thoughtflow 0.2.0__tar.gz → 0.2.2__tar.gz

{thoughtflow-0.2.0 → thoughtflow-0.2.2}/CHANGELOG.md

@@ -5,16 +5,42 @@ All notable changes to ThoughtFlow will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [Unreleased]
+## [0.2.2] - 2026-07-10
+
+Production-hardening release: THOUGHT/PLAN parsing improvements (Foundation Agent feedback) plus unified LLM param pipeline, prompt caching, and native provider convenience classes.
 
 ### Added
-- Nothing yet
+- THOUGHT `on_failure=` hook and structured failure events on MEMORY (`{name}_last_error`, `{name}_attempts`)
+- THOUGHT opt-in production lane: `validation_mode='rescue'`, `strict_schema=True`, `hold_msg=True`, `repair_target=`
+- `valid_extract` optional `oneOf` schema dialect for branch-shaped responses
+- PLAN `normalize_plan()` subclass hook and aggregate missing-param error messages
+- LLM `last_error`, `last_usage`, opt-in `request_timeout=`, `on_usage=` callback
+- Unified model-param pipeline (`split_call_params`) — transport keys never leak into JSON bodies
+- Prompt caching: `cached_system()`, `cached_message()`, `cache=True` on messages; Anthropic `cache_control` wire adapter
+- Native provider classes: `GroqLLM`, `CerebrasLLM`, `OpenRouterLLM`
+- `docs/concepts/thought.md` — parse-validate pipeline guide
+- Regression tests: `test_foundation_agent_response.py`, `test_llm_provider_integration.py`
+- Example: `examples/scripts/15_thought_integration_test.py`
+- Integration smoke tests: `tests/integration/test_llm_providers.py`, `test_embed_providers.py`
 
 ### Changed
-- Nothing yet
+- Validation callables may return bare `bool` (coerced to tuple automatically)
+- Parse-failure and repair-suffix messages are model-friendly
+- `parser='json'` uses balanced JSON extraction
+- Execution events include `failed: true` when retries exhausted with error
+- OpenAI-compatible choice extraction falls back to `message.reasoning` when `content` is empty
+- Expanded `docs/concepts/llm.md`, `primitives/LLM.md`, `primitives/THOUGHT.md`, `primitives/PLAN.md`
 
 ### Fixed
-- Nothing yet
+- Bare-bool `validation=` lambdas no longer crash unpack
+- PLAN param validation reports all missing params at once
+- OpenRouter `referer`/`title` no longer leak into request JSON body
+- EMBED Groq calls include User-Agent (Cloudflare HTTP 1010 without it)
+- EMBED `text-embedding-004` aliases to `gemini-embedding-001`; `last_error` surfaced on failures
+
+## [Unreleased]
+
+_Nothing yet._
 
 ---
 
@@ -171,7 +197,8 @@ grown from MEMORY itself, with zero new concepts.
 ---
 
 <!-- Release links -->
-[Unreleased]: https://github.com/jrolf/thoughtflow/compare/v0.2.0...HEAD
+[Unreleased]: https://github.com/jrolf/thoughtflow/compare/v0.2.2...HEAD
+[0.2.2]: https://github.com/jrolf/thoughtflow/compare/v0.2.0...v0.2.2
 [0.2.0]: https://github.com/jrolf/thoughtflow/compare/v0.1.3...v0.2.0
 [0.1.3]: https://github.com/jrolf/thoughtflow/compare/v0.1.2...v0.1.3
 [0.1.2]: https://github.com/jrolf/thoughtflow/compare/v0.1.1...v0.1.2

{thoughtflow-0.2.0 → thoughtflow-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: thoughtflow
-Version: 0.2.0
+Version: 0.2.2
 Summary: Powerful AI systems from simple parts. Composable cognitive primitives, event-sourced memory, and autonomous agents in pure Python with zero dependencies.
 Project-URL: Homepage, https://github.com/jrolf/thoughtflow
 Project-URL: Documentation, https://thoughtflow.dev
@@ -164,7 +164,7 @@ supply-chain anxiety. The core library uses only Python's standard library —
 pip install --upgrade thoughtflow
 
