When Claude Code loses its API connection or is refused, it retries on its own
for up to several minutes. The session showed only "Working…" meanwhile. It
now shows "API request failed, retrying in Ns (attempt 3/10)" with the error
in the tooltip, cleared by the agent's next output or the end of the turn.
Background tasks often couldn't be cancelled, or cancelling didn't stop them:
- A cancel that arrived before the backend registered the request was dropped (build while saving files, multi-repository fetch/pull between repositories). The host now remembers early cancels (bounded, 2 min), and request ids start at a random value per page.
- There was no queue and no common cancel path: backend-reported tasks (LSP progress, deployment transfers) couldn't be cancelled, and tool installs, VCS operations and update checks had no cancel button in Background Tasks.
- Only the direct child process was killed, and JDK / tool archives were extracted without checking for cancellation.
Tool installs and JDK downloads now share a downloads queue; cancelling a queued task removes it without starting any work. Every task cancels through one path that shows "Cancelling…" and then removes it; tasks that can't be interrupted show a disabled cancel button with the reason. IdeEvent::Task carries `cancellable` with a cancel_task command, cancellable LSP progress sends window/workDoneProgress/cancel, installers, curl and tar run in their own process group and are killed as a group, and extraction checks for cancellation between entries.
Claude-Session: https://claude.ai/code/session_01Dq6jQQn8ZhRfkkyj3PRutA
Agents can run through native drivers or any ACP agent, speak ACP v1 or v2, and be installed from the ACP Registry. Tasks gain fork, in-place edit-and-rerun, steering and usage-limit handling where the agent supports them.
- Native Claude Code (`claude` stream-json) and Codex (`codex app-server`) drivers, each served to the task machinery as an in-process ACP shim, using the newest install on PATH.
- ACP v2 with v1 fallback, negotiated on initialize; per-agent capabilities (fork, rollback, steer) so the UI only shows supported actions; usage-limit errors put the task in a `limited` status with the reset time.
- Settings › Tools › Agents lists Registry agents (featured first) to install, update and remove; installed agents update when a project opens. "More agents…" in the provider picker opens it. Registry index and downloads are HTTPS-only, release archives are checked against their sha256, and entries without one show "No checksum provided" and ask for confirmation before installing. Package names from the registry can't be passed as options to npm/uv.
Claude-Session: https://claude.ai/code/session_01Dq6jQQn8ZhRfkkyj3PRutA
* feat(agents): Show subagents as their own conversations
Agents that support ACP subagent sessions (Claude Code, Codex) now run each
subagent as a child session. Its tool calls and messages no longer mix into
the task's conversation: the parent shows a card per subagent with its state,
tool call count, latest step, task and result, and the card opens the
subagent's own conversation.
- Route session/update by session id; subagent settings and usage stay out
of the task.
- Subagent permission requests are answered in the task's conversation,
named after the subagent.
- Subagent conversations are saved with the task; a running subagent is
marked stopped after a restart.
- A running subagent can be stopped on its own.
* fix(agents): Save a subagent's state when it changes after the turn
* fix(agents): Leave changes made before a local checkout task out of its changes
A task in the project's own checkout listed every uncommitted and untracked
file as its own. The files already changed when the task starts are now
recorded with their size and modification time, and drop out of the task's
changes until they are edited again.
* feat(agents): Show the current branch for a local checkout task
A local-checkout task shows the repository's current branch in its header, and
Changes compares against it; both follow branch switches. The header also shows
the time the agent has spent on the task's turns.
A usage update without a cost (Claude sends one for its rate limits) no longer
clears the session cost.
The task title starts as the first line of the prompt. When the agent sends
a session_info_update with a title, the task takes that title, saves it and
shows it in the Agents list.
* feat(agents): Select, copy, quote and search the conversation
- Conversation text can be selected and copied.
- Selected text gets a Quote button that replies to it in the composer.
- Messages have Copy on hover; your own messages also have Edit and
Resend and Send Again.
- Code blocks are highlighted and have a Copy button.
- File paths in replies (`src/lib.rs:12`) open in the editor when the
file exists.
- Find in Conversation (⌘F / Ctrl+F) highlights every match and steps
through them, opening collapsed tool output.
- Jump to Bottom appears after scrolling up.
- Your messages show when they were sent and how long the agent took;
a running turn shows its elapsed time.
