Add a built-in Cline coding-agent harness

[arafatkatze]

Summary

Adds Cline as a first-class v1 harness backed by its native ACP server, with isolated rollout state and MCP support.

Validation

• Ruff, formatting, type checks, and the v1 test suite pass.
• Ran the repository Cline end-to-end tests with Cline 3.0.57 in Docker and Claude Sonnet 4.6 through OpenRouter and Verifiers model interception.
• Verified native shell tool execution and artifact collection, plus a persistent two-turn ACP session with MCP tool use after resume.

Note

Add built-in ClineHarness ACP coding-agent harness

• Introduces ClineHarness and ClineHarnessConfig in harness.py, installing a pinned Cline CLI version (3.0.57) into the runtime via npm and running it in ACP mode with provider-bound config, disabled tools, and MCP servers
• Per-trace data directories under CLINE_DATA_ROOT store global settings and MCP config; cleanup removes them with rm -rf
• Exports the new types from verifiers/v1/harnesses/init.py and cline/init.py so they are available for harness selection by id
• Adds cline entries to the AGENTIC_PLACEMENTS and ACP_RESUME_PLACEMENTS test matrices in test_e2e.py and a cline pytest marker in pyproject.toml
• Risk: ClineHarnessConfig.version is regex-constrained; invalid versions fail at config validation. setup runs npm install inside the runtime and aborts on non-zero exit with an error tail

^{Macroscope summarized d539ef0.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 87dcc418bc

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[arafatkatze]

@xeophon I tested this with an OR key and it works.

Would love a review.

Thank you!

[arafatkatze]

@xeophon @hallerite @mikasenghaas Would love a review!

[xeophon]

hey @arafatkatze! does Cline offer native ACP support? we are moving all harnesses over to ACP-backed if possible.

also, what about pre + post tool hooks?

[arafatkatze]

@xeophon Yes, Cline supports native ACP. I updated the PR to use it and added end-to-end coverage for persistent sessions, resume, and MCP tool use. I also verified native tool execution and artifact collection end to end through Verifiers model interception. Also PreToolUse and PostToolUse hooks both work in ACP mode and I have verified it in this PR too.

[xeophon]

added some comments @arafatkatze

have you run cline + any model with terminalbench2.1 (https://github.com/PrimeIntellect-ai/prime-envs/tree/main/environments/terminal/terminal_bench_2) in verifiers?

[arafatkatze]

have you run cline + any model with terminalbench2.1 (https://github.com/PrimeIntellect-ai/prime-envs/tree/main/environments/terminal/terminal_bench_2) in verifiers?

@xeophon Yes, since your comment I ran Terminal-Bench 2.1 through this Verifiers ACP harness with latest Cline and anthropic/claude-sonnet-5 via OpenRouter. Both Docker and Modal smoke runs solved successfully, and the full 89-task evaluation completed. The traces contain 2,396 intercepted model calls and tool use across every task.