fix: bump agent-client-protocol to >=0.8.1

[simonrosenberg]

Summary

• Bumps agent-client-protocol from >=0.7.0,<0.8.0 to >=0.8.1
• Resolves dependency conflict with openhands-sdk v1.12.0 which requires agent-client-protocol>=0.8.1
• The previous upper bound (<0.8.0) prevented the CLI from upgrading to newer SDK versions, blocking releases that include other SDK fixes

Context

PR #2133 added agent-client-protocol>=0.8.1 as a required dependency of openhands-sdk. This created a mutual exclusion with the CLI's <0.8.0 constraint, making it impossible to release CLI versions that include SDK v1.12.0+.

What changed in ACP 0.8.1

All code changes in this PR are direct consequences of the agent-client-protocol 0.7.x → 0.8.1 upgrade:

Renamed symbols (runtime ImportError without these):

• StdioMcpServer → McpServerStdio (1 production file, 2 test files)
• SessionUpdate2–SessionUpdate6 → named types: AgentMessageChunk, AgentThoughtChunk, ToolCallStart, ToolCallProgress, AgentPlanUpdate (1 test file)

Protocol signature changes (pyright errors without these):

• new_session: mcp_servers parameter changed from list[Any] to list[...] | None = None
• load_session: parameter order changed from (cwd, mcp_servers, session_id) to (cwd, session_id, mcp_servers), and mcp_servers became optional
• All callers use keyword arguments, so the reorder is safe

New Protocol methods (pyright errors without these):

• fork_session, resume_session, set_config_option were added to the Agent Protocol in 0.8.1 — stub implementations added to BaseOpenHandsACPAgent

Test plan

• Verify uv lock resolves without conflicts
• Verify pre-commit (pyright) passes
• Verify unit tests pass
• Verify E2E ACP server test passes in build

🤖 Generated with Claude Code

---

🚀 Try this PR

uvx --python 3.12 git+https://github.com/OpenHands/OpenHands-CLI.git@fix/bump-acp-version

[github-actions]

Coverage Report •

File	Stmts	Miss	Cover	Missing
openhands_cli/acp_impl/agent
base_agent.py	229	30	86%	132, 175, 216, 287–289, 349, 359, 369, 384–387, 391–398, 423, 505–506, 509–510, 586, 656–658
local_agent.py	84	4	95%	80, 144, 154, 168
remote_agent.py	147	63	57%	85–86, 92–93, 97–98, 102, 104–108, 111–113, 123–124, 127–129, 133–135, 145–146, 161–163, 167, 169–170, 176, 180–181, 183, 192–193, 198–200, 211–212, 217, 219–220, 224, 231–232, 234–235, 238–240, 242, 250, 252–253, 284, 346–348, 354–355
openhands_cli/acp_impl/utils
mcp.py	25	4	84%	69–72
TOTAL	6647	931	85%

[jpshackelford]

I have run simple manual tests (single prompt) against the CLI, acp via Zed and acp via JetBrains IntelliJ IDEA. All looks good.

[jpshackelford]

🟢 Good Taste - Clean, Mechanical Dependency Upgrade

Linus-Style Analysis:

This PR is boring in the best possible way. It's a dependency version bump that does exactly what it needs to do and nothing more. The changes fall into three categories:

1. Symbol Renames (Mechanical, correct)

• StdioMcpServer → McpServerStdio
• SessionUpdate2–SessionUpdate6 → properly named types (AgentMessageChunk, AgentThoughtChunk, etc.)

These are upstream API renames. You have no choice. The new names are better anyway — SessionUpdate6 was garbage naming from the library author.

2. Signature Compliance (Mechanical, correct)

• mcp_servers: list[Any] → mcp_servers: list[Any] | None = None
• Parameter reorder in load_session (callers use kwargs, so safe)

Again, upstream dictates this. The optional None default is fine — MCP servers are genuinely optional.

3. Protocol Stub Implementations (Correct approach)

async def fork_session(...) -> ForkSessionResponse:
    """Fork a session (not supported)."""
    raise RequestError.method_not_found("session/fork")

This is exactly what you should do when a Protocol forces you to implement methods you don't support. Raise method_not_found. Don't silently return garbage. Don't pretend you support it.

The only slight smell is set_config_option returning None instead of raising, but looking at the ACP Protocol, that seems to be the expected behavior for "not implemented" in that particular method.

What's not here (good):

• No "while I'm here, let me also..." scope creep
• No refactoring
• No new features hiding in a dependency bump
• Tests updated to match the renames (necessary, done)

---

✅ VERDICT: Worth merging

This is a dependency upgrade PR that doesn't try to be anything else. The code changes are 100% mechanical consequences of the upstream API changes. All CI checks pass. The PR description explains why every change exists.

KEY INSIGHT:
The author understood that dependency upgrades should be boring — the diff should be predictable from reading the upstream changelog, and this one is.

---

One Minor Note (Non-blocking):

[pyproject.toml, Line 26] The constraint changed from >=0.7.0,<0.8.0 to >=0.8.1 with no upper bound. This is fine for now since you're actively tracking ACP releases, but be aware that a future ACP 0.9.0 with breaking changes will auto-install and break things. Consider >=0.8.1,<0.9.0 if you want to be defensive. But that's a policy decision, not a code quality issue.

[jpshackelford]

@simonrosenberg Since this agent-client-protocol does tend to be somewhat reckless, the suggestion of pinning with >=0.8.1,<0.9.0 or something more restrictive is a good idea.