feat: streaming dep optional, non streaming path, streaming flag, rel…

CHANGELOG.md

@@ -7,158 +7,87 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/2.0.0.
 
 ## [Unreleased]
 
-## [0.2.0] - 2026-08-01
+## [0.2.0] - 2026-08-02
 
 ### Added
 
-- **Sans-IO state machine** (`engine::machine::LoopMachine`): serializable,
-  owns every agent-loop decision (turn counting, max-turn enforcement,
-  tool-call validity, compaction trigger, history, cancellation). Exposed via
-  `BareLoop::machine()` / `into_machine()` / `from_machine()` for inspection
-  and serialize-and-resume. `BareLoop::run()` now drives the machine internally.
-- **Session/Run/Turn lifetime model** (`engine::core`): one `Session` spans the
-  process, one `Run` per `run()` call, one `Turn` per loop iteration. Session
-  derives per-session totals; `Run` carries per-run turns, tokens, and error.
-  Construction splits into `SessionConfig` (session-scoped) and `RunConfig`
-  (per-run budgets).
-- **Reasoning-model support** (`DeltaPart::Thinking` + `on_thinking_delta`):
-  reasoning tokens (Claude extended-thinking, OpenAI o-series, Gemini 2.5+) are
-  routed as their own stream kind. Stream-only — not accumulated into `Message`.
-- **Structured output** (`structured` module): `StructuredOutput` trait,
-  `ResponseFormat`, `request_structured::<T>()`. All three providers override
-  `stream_messages_with_options` / `create_message_with_options` to inject the
-  schema natively.
-- **Tool constraints** (`ToolConstraint` enum): `Strict` tightens tool schemas
-  via the provider's native strict mode; `Grammar` compiles schemas into a
-  grammar for vLLM-style samplers (`grammar` feature).
-- **Tool reflection** (`LlmReflector`): asks the model to classify failed tool
-  calls and suggest corrections via `request_structured`.
-- **Parallel tool dispatch** (`ParallelDispatchConfig`): independent,
-  concurrency-safe calls within a single turn run concurrently. Sequential by
-  default.
-- **Stream resilience** (`StreamHandler`): retries, timeouts, rate-limit
-  backoff, and non-streaming fallback. Configurable via
-  `with_timeout_config` / `with_retry_config` / `with_rate_limit_config`.
-  `HandlerEvent` enum provides real-time observability during streaming.
-- **Client-side rate limiting** (`TokenBucket` / `RateLimiter`): proactive
-  per-provider token-bucket. One bucket per `base_url`.
-- **Context contributors** (`ContextContributor` trait): turn-boundary hook for
-  injecting messages before each model call.
-- **Agent memory** (`LoopMemory` trait, now wired): stores tool-call
-  trajectories, retrieves relevant entries before each turn, consolidates on
-  successful runs. Object-safe; configure via `BareLoop::set_memory`.
-- **Display hints** (`DisplayHint` on `ToolOutput`): advisory rendering hints
-  (Text, Diff, Json, Code, Suppress, Markdown) for presentation layers.
-- **Middleware**: `VerifyMiddleware` (post-execution verification),
-  `MemoizingMiddleware` (tool-call result caching with path-aware invalidation).
-- **Presets** (`ConstrainedProfile`, `FrontierProfile`, `GoalReminder`): named
-  runtime profiles for small-model-tuned and frontier configurations.
-- **`StreamRequest`**: bundles `(messages, system, tools)` into one parameter
-  for all `ApiClient` methods.
-- **`Role::System`** variant: framework-injected system context. Providers map
-  to their native representation.
-- **Pluggable `TokenCounter`**: `HeuristicTokenCounter` (4 chars/token) default;
-  swap in a real tokenizer. Synced bidirectionally with `ContextManager`.
-- **`OpenAiClientBuilder::with_stream_usage(bool)`**: controls
-  `stream_options.include_usage`. `ollama()` disables it automatically.
-- **`with_tcp_nodelay(bool)`** on all three provider builders + `HttpClientConfig`.
-- **HTTP connection-pool injection**: shared `reqwest::Client`, pool knobs
-  (`pool_max_idle_per_host`, `pool_idle_timeout`, `tcp_keepalive`).
-- **Fluent `with_*()` builders** on `LoopConfig`, `BareLoop`, and all provider
-  builders. `CancelSignal::reset()` for multi-run agents.
-- `LoopError` now derives `Serialize`, `Deserialize`, `PartialEq`, `Eq`.
+- `TurnMode` enum: engine runs non-streaming (`create_message`) or streaming (`StreamHandler`), selectable at runtime via `set_turn_mode`. Default is feature-dependent.
+- `streaming` feature: gates `StreamHandler`, per-delta callbacks, `async-stream`. With `default = []`, no streaming code is compiled.
+- Sans-IO `LoopMachine`: serializable, owns all loop decisions. Exposed via `machine()` / `into_machine()` / `from_machine()`.
+- Session/Run/Turn lifetime model: `SessionConfig` (session-scoped) + `RunConfig` (per-run budgets).
+- Reasoning-model support: `DeltaPart::Thinking` + `on_thinking_delta`.
+- Structured output: `StructuredOutput` trait, `ResponseFormat`, `request_structured::<T>()`.
+- Tool constraints (`ToolConstraint`): `Strict` and `Grammar` modes.
+- Tool reflection (`LlmReflector`): model classifies failed tool calls and suggests corrections.
+- Parallel tool dispatch (`ParallelDispatchConfig`).
+- `StreamHandler`: retry, timeout, rate-limit backoff, non-streaming fallback.
+- Client-side rate limiting: `TokenBucket` / `RateLimiter`, one bucket per `base_url`.
+- `ContextContributor` trait: turn-boundary message injection.
+- `LoopMemory` trait wired: stores trajectories, retrieves before turns, consolidates on success.
+- `DisplayHint` on `ToolOutput`: rendering hints (Text, Diff, Json, Code, Suppress, Markdown).
+- Middleware: `VerifyMiddleware`, `MemoizingMiddleware`.
+- Presets: `ConstrainedProfile`, `FrontierProfile`, `GoalReminder`.
+- `StreamRequest`: bundles `(messages, system, tools)` for all `ApiClient` methods.
+- `Role::System` variant.
+- Pluggable `TokenCounter` (`HeuristicTokenCounter` default).
+- `OpenAiClientBuilder::with_stream_usage(bool)`.
+- `with_tcp_nodelay(bool)` on all provider builders.
+- HTTP connection-pool injection: shared `reqwest::Client`, pool knobs.
+- Fluent `with_*()` builders on `BareLoop` and provider builders.
+- `LoopError` derives `Serialize`, `Deserialize`, `PartialEq`, `Eq`.
 
