cdp-toolkit 1.9.5 → 1.10.0

package/CHANGELOG.md

@@ -5,6 +5,60 @@ All notable changes to cdp-toolkit are documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [1.10.0] - 2026-08-18
+
+**MCP context-rot reduction: the `tools/list` payload every consuming agent loads is cut ~24% (≈26,600 → ≈20,150 tokens), with NO change to tool behavior, names, parameters, or schema shapes** — the trim touched `description` strings only and was proven shape-identical to the prior manifest (a strip-descriptions deep-equal check gates the regeneration).
+
+### Added
+
+- **Server `instructions` block** (MCP `initialize` result). The conventions previously re-stated inside every tool schema — the target-selector grammar, the lease-token model, the Chrome-only MV3 `worker:`/`wake` arm, and the `origin` vocabulary — are now stated once here, loaded a single time at session start. Front-loaded and kept under Claude Code's 2 KB instructions cap so nothing critical is clipped.
+
+### Changed
+
+- **Deduplicated the shared parameter blurbs.** The 453-character `lease` description was copy-pasted onto 42 tools (~5,900 tokens of pure duplication) and the `target` selector was written 17 different ways; both now collapse to short, consistent pointers at the server `instructions`. Every per-tool description was tightened for signal density with each tool's load-bearing footgun front-loaded — the "declared-not-discovered" warnings are all preserved, just terser (e.g. `list_pages` 2,702 → 1,188 chars, `claim_page` 2,058 → 1,198).
+
+### Fixed
+
+- **`list_pages`'s description exceeded Claude Code's 2 KB per-description cap and was being silently truncated from the end** (dropping its `responsive`/`humanActiveMs` tail). Every tool description is now ≤ 2 KB.
+
+## [1.9.7] - 2026-08-14
+
+### Fixed
+
+- **A dropped Firefox WebDriver BiDi connection is now re-dialed instead of erroring
+  permanently.** The backend memoizes one `BidiConnection` per endpoint, and `getConnection`
+  (`src/bidi/driver.ts`) returned the cached one unconditionally; once its socket dropped —
+  a Firefox restart, an idle/network drop, or a lost BiDi session slot — every later call
+  rejected forever with `connection not open`, for the entire life of the MCP server.
+  `getConnection` now evicts a dead cached connection (via a new `BidiConnection.isOpen`
+  getter) and re-creates the session on the next call, so Firefox tools recover on their own
+  instead of needing a server restart.
+- **Firefox launch mode no longer returns the hardcoded macOS Firefox path on Linux.**
+  `resolveBinary` (`src/bidi/launch.ts`) returned the first truthy candidate without checking
+  it existed, so on Linux it handed back `/Applications/Firefox.app/Contents/MacOS/firefox`
+  and never fell back to `firefox` on `PATH` — `ENOENT` at spawn, in launch mode only. An
+  absolute candidate (`FIREFOX_BIN`, the macOS path) is now only returned if it exists on this
+  box; the bare `firefox` name is left for `spawn` to resolve against `PATH`.
+- **Recorder artifact directory is now resolved per call, so `CDP_ARTIFACT_DIR` is honored at
+  runtime instead of being frozen at import time.** The buffer directory in `src/tools/recorder.ts`
+  was a module-level `const` bound the moment the module loaded, so a caller (or a test) that set
+  `CDP_ARTIFACT_DIR` afterward was ignored and `recFile`/`captureFile` still pointed at
+  `/tmp/cdp-toolkit`. On a clean CI runner that path does not exist, so the first buffer
+  `writeFile` threw `ENOENT` and the worker-capture suite failed there while passing on any box
+  where `/tmp/cdp-toolkit` happened to exist. It now reads the env var per call through a local
+  `artifactDir()` helper — the same "read per call, not at module load" pattern as `leaseDir()` in
+  `src/leases.ts` — so the directory is resolved at write time and a redirect takes effect.
+
+## [1.9.6] - 2026-08-13
+
+**Fix: `upload_file`'s `files` parameter is now declared array-only in the tool manifest** (was
+`"type": ["string","array"]`). Google's Gemini function-calling API rejects union-typed parameter
+schemas — the `any_of` branch's untyped `items` fails its `$type == Type.ARRAY` predicate — which
+hard-fails the entire `GenerateContentRequest` and kills any Gemini-backed agent session the
+moment this manifest is included in its toolset (observed as subagents returning empty results on
+every dispatch). Runtime behavior is unchanged: the handler still accepts and wraps a bare string
+for direct CDP/CLI callers; MCP/model callers now pass `[path]` per the declared schema.
+
 ## [1.9.5] - 2026-08-11
 
 **Documentation and stale code-comment clarifications; NO behavior changes since 1.9.4.** Two