diff --git a/.claude/commands/implement-unblocked.md b/.claude/commands/implement-unblocked.md
new file mode 100644
index 0000000..51b993f
--- /dev/null
+++ b/.claude/commands/implement-unblocked.md
@@ -0,0 +1,193 @@
+---
+description: Assign every unblocked, unassigned issue not awaiting a decision or a major version to the gh user, move it to In Progress, and implement each one as a PR from its own worktree via sub-agents.
+argument-hint: "[issue numbers to restrict to] [--dry-run]"
+allowed-tools: Bash(gh:*), Bash(git:*), Agent
+---
+
+# Implement unblocked issues
+
+You are the **fanning agent** described under *Worktrees* in `AGENTS.md`. You provision one worktree
+per issue up front, then hand each sub-agent a path that already exists. Sub-agents only do the
+work. They never provision.
+
+Arguments: `$ARGUMENTS`
+
+* Bare numbers restrict the run to those issues. They must still pass the filter in step 2.
+* `--dry-run` stops after step 2. Print the issues that would be picked up and change nothing.
+
+The repository is `PaulTrampert/PTrampert.SimplePatch`.
+
+## 1. Preflight
+
+If any check fails, stop and tell the user how to fix it. Don't work around a failure.
+
+* `gh auth status` must list the `project` scope, which moving an issue on a project board needs.
+  If it's missing, ask the user to run `! gh auth refresh -s project`.
+* Bring `main` up to date so every worktree starts from the latest `main`, not from whatever this
+  clone last fetched. Find the primary checkout from the shared git directory so this works from
+  any worktree. Then fetch and pin the commit the whole batch branches from:
+
+  ```bash
+  PRIMARY="$(dirname "$(git rev-parse --path-format=absolute --git-common-dir)")"
+  git -C "$PRIMARY" fetch origin main
+  BASE="$(git -C "$PRIMARY" rev-parse origin/main)"
+  git -C "$PRIMARY" fetch . origin/main:main   # fast-forward local main; refuses if it has diverged
+  ```
+
+  If the fetch fails, stop. If git refuses the fast-forward because local `main` has commits that
+  aren't on `origin/main`, stop and report those commits. Never reset `main`. If git refuses only
+  because `main` is checked out in a worktree, and that tree is clean, run
+  `git merge --ff-only origin/main` there. If it isn't clean, leave it and note it for the final
+  report. Either way, the worktrees branch from `$BASE`.
+
+## 2. Find the issues
+
+An issue qualifies when all of the following hold:
+
+* It is **open**.
+* It has **no assignees**.
+* It has **no open blocking issue** (GitHub issue dependencies). A closed blocker no longer blocks.
+* It has **no open pull request** that will close it. Someone is already working on it.
+* It isn't labelled **`needs decision`**, meaning it isn't defined well enough to implement yet.
+* It isn't labelled **`Breaking Change`**, meaning it has to wait for a major version.
+
+Naming an issue in the arguments doesn't override the label rules. The label has to be removed
+first.
+
+```bash
+gh api graphql --paginate -f query='
+query($endCursor: String) {
+  repository(owner: "PaulTrampert", name: "PTrampert.SimplePatch") {
+    issues(states: OPEN, first: 100, after: $endCursor) {
+      pageInfo { hasNextPage endCursor }
+      nodes {
+        number
+        title
+        labels(first: 20) { nodes { name } }
+        assignees(first: 1) { totalCount }
+        blockedBy(first: 50) { nodes { number state } }
+        closedByPullRequestsReferences(first: 10, includeClosedPrs: false) { nodes { number state } }
+      }
+    }
+  }
+}' --jq '.data.repository.issues.nodes[]
+  | select(.assignees.totalCount == 0)
+  | select([.labels.nodes[].name] | (index("needs decision") or index("Breaking Change")) | not)
+  | select([.blockedBy.nodes[] | select(.state == "OPEN")] | length == 0)
+  | select([.closedByPullRequestsReferences.nodes[] | select(.state == "OPEN")] | length == 0)
+  | {number, title, labels: [.labels.nodes[].name]}'
+```
+
+Print the list with each issue's number and title. If the list is empty, say so and stop. If
+`--dry-run` was given, stop here.
