feat: 0.3.0 add Vercel AI SDK adapter, streaming, hybrid fix, polish

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -0,0 +1,46 @@
+name: Bug report
+description: Report something that isn't working
+labels: [bug]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to file a bug report. The more specific you
+        can be, the faster I can fix it.
+  - type: input
+    id: version
+    attributes:
+      label: ctx-opt version
+      description: Output of `npm ls ctx-opt`
+      placeholder: e.g. 0.3.0
+    validations:
+      required: true
+  - type: input
+    id: node
+    attributes:
+      label: Node version
+      placeholder: e.g. 20.10.0
+    validations:
+      required: true
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened?
+      description: A clear description of the bug.
+    validations:
+      required: true
+  - type: textarea
+    id: reproduction
+    attributes:
+      label: Minimal reproduction
+      description: A short, self-contained code snippet that triggers the bug.
+      render: ts
+    validations:
+      required: true
+  - type: textarea
+    id: expected
+    attributes:
+      label: Expected behavior
+      description: What you expected to happen instead.
+    validations:
+      required: true

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -0,0 +1,23 @@
+name: Feature request
+description: Suggest a new feature or improvement
+labels: [enhancement]
+body:
+  - type: textarea
+    id: problem
+    attributes:
+      label: The problem
+      description: What are you trying to do that ctx-opt doesn't currently support? Or what's painful about how it works today?
+    validations:
+      required: true
+  - type: textarea
+    id: proposal
+    attributes:
+      label: Proposed solution
+      description: How would you like it to work?
+    validations:
+      required: true
+  - type: textarea
+    id: alternatives
+    attributes:
+      label: Alternatives considered
+      description: Other approaches you've thought about or workarounds you're using.

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,21 @@
+## Summary
+
+<!-- One or two sentences describing the change and why. -->
+
+## Type of change
+
+- [ ] Bug fix
+- [ ] New feature
+- [ ] Breaking change
+- [ ] Documentation
+- [ ] Refactor / chore
+
+## Test plan
+
+<!-- How did you verify this? -->
+
+- [ ] `npm run typecheck` passes
+- [ ] `npm test` passes
+- [ ] `npm run build` succeeds
+- [ ] Added or updated tests
+- [ ] Updated `CHANGELOG.md` (for user-facing changes)

CHANGELOG.md

@@ -5,6 +5,39 @@ All notable changes to this project 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).
 
+## [0.3.0] - 2026-05-14
+
+### Added
+- **Vercel AI SDK adapter** at `ctx-opt/ai-sdk`. `withOptimizer(fn, config)`
+  wraps any AI SDK function (`generateText`, `streamText`, `generateObject`,
+  `streamObject`) so its `messages` array is trimmed before forwarding.
+  Also ships `trimMessages(messages, config)` for callers that prefer a
+  one-shot preprocessor.
+- **Streaming support** in the OpenAI and Anthropic adapters. `stream: true`
+  on `chat.completions.create` / `messages.create` now correctly returns
+  the SDK's async-iterable stream after optimization. The Anthropic
+  adapter also wraps `client.messages.stream()`.
+- **Native Anthropic token counting** via `countMessageTokensWithAnthropic`,
+  which delegates to `client.messages.countTokens` for exact counts on
+  `claude-*` models. Tradeoff is a network round-trip per call versus
+  tiktoken's free local approximation.
+- **Per-strategy `recentWindow` overrides**. `relevance.recentWindow` and
+  `summarizer.recentWindow` can now be set independently of the top-level
+  `recentWindow`. This unblocks the hybrid strategy (see below).
+- `CONTRIBUTING.md`, GitHub issue templates, and a PR template.
+
+### Fixed
+- **Hybrid strategy now actually summarizes**. Previously the relevance
+  and summarizer phases shared `recentWindow`, so the post-relevance set
+  was entirely "system + recent" with nothing left for the summarizer to
+  compress. Setting `relevance.recentWindow` larger than
+  `summarizer.recentWindow` gives the summarizer real material. The
+  known-limitation note from 0.2.0 is resolved.
+
+### Changed
+- New peer dep: `ai` (optional, for the Vercel AI SDK adapter).
+- `tsup` now builds four entry points instead of three.
+
 ## [0.2.0] - 2026-05-14
 
 ### Added
@@ -38,8 +71,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - In the `hybrid` strategy, `applyRelevance` and `applySummarizer`
   share `recentWindow`, so the post-relevance set rarely contains
   "compressible" messages for the summarizer to operate on, and the
-  pipeline falls through to a sliding-window pass. A future release
-  will let each phase have its own window.
+  pipeline falls through to a sliding-window pass. *Fixed in 0.3.0
+  via per-strategy `recentWindow` overrides.*
 
 ## [0.1.0] - 2026-05-01
 
@@ -52,5 +85,6 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Tool-pair preservation across boundary trims.
 - ESM + CJS builds, TypeScript types, Node 18+.
 