 # Pin to a specific version for stability
-pip install thoughtflow==0.2.0
+pip install thoughtflow==0.2.2
 
 # Check your installed version
 python -c "import thoughtflow; print(thoughtflow.__version__)"
@@ -452,8 +452,18 @@ ThoughtFlow works with **any LLM provider** through a unified interface:
 | **Groq** | `groq:model` | `groq:llama-3.1-70b-versatile` | Fast inference for open models |
 | **Google Gemini** | `gemini:model` | `gemini:gemini-1.5-pro` | Gemini Pro, Flash, etc. |
 | **OpenRouter** | `openrouter:model` | `openrouter:anthropic/claude-3-opus` | Access any model via OpenRouter |
+| **Cerebras** | `openai:model` + `OpenAICompatibleLLM` | `CerebrasLLM("gpt-oss-120b", key=...)` | OpenAI-compatible; Cloudflare-safe defaults |
 | **Ollama** | `ollama:model` | `ollama:llama3.2` | Local models, no API key needed |
 
+**Native convenience classes** (v0.2.2+) wrap production edge cases while keeping the same `call()` API:
+
+| Class | When to use |
+|-------|-------------|
+| `GroqLLM` | Qwen models on Groq — sets `reasoning_format=hidden` by default |
+| `CerebrasLLM` | Cerebras inference — safe `User-Agent`, optional auth-key fallback |
+| `OpenRouterLLM` | OpenRouter gateway — sets `HTTP-Referer` and `X-Title` headers |
+| `OpenAICompatibleLLM` | Any OpenAI-compatible server (vLLM, MLX, LM Studio, Ollama `/v1`) |
+
 **Switching providers is a one-line change:**
 
 ```python
@@ -585,6 +595,28 @@ recording.to_json("session.json")
 
 replay_llm = LLM.replay(MEMORY.from_json("session.json"))
 # replay_llm is a drop-in LLM: same flows, no network, identical outputs
+
+# ═══════════════════════════════════════════════════════════════════════════
+# PROMPT CACHING — stable prefixes for agentic loops (v0.2.2+)
+# ═══════════════════════════════════════════════════════════════════════════
+
+from thoughtflow import cached_system, cached_message
+
+llm.call([
+    cached_system("You are a helpful assistant. Follow these rules exactly..."),
+    {"role": "user", "content": "Hello"},
+])
+# Anthropic: cache_control blocks on marked messages
+# OpenAI-compat providers: markers stripped from bodies; opt-in wire params apply
+
+# ═══════════════════════════════════════════════════════════════════════════
+# OBSERVABILITY — last_error, last_usage, request_timeout (v0.2.2+)
+# ═══════════════════════════════════════════════════════════════════════════
+
+llm = LLM("openai:gpt-4o", key="sk-...", request_timeout=30)
+response = llm.call([{"role": "user", "content": "Hi"}])
+print(llm.last_usage)   # normalized token counts when the provider returns them
+print(llm.last_error)   # None on success, error string on provider failure
 ```
 
 **Key features:**
@@ -595,6 +627,9 @@ replay_llm = LLM.replay(MEMORY.from_json("session.json"))
 - **Structured output** — JSON Schema enforcement via each provider's native mechanism
 - **Streaming** — SSE/NDJSON streaming with a plain generator interface
 - **Record/replay** — Capture exchanges into MEMORY; replay them deterministically
+- **Prompt caching** — `cached_system()` / `cache=True` for stable agent prefixes
+- **Production provider classes** — `GroqLLM`, `CerebrasLLM`, `OpenRouterLLM`
+- **Unified param pipeline** — Transport keys (`base_url`, `referer`) never leak into JSON bodies
 
 ---
 
@@ -979,6 +1014,23 @@ memory = thought(memory)
 # Tokens stream through your hook in real time; the complete text still
 # flows through parsing/validation and lands in memory as usual.
 