 ### Changed
 
-- **Breaking (`create_message` returns typed `NonStreamingResponse`):** no longer
-  returns raw `serde_json::Value`. The struct carries `message: Message`,
-  `stop_reason: StreamStopReason`, `usage: Option<Usage>`. Migration: read fields
-  from the struct.
-- **Breaking (`extract_structured` takes `&Message`):** default implementation
-  derives from the message; per-provider overrides removed. Migration: change
-  parameter type or delete override.
-- **Breaking (`run()` signature):** `run(&mut self, input: &str, &RunConfig)`
-  returns `Result<Run, LoopError>`. `Loop::initialize` / `config()` removed.
-- **Breaking (session→run lifecycle rename):** `on_session_start` / `on_session_end`
-  → `on_run_start` / `on_run_end`. Fire on every `run()` call. Migration: rename
-  methods and context types.
-- **Breaking (`LoopConfig` removed):** replaced by `SessionConfig` + `RunConfig`.
-  The `model` field lives on `ApiClient`.
-- **Breaking (compaction thresholds → percentages):** `f64` fractions → `u8`
-  percentages (0–100). `0.80 → 80`.
-- **Breaking (builder renames):** all consuming builders uniformly `with_`-prefixed.
-  `Option<T>` builders take `Option<T>` (no `_opt` suffix). Migration: add prefix,
-  wrap literals in `Some(...)`.
-- **Breaking (`StreamHandler::stream_turn`):** returns `impl Stream<Item =
-  Result<HandlerEvent, StreamHandlerError>>` instead of a future.
-- **Breaking (`StreamRequest` parameter):** all `ApiClient` streaming/creation
-  methods take `&StreamRequest` instead of positional params.
-- **Breaking (`DeltaPart` non-exhaustive):** add `_ =>` arm to downstream matches.
-- **Breaking (`Role::System` variant):** add `System =>` arm to downstream matches.
-- **Breaking (`ToolOutput` / `ToolDispatchResult` / `ToolPostContext`
-  non-exhaustive):** use named constructors or `..Default::default()`.
-- **Breaking (`Reflector::analyze`):** gains `tool_schema: Option<&ToolSchema>`
-  parameter.
-- **Breaking (`MessagePart::ToolResult` gains `name` field):**
-  `tool_result()` gains a `name` argument (second param). `#[serde(default)]`
-  allows deserializing older data.
-- **Breaking (`NonStreamingResponse.usage` is `Option<Usage>`):** symmetric with
-  streaming. `None` = provider omitted usage.
-- **OpenAI streaming usage:** now sets `stream_options.include_usage` and captures
-  token counts from the final chunk. All three providers report usage on both paths.
-  All-zero usage collapses to `None`.
-- **OpenAI malformed arguments:** non-empty `function.arguments` that fail JSON
-  parse now return an `ApiError` instead of silently defaulting to `{}`.
-- **Gemini tool-call ids:** `functionCall.id` is now parsed (Gemini 3) and echoed
-  in `functionResponse.id`. `functionResponse` sends both `name` and `id`.
-- **SSE invalid-UTF-8:** `take_line` surfaces invalid UTF-8 as a protocol error
-  instead of silent `U+FFFD` replacement.
-- **`StreamStopReason::from_api_str`** accepts `"tool_use"` as alias for
-  `"tool_call"` (Anthropic).
-- **`set_token_counter`** now propagates to the `ContextManager` if one is set,
-  regardless of setter order.
-- **Memory injection** uses `Role::User` (not `Role::System`) with explicit
-  "reference only" delimitation.
-- **Context token estimate** includes the model response message before counting.