+[0.3.0]: https://github.com/EvanPaules/ctx-opt/releases/tag/v0.3.0
 [0.2.0]: https://github.com/EvanPaules/ctx-opt/releases/tag/v0.2.0
 [0.1.0]: https://github.com/EvanPaules/ctx-opt/releases/tag/v0.1.0

CONTRIBUTING.md

@@ -0,0 +1,62 @@
+# Contributing to ctx-opt
+
+Thanks for your interest. This is a small library, so I try to keep the
+contribution surface simple.
+
+## Quick start
+
+```bash
+git clone https://github.com/EvanPaules/ctx-opt.git
+cd ctx-opt
+npm install
+npm test
+```
+
+## Useful commands
+
+| Command | What it does |
+|---|---|
+| `npm test` | Run the vitest suite once. |
+| `npm run test:watch` | Run vitest in watch mode while iterating. |
+| `npm run typecheck` | Strict TS typecheck without emitting. |
+| `npm run build` | Produce the ESM + CJS + .d.ts dist. |
+| `npm run bench` | Run the strategy benchmark and write `benchmarks/RESULTS.md`. |
+
+CI runs typecheck, test, and build on Node 18, 20, and 22. All three must
+pass before a PR is mergeable.
+
+## What I'm looking for
+
+- **Bug fixes** with a regression test.
+- **New strategies** that fit the existing `Strategy` interface and come
+  with their own test file.
+- **SDK adapters** for new ecosystems (LangChain, Mastra, etc). Use the
+  existing `src/adapters/openai.ts` and `src/adapters/anthropic.ts` as
+  templates and add a subpath export to `package.json`.
+- **Docs and examples** that make a real workflow easier to get started with.
+
+## What I'm probably going to push back on
+
+- Refactors with no user-facing benefit.
+- Optional features behind flags that aren't requested by users.
+- Removing the zero-required-dependency property of the core package.
+
+## Style
+
+- TypeScript strict mode is on. No `any` without an explanatory comment.
+- Prefer named exports over default exports (except `ContextOptimizer`).
+- Tests use vitest. Mock SDKs over real network calls.
+- Keep public API surface small. Add to `src/index.ts` deliberately.
+
+## Releasing (maintainer notes)
+
+1. Land all changes on `main` via PR.
+2. Bump `version` in `package.json`.
+3. Update `CHANGELOG.md`.
+4. Tag and push: `git tag v0.x.0 && git push --tags`.
+5. `npm publish`.
+
+## License
+
+By contributing you agree your contributions are licensed under MIT,
+matching the project license.

README.md

@@ -79,8 +79,23 @@ const res = await ai.messages.create({
 });
 ```
 
-`openai` and `@anthropic-ai/sdk` are **optional peer deps** — install only
-the one you use. See [`examples/`](./examples) for OpenAI, Anthropic,
+```ts
+import { generateText } from 'ai';
+import { withOptimizer } from 'ctx-opt/ai-sdk';
+
+const trimmedGenerate = withOptimizer(generateText, {
+  maxTokens: 8_000,
+  strategy: 'sliding-window',
+});
+
+const { text } = await trimmedGenerate({
+  model: openai('gpt-4o'),
+  messages: longHistory,
+});
+```
+
+`openai`, `@anthropic-ai/sdk`, and `ai` are **optional peer deps** — install
+only the one(s) you use. See [`examples/`](./examples) for OpenAI, Anthropic,
 summarizer-with-real-LLM, and LangChain.js integrations.
 
 ## Strategies
@@ -240,8 +255,20 @@ Every call to `optimize()` returns a `meta` describing what happened:
 - A **per-message overhead** of 4 tokens is added to each message to approximate the
   role and formatting tokens (per OpenAI's chat-completion cookbook formula).
 
-For exact Anthropic counts, call Anthropic's `messages.countTokens` API and pass that
-through your own wrapper.
+For exact Anthropic counts, use the built-in helper that delegates to
+Anthropic's `messages.countTokens` endpoint:
+
+```ts
+import Anthropic from '@anthropic-ai/sdk';
+import { countMessageTokensWithAnthropic } from 'ctx-opt';
+
+const client = new Anthropic();
+const tokens = await countMessageTokensWithAnthropic(
+  client,
+  messages,
+  'claude-haiku-4-5-20251001'
+);
+```
 
 ## Changelog