+# ═══════════════════════════════════════════════════════════════════════════
+# PRODUCTION LANE — opt-in hardening (v0.2.2+, defaults unchanged)
+# ═══════════════════════════════════════════════════════════════════════════
+
+thought = THOUGHT(
+    name="extract_plan",
+    llm=llm,
+    prompt="Return JSON for: {task}",
+    parser="json",
+    validation_mode="rescue",   # try repair before failing
+    strict_schema=True,         # enforce schema shape strictly
+    hold_msg=True,              # keep failed assistant msg out of history
+    on_failure=lambda m, err: m.add_log("extract_plan failed: " + str(err)),
+    max_retries=2,
+)
+# Failures surface on MEMORY: {name}_last_error, {name}_attempts, failed: true events
+
 # ═══════════════════════════════════════════════════════════════════════════
 # SERIALIZATION — save and restore thoughts
 # ═══════════════════════════════════════════════════════════════════════════
@@ -1015,6 +1067,8 @@ thought.last_response    # Raw LLM response
 - **Schema-based parsing** — Via `valid_extract` for bulletproof extraction
 - **Multiple validators** — Built-in or custom, via the `validation=` parameter
 - **Token streaming** — `on_token=` hook streams chunks without breaking the contract
+- **Production lane (opt-in)** — `validation_mode`, `strict_schema`, `hold_msg`, `on_failure`
+- **Structured failure visibility** — `{name}_last_error` and `{name}_attempts` on MEMORY
 - **Four operations** — `llm_call`, `memory_query`, `variable_set`, `conditional`
 - **Pre/post hooks** — Custom processing before and after execution
 - **Full serialization** — Save, restore, and copy thoughts
@@ -1627,8 +1681,15 @@ vectors = embed.call(["Hello", "World"])
 # Record/replay works here too — same seam as LLM
 embed.record(memory)
 replay_embed = EMBED.replay(memory)
+
+# Gemini: text-embedding-004 aliases to gemini-embedding-001 automatically
+embed = EMBED("gemini:text-embedding-004", key="...")
+vector = embed.call("Hello world")
+print(embed.last_error)  # None on success; error string on provider failure
 ```
 
+Supported embedding providers: OpenAI, Gemini, Groq, OpenRouter, and local Ollama (`ollama:all-minilm:33m`, etc.).
+
 ---
 
 ## 🔧 Utilities
@@ -2236,6 +2297,18 @@ deterministic: no API keys in CI, no flaky outputs, no spend. The complete
 lifecycle is demonstrated in
 [`examples/scripts/13_record_replay.py`](examples/scripts/13_record_replay.py).
 
+**Integration smoke tests** (optional, real API calls) live under `tests/integration/`.
+Enable with `THOUGHTFLOW_INTEGRATION_TESTS=1` and provider API keys in your environment:
+
+```bash
+THOUGHTFLOW_INTEGRATION_TESTS=1 pytest tests/integration/test_llm_providers.py \
+  tests/integration/test_embed_providers.py -v
+```
+
+These cover base `LLM` routing, native provider classes (`GroqLLM`, `CerebrasLLM`,
+`OpenRouterLLM`), Ollama, prompt caching, record/replay, and `EMBED` providers.
+Tests skip automatically when a key or account lacks access to a provider.
+
 ---
 
 ## 📁 Project Structure
@@ -2245,6 +2318,9 @@ thoughtflow/