-- MSRV bumped to 1.94 (let-chain syntax). Removed `parking_lot` dependency.
+- **Breaking:** `default = []` no longer pulls `async-stream`; streaming is opt-in via `streaming`. HTTP providers imply it, so `features = ["openai"]` is unchanged. Migration: add `streaming` if you used `providers` alone.
+- **Breaking:** `TurnMode` no longer implements `Default`. Migration: use `turn_mode()` / `set_turn_mode()`.
+- **Breaking:** `create_message` returns typed `NonStreamingResponse` instead of `serde_json::Value`.
+- **Breaking:** `extract_structured` takes `&Message`; per-provider overrides removed.
+- **Breaking:** `run()` is `run(&mut self, &str, &RunConfig) -> Result<Run, LoopError>`. `Loop::initialize` / `config()` removed.
+- **Breaking:** `on_session_start`/`on_session_end` renamed to `on_run_start`/`on_run_end`.
+- **Breaking:** `LoopConfig` removed; split into `SessionConfig` + `RunConfig`.
+- **Breaking:** compaction thresholds are `u8` percentages (0–100) instead of `f64`.
+- **Breaking:** builders uniformly `with_`-prefixed; `Option<T>` builders take `Option<T>`.
+- **Breaking:** `StreamHandler::stream_turn` returns `impl Stream<Item = Result<HandlerEvent, StreamHandlerError>>`.
+- **Breaking:** `ApiClient` methods take `&StreamRequest` instead of positional params.
+- **Breaking:** `DeltaPart`, `Role`, `ToolOutput`, `ToolDispatchResult`, `ToolPostContext` are `#[non_exhaustive]`.
+- **Breaking:** `Reflector::analyze` gains `tool_schema: Option<&ToolSchema>`.
+- **Breaking:** `MessagePart::ToolResult` gains `name` field (serde-defaulted for old data).
+- **Breaking:** `NonStreamingResponse.usage` is `Option<Usage>`.
+- Cancellation no longer trips the fallback breaker (`record_turn_failure` guards `Cancelled`).
+- OpenAI streaming sets `stream_options.include_usage`; all providers report usage on both paths.
+- OpenAI malformed `function.arguments` returns `ApiError` instead of defaulting to `{}`.
+- Gemini parses `functionCall.id` (Gemini 3) and echoes it in responses.
+- SSE invalid UTF-8 surfaced as protocol error instead of `U+FFFD` replacement.
+- `StreamStopReason::from_api_str` accepts `"tool_use"` (Anthropic alias).
+- MSRV bumped to 1.94.
 
 ### Removed
 
-- `StreamHandler::with_config(timeout, retry)` — use `with_timeout_config` /
-  `with_retry_config`.
-- `FallbackManager::record_api_failure` / `record_model_failure` — merged into
-  `record_failure(FailureKind)`.
-- `Loop::process_turn`, `BareLoop::run_turn_body` — replaced by machine-driven
-  `run()`.
-- `StreamTurnResult` — engine assembles result from event stream.
-- `StreamHandler::with_request_options` — use `BareLoop::set_request_options`.
-- `StreamHandlerError::RateLimitEscalation.prior` field (never read).
+- `record_stream_success` / `record_stream_failure` (renamed to `record_turn_*`).
+- `StreamCapable` trait now requires the `streaming` feature.
+- `StreamHandler::with_config(timeout, retry)` — use `with_timeout_config` / `with_retry_config`.
+- `FallbackManager::record_api_failure` / `record_model_failure` — merged into `record_failure(FailureKind)`.
+- `Loop::process_turn`, `BareLoop::run_turn_body` — replaced by machine-driven `run()`.
+- `StreamTurnResult`, `StreamHandler::with_request_options`, `StreamHandlerError::RateLimitEscalation.prior`.
+- `parking_lot` dependency.
 