+
+## 3. Claim them
+
+For each issue:
+
+1. Assign it to the account `gh` is logged in as:
+   `gh issue edit <n> --repo PaulTrampert/PTrampert.SimplePatch --add-assignee @me`.
+2. Move it to **In Progress** on every project board it belongs to. Look up the issue's project
+   items and each project's `Status` field:
+
+   ```bash
+   gh api graphql -f query='
+   query($n: Int!) {
+     repository(owner: "PaulTrampert", name: "PTrampert.SimplePatch") {
+       issue(number: $n) {
+         projectItems(first: 10) {
+           nodes {
+             id
+             project {
+               id
+               title
+               field(name: "Status") {
+                 ... on ProjectV2SingleSelectField { id options { id name } }
+               }
+             }
+           }
+         }
+       }
+     }
+   }' -F n=<n>
+   ```
+
+   Then set the option whose name is `In Progress`, compared case-insensitively:
+
+   ```bash
+   gh project item-edit --id <item id> --project-id <project id> \
+     --field-id <field id> --single-select-option-id <option id>
+   ```
+
+   If the issue isn't on any project, or a project has no `In Progress` option, note it for the
+   final report and carry on. Don't add the issue to a project or invent a status.
+
+Claim every issue before starting any implementation, so the board shows the whole batch at once.
+
+## 4. Provision the worktrees
+
+Worktrees go inside the primary checkout, under `.claude/worktrees/`, which git ignores. Never
+create one beside the checkout. `$PRIMARY` and `$BASE` come from step 1:
+
+```bash
+WORKTREES="$PRIMARY/.claude/worktrees"
+```
+
+For each issue, pick a branch name. Use `bugfix/<n>-<slug>` if the issue has the `bug` label, and
+`feature/<n>-<slug>` otherwise. `<slug>` is the title in lower case, reduced to `[a-z0-9-]`, and
+cut to a few words. Then run:
+
+```bash
+git -C "$PRIMARY" worktree add "$WORKTREES/issue-<n>" -b <branch> "$BASE"
+```
+
+If the branch or the directory already exists, don't reuse or overwrite it. Skip that issue and
+report it. Never switch the branch of an existing worktree.
+
+## 5. Implement, one sub-agent per issue
+
+Launch every sub-agent **in a single message** so they run concurrently. Use the `general-purpose`
+agent type. Don't pass `isolation`, because the worktree already exists. Give each one this
+prompt, filled in:
+
+> You are implementing GitHub issue #<n> ("<title>") in `PaulTrampert/PTrampert.SimplePatch`.
+>
+> Work only in the worktree at `<absolute worktree path>`. It is already checked out on branch
+> `<branch>`, from `main` at `<base commit>`. Use absolute paths or `git -C` for everything. Don't
+> create another worktree and don't switch branches. If `git branch --show-current` there isn't
+> `<branch>`, stop and report that.
+>
+> 1. Read `AGENTS.md` in the worktree and follow it. It holds the project's coding standards, test
+>    policy, and PR conventions.
+> 2. Read the issue in full: `gh issue view <n> --repo PaulTrampert/PTrampert.SimplePatch --comments`.
+>    If it points to a design document under `docs/`, read that too. A design document on `main`
+>    is settled, so implement it as written. If the issue can't be delivered as written, or needs a
+>    decision that neither the issue nor the design doc makes, stop and report why. Don't guess,
+>    and don't deviate from the design.
+> 3. Implement the issue, with a regression test that fails before the fix. Keep the change scoped
+>    to this issue. Any change to the public API must be additive. If the issue can't be done
+>    without a breaking change, stop and report that.
