **uncaught** through ``(RuntimeError, ValueError)``
+    handlers and halt the build with a clear failure mode.
+
+    Callers must not catch this exception in their existing fallback
+    blocks. If a caller genuinely needs to handle Anthropic failures
+    (e.g. retry logic), it should catch ``AnthropicProviderError``
+    explicitly.
+
+    **Raised by:** ``LLMClient.complete_json()`` when
+    ``provider == "anthropic"``. This means ``complete_json``'s
+    documented ``(RuntimeError, ValueError)`` escape contract no
+    longer covers every path — on the Anthropic path,
+    ``AnthropicProviderError`` can also escape. The "anything else
+    escaping from here is a real bug" rule in ``complete_json``'s
+    docstring applies only to the Ollama path.
+    """
+
+
+# Allowed provider identifiers. Anything else fails fast at construction
+# time so a typo in `RESUME_BUILDER_LLM_PROVIDER` doesn't silently route
+# requests to the Ollama endpoint while the user thinks Anthropic is
+# active.
+_VALID_PROVIDERS = frozenset({PROVIDER_OLLAMA, PROVIDER_ANTHROPIC})
+
+
+# Errno values that indicate the LLM endpoint is not reachable at all
+# (connection refused, host unreachable, DNS resolution failure). When
+# urllib raises one of these the request returns essentially instantly,
+# so we treat them as a single "endpoint down" signal rather than as a
+# per-stage timeout.
+_UNREACHABLE_ERRNOS = {
+    errno.ECONNREFUSED,
+    errno.EHOSTUNREACH,
+    errno.ENETUNREACH,
+    errno.EHOSTDOWN,
+}
+
+
+def _is_endpoint_unreachable(exc: BaseException) -> bool:
+    """Return True when ``exc`` signals an unreachable LLM endpoint.
+
+    Covers both urllib's ``URLError`` (which wraps the underlying socket
+    failure in ``.reason``) and the raw socket/OSError shapes. DNS
+    resolution failure surfaces as ``socket.gaierror`` (an OSError
+    subclass) and we treat it the same as connection-refused: the
+    operator's local LLM is not answering.
+    """
+    inner: BaseException = exc
+    if isinstance(exc, URLError):
+        reason = exc.reason
+        if isinstance(reason, BaseException):
+            inner = reason
+        else:
+            return False
+    if isinstance(inner, socket.gaierror):
+        return True
+    if isinstance(inner, ConnectionRefusedError):
+        return True
+    err_no = getattr(inner, "errno", None)
+    return err_no in _UNREACHABLE_ERRNOS
+
+
+class LLMClient:
+    """OpenAI-compatible chat client with file cache and fixture mode."""
+
+    def __init__(
+        self,
+        *,
+        endpoint: str,
+        model: str,
+        cache_dir: Path,
+        fixture_mode: bool,
+        timeout_seconds: float = _DEFAULT_TIMEOUT_SECONDS,
+        provider: str = _DEFAULT_PROVIDER,
+        api_key: str | None = None,
+        max_tokens: int = _DEFAULT_MAX_TOKENS,
+    ) -> None:
+        if provider not in _VALID_PROVIDERS:
+            raise ValueError(
+                f"unsupported LLM provider {provider!r}; "
+                f"expected one of {sorted(_VALID_PROVIDERS)}. "
+                f"Set RESUME_BUILDER_LLM_PROVIDER to a supported value."
+            )
+        self._endpoint = endpoint
+        self._model = model
+        self._cache_dir = cache_dir
+        self._fixture_mode = fixture_mode
+        self._timeout_seconds = timeout_seconds
+        self._provider = provider
+        self._api_key = api_key
+        self._max_tokens = max_tokens
+
+    @property
+    def endpoint(self) -> str:
+        """Configured chat endpoint URL. Public for diagnostic / warning text."""
+        return self._endpoint
+
+    @property
+    def model(self) -> str:
+        """Configured model name. Public for diagnostic / warning text."""