 ├── src/thoughtflow/
 │   ├── __init__.py      # Public API exports
 │   ├── llm.py           # LLM — multi-provider model interface
+│   ├── llm_params.py    # Unified model-param pipeline (transport vs wire)
+│   ├── llm_cache.py     # Prompt caching helpers (cached_system, etc.)
+│   ├── providers/       # GroqLLM, CerebrasLLM, OpenRouterLLM
 │   ├── embed.py         # EMBED — multi-provider embeddings
 │   ├── memory.py        # MEMORY — event-sourced state container
 │   ├── thought.py       # THOUGHT — atomic cognitive unit
@@ -2313,9 +2389,10 @@ See [developer/](developer/) for comprehensive development documentation.
 | **Orchestration** | ✅ Stable | WORKFLOW, CHRON |
 | **Action Subclasses** | ✅ Stable | 16 elemental operations (SEARCH, SCRAPE, FETCH, etc.) |
 | **Record/Replay** | ✅ Shipped | LLM/EMBED recording, deterministic replay, eval harness |
+| **Provider convenience classes** | ✅ Shipped (0.2.2) | GroqLLM, CerebrasLLM, OpenRouterLLM, prompt caching |
 | **API Stability** | 🟡 Alpha | May evolve based on feedback |
 | **Documentation** | ✅ Per-primitive docs | `primitives/` folder with Markdown per class |
-| **Test Coverage** | ✅ ~700 unit tests | Plus integration tests gated by API keys |
+| **Test Coverage** | ✅ ~744 unit tests | Plus optional integration smoke tests (API keys) |
 | **Serverless Ready** | ✅ Yes | Zero deps, deployable example in `examples/serverless/` |
 
 See [CHANGELOG.md](CHANGELOG.md) for version history.

{thoughtflow-0.2.0 → thoughtflow-0.2.2}/README.md

@@ -93,7 +93,7 @@ supply-chain anxiety. The core library uses only Python's standard library —
 pip install --upgrade thoughtflow
 
 # Pin to a specific version for stability
-pip install thoughtflow==0.2.0
+pip install thoughtflow==0.2.2
 
 # Check your installed version
 python -c "import thoughtflow; print(thoughtflow.__version__)"
@@ -381,8 +381,18 @@ ThoughtFlow works with **any LLM provider** through a unified interface:
 | **Groq** | `groq:model` | `groq:llama-3.1-70b-versatile` | Fast inference for open models |
 | **Google Gemini** | `gemini:model` | `gemini:gemini-1.5-pro` | Gemini Pro, Flash, etc. |
 | **OpenRouter** | `openrouter:model` | `openrouter:anthropic/claude-3-opus` | Access any model via OpenRouter |
+| **Cerebras** | `openai:model` + `OpenAICompatibleLLM` | `CerebrasLLM("gpt-oss-120b", key=...)` | OpenAI-compatible; Cloudflare-safe defaults |
 | **Ollama** | `ollama:model` | `ollama:llama3.2` | Local models, no API key needed |
 
+**Native convenience classes** (v0.2.2+) wrap production edge cases while keeping the same `call()` API:
+
+| Class | When to use |
+|-------|-------------|
+| `GroqLLM` | Qwen models on Groq — sets `reasoning_format=hidden` by default |
+| `CerebrasLLM` | Cerebras inference — safe `User-Agent`, optional auth-key fallback |
+| `OpenRouterLLM` | OpenRouter gateway — sets `HTTP-Referer` and `X-Title` headers |
+| `OpenAICompatibleLLM` | Any OpenAI-compatible server (vLLM, MLX, LM Studio, Ollama `/v1`) |
+
 **Switching providers is a one-line change:**
 