 ### Fixed
 
-- OpenAI streaming dropped multi-chunk tool-call argument fragments after the
-  first; re-opened tool-call parts on every chunk carrying `function`.
-- `StreamAccumulator` dropped parallel tool calls whose arguments arrived
-  interleaved.
-- `BareLoop` was permanently dead after a single cancellation (`CancelSignal`
-  not re-armed). Now resets in `finalize()`.
-- Tool results were split across multiple user messages instead of merged into
-  one per turn.
-- `StreamHandler` accepted invalid timeout/retry configs (validation never
-  called); `jitter_factor` was validated but never applied.
-- Empty-stream fast-fail: zero-event streams could hang for ~20 min before
-  failing.
-- `ToolHealthRegistry::is_tool_available` consumed the HalfOpen recovery probe
-  as a side effect of a read.
-- Anthropic provider hardcoded text-block index to 0, ignoring server index.
-- Per-run manager reset wiped session-scoped state between runs.
+- Release profile no longer uses `panic = "abort"` (it disabled `catch_unwind` tool-panic isolation).
+- OpenAI streaming dropped multi-chunk tool-call argument fragments.
+- `StreamAccumulator` dropped parallel tool calls with interleaved arguments.
+- `BareLoop` was dead after one cancellation (`CancelSignal` not re-armed; now resets in `finalize()`).
+- Tool results split across multiple user messages instead of merged per turn.
+- `StreamHandler` accepted invalid timeout/retry configs; `jitter_factor` was validated but never applied.
+- Zero-event streams could hang ~20 min before failing.
+- `ToolHealthRegistry::is_tool_available` consumed the HalfOpen probe as a read side effect.
+- Anthropic provider hardcoded text-block index to 0.
+- Per-run manager reset wiped session-scoped state.
 
 ### Security
 
-- Auto-commit hook's `git add -A` on empty file list staged the entire working
-  tree. Now refuses with an error.
-- Non-streaming response body size guard (10 MB) fired after full
-  materialization. Now pre-checks `Content-Length` and caps streaming reads.
+- Auto-commit hook's `git add -A` on empty file list staged the whole working tree; now refuses.
+- Response body size guard (10 MB) now pre-checks `Content-Length` instead of firing after full materialization.
 
 ## [0.1.0] - 2025-07-01
 
@@ -181,4 +110,5 @@ Initial crates.io release.
 - Built-in testing utilities for writing LLM loop tests
 - Example CLIs: hello, REPL, echo tool, and multi-provider chat
 
+[0.2.0]: https://github.com/dch-labs/loopctl/releases/tag/v0.2.0
 [0.1.0]: https://github.com/dch-labs/loopctl/releases/tag/v0.1.0

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "loopctl"
-version = "0.1.0"
+version = "0.2.0"
 edition = "2024"
 license = "MIT OR Apache-2.0"
 description = "A trait-based framework for building agent loops with pluggable LLM clients, tools, and memory"
@@ -30,7 +30,7 @@ tracing = "0.1"
 
 reqwest = { version = "0.13", default-features = false, features = ["json", "stream", "rustls"], optional = true }
 bytes = { version = "1", optional = true }
-async-stream = "0.3"
+async-stream = { version = "0.3", optional = true }
 httpdate = { version = "1", optional = true }
 jsonschema = { version = "0.49", optional = true }
 
@@ -45,15 +45,18 @@ testing = []
 tool_health = []
 tool_shield = ["tool_health"]
 
+# Streaming engine path (StreamHandler, per-delta callbacks, async-stream codegen).
+streaming = ["dep:async-stream"]
+
 # Providers
 providers = ["dep:reqwest", "dep:httpdate", "dep:bytes"]
-openai = ["providers"]
-anthropic = ["providers"]
+openai = ["providers", "streaming"]
+anthropic = ["providers", "streaming"]
 ollama = ["providers", "openai"]
 deepseek = ["providers", "openai"]
 grok = ["providers", "openai"]
 xai = ["grok"]
-gemini = ["providers"]
+gemini = ["providers", "streaming"]
 zai = ["providers", "anthropic"]
 grammar = ["providers"]
 schema_validation = ["dep:jsonschema"]
@@ -92,3 +95,9 @@ let_underscore_must_use = "deny"
 [package.metadata.docs.rs]
 all-features = true
 rustdoc-args = ["--cfg", "docsrs"]
+
+[profile.release]
+lto = "fat"
+codegen-units = 1
+strip = true
+