chore(runtime): bump Claude, Codex, DeepSeek and browser pins

[agentconnect-md-test]

Cloud pool sandboxes install fixed runtime versions, so published ACP updates do not reach Cloud agents until the image pins change. Update the four application pins in docker/runtime-sandbox.Dockerfile:

Package	Previous	New
agent-browser	0.38.1	0.38.2
Claude ACP	0.84.0	0.85.1
AgentConnect Codex ACP fork	2.1.0-agentconnect.1	2.1.1-agentconnect.1
DeepSeek Harness ACP	0.4.35	0.4.36

The executable names remain agent-browser, claude-agent-acp, codex-acp, and dsh-acp. Source comparison found no permission mode ID changes: Codex keeps read-only, workspace-write, agent, agent-full-access; DeepSeek keeps read-only, workspace-write, danger-full-access; Claude's advertised modes are unchanged. Codex and Claude have no model ID mapping changes.

Model ID change: DeepSeek 0.4.36 changes its default and selectable model ID from deepseek-v4-flash to deepseek-flash as its bundled DSH moves from 0.1.7-rc.2 to 0.2.0-rc.2. The adapter includes a live-catalog substitution for persisted deepseek-v4-flash sessions, and falls back to an available model for unavailable IDs. deepseek-v4-pro remains selectable. The pool's runtime table is generated from the installed adapter at image build time.

Claude ACP 0.85.1 also restores background-task stops on replay, fails turns that leave foreground tools unfinished, and adds an opt-in experimental ACP v2 handshake. Its default remains ACP v1. Codex 2.1.1 changes attachment replay, custom answers, and service error messages; the AgentConnect fork patches were reapplied and published.

Validation: the Codex fork's tag-triggered CI passed typecheck, tests, build, and npm publish. Locally, npm installed Claude ACP 0.85.1 and DeepSeek ACP 0.4.36. DeepSeek's bundled DSH 0.2.0-rc.2 unpacked, its ACP initialize probe returned protocol version 1, and the image's no-search preset script generated the expected bundle. git diff --check passed. Docker is unavailable on this runner, so the image was not built locally.

Created by Codex . GPT-6

[agentconnect-md-test]

I approve PR #2771 with no blocking findings, but the formal review didn't go through: GitHub rejected the submitCodeReview call with a 422 error. Re-requesting the review will retry it; until then, this comment is the review.

The PR changes three version pins in docker/runtime-sandbox.Dockerfile. I checked each new version against npm and against the previous version's source.

Codex fork 2.1.0-agentconnect.1 → 2.1.1-agentconnect.1

• The publish run 36951875038 passed both its verify and publish jobs.
• npm returned 404 for the new version for a few minutes while it processed the release. It is now available, and the agentconnect tag points to it.
• The package hash matches the publish log.
• The executable is still codex-acp -> dist/index.js, and the @openai/codex ^0.159.1 and SDK ^1.5.0 ranges are unchanged.
• The permission-profile config (CODEX_ACP_PERMISSION_PROFILE_CONFIG) is still in the bundle, as are the four permission modes: read-only, workspace-write, agent, agent-full-access.
• Between the two fork tags, the only changes are:
  • how attachments appear when a session is reloaded
  • a legacy key for custom answers
  • more readable service error messages
  • tests
• On reload, the user's attached images and file mentions now come back as resource_link blocks. The daemon ignores those user-message updates, so this doesn't affect AgentConnect.

Claude ACP 0.84.0 → 0.85.0

• The executable, the Node version requirement (22 or later), and the permission modes are unchanged. The bundled Claude SDK moves from 0.3.284 to 0.3.286.
• Changing effort now keeps an ultracode setting configured in settings.
• Reloading a session now hides messages that came from other agents, sessions or channels.
• No model ID mapping changed.

agent-browser 0.38.1 → 0.38.2

• This is a patch release with the same executable and Node requirement.

No other file in the repo pins these versions, and the image rebuilds its own runtime table from the installed adapters, so nothing else needed to change.

I couldn't build the runtime image because this runner has no Docker. The PR's CI, including the Linux sandbox job, was still running when I reviewed.

_{sent by review-bot (Claude Agent · opus) · open in session}

[agentconnect-md-test]

I approve PR #2771 at 44b4fc7e with no blocking findings, but the formal review didn't go through: GitHub rejected the submitCodeReview call with a 422 error. Re-requesting the review will retry it; until then, this comment is the review.

The new commit raises two more pins in docker/runtime-sandbox.Dockerfile:

Package	Old	New
Claude ACP	0.84.0	0.85.1
Codex ACP fork	2.1.0-agentconnect.1	2.1.1-agentconnect.1 (no change since last review)
DeepSeek Harness ACP	0.4.35	0.4.36
agent-browser	0.38.1	0.38.2 (no change since last review)

Claude ACP 0.85.0 → 0.85.1

• This is the current latest release on npm. The executable name, Node requirement, permission modes and bundled Claude SDK are unchanged. The ACP SDK goes from 1.5.1 to 1.6.0.
• The new ACP v2 code only runs when CLAUDE_AGENT_ACP_EXPERIMENTAL_V2=1 is set; without it, the adapter still speaks ACP v1.
• One behavior change worth watching after rollout: if Claude ends a turn while a tool call still has no result, the adapter now marks that tool call as failed and fails the turn with an internal_error. Before, the turn just ended quietly. Users will see these turns fail.

DeepSeek Harness ACP 0.4.35 → 0.4.36

• This is also the current latest release. The executable name, the --bundle handling and the dsh binary that the install script links are unchanged.
• The bundled DeepSeek runtime moves from 0.1.7-rc.2 to 0.2.0-rc.2, but its four shipped preset files are byte-identical to the old version.
• I ran the image's no-search bake script (bake-dsh-preset.mjs) against the presets from both versions, and both builds succeeded.
• Adapter changes:
  • Live streaming now uses dsh 0.2's new streaming event.
  • Questions to the user now handle being aborted and answers that arrive late.
  • Sessions using the removed deepseek-v4-flash model are switched to deepseek-flash.
• The control plane's default model is already deepseek-flash; only web test fixtures still mention the old ID.

Nothing else in the repo pins these versions.

I couldn't build the image because this runner has no Docker. On the new head, three CI checks have passed (Scope, Conventional PR and Daemon Store). Build, Unit, Windows, Integration, Evaluation and Linux sandbox were still running when I reviewed.

_{sent by review-bot (Claude Agent · opus) · open in session}