 ```python
@@ -514,6 +524,28 @@ recording.to_json("session.json")
 
 replay_llm = LLM.replay(MEMORY.from_json("session.json"))
 # replay_llm is a drop-in LLM: same flows, no network, identical outputs
+
+# ═══════════════════════════════════════════════════════════════════════════
+# PROMPT CACHING — stable prefixes for agentic loops (v0.2.2+)
+# ═══════════════════════════════════════════════════════════════════════════
+
+from thoughtflow import cached_system, cached_message
+
+llm.call([
+    cached_system("You are a helpful assistant. Follow these rules exactly..."),
+    {"role": "user", "content": "Hello"},
+])
+# Anthropic: cache_control blocks on marked messages
+# OpenAI-compat providers: markers stripped from bodies; opt-in wire params apply
+
+# ═══════════════════════════════════════════════════════════════════════════
+# OBSERVABILITY — last_error, last_usage, request_timeout (v0.2.2+)
+# ═══════════════════════════════════════════════════════════════════════════
+
+llm = LLM("openai:gpt-4o", key="sk-...", request_timeout=30)
+response = llm.call([{"role": "user", "content": "Hi"}])
+print(llm.last_usage)   # normalized token counts when the provider returns them
+print(llm.last_error)   # None on success, error string on provider failure
 ```
 
 **Key features:**
@@ -524,6 +556,9 @@ replay_llm = LLM.replay(MEMORY.from_json("session.json"))
 - **Structured output** — JSON Schema enforcement via each provider's native mechanism
 - **Streaming** — SSE/NDJSON streaming with a plain generator interface
 - **Record/replay** — Capture exchanges into MEMORY; replay them deterministically
+- **Prompt caching** — `cached_system()` / `cache=True` for stable agent prefixes
+- **Production provider classes** — `GroqLLM`, `CerebrasLLM`, `OpenRouterLLM`
+- **Unified param pipeline** — Transport keys (`base_url`, `referer`) never leak into JSON bodies
 
 ---
 
@@ -908,6 +943,23 @@ memory = thought(memory)
 # Tokens stream through your hook in real time; the complete text still
 # flows through parsing/validation and lands in memory as usual.
 
+# ═══════════════════════════════════════════════════════════════════════════
+# PRODUCTION LANE — opt-in hardening (v0.2.2+, defaults unchanged)
+# ═══════════════════════════════════════════════════════════════════════════
+
+thought = THOUGHT(
+    name="extract_plan",
+    llm=llm,
+    prompt="Return JSON for: {task}",
+    parser="json",
+    validation_mode="rescue",   # try repair before failing
+    strict_schema=True,         # enforce schema shape strictly
+    hold_msg=True,              # keep failed assistant msg out of history
+    on_failure=lambda m, err: m.add_log("extract_plan failed: " + str(err)),
+    max_retries=2,
+)
+# Failures surface on MEMORY: {name}_last_error, {name}_attempts, failed: true events
+
 # ═══════════════════════════════════════════════════════════════════════════
 # SERIALIZATION — save and restore thoughts
 # ═══════════════════════════════════════════════════════════════════════════
@@ -944,6 +996,8 @@ thought.last_response    # Raw LLM response
 - **Schema-based parsing** — Via `valid_extract` for bulletproof extraction
 - **Multiple validators** — Built-in or custom, via the `validation=` parameter
 - **Token streaming** — `on_token=` hook streams chunks without breaking the contract
+- **Production lane (opt-in)** — `validation_mode`, `strict_schema`, `hold_msg`, `on_failure`
+- **Structured failure visibility** — `{name}_last_error` and `{name}_attempts` on MEMORY
 - **Four operations** — `llm_call`, `memory_query`, `variable_set`, `conditional`
 - **Pre/post hooks** — Custom processing before and after execution
 - **Full serialization** — Save, restore, and copy thoughts
@@ -1556,8 +1610,15 @@ vectors = embed.call(["Hello", "World"])
 # Record/replay works here too — same seam as LLM
 embed.record(memory)
 replay_embed = EMBED.replay(memory)
+
+# Gemini: text-embedding-004 aliases to gemini-embedding-001 automatically
+embed = EMBED("gemini:text-embedding-004", key="...")
+vector = embed.call("Hello world")
+print(embed.last_error)  # None on success; error string on provider failure
 ```
 
+Supported embedding providers: OpenAI, Gemini, Groq, OpenRouter, and local Ollama (`ollama:all-minilm:33m`, etc.).
+
 ---
 
 ## 🔧 Utilities
@@ -2165,6 +2226,18 @@ deterministic: no API keys in CI, no flaky outputs, no spend. The complete
 lifecycle is demonstrated in
 [`examples/scripts/13_record_replay.py`](examples/scripts/13_record_replay.py).
 