+        return self._model
+
+    @property
+    def provider(self) -> str:
+        """Configured LLM provider (``ollama`` or ``anthropic``). Public for
+        diagnostic / warning text."""
+        return self._provider
+
+    @classmethod
+    def from_env(cls: type[_LLMClientT]) -> _LLMClientT:
+        fixture_mode = os.getenv(_FIXTURE_ENV, "0").strip() == "1"
+        default_cache_dir = "tests/fixtures/llm_cache" if fixture_mode else ".llm_cache"
+        cache_dir = Path(
+            os.getenv(_CACHE_DIR_ENV, default_cache_dir).strip() or default_cache_dir
+        )
+        provider = (
+            os.getenv(_PROVIDER_ENV, _DEFAULT_PROVIDER).strip() or _DEFAULT_PROVIDER
+        ).lower()
+        # Per-provider endpoint default. An explicit
+        # RESUME_BUILDER_LLM_API_URL override still wins; this only
+        # matters when the env var is unset. Picking the right default
+        # keeps cache keys and diagnostic output from advertising the
+        # Ollama URL when the actual provider is Anthropic.
+        default_endpoint = (
+            _DEFAULT_ANTHROPIC_ENDPOINT
+            if provider == PROVIDER_ANTHROPIC
+            else _DEFAULT_CHAT_ENDPOINT
+        )
+        endpoint = os.getenv(_ENDPOINT_ENV, default_endpoint).strip()
+        # Per-provider model default. The user can always override via
+        # RESUME_BUILDER_LLM_MODEL — this just picks the right "no env
+        # var set" default for whichever provider is active.
+        default_model = (
+            _DEFAULT_ANTHROPIC_MODEL
+            if provider == PROVIDER_ANTHROPIC
+            else _DEFAULT_MODEL
+        )
+        model = os.getenv(_MODEL_ENV, default_model).strip() or default_model
+        timeout_seconds = _parse_timeout_env(os.getenv(_TIMEOUT_ENV))
+        # API-key precedence: repo-convention env first, fall through to
+        # Anthropic's own convention. None is fine for the Ollama path.
+        api_key = os.getenv(_API_KEY_ENV) or os.getenv("ANTHROPIC_API_KEY") or None
+        # Mirror `_parse_timeout_env`: unset / empty / non-numeric /
+        # zero / negative all fall back to the default. A value <= 0
+        # would otherwise round-trip into the Anthropic SDK and fail at
+        # request time with a confusing "max_tokens must be positive"
+        # error far from the misconfiguration site.
+        max_tokens_raw = os.getenv(_MAX_TOKENS_ENV, "").strip()
+        if max_tokens_raw:
+            try:
+                parsed_max = int(max_tokens_raw)
+            except ValueError:
+                max_tokens = _DEFAULT_MAX_TOKENS
+            else:
+                max_tokens = parsed_max if parsed_max > 0 else _DEFAULT_MAX_TOKENS
+        else:
+            max_tokens = _DEFAULT_MAX_TOKENS
+        return cls(
+            endpoint=endpoint,
+            model=model,
+            cache_dir=cache_dir,
+            fixture_mode=fixture_mode,
+            timeout_seconds=timeout_seconds,
+            provider=provider,
+            api_key=api_key,
+            max_tokens=max_tokens,
+        )
+
+    def complete_json(
+        self,
+        *,
+        namespace: str,
+        system_prompt: str,
+        user_payload: dict[str, object],
+        response_schema: dict[str, Any] | None = None,
+    ) -> dict[str, Any]:
+        """Return a parsed JSON object from the LLM.
+
+        When ``response_schema`` is provided, the underlying request uses
+        OpenAI-compatible strict json_schema mode
+        (``response_format: {"type": "json_schema", "json_schema": {...},
+        "strict": true}``) so the model is constrained to produce output
+        matching the schema. Ollama's OpenAI-compat layer honors this
+        (verified live during #436 investigation). Without a schema, the
+        request falls back to the prior weak ``{"type": "json_object"}``
+        mode which only enforces valid-JSON structure.
+
+        The schema is hashed into the cache key alongside the messages —
+        the same prompt with a different schema produces a different
+        cached response, since the model's output can legitimately differ.
+
+        Error contract (callers may rely on this for narrowed fallback
+        handlers):
+          - ``ValueError`` for response-shape failures: invalid JSON,
+            non-object JSON, or malformed cache payload.
+          - ``RuntimeError`` for **provider-agnostic** failures raised
+            by ``_chat`` itself (outside the upstream-call branch):
+            cache filesystem I/O errors (read or write) and fixture-mode
+            misses. Cache reads and fixture-mode misses fire before the
+            upstream call; cache writes fire after a successful upstream
+            response. All three surface on **either** path — including
+            the Anthropic path — because they're independent of which
+            provider was configured.
+          - ``RuntimeError`` for **Ollama-path-only** failures during
+            or after the upstream call: transport errors (network /
+            DNS / timeout / connection reset) and malformed upstream
+            chat-completion responses.
+          - ``AnthropicProviderError`` (NOT a ``RuntimeError`` subclass)
+            for **Anthropic-path** failures at any point in the Anthropic
+            path (before, during, or after the upstream call): transport
+            errors, API status errors (rate limit, 4xx, 5xx), missing API
+            key (validated before the upstream call), empty response. This
+            intentionally escapes the ``(RuntimeError, ValueError)``
+            fallback so a paid Claude call halts the build loudly instead
+            of silently degrading to deterministic baseline (which is far
+            worse than what the user paid for). See
+            ``AnthropicProviderError``'s docstring for the full rationale.
+
+        Callers on the **Ollama path** should catch
+        ``(RuntimeError, ValueError)`` to degrade gracefully; anything
+        else escaping from there is a real bug.
+
+        Callers on the **Anthropic path** should NOT catch
+        ``AnthropicProviderError`` in their fallback blocks — let it
+        propagate. If a caller genuinely needs to handle Anthropic
+        failures (e.g. for retry logic), it should catch
+        ``AnthropicProviderError`` explicitly.
+        """
+        messages = [
+            {"role": "system", "content": system_prompt},
+            {"role": "user", "content": json.dumps(user_payload, sort_keys=True)},
+        ]
+        response = self._chat(
+            messages=messages,
+            namespace=namespace,
+            response_schema=response_schema,
+        )
+        try:
+            parsed = json.loads(response.content)
+        except json.JSONDecodeError as exc:
+            raise ValueError(
+                f"LLM response was not valid JSON for namespace={namespace} "
+                f"cache_key={response.cache_key}"
+            ) from exc
+        if not isinstance(parsed, dict):
+            raise ValueError(
+                f"LLM JSON response must be an object for namespace={namespace} "
+                f"cache_key={response.cache_key}"
+            )
+        return parsed
+
+    def _chat(
+        self,
+        *,
+        messages: list[dict[str, str]],
+        namespace: str,
+        response_schema: dict[str, Any] | None = None,
+    ) -> LLMResponse:
+        cache_key = self._cache_key(
+            messages=messages, namespace=namespace, response_schema=response_schema
+        )
+        cache_path = self._cache_dir / f"llm_response_{cache_key}.json"
+        pass  # runtime-guard call dropped in kit fork (host-app concern)
+
+        try:
+            cached = self._read_cache(cache_path)
+        except OSError as exc:
+            # Normalize cache-read filesystem failures (permission denied,
+            # unreadable file, etc.) into RuntimeError so callers that catch
+            # (RuntimeError, ValueError) around complete_json continue to get
+            # the documented fallback behavior instead of an unexpected
+            # OSError propagating out.
