Skip to content

v1.12.5.md

Latest commit

 

History

History
241 lines (175 loc) · 9.03 KB

v1.12.5.md

File metadata and controls

241 lines (175 loc) · 9.03 KB
Raw

OpenChrome v1.12.5

Release date: 2026-05-25

Summary

v1.12.5 is the first release cut after the OpenChrome direction was clarified in SSOT issue #1359: OpenChrome is a host-neutral MCP browser harness. The release keeps the package focused on real Chrome control, deterministic MCP semantics, and portable evidence that can run under Codex, Claude Code, Hermes Agent, OpenClaw, and other MCP-capable hosts without assuming a specific agent runtime.

This patch integrates the full open PR stack that was pending after v1.12.4:

  • startup Chrome lifecycle polish for real browser sessions;
  • a state-graph and verified-memory pilot foundation;
  • automatic skill/episode curation from successful task contracts;
  • benchmark smoke and release-gate hardening that avoids overstated competitor claims;
  • documentation/runbook updates for reproducible benchmark operation.
v1.12.4  ->  v1.12.5   (this release)

  12 PRs integrated (#1347 through #1358)
  1 host-neutral direction SSOT used as the merge gate (#1359)
  1 post-merge develop compatibility repair for lifecycle API reuse
  0 MCP tool removals
  0 npm publish performed by this release-prep step

Upgrade

npm install -g openchrome-mcp@1.12.5
openchrome check
openchrome doctor

Fresh install:

npm install -g openchrome-mcp@latest
openchrome check
openchrome info

If you use OpenChrome as an MCP server, CLI, or embedded harness, this is a drop-in patch upgrade from v1.12.4. The new pilot/benchmark pieces are additive and are gated behind their existing commands, flags, and test contracts.

Breaking changes

None. v1.12.5 does not remove or rename MCP tools, transports, CLI commands, or public configuration options.

Operational notes:

  • The release does not publish npm from the release-prep automation. Maintainers should run npm publish separately after reviewing the tag and release notes.
  • Pilot memory/curator behavior remains conservative and evidence-backed. It is intended to build portable host-neutral memory artifacts, not host-specific automation shortcuts.
  • Benchmark additions improve reproducibility and release gating; they should not be read as a claim that every full live competitor benchmark is now published.

Direction: host-neutral MCP browser harness

Issue #1359 was used as the final product-direction gate for this release. The merged work supports the same core position:

  1. Real browser control first. OpenChrome should remain a reliable MCP layer over actual Chrome/CDP behavior rather than becoming a synthetic agent runner.
  2. Host neutrality. Features should work from different MCP hosts and coding agents without depending on Codex-specific, Claude-specific, or local orchestration assumptions.
  3. Evidence over claims. Benchmarks, memory, and replay artifacts must be backed by explicit contracts, traces, and failure modes.
  4. Portable learning loops. Pilot memory should accrete reusable skills from verified task outcomes while keeping deterministic schemas and clear provenance.

Browser lifecycle and CDP reliability

Reuse the startup blank tab on first createPage() (#1347)

When OpenChrome launches an isolated Chrome instance, Chrome commonly starts with an initial blank/new-tab target. v1.12.5 teaches the first createPage() path to reuse that startup blank target when it is safe to do so. This avoids an extra blank tab in isolated launch flows while preserving attach-mode safety.

The behavior is guarded so it does not reuse a target when:

  • OpenChrome attached to a user-owned Chrome instance;
  • a target is already indexed;
  • multiple untracked page candidates exist;
  • the candidate target is a real site rather than a blank/new-tab URL;
  • a non-default browser context is requested;
  • resolving the startup target fails.

A post-merge compatibility repair keeps this behavior on the existing public CDPClient.getChromeLifecycleMode() API instead of adding a second private lifecycle method. This preserves the same lifecycle ownership contract used by session cleanup and other CDP flows.

Pilot state graph and verified memory foundation

State graph hash anchor (#1348)

The pilot stack now has a minimal state-graph anchor over task transaction records. This gives later memory and replay features a stable way to name page state without coupling to a particular host process.

DOM skeleton v2 hashing (#1353)

A second state hash mode folds a bounded DOM skeleton into the page-state hash while keeping the original v1 hash byte-stable. The v2 path is bounded and has fallback protection so hashing cannot expand without limits.

Pilot bootstrap lifecycle (#1354)

Pilot startup and shutdown hooks are now idempotent. Embedded server tests and host integrations can start/stop the server repeatedly without leaking pilot subscriptions or timers.

Automatic skill extraction from successful contracts (#1355)

Successful contract-backed task runs can now feed a conservative curator runner that extracts reusable skill candidates. This is the first step toward verified memory that is created from actual task evidence rather than free-form agent summaries.

Sidecar-backed curator stats and failed-run evidence (#1356)

Curator sidecars now expose stats and record ok=false outcomes. Failed-run artifacts are hardened so negative evidence is available to future selection and review logic.

Auto-recall on task start (#1357)

Pilot auto-recall can promote curated skills when a new task starts. The recall path is gated before pilot imports so non-pilot users do not pay unnecessary startup or dependency costs.

Deterministic SKILL.md body generation (#1358)

Curated journal slices can be rendered into deterministic SKILL.md bodies. Sensitive body-builder arguments are redacted so generated memory artifacts do not accidentally persist credentials or private payloads.

Selector confidence accretion

The auto-memory layer can accrete selector confidence from task contracts. Exact selector decay was corrected so repeated evidence adjusts confidence predictably.

Benchmark and release-gate hardening

v1.12.5 integrates the benchmark PR ladder that was planned after v1.12.4. The focus is still safe measurement infrastructure and reproducibility, not inflated headline claims.

Contract and smoke hardening

  • Stress thresholds for loaded runners are stabilized.
  • Competitor smoke provenance is pinned and aligned with COMPETITORS.md.
  • Empty-payload smoke rows are rejected.
  • Operator CDP sharing is documented for repeatable smoke runs.

Measurement axes and live-run gates

  • Non-LLM benchmark axes are measured before live promotion.
  • Real-world corpus contracts are locked before live episodes.
  • Live LLM samples are gated before native execution.
  • Native competitors are routed without OpenChrome fallback.
  • Real-world tasks are stressed through final postconditions.
  • Full benchmark release gates derive task counts from the configured suite and perform preflight checks.

Operator runbook

The smoke runbook now documents the six-way competitor matrix and clarifies CDP setup expectations for operators. This reduces the chance of accidentally mixing mock/scaffold evidence with live benchmark evidence.

PRs included

PR Area Result
#1347 CDP lifecycle Reuse isolated startup NTP target on the first safe createPage() call.
#1348 Pilot state graph Add the initial transaction-record node hash anchor.
#1349 Benchmark stack Merge the full benchmark release-gate ladder into develop.
#1350 Benchmark docs Add the competitor smoke runbook.
#1351 Benchmark pins Align Playwright smoke registry pins with COMPETITORS.md.
#1352 Benchmark sanity Share operator CDP and reject empty-payload smoke rows.
#1353 State graph v2 Fold bounded DOM skeleton data into v2 hashes while preserving v1.
#1354 Pilot lifecycle Make bootstrap/cleanup idempotent.
#1355 Curator Auto-extract skills from successful contract evidence.
#1356 Curator stats Add sidecar-backed stats and failed-run logging.
#1357 Auto-recall Recall promoted curated skills when a task starts.
#1358 Skill body builder Generate deterministic SKILL.md bodies from journal slices.

Verification

Release preparation verified locally after all PR merges and the post-merge compatibility repair:

gh pr list --state open --limit 100 --json number,title,url | jq length
npm run build
npm test -- --runInBand tests/cdp/create-page-startup-reuse.test.ts tests/integration/embedded-server.test.ts
npm run lint:tier

Additional pre-merge verification covered:

  • pilot state graph, runtime events, curator, auto-memory, and harness flag test suites;
  • benchmark competitor smoke test and benchmark build path;
  • #1347 startup-tab reuse tests and build;
  • #1348 embedded-server test after aligning the pilot cleanup mock.

At release-prep time, the latest develop push CI for the final compatibility repair was queued on GitHub Actions, with no known failing local verification.