Agentic Pipelines - Outdated Claude runtime prevents now Opus 5/5.5 and Sonnet 5 models

Question

I'm trying to use Bitbucket Agentic Pipelines, but the current outdated runtime is preventing usage of Opus 5.5

We run a code review in Bitbucket Agentic Pipelines with provider: claude. The agent step can't use current Claude models. As far as we can tell the cause is the Claude Code version the installer bundles, not our configuration.

What we see

* claude-sonnet-5 fails with a 400 that asks for thinking.type: "adaptive".
* claude-opus-5-5 fails the same way. The client sends thinking.type: "enabled", and current models reject that at every effort level.
* Get past that, and the API refuses the client itself: "version 2.1.280 or newer is required".

Where it comes from

The Claude step runs bitbucket-pipelines-agent-installer:7-claude. The ACP handshake in the step log reports claude-agent-acp 0.25.3. On npm, that version pins @anthropic-ai /claude-agent-sdk 0.2.91, which carries Claude Code 2.1.91. The old client picks thinking.type: "enabled" from its own model table, and it is below the 2.1.280 minimum that current models need.

The current release, claude-agent-acp 0.81.2, pins claude-agent-sdk 0.3.280, which carries Claude Code 2.1.280.

What we tried

* Installing @agentclientprotocol /claude-agent-acp@0 .81.1 in the step before the agent: line. The handshake still reported 0.25.3, so the installer's own copy is what runs.
* A proxy that rewrote thinking.type: "enabled" to "adaptive". That got past the 400, and the next request was refused on the Claude Code version.

The same pipeline with provider: codex runs current OpenAI models without problems.

Questions

* Is there a timeline for updating the Claude installer to a current claude-agent-acp?
* Is there a supported way to choose the installer version, or to run our own claude-agent-acp, in a pipeline?
* Could the installer release notes list which claude-agent-acp and Claude Code version each tag bundles? Then we could tell from the tag when current models become usable.

Abes Dabir · Answer

Thanks for raising this and sharing the support answer. My interaction with support for the same issue didn't include the two weeks timeline, so I'm happy to hear that.

Interestingly enough, I'm using Claude Code backed by AWS bedrock instead of Anthropic and it does work with Opus 5. I only hit a wall when I tried using Opus 5.5 and as expected, the old version blocked me. I also want to use built-in skills like "/code-review" and "/security-review" which the old version don't have.

René Dyhr · Answer

Thank you for the answer!

So we can actually use the newer models by using the MAX_THINKING_TOKENS variable to override some of the settings, kinda, if I understand it correctly?

barryvdh · Answer

Update from Support:

Good day.

I reviewed this with the Pipelines team, and they confirmed that the following upgrades are planned:

* @agentclientprotocol /claude-agent-acp from 0.25.3 to 0.75.1
* Claude Code from 2.1.92 to 2.1.136 (Note: I think they mean 2.1.236)

These upgrades are expected to be rolled out within the next two weeks. Once the upgrade is available, you should be able to use Sonnet 5, Opus 5, and Sonnet 5.5 without encountering the current compatibility errors.

Until then, to avoid the thinking-related errors, you can use:

MAX_THINKING_TOKENS=0

Regarding Opus 5.5 specifically, I have also raised the following feature request for tracking:

https://jira.atlassian.com/browse/BCLOUD-24256

Agent upgrades are rolled out in a controlled manner to minimize the impact, as the Agentic Pipelines functionality is used by many customers. An agent upgrade can also introduce backward compatibility considerations with previously supported model versions, so the team is validating the upgrade carefully.

I hope this helps clarify the current status.