+            raise RuntimeError(
+                f"LLM cache read failed for namespace={namespace} "
+                f"cache_key={cache_key}: {exc}"
+            ) from exc
+        if cached is not None:
+            return LLMResponse(content=cached, cache_key=cache_key, from_cache=True)
+
+        if self._fixture_mode:
+            raise RuntimeError(
+                f"LLM fixture missing for namespace={namespace} cache_key={cache_key}"
+            )
+
+        content = self._request_chat_completion(
+            messages, namespace=namespace, response_schema=response_schema
+        )
+        try:
+            self._write_cache(cache_path=cache_path, content=content)
+        except OSError as exc:
+            # Same rationale as the read path above: a disk/permission failure
+            # writing the cache must not bypass callers' (RuntimeError,
+            # ValueError) handlers.
+            raise RuntimeError(
+                f"LLM cache write failed for namespace={namespace} "
+                f"cache_key={cache_key}: {exc}"
+            ) from exc
+        return LLMResponse(content=content, cache_key=cache_key, from_cache=False)
+
+    def _cache_key(
+        self,
+        *,
+        messages: list[dict[str, str]],
+        namespace: str,
+        response_schema: dict[str, Any] | None = None,
+    ) -> str:
+        payload: dict[str, Any] = {
+            "namespace": namespace,
+            "model": self._model,
+            "messages": messages,
+            "endpoint": self._endpoint,
+            # Provider is part of the key so Ollama and Anthropic don't
+            # share cache entries — even when the prompts and model name
+            # collide by accident, their output shapes and JSON-mode
+            # enforcement differ.
+            "provider": self._provider,
+        }
+        # Different schemas legitimately produce different model output for
+        # the same prompt, so include the schema in the cache key.
+        if response_schema is not None:
+            payload["response_schema"] = response_schema
+        serialized = json.dumps(payload, sort_keys=True, separators=(",", ":"))
+        return hashlib.sha256(serialized.encode("utf-8")).hexdigest()
+
+    def _resolve_timeout(self, namespace: str) -> float:
+        """Pick the timeout for a single request.
+
+        Resolution order (first non-empty wins):
+          1. Per-stage env override ``RESUME_BUILDER_LLM_TIMEOUT_<NAMESPACE>``
+          2. The client's configured timeout (which itself came from
+             ``RESUME_BUILDER_LLM_TIMEOUT_SECONDS`` via :meth:`from_env`,
+             or the default when that env var was unset).
+        """
+        raw = os.getenv(_per_stage_timeout_env(namespace))
+        if raw is None:
+            return self._timeout_seconds
+        try:
+            value = float(raw.strip())
+        except (ValueError, AttributeError):
+            return self._timeout_seconds
+        if value <= 0:
+            return self._timeout_seconds
+        return value
+
+    def _request_chat_completion(
+        self,
+        messages: list[dict[str, str]],
+        *,
+        namespace: str = "",
+        response_schema: dict[str, Any] | None = None,
+    ) -> str:
+        if self._provider == PROVIDER_ANTHROPIC:
+            return self._request_anthropic(
+                messages, namespace=namespace, response_schema=response_schema
+            )
+        # Default: weak json mode (valid JSON, any keys). When a schema is
+        # supplied: OpenAI-compatible strict json_schema mode — Ollama
+        # rejects model output that doesn't match the schema. Issue #436
+        # diagnosis: weak json_object mode let the model invent or drop
+        # keys when the JD content grew past ~1KB, collapsing the
+        # cover-letter body schema entirely.
+        if response_schema is None:
+            response_format: dict[str, Any] = {"type": "json_object"}
+        else:
+            response_format = {
+                "type": "json_schema",
+                "json_schema": {
+                    "name": response_schema.get("name", "response"),
+                    "schema": response_schema.get("schema", response_schema),
+                    "strict": True,
+                },
+            }
+        payload = {
+            "model": self._model,
+            "messages": messages,
+            "temperature": 0,
+            "response_format": response_format,
+        }
+        request = Request(
+            self._endpoint,
+            data=json.dumps(payload).encode("utf-8"),
+            headers={"Content-Type": "application/json"},
+            method="POST",
+        )
+        request_timeout = (
+            self._resolve_timeout(namespace) if namespace else self._timeout_seconds
+        )
+        try:
+            with urlopen(request, timeout=request_timeout) as response:  # nosec B310
+                body = response.read().decode("utf-8", errors="replace")
+        except (OSError, TimeoutError) as exc:
+            # urllib.error.URLError and socket.timeout both inherit from
+            # OSError, and response.read() can also raise raw OSError /
+            # TimeoutError mid-stream. Normalize all transport-level
+            # failures to RuntimeError so callers that catch
+            # (RuntimeError, ValueError) around complete_json (per its
+            # docstring contract) keep the documented fallback behavior
+            # instead of being aborted by an unexpected OSError.