+> 4. `dotnet build` must pass with no new warnings, and `dotnet test` must pass.
+> 5. Commit in meaningful steps. End each commit message with
+>    `Co-Authored-By: Claude <noreply@anthropic.com>`.
+> 6. Push with `git -C <path> push -u origin <branch>`, then open a PR against `main` with
+>    `gh pr create`. Start the title with `(PATCH): `, `(MINOR): ` or `(MAJOR): `, as `AGENTS.md`
+>    describes. The body explains what the diff doesn't make obvious, gives the test results,
+>    includes `Closes #<n>`, and ends with
+>    `🤖 Generated with [Claude Code](https://claude.com/claude-code)`.
+>
+> Finish with a short report: the PR URL (or why there isn't one), what you tested and how, and
+> anything the reviewer should look at first.
+
+If your own system prompt specifies a co-author attribution, use it in place of `Claude` in the
+co-author line.
+
+## 6. Report
+
+When every sub-agent has finished, give the user one table showing each issue, its branch, and its
+PR link or the reason there isn't one. Also list any issue that couldn't be moved to *In Progress*.
+Leave the worktrees in place, because they hold the branches under review. Give the command to
+remove one once its PR merges: `git worktree remove <path>`.
+
+If a sub-agent fails, don't unassign its issue or move it back on the board. Report the failure and
+leave the decision to the user.
diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 0000000..2139d10
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,112 @@
+# AGENTS.md
+
+Guidance for coding agents working in this repository. `CLAUDE.md` is a symlink to this file, so
+edit this one.
+
+## What this is
+
+PTrampert.SimplePatch is a .NET library for flat PATCH request bodies. It tells a property that was
+omitted apart from one that was explicitly set to `null` or to a value. A controller takes
+`[FromBody] IPatchObject<TWriteModel>`, and the library builds the concrete class at runtime.
+
+## Layout
+
+| Project | Target | Purpose |
+| --- | --- | --- |
+| `PTrampert.SimplePatch` | net8.0 | Core package: `Optional<T>`, `IPatchObject<T>`, `PatchClassBuilder`, JSON converters, validation |
+| `PTrampert.SimplePatch.Schema` | net8.0 | `PatchSchemaTransform`: the OpenAPI schema rewrite shared by both integration packages |
+| `PTrampert.SimplePatch.Swashbuckle` | net8.0 | Swashbuckle schema filter |
+| `PTrampert.SimplePatch.OpenApi` | net10.0 | `Microsoft.AspNetCore.OpenApi` schema transformer. It is net10.0 only because it needs `GetOrCreateSchemaAsync`. |
+| `*.Test` | match their subject | NUnit test projects, one per shipped package (except `Schema`, which the integration tests cover) |
+| `PTrampert.SimplePatch.Sample` | net8.0 | Sample web API, not packed |
+
+Docs are built with docfx (`docfx.json`, `index.md`, `docs/`). The API reference is generated
+into `api/` from XML doc comments. `README.md` is packed into every NuGet package, so it is the
+public face of the library on nuget.org.
+
+## How it works
+
+- `PatchClassBuilder.GetPatchClassFor(type)` generates C# source with CodeDom, compiles it with
+  Roslyn into its own in-memory assembly, and caches the result in a **static** dictionary.
+  `PatchClassBuilder.Instance` is the only instance to use. The public constructor is obsolete.
+- The generated class has one `Optional<T>` property per patchable source property and a `Patch`
+  method. Records are patched with a `with` expression. Other types go through constructor binding
+  and an object initializer.
+- Because the generated assembly is separate, it can only reference **public** types and public
+  setters or init accessors. Non-public source types throw `NotSupportedException`.
+- Source property attributes are carried over: `[JsonConverter]` becomes
+  `[OptionalConverter]`, `[JsonPropertyName]` is copied, and each `ValidationAttribute` becomes an
+  `[OptionalValidation(type, index)]` that runs only when the property is present.
