alibaba/open-code-review — hybrid deterministic + LLM-agent code review CLI, battle-tested at Alibaba's scale

Install (Go 1.21+)

go install github.com/alibaba/open-code-review@latest

Review staged + unstaged diff in current repo

ocr review

Scan entire files rather than diff hunks

ocr scan ./path/to/changed/files

Gemini API: `deep-research-pro-preview-12-2025` deprecated — shutdown October 23, 2026

Anthropic Managed Agents: Dynamic Workflows (beta) — agent writes and runs its own fan-out orchestration programs

{

"type": "multiagent_20261001",

"workflows": {"type": "enabled"}

}

Tell the agent in its system prompt when to initiate a run and what tasks warrant one. You then follow the run via workflow_run.* events on the session's primary event stream — the sequence is: workflow_run.created (with phases[] declared), workflow_run.status_running, workflow_run.phase_started/ended per phase, then workflow_run.status_ended with a result field. Each agent within the workflow runs in its own session thread (visible in the thread list, grouped by workflow_run_id). A run's threads are exempt from the session's child-thread limit. The default run lifetime is 24 hours; the agent can set a shorter one. Important edge cases: archive a session to cancel open runs; threads within a run cannot themselves start nested runs; inline agents (those the workflow defines) inherit the session agent's model, tools, and MCP servers. You can ask the agent to print its own workflow code after a run completes — useful for debugging.

Claude Code v2.1.296 — workflow model env var, managed policy for Code tab, non-UTF-8 file protection, hooks fix

• CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL: New env var routes all Dynamic Workflow sub-agents to a specific model while other subagents keep theirs — direct companion to today's server-side Dynamic Workflows beta.

• managed.policies[] gains code key: Same configuration options as cli; the policy now applies in Claude Desktop's Code tab too.

• Non-UTF-8 file protection: Edit and NotebookEdit now refuse to edit files encoded in Windows-1252, Shift-JIS, or GBK, which they previously silently corrupted.

• allow_large option for Read: Read tool can now read a text file past the usual size limit in a single call when explicitly opted in.

• Sonnet 5.5 cache-read pricing fix: /cost, the status line, --max-budget-usd, and SDK cost figures now price Sonnet 5.5 cache reads at $0.10/M tokens (was displaying $0.20/M — per the platform's Oct 7 price cut).

• Managed hook behavior fix: Managed PreToolUse and prompt hooks that deny a call with "continue": false now refuse the call without ending the session turn (previously ended the turn).

• Windows PowerShell: Commands up to 32 KB now go through allow rules and read-only detection; the old threshold was ~1 KB, so long PowerShell calls no longer always prompt.

• autoCompactWindow for subagents: Frontmatter and --agents definitions can now set this to let a subagent compact earlier than the main conversation.

The agent-orchestration layer is now the fastest-moving layer in the stack