+            #
+            # Endpoint-down failures (connection refused / DNS failure /
+            # host unreachable) are surfaced as the typed
+            # ``LLMEndpointUnreachableError`` subclass so the build
+            # pipeline can emit a single actionable warning and
+            # short-circuit subsequent stages instead of attempting each
+            # one and burning the per-stage timeout budget.
+            if _is_endpoint_unreachable(exc):
+                raise LLMEndpointUnreachableError(
+                    f"LLM endpoint unreachable at {self._endpoint}: {exc}"
+                ) from exc
+            raise RuntimeError(f"LLM request failed: {exc}") from exc
+
+        try:
+            decoded = json.loads(body)
+        except json.JSONDecodeError as exc:
+            raise RuntimeError("LLM response was not valid JSON") from exc
+
+        choices = decoded.get("choices", [])
+        if not isinstance(choices, list) or not choices:
+            raise RuntimeError("LLM response missing choices")
+        first = choices[0]
+        if not isinstance(first, dict):
+            raise RuntimeError("LLM response choice was malformed")
+        message = first.get("message", {})
+        if not isinstance(message, dict):
+            raise RuntimeError("LLM response message was malformed")
+        content = message.get("content", "")
+        if not isinstance(content, str) or not content.strip():
+            raise RuntimeError("LLM response content was empty")
+        return content
+
+    def _request_anthropic(
+        self,
+        messages: list[dict[str, str]],
+        *,
+        namespace: str = "",
+        response_schema: dict[str, Any] | None = None,
+    ) -> str:
+        """Single Anthropic chat-completion call via the official SDK.
+
+        Returns the model's raw response string (typically JSON when
+        `response_schema` is set via Anthropic's native
+        `output_config.format` structured-output mode). Caller parses +
+        validates.
+
+        Caching of the system prompt is on by default: every system
+        message in `messages` is hoisted out of the message array into
+        the top-level Anthropic `system` field as a text block, with a
+        `cache_control: {"type": "ephemeral"}` block attached so the
+        prompt is cached across calls within the 5-minute TTL. The
+        cover-letter body system prompt is ~4K tokens — well above
+        Sonnet 4.6's 2048-token minimum cache prefix — and is reused
+        across every call in a multi-JD compare session, so this
+        materially cuts cost on repeat calls.
+
+        **Error contract — distinct from `_request_chat_completion`.**
+        All failures of the Anthropic path raise `AnthropicProviderError`,
+        which does NOT inherit from `RuntimeError`. The pipeline's
+        existing `(RuntimeError, ValueError)` fallback handlers will
+        not catch these — the build halts loudly instead of silently
+        degrading to deterministic non-LLM output. Rationale: when a
+        user opted into the paid Anthropic provider, falling back to
+        baseline produces output far worse than what they paid for,
+        and burns API latency + tokens producing nothing useful.
+        Specific cases:
+          - `APIConnectionError` (transport unreachable) →
+            `AnthropicProviderError`.
+          - `APIStatusError` (rate limit, 4xx, 5xx) →
+            `AnthropicProviderError`.
+          - Missing API key / empty response → `AnthropicProviderError`.
+
+        The `anthropic` SDK is lazy-imported so `import llm_client`
+        succeeds in environments that haven't installed it (e.g. CI's
+        lint-only image, the default Ollama path). Same pattern as the
+        playwright lazy imports in `playwright_stealth_kit.launch`.