+- `JsonOptionsExtensions.AddSimplePatchConverters` registers `OptionalJsonConverterFactory` and
+  `PatchJsonConverterFactory`.
+- The OpenAPI packages build the patch schema from the **source model's** schema, not from the
+  generated class, so the two contracts can't drift. See
+  `docs/proposals/openapi-contract-generation.md` for the design record.
+
+## Build and test
+
+The SDK is pinned by `global.json` (10.0.x, rolling forward). Run from the repository root:
+
+```bash
+dotnet build
+dotnet test
+```
+
+To build the docs, run `dotnet tool restore`, then `dotnet docfx docfx.json`. Output goes to
+`_site/`, which git ignores.
+
+CI (`.github/workflows/dotnet-library.yml`) uses a shared workflow from
+`PaulTrampert/github-workflows` to build, test, and publish to NuGet on merge to `main`.
+
+## Conventions
+
+- **PR titles must start with `(MAJOR)`, `(MINOR)` or `(PATCH)`.** A CI check enforces this, and
+  the prefix drives the released version. Use the form `(PATCH): Short imperative summary`.
+  Choose the level by the change's effect on the public API of the shipped packages.
+- PR descriptions explain cause, fix, and the alternatives that were rejected, and finish with test
+  results. Reference the issue with `Closes #N`.
+- Public API changes should be additive within a major version. Mark a member `[Obsolete]` with a
+  link to the tracking issue before removing it.
+- Every public member has XML doc comments. `GenerateDocumentationFile` is on, and the docs site is
+  built from them.
+- Comments explain *why*, especially constraints that aren't obvious from the code (for example,
+  why the generated assembly can only see public types). Match the comment density of the
+  surrounding code.
+- Tests use NUnit 4 with the constraint model (`Assert.That`). Test-only types go in the test
+  project's `TestObjects/` folder. Add a regression test that fails before the fix.
+- Don't let the build's warning set grow.
+- Keep `README.md` and `docs/` in step with behaviour changes.
+- Name branches `bugfix/<issue>-<slug>` for bugs and `feature/<issue>-<slug>` otherwise. Never commit
+  directly to `main`.
+- Deliver one issue per PR, small enough for a reviewer to hold in their head.
+- An agent never merges a PR unless the user directly asks it to.
+
+## Worktrees
+
+- **Worktrees go under `.claude/worktrees/` inside the primary checkout**, which git ignores.
+  Never create one beside the checkout. Use
+  `git worktree add .claude/worktrees/<name> -b <branch> origin/main`. `EnterWorktree` and
+  sub-agent isolation already put them there. Remove a worktree with `git worktree remove` once
+  its branch is merged.
+- **A top-level agent** creates its own worktree before making any edits and works there, not in
+  the primary checkout.
+- **A sub-agent** works in the worktree it was handed and doesn't provision another. A sub-agent of
+  a worktree-isolated session usually can't use a new worktree anyway.
+- **Never switch the branch of a worktree you didn't create.** If a worktree is on the wrong
+  branch for your task, say so and stop. Check `git branch --show-current`, not the directory name.
+- **When fanning work out across several issues**, the fanning agent creates one worktree per
+  issue up front, each already on the right branch. It then hands each sub-agent a path that
+  already exists.
+
+## Repository metadata
+
+- `CLAUDE.md` is a symlink to this file. Edit `AGENTS.md`, and never replace the symlink with a
+  copy.
+- `.claude/commands/implement-unblocked.md` is the `/implement-unblocked` slash command. It takes
+  every open, unassigned issue that meets all of these conditions:
+  - no open blocker;
+  - no open PR that will close it;
+  - no `needs decision` or `Breaking Change` label.
+
+  It assigns each issue to the `gh` user and moves it to *In Progress*. It then fans the batch out
+  to sub-agents, one worktree and one PR per issue, following [Worktrees](#worktrees).
diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 120000
index 0000000..47dc3e3
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1 @@
+AGENTS.md
\ No newline at end of file