Files
zed/crates/eval_cli
Ben Brandt 4ec98df500 acp: Update agent-client-protocol to 5.0.0 (#66425)
Upgrades `agent-client-protocol` from 3.6.0 (schema 1.13.0) to 5.0.0
(schema 1.15.1), and brings Zed's v1 and v2 ACP clients in line with the
spec and schema changes made since then.

## SDK 5 port

- Schema types move from `schema::v1`/`schema::v2` to
`v1::schema`/`v2::schema`, builders to `Client.v1()`, and v2 handle
types to `agent_client_protocol::v2`.
- `agent_client_protocol::Error` is now the SDK's own type, not the v1
schema error. Zed uses it at the SDK boundary, and `ProtocolErrorInfo`
also recognizes it so auth-required detection keeps working.
- Foreground dispatch helpers are generic over the protocol marker, and
v2 request IDs come straight from the handles.
- SDK 5 only matches transport-closed errors that the SDK itself
produced, so peers can't forge them. A test that built one by hand now
covers only the local disconnect case. The `agent_servers` tests still
cover real transport loss.

## Spec changes

- `session/close` and `session/delete`: a `-32002` response counts as
success, on both v1 and v2.
- Permission and elicitation requests that name an unknown session get
`-32002` with `data.sessionId` instead of `-32603`. The v2 migration
guide makes this required for Clients too.
- v2 `usage_update.cost` is a patch field: omitted leaves it unchanged,
`null` clears it. A v1 update without a cost still leaves it unchanged.
- A `plan_update` whose type Zed can't render (`file`, `markdown`, or an
unknown type) still replaces the plan with that `planId`, so the old
entries stop showing.
- A diff `patch` with an unrecognized `format` renders from `changes`,
as if `patch` were omitted.
- v2 never sends an empty prompt.
- On v2, `session/cancel` no longer answers pending permission requests
locally. They stay open until the agent cancels them with
`$/cancel_request`, because requests for background work survive a
cancel. v1, including v1 subagents, keeps answering them with
`cancelled`.

## Follow-ups

These are left for separate PRs:
- Background-work UI: the `paused` status, the `background` list, and
cancel buttons for tool calls and terminals. Until then, `paused`
displays as in progress, which the spec allows.
- Subagent `cwd` and `additionalDirectories`.
- The new unstable methods: `auth/status`, `session/set_title`,
`session/status`, `session/archive`.

Release Notes:

- Improved compatibility with the latest Agent Client Protocol,
including handling of missing sessions, plan updates, diff formats, and
cancellation of permission requests.
2026-11-02 15:26:50 +00:00
..
…
2026-09-11 09:13:44 +00:00
…

eval-cli

Headless Rust binary for running Zed's agent in evaluation and benchmark environments. It is designed for containerized harnesses such as Harbor and Pier, where the repository is already checked out and model API keys are provided via environment variables.

eval-cli uses the same NativeAgent + AcpThread pipeline as the production Zed editor: a full agentic loop with tool calls, subagents, and retries, without a GUI.

This directory also contains zed_eval/, the Python zed-eval package used to build this binary, launch remote benchmark runs on Modal/Harbor/Pier, and fetch results. For normal benchmark orchestration, start with zed_eval/README.md.

Building

Native, for local testing on the same OS

cargo build --release -p eval_cli

Linux x86_64, for Harbor/Pier sandboxes

Harbor and Pier containers run Linux x86_64. From the repository root, use the Docker-based build script:

crates/eval_cli/script/build-linux

This produces target/eval-cli, an x86_64 Linux ELF binary. You can also specify a custom output path:

crates/eval_cli/script/build-linux --output ~/bin/eval-cli-linux

Standalone usage

eval-cli \
  --workdir /testbed \
  --model anthropic/claude-sonnet-4-6 \
  --instruction "Fix the bug described in..." \
  --timeout 600 \
  --output-dir /logs/agent

eval-cli reads provider API keys from environment variables such as ANTHROPIC_API_KEY and OPENAI_API_KEY. It writes result.json, thread.md, and thread.json to the output directory.

Exit codes

Code Meaning
0 Agent finished
1 Error, such as model/auth/runtime failure
2 Timeout
3 Interrupted by SIGTERM or SIGINT

Running benchmarks

Most benchmark runs should use the Python zed-eval CLI instead of invoking eval-cli directly. From the repository root:

crates/eval_cli/script/install-zed-eval
zed-eval doctor --create-volume
zed-eval run rf --from local --n-tasks 2

For one-off source runs without installing the tool globally, use crates/eval_cli/script/zed-eval <args>.

See zed_eval/README.md for supported benchmarks, remote builds, Modal setup, reporting, rejudging, baselines, and Harbor/Pier installed agent usage.