+        ImportError surfaces directly (not wrapped) since it's a setup
+        failure, not a runtime degradation candidate.
+        """
+        # API-key validation goes BEFORE the SDK import: a missing key
+        # is a configuration error and should surface as an
+        # AnthropicProviderError regardless of whether the SDK is
+        # installed. CI's test-fast / test-slow images don't carry the
+        # `anthropic` dep, so importing first would mask the api-key
+        # check behind an ImportError on those images.
+        if not self._api_key:
+            raise AnthropicProviderError(
+                "anthropic provider requires an API key via "
+                "RESUME_BUILDER_LLM_API_KEY or ANTHROPIC_API_KEY"
+            )
+
+        try:
+            import anthropic
+            from anthropic import (
+                APIConnectionError,
+                APIStatusError,
+            )
+        except ImportError as exc:
+            raise ImportError(
+                "anthropic SDK is required when "
+                "RESUME_BUILDER_LLM_PROVIDER=anthropic. Install: "
+                "pip install anthropic"
+            ) from exc
+
+        # Split out the system message (Anthropic puts it at the top
+        # level, not in the messages array) and pull off a cache_control
+        # breakpoint so the system prompt is cached across calls.
+        system_blocks: list[dict[str, Any]] = []
+        user_messages: list[dict[str, Any]] = []
+        for m in messages:
+            role = m.get("role", "")
+            content = m.get("content", "")
+            if role == "system":
+                system_blocks.append(
+                    {
+                        "type": "text",
+                        "text": content,
+                        "cache_control": {"type": "ephemeral"},
+                    }
+                )
+            else:
+                user_messages.append({"role": role, "content": content})
+
+        kwargs: dict[str, Any] = {
+            "model": self._model,
+            "max_tokens": self._max_tokens,
+            "messages": user_messages,
+        }
+        if system_blocks:
+            kwargs["system"] = system_blocks
+        if response_schema is not None:
+            # Anthropic's native structured-output mode. The caller's
+            # schema dict matches the OpenAI shape `{name, schema}`; the
+            # Anthropic API wants the inner schema only.
+            schema = response_schema.get("schema", response_schema)
+            kwargs["output_config"] = {
+                "format": {"type": "json_schema", "schema": schema}
+            }
+
+        timeout = (
+            self._resolve_timeout(namespace) if namespace else self._timeout_seconds
+        )
+        client = anthropic.Anthropic(api_key=self._api_key, timeout=timeout)
+        try:
+            response = client.messages.create(**kwargs)
+        except APIConnectionError as exc:
+            raise AnthropicProviderError(f"Anthropic API unreachable: {exc}") from exc
+        except APIStatusError as exc:
+            # Covers RateLimitError, BadRequestError, AuthenticationError,
+            # OverloadedError, etc. — all surface as AnthropicProviderError
+            # so the build halts loudly instead of silently falling back
+            # to deterministic baseline (which would be far worse than
+            # what the user paid for).
+            raise AnthropicProviderError(
+                f"Anthropic API error {exc.status_code}: {exc.message}"
+            ) from exc
+
+        # Response shape: `content` is a list of typed blocks; structured
+        # output guarantees the first text block holds the JSON payload.
+        for block in response.content:
+            if getattr(block, "type", "") == "text":
+                text = getattr(block, "text", "")
+                if not isinstance(text, str) or not text.strip():
+                    raise AnthropicProviderError(
+                        "Anthropic response text block was empty"
+                    )
+                return text
+        raise AnthropicProviderError("Anthropic response had no text block")
+
+    def _read_cache(self, path: Path) -> str | None:
+        if not path.exists():
+            return None
+        pass  # runtime-guard call dropped in kit fork (host-app concern)
+        raw = path.read_text(encoding="utf-8")
+        payload = json.loads(raw)
+        if not isinstance(payload, dict):
+            raise ValueError("LLM cache payload must be an object")
+        content = payload.get("content")
+        if not isinstance(content, str):
+            raise ValueError("LLM cache payload missing content string")
+        return content
+
+    def _write_cache(self, *, cache_path: Path, content: str) -> None:
+        pass  # runtime-guard call dropped in kit fork (host-app concern)
+        cache_path.parent.mkdir(parents=True, exist_ok=True)
+        payload = {
+            "model": self._model,
+            "endpoint": self._endpoint,
+            "content": content,
+        }
+        cache_path.write_text(json.dumps(payload, indent=2) + "\n", encoding="utf-8")

jsmithpkp_llm_client_kit-0.1.3/tests/test_smoke.py

@@ -0,0 +1,89 @@
+"""Smoke tests for the v0.1.0 carve-out.
+
+Full behavioral coverage lives in resume-builder's
+``tests/scripts/test_llm_client.py`` (1000+ lines) and is preserved
+there; this file only exercises that the kit's public API is
+importable + constructs without side effects. The richer test suite
+will migrate over once we genericize env vars in a follow-up release.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+from llm_client_kit import (
+    PROVIDER_ANTHROPIC,
+    PROVIDER_OLLAMA,
+    AnthropicProviderError,
+    LLMClient,
+    LLMEndpointUnreachableError,
+    LLMResponse,
+)
+
+
+def test_public_api_importable() -> None:
+    """Every name in __all__ resolves to something usable."""
+    assert LLMClient is not None
+    assert LLMResponse is not None
+    assert LLMEndpointUnreachableError is not None
+    assert AnthropicProviderError is not None
+    assert PROVIDER_OLLAMA == "ollama"
+    assert PROVIDER_ANTHROPIC == "anthropic"
+
+
+def test_exception_hierarchy() -> None:
+    """Endpoint-unreachable IS a RuntimeError (for legacy fallback
+    handlers); Anthropic-provider error is NOT (intentional, see
+    AnthropicProviderError docstring)."""
+    assert issubclass(LLMEndpointUnreachableError, RuntimeError)
+    assert not issubclass(AnthropicProviderError, RuntimeError)
+
+
+def test_client_constructs_with_explicit_kwargs(tmp_path: Path) -> None:
+    """Direct construction is the non-env path consumers can use to
+    avoid the resume-builder-flavored env var convention."""
+    client = LLMClient(
+        endpoint="http://localhost:11434/v1/chat/completions",
+        model="llama3.2:3b",
+        cache_dir=tmp_path / "cache",
+        fixture_mode=True,
+        timeout_seconds=5.0,
+        provider=PROVIDER_OLLAMA,
+    )
+    assert client.endpoint == "http://localhost:11434/v1/chat/completions"
+    assert client.model == "llama3.2:3b"
+    assert client.provider == PROVIDER_OLLAMA
+
+
+def test_client_rejects_unknown_provider(tmp_path: Path) -> None:
+    """Typo in the provider arg fails fast at construction so the user
+    sees the misconfiguration immediately, not silently routed to the
+    Ollama path."""
+    with pytest.raises(ValueError, match="unsupported LLM provider"):
+        LLMClient(
+            endpoint="http://x",
+            model="m",
+            cache_dir=tmp_path,
+            fixture_mode=False,
+            provider="gemini",  # type: ignore[arg-type] — intentional bad value
+        )
+
+
+def test_fixture_mode_misses_raise_runtime_error(tmp_path: Path) -> None:
+    """In fixture mode, calling complete_json without a cached entry
+    must raise — never silently issue a network request."""
+    client = LLMClient(
+        endpoint="http://localhost:99999",  # nothing listening, would explode if called
+        model="llama3.2:3b",
+        cache_dir=tmp_path / "empty-cache",
+        fixture_mode=True,
+        timeout_seconds=1.0,
+    )
+    with pytest.raises(RuntimeError, match="LLM fixture missing"):
+        client.complete_json(
+            namespace="test",
+            system_prompt="x",
+            user_payload={"q": "y"},
+        )