Skip to content

feat(acp): move codex-acp to 1.10.0 #655

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Adam-Dalloul wants to merge 1 commit into
base: main
Choose a base branch
from
+124 −23
Open

feat(acp): move codex-acp to 1.10.0 #655

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
50 changes: 32 additions & 18 deletions src-tauri/src/acp/connection.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3832,20 +3832,31 @@ fn build_client_capabilities(
meta.insert("subagent-transcript".to_string(), serde_json::Value::Bool(true));
}
// claude-agent-acp 0.73.0 added "asyncTasks", and it is advertised — to
// claude ONLY, because codex-acp 1.8.0 does not implement the channel
// (its bundle contains no `asyncTasks` string at all). It publishes the
// lifecycle of Claude's NON-AGENT background work (background shells,
// workflows, monitors) as `async_task_spawned` / `_progress` /
// `_state_update`, all on the parent session id. Unlike the two capabilities
// below, this one adds something codeg cannot get anywhere else: the
// transcript watcher (`background_watch`) can see that a task was launched
// but explicitly CANNOT tell a still-running task from one whose CLI died,
// it never sees workflow/monitor tasks at all (they produce no tool call),
// and there is no way to stop a task from outside. This channel carries a
// real terminal edge, a liveness boundary, an output file path, and the
// `_session/async_task/stop` control. Sub-agent tasks stay out of it by the
// adapter's own filter (`taskType: "local_agent"` is marked ignored), so
// advertising this does not disturb the sub-agent surfaces.
// claude ONLY. It publishes the lifecycle of Claude's NON-AGENT background
// work (background shells, workflows, monitors) as `async_task_spawned` /
// `_progress` / `_state_update`, all on the parent session id. Unlike the
// two capabilities below, this one adds something codeg cannot get
// anywhere else: the transcript watcher (`background_watch`) can see that a
// task was launched but explicitly CANNOT tell a still-running task from
// one whose CLI died, it never sees workflow/monitor tasks at all (they
// produce no tool call), and there is no way to stop a task from outside.
// This channel carries a real terminal edge, a liveness boundary, an output
// file path, and the `_session/async_task/stop` control. Sub-agent tasks
// stay out of it by the adapter's own filter (`taskType: "local_agent"` is
// marked ignored), so advertising this does not disturb the sub-agent
// surfaces.
//
// codex-acp 1.10.0 implements the same capability name (it was absent
// through 1.8.0), so "codex cannot answer" is no longer why it is withheld
// there. The reason is now that it buys less: codex publishes background
// SHELLS only, with no `async_task_progress`, no `outputFilePath` and no
// `usage`, and it sets `showInTranscript: false` because the terminal
// already has a tool call — so the frames add a terminal-state edge and a
// stop control to a card codeg draws anyway, rather than revealing work it
// otherwise cannot see. Turning it on is a behavior change that should
// carry its own reasoning and verification, not ride a pin bump;
// `air_async_task_delta` is already agent-agnostic and reads both variants
// codex emits, so only this advertisement stands in the way.
//
// The remaining two AIR capabilities are deliberately still out.
// claude-agent-acp 0.69.0 and codex-acp 1.4.0 added
Expand Down Expand Up @@ -14174,10 +14185,13 @@ mod tests {
// And exactly this much. Adding a capability here is not free — it
// is what turns the corresponding behavior on.
//
// "asyncTasks" (claude-agent-acp 0.73.0) IS wanted, and only claude
// has it: codex-acp 1.8.0 contains no async-task code at all, so
// advertising it there would be a promise about a channel that
// cannot answer.
// "asyncTasks" (claude-agent-acp 0.73.0) IS wanted, and stays
// claude-only. codex-acp 1.10.0 does implement the name, so this is
// no longer a promise codex cannot answer — its channel is just
// narrower (background shells only, no progress frames, and
// `showInTranscript: false` because the terminal already has a tool
// call). Enabling it for codex is its own change; see the
// advertisement site.
//
// The other two stay out. "agentFileChangeReport"
// (claude-agent-acp 0.69.0 / codex-acp 1.4.0) buys an extra model
Expand Down
2 changes: 1 addition & 1 deletion src-tauri/src/acp/preflight.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -738,7 +738,7 @@ mod adapter_tests {
#[test]
fn codex_adapter_info_uses_codex_home() {
let info = info_for(AgentType::Codex, None, true);
assert_eq!(info.adapter_package, "@agentclientprotocol/codex-acp@1.8.0");
assert_eq!(info.adapter_package, "@agentclientprotocol/codex-acp@1.10.0");
assert_eq!(info.adapter_cmd, "codex-acp");
assert!(info.adapter_installed);
assert_eq!(info.native_cmd, "codex");
Expand Down
95 changes: 91 additions & 4 deletions src-tauri/src/acp/registry.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -895,9 +895,96 @@ pub fn get_agent_meta(agent_type: AgentType) -> AcpAgentMeta {
// three names). Steering still ships no `promptRequired` opt-in
// (tarball grep: zero hits), and there is still no `engines.node`,
// so the 20.0.0 floor is retained.
//
// 1.9.0 and 1.10.0 are both small, and neither moves a surface
// codeg reads today: `sessionCapabilities` is byte-identical to
// 1.8.0's, so `supports_fork` and the rest are unchanged, and the
// entire `initialize` delta across the two releases is one new
// agent capability (`_meta.authStatus`) plus a fourth name in the
// AIR array. The reason to take them is `@openai/codex`.
//
// (a) 1.9.0 reports the agent's auth identity (#467). A new
// presence-only `agentCapabilities._meta.authStatus = {}` means
// "this agent pushes `_auth/status_update`", and it then pushes
// `{authStatus: {kind, label, detail?, account?}}` — `kind` one of
// `account` / `api_key` / `external` / `gateway` / `none`, `label`
// already presentable ("ChatGPT Pro", "OpenAI API key", "AWS
// Bedrock", "Not logged in"). Deliberately NOT bilateral: there is
// no client capability to advertise and no request to make, and it
// pushes on every connection, deduped by payload equality with the
// first push always sent. It reports the AGENT-owned login only —
// `applyGatewayConfig` now records whether routing came from the
// `gateway` auth method or from ACP `providers/set`, and the latter
// is excluded on purpose.
//
// codeg drops it. `_auth/status_update` reaches
// `maybe_emit_ext_notification`, no mapper claims it, and
// `is_known_ext_method` is false, so it leaves on the
// unrecognized-ext-notification `debug!`. That is the correct
// no-op and exactly the tolerance that arm exists for. Worth a
// second look only if the version card ever grows an identity
// line: this is the one frame either adapter sends that names the
// account paying for the session.
//
// (b) 1.9.0 also completes `/status` usage and limits (#463) via
// `src/RateLimitsMap.ts` and the app-server `account/rateLimits/read`
// + `account/rateLimits/updated`. Not a wire shape at all — the map
// is rendered by `buildStatusMessage` into the `/status` markdown
// and arrives as an ordinary `agent_message_chunk`.
//
// (c) 1.10.0 exposes background terminals as async tasks (#460),
// and this is the one that could have bitten. It adds
// `src/async-tasks/{AsyncTaskExtension,CodexBackgroundTerminalTasks}.ts`
// and `src/AcpSessionExtensions.ts`, whose whole body is a cast
// that lets a notification OUTSIDE the SDK schema typecheck —
// which is precisely what `async_task_spawned` and
// `async_task_state_update` are. Neither variant can deserialize
// into `agent-client-protocol-schema` 0.11.7's `SessionUpdate`,
// the same wall the subagent RFD hit.
//
// They stay off the wire, and not by luck. `createAsyncTasks`
// passes `clientSupportsAirCapability(caps, "asyncTasks")` into the
// manager as `enabled`, `isActive()` is `enabled && !disposed`, and
// every entry point (`handleNotification`, `observeCommandStarted`,
// `observeCommandCompleted`, `sync`, `stop`) returns early on it.
// `build_client_capabilities` sends Codex `["sessionFailure"]`, so
// `enabled` is false for the whole connection and neither variant —
// nor the paired `tool_call_update` carrying
// `_meta.jetbrains.air.asyncTasks.backgrounded` — is ever emitted.
// `_session/async_task/stop` is client-initiated, so it costs
// nothing either.
//
// Two comments asserting codex has no async-task code are corrected
// with this bump, because the channel is real now. Advertising it
// to Codex is a separate change, not a pin bump, and it buys less
// than Claude's did: the vocabulary is background SHELLS only,
// there is no `async_task_progress`, no `outputFilePath` and no
// `usage`, and codex sets `showInTranscript: false` because the
// terminal ALREADY has a tool call — the frame adds a terminal-state
// edge and a stop control to a card codeg draws anyway, where for
// Claude it was the only sight of workflow and monitor tasks at all.
// `air_async_task_delta` is not gated on `agent_type` and already
// reads both variants and every field codex populates, so the
// consumer side is done whenever that case gets made.
//
// (d) `@openai/codex` moves ^0.152.0 → ^0.153.3 across the two
// releases (#469, #476), and this is the actual reason to bump. The
// adapter bundle names exactly one model, `gpt-5.6-luna` for
// `TitleGenerator`; every id the picker offers comes from the
// pinned core. Caret on a `0.x` pins the MINOR, so 1.8.0's
// ^0.152.0 resolves no higher than 0.152.1 and cannot reach a 0.153
// core however long it sits. That is why 1.8.0 cannot list
// `gpt-6-astra` and 1.10.0 can.
//
// Re-checked against both tarballs and unchanged: native subagent
// sessions and `agentFileChangeReport` still not adopted, the
// `SessionFork` fork-point resolution (`messageId` /
// `messageFingerprint` / `messageOccurrence`) untouched, still no
// `promptRequired` steering opt-in (tarball grep: zero hits), and
// still no `engines.node`, so the 20.0.0 floor is retained.
distribution: AgentDistribution::Npx {
version: "1.8.0",
package: "@agentclientprotocol/codex-acp@1.8.0",
version: "1.10.0",
package: "@agentclientprotocol/codex-acp@1.10.0",
cmd: "codex-acp",
args: &[],
env: &[],
Expand Down Expand Up @@ -1809,8 +1896,8 @@ mod tests {
);
assert_npx_version(
AgentType::Codex,
"1.8.0",
"@agentclientprotocol/codex-acp@1.8.0",
"1.10.0",
"@agentclientprotocol/codex-acp@1.10.0",
Some("20.0.0"),
);
assert_npx_version(AgentType::Pi, "0.0.33", "pi-acp@0.0.33", Some("22.0.0"));
Expand Down
Loading