Fix the cache key source and provenance; stop writing conversation-identity headers

.github/workflows/test.yml

@@ -0,0 +1,20 @@
+name: test
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        node: ["20", "22"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-node@v4
+        with:
+          node-version: ${{ matrix.node }}
+      # No install step: this project has no dependencies.
+      - run: npm test

.gitignore

@@ -0,0 +1,3 @@
+node_modules/
+context-cache.log
+*.log

CHANGELOG.md

@@ -0,0 +1,63 @@
+# Changelog
+
+## 0.2.0
+
+### Breaking
+
+- The plugin no longer writes the `x-session-id`, `conversation_id` or
+  `session_id` headers. Those names identify a *conversation*, and a
+  project-stable value is wrong in them: on opencode's OpenAI/Codex path,
+  `x-session-affinity` keys a WebSocket connection pool whose `busy` and
+  `fallback` state would then be shared by every concurrent session in the
+  project. opencode core already sends `x-session-affinity` and `X-Session-Id`
+  derived from the real session ID. A gateway that parsed the underscore names
+  must be reconfigured to read core's headers instead.
+
+  The plugin also previously wrote `x-session-id` while core writes
+  `X-Session-Id`. Those are distinct keys in a JavaScript object and only
+  collapse at the HTTP layer, so both values were reaching the wire.
+
+### Fixed
+
+- The cache key is derived from `PluginInput.worktree` rather than
+  `process.cwd()`. One `opencode serve` process serving several projects
+  previously gave all of them the same key, because the plugin factory is
+  invoked per project while `process.cwd()` stays the server's launch
+  directory.
+- `prompt_cache_key` (deepinfra, cerebras) is now handled. Only the camelCase
+  spelling was written before, so those providers were unaffected by the plugin.
+- The key is replaced only when it still holds opencode's own session-ID
+  default, so an explicit operator setting, a model or agent option, or another
+  plugin's value is no longer overwritten.
+- Explicit overrides longer than 64 characters or containing non-printable
+  characters are hashed rather than sent verbatim.
+- A degenerate `/` worktree falls back to the session directory, mirroring
+  opencode's own guard, instead of collapsing every project onto one key.
+- The debug log moved out of the plugin directory to
+  `$XDG_STATE_HOME/opencode/context-cache.log`, so the documented
+  `./plugins/...` install no longer writes a log file into your repository.
+
+### Added
+
+- `OPENCODE_CONTEXT_CACHE_SCOPE` (`worktree` default, `directory`, `session`).
+  `session` is a full opt-out and takes precedence over an explicit key.
+- `OPENCODE_CONTEXT_CACHE_LOG` to relocate the debug log.
+- Plugin `options` support, so scope and key can be set from `opencode.jsonc`.
+  Environment variables take precedence over options.
+- Always-on, deduplicated operator warnings for compatibility failures, so a
+  renamed upstream field surfaces without the debug flag being on first. A
+  provider that simply has no cache key field is **not** one of those failures
+  and is silent: Anthropic and every `@ai-sdk/openai-compatible` provider
+  (DeepSeek among them) have no such setting, so warning there would fire on a
+  routine configuration and teach operators to tune out the channel. That case
+  is recorded in the debug log as `reason=no-fields`.
+- A README section stating exactly which providers the plugin does and does not
+  affect, so its scope is knowable without reading the source.
+- A `package.json`, so the plugin can be installed by npm identifier.
+- A test suite and CI.
+
+### Removed
+
+- The "digest detection to avoid double-hashing" behavior, which could never
+  fire: it was only reachable from a precedence level that was itself
+  unreachable. Explicit overrides are now used verbatim when they are safe.