+**Integration smoke tests** (optional, real API calls) live under `tests/integration/`.
+Enable with `THOUGHTFLOW_INTEGRATION_TESTS=1` and provider API keys in your environment:
+
+```bash
+THOUGHTFLOW_INTEGRATION_TESTS=1 pytest tests/integration/test_llm_providers.py \
+  tests/integration/test_embed_providers.py -v
+```
+
+These cover base `LLM` routing, native provider classes (`GroqLLM`, `CerebrasLLM`,
+`OpenRouterLLM`), Ollama, prompt caching, record/replay, and `EMBED` providers.
+Tests skip automatically when a key or account lacks access to a provider.
+
 ---
 
 ## 📁 Project Structure
@@ -2174,6 +2247,9 @@ thoughtflow/
 ├── src/thoughtflow/
 │   ├── __init__.py      # Public API exports
 │   ├── llm.py           # LLM — multi-provider model interface
+│   ├── llm_params.py    # Unified model-param pipeline (transport vs wire)
+│   ├── llm_cache.py     # Prompt caching helpers (cached_system, etc.)
+│   ├── providers/       # GroqLLM, CerebrasLLM, OpenRouterLLM
 │   ├── embed.py         # EMBED — multi-provider embeddings
 │   ├── memory.py        # MEMORY — event-sourced state container
 │   ├── thought.py       # THOUGHT — atomic cognitive unit
@@ -2242,9 +2318,10 @@ See [developer/](developer/) for comprehensive development documentation.
 | **Orchestration** | ✅ Stable | WORKFLOW, CHRON |
 | **Action Subclasses** | ✅ Stable | 16 elemental operations (SEARCH, SCRAPE, FETCH, etc.) |
 | **Record/Replay** | ✅ Shipped | LLM/EMBED recording, deterministic replay, eval harness |
+| **Provider convenience classes** | ✅ Shipped (0.2.2) | GroqLLM, CerebrasLLM, OpenRouterLLM, prompt caching |
 | **API Stability** | 🟡 Alpha | May evolve based on feedback |
 | **Documentation** | ✅ Per-primitive docs | `primitives/` folder with Markdown per class |
-| **Test Coverage** | ✅ ~700 unit tests | Plus integration tests gated by API keys |
+| **Test Coverage** | ✅ ~744 unit tests | Plus optional integration smoke tests (API keys) |
 | **Serverless Ready** | ✅ Yes | Zero deps, deployable example in `examples/serverless/` |
 
 See [CHANGELOG.md](CHANGELOG.md) for version history.

{thoughtflow-0.2.0 → thoughtflow-0.2.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "thoughtflow"
-version = "0.2.0"
+version = "0.2.2"
 description = "Powerful AI systems from simple parts. Composable cognitive primitives, event-sourced memory, and autonomous agents in pure Python with zero dependencies."
 readme = "README.md"
 requires-python = ">=3.9"

{thoughtflow-0.2.0 → thoughtflow-0.2.2}/src/thoughtflow/__init__.py

@@ -43,6 +43,8 @@ from __future__ import annotations
 
 # Core primitives
 from thoughtflow.llm import LLM, OpenAICompatibleLLM, ReplayLLM, ReplayMissError
+from thoughtflow.llm_cache import cached_system, cached_message
+from thoughtflow.providers import GroqLLM, CerebrasLLM, OpenRouterLLM
 from thoughtflow.embed import EMBED, ReplayEMBED
 from thoughtflow.memory import MEMORY
 from thoughtflow.thought import THOUGHT
@@ -103,7 +105,7 @@ try:
 
     __version__ = _get_version("thoughtflow")
 except Exception:
-    __version__ = "0.0.0"
+    __version__ = "0.2.2"
 
 # Public API
 __all__ = [
@@ -112,6 +114,11 @@ __all__ = [
     "OpenAICompatibleLLM",
     "ReplayLLM",
     "ReplayMissError",
+    "GroqLLM",
+    "CerebrasLLM",
+    "OpenRouterLLM",
+    "cached_system",
+    "cached_message",
     "EMBED",
     "ReplayEMBED",
     "MEMORY",