Release Notes: 0.1.148

Version 0.1.148 is a large release spanning the full 0.1.119–0.1.148 development series. Its pillars are: distribution (a one-command npx install with prebuilt cross-platform bridge binaries, an in-repo storage scaffold); a complete offline production pipeline for turning tested site maps into reviewed, promotable artifacts; a stronger workflow engine (state-machine workflows, failure classification, oversized-result spill); multi-tab routing so an agent drives several tabs at once; and cost observability with cloud persistence and an agent event stream for supervision — plus authoring and reliability improvements throughout.

What Changed

Distribution and install

  • npx @actions-json/bridge — register the MCP bridge with a coding agent in one command. No build step, no manual setup.
  • Prebuilt cross-platform bridge binaries for linux-x64, macos-x64, macos-arm64, and win-x64, produced in CI and downloaded on first run.
  • Bundled tool catalog and primitive dictionary in the npx wrapper, so the bridge advertises the full browser-control surface out of the box.
  • Storage auto-created on first run at ~/.actions-json/storage — the wrapper scaffolds it for you (the old init-storage.sh step is gone).

Storage and scope model

  • In-repo storage scaffold with the private / shared / public scope model, so site maps, observations, overlays, and hosted-agent context have a defined home and a defined visibility.
  • Read-only storage file primitive for authoring flows that need to read a stored file without mutating it.

Production pipeline

An offline toolchain for taking a tested site map to a reviewed, promotable artifact:

  • Offline loader CLI — load and validate a site map without a running browser.
  • Deterministic audit checks — catch broad selectors, weak postconditions, oversized payloads, and cross-surface drift before promotion.
  • Readiness scorecard — a repeatable score for whether a map is ready to promote.
  • Proof package writer and promotion review bundle — assemble the evidence and the diff a reviewer needs to approve a promotion.

Workflow engine and authoring

  • State-machine workflows (x_state_machine) — a workflow can declare a state machine (e.g. the reschedule editor’s date/time axes) so the engine reasons about which state it is in and what transition to make, instead of running a flat step list blindly.
  • Workflow failure classification — each workflow failure is classified with a retryability flag and a concrete safe-recovery hint (e.g. “hide the overlay and retry the mutation,” “read the state projection first — the mutation may have partially landed”), so agents recover deliberately rather than blindly re-running a mutation.
  • bridge.payloads.configure / oversized-result spill — MCP tool results that exceed the inline limit spill to disk and return a path, so a large DOM snapshot or session log no longer blows the response budget.
  • actions.site timeout control — per-call dispatch budget for site actions.
  • Quality-score authoring skill — a scored self-review for authored maps, plus documented failure patterns (overlay occlusion, specific postconditions, two-axis scroll) and new-primitive wiring gotchas.

Multi-tab routing and observability

  • browser.active_tab.set and multi-runtime routing — with several tabs connected, calls route to the chosen tab by runtime id or URL substring, with a designated active tab as the default, instead of erroring on ambiguity.
  • browser.claimed_tabs / browser.tabs — enumerate and manage the tabs the agent has claimed.
  • dom.snapshot_text — a text snapshot primitive for reading page content.

Hosted agent and runtime

  • Side panel shell for the hosted agent, alongside the in-page overlay.
  • Interaction transfer primitives for moving authored interaction state between surfaces.
  • Tab-lifecycle primitives: browser.navigate, browser.open_tab, browser.close_tab, browser.dismiss_dialog — the agent manages tabs and dialogs as standalone steps between actions.
  • Agent memory controls: runtime.agent.memory_clear, with a fix to session log retention so diagnostics are not lost mid-session.
  • Primitive-composition competence guidance so the hosted agent composes primitives to accomplish unmapped operations instead of treating the catalog as a hard boundary.

Cost observability and cloud persistence

  • Expenditure tracking — live per-response cost estimation for the hosted gpt-realtime-2 session (per-modality pricing, cached vs. uncached input), a whole-day total in the overlay, and the cache-miss drain signature (input_cached = 0) that causes surprise overcharges. See Expenditure Tracking.
  • Configurable cloud persistence of usage records to your own storage (S3 first), with an on-device spool that survives offline windows.
  • Tracking-error reconciliation against the provider’s reported usage when an OpenAI key is present.

Supervision

  • runtime.agent.await_event — a supervising agent learns of each hosted-agent output event (responses, tool calls and results, refusals, lifecycle changes) event-driven instead of polling; a timeout returns an explicit idle result so a stall is observable. See Awaiting Agent Events.

Bridge and observability

  • Runtime connect/disconnect lifecycle log persisted on the bridge, so runtime churn is visible after the fact.
  • Dev bridge credential hydration so a development bridge can supply the hosted agent’s credentials.

What To Test

  • Fresh install: npx @actions-json/bridge mcp, register it, load the extension, and confirm storage is created and the agent can drive tabs without manual setup.
  • Run a hosted-agent session and watch the overlay cost meter update per response with a running whole-day total.
  • Configure S3 credentials and confirm per-response usage records land in your bucket; interrupt connectivity mid-session and confirm the spool flushes on reconnect.
  • From a coding agent, drive the hosted agent and call runtime.agent.await_event in a loop; confirm each event arrives and a deliberate silence returns idle.
  • Run a site map through the offline pipeline: load, audit, score, and build a proof/promotion bundle, and confirm the audit flags a deliberately broad selector.