* feat(agents): Quote a selection from the context menu
* feat(agents): Drop Jump to Bottom from the conversation
* feat(agents): Drop the hover Copy button from agent messages
* fix(agents): Highlight only the find matches on screen
WebKit paints a highlight range inside a collapsed tool output over the
text next to it. Matches in collapsed output still count, and stepping
to one expands it.
* fix(agents): Repaint find highlights as the search changes
WebKit doesn't repaint where a replaced highlight object was, so matches
of an earlier search stayed painted; each highlight is now one object
changed in place.
`ide-lsp` tests failed intermittently (about 1 run in 4, more under load).
The main cause, requests from the other side starting before earlier notifications, is fixed on main by b5bc150; this adds a regression test for it in `ide-jsonrpc`, documents that `on_notification` must not wait for the remote side, and drops the 200 ms polling workaround in `ide-agent` that papered over the same race for `session/request_permission`.
Test-only races fixed: mock replies are held with a gate instead of fixed sleeps and timing asserts; the indexing-retry test ends indexing after the first reply; the crash test counts Ready/Crashed transitions instead of reading the last status; the mock records a notification and applies its effect under one lock; `eventually` waits up to a deadline; `tcp_server` reserves its port with a bound socket and waits for the server's start marker.
Claude-Session: https://claude.ai/code/session_01VUYsdXA2nkF5v6zsC93dpn
A save that ran after delete (from the turn ending or the agent exiting)
wrote the task file back, so the task returned on the next load. Deleting
now marks the task under the lock that saving holds.
The new-task composer only offered the slash commands the agent sent in its last session, stored once globally, so a project that hadn't run a session (or whose last session ran elsewhere) showed none of its skills; matching was substring-only, so `/trig` didn't find `/triage` either.
Offered commands are now remembered per project root. For Claude Code, skills (`.claude/skills/*/SKILL.md`) and commands (`.claude/commands/**/*.md`) are also discovered on disk, from the working directory up to home plus `~/.claude`, and merged with the agent's list (agent entries win). New `agent_commands` IPC serves the home composer (project root) and session composer (the task's cwd). Matching ranks prefix (also after `plugin:`), then substring, then subsequence. Picking a completion inserts `/name ` and the prompt is sent unchanged.
Not covered: a new worktree of a project that gitignores `.claude/` doesn't contain the project's skills.
Claude-Session: https://claude.ai/code/session_01VUYsdXA2nkF5v6zsC93dpn
@zed-industries/codex-acp is archived; its last release (0.16.0) bundles an
older Codex core that doesn't know newer models, so a config.toml model such as
gpt-6.1-sol fell back to generic metadata with a warning. The maintained adapter,
@agentclientprotocol/codex-acp, runs the Codex App Server from a current
@openai/codex and lists the account's models.
## What
An Agents feature: hand a task to a coding agent and follow it from the IDE. Agents speak the [Agent Client Protocol](https://agentclientprotocol.com). Built in: Claude Code, Codex, Gemini CLI, GitHub Copilot, Cursor, OpenCode, Qwen Code, Kimi CLI, Goose, Auggie, Kiro, Mistral Vibe. You can add your own in `<config>/agents/<id>.toml` (`name`, `command`, `args`, `[env]`). An agent whose command isn't on `PATH` is listed under "Not installed".
- **New Task page.** An editor page with a prompt box. You pick the agent, model, mode and thinking effort, and whether it runs in a new worktree (from a chosen branch) or in the checkout.
- **Agents tool window.** Tasks are grouped by what they need: Needs Input, Failed, Running, Ready to Review, Archived. It also shows a stripe badge, a status bar item and notifications.
- **Session tab.**
- The conversation: Markdown, thinking, tool calls with diff stats, the plan.
- Permission requests appear as an editor banner.
- The changed files are compared against the task's base commit and open in the diff viewer.
- A context-usage meter.
- **Composer.**
- `@` references files from the project index. A referenced file's text is sent when the agent accepts embedded context, otherwise a `resource_link`.
- `/` lists the agent's commands; ↑/↓ recall earlier messages.
- Images can be pasted, dropped or picked; Shift+Tab cycles the mode.
- While the agent works, Enter queues the message and ⌘↵ interrupts the turn and sends it now. Queued messages can be sent now, edited or removed.