Skip to content

fix(harness): The cached Claude ACP adapter bundles a CLI too old for the configured model, so every Harness turn fails silently to Static #514

Closed

Description

@omesser

Found while running probe 1 of #508.

What happens

The claude launch row is npx -y @agentclientprotocol/claude-agent-acp (src-tauri/src/harness.rs:81). npx reuses whatever it cached the first time. On this machine the cache holds adapter 0.49.0, which pins @anthropic-ai/claude-agent-sdk 0.3.185 and bundles Claude Code 2.1.185. The registry is at adapter 0.75.1 / SDK 0.3.257. ~/.claude/settings.json sets "model": "opus[1m]", and 2.1.185 refuses it:

API Error: 400 Claude Code 2.1.185 does not support this model; version 2.1.251 or newer is required.

Every assistant record in the Director's session file today (19 of 19, 12:58–17:06) is that text. The action log for the same window has 154 turn events carrying it as error, each followed by parsed … result: failed.

What the user sees

Nothing, which is the severity. ADR-0008 routes a failed turn to the Static Director, so the buddy keeps moving on static weights and the Settings source row says static weights are answering. The Harness is attached, signed in, and answering the protocol; it never once proposed a Behavior or said a line. A Poke, Throw, Summon, or Perch drop gets the same static pick it would get with no Harness at all, and the ambient backoff keeps spending a Harness process on wakes that cannot succeed.

The Chat surface shows the same failure as No answer came back. (src/chat.js:332) — the copy for "static weights took over", with no hint that the reason is a version, or that claude update does not fix it because the adapter bundles its own CLI.

So a user who set up the Harness and picked a current model gets a buddy indistinguishable from one with the Harness Off, and nothing in the product says why. It has been that way on this machine for the whole of today's session.

Fix

The row should not trust a stale cache: npx -y @agentclientprotocol/claude-agent-acp@latest re-resolves on each spawn, at the cost of one registry round-trip per attach. Pinning a version drifts the same way in slower motion.

Separately, a turn whose text starts API Error: is not a proposal near-miss. Surfacing that line in the Chat surface and the Settings source row, instead of No answer came back., would have made this visible on the first wake.

— Claude, in a session with @omesser

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    • Status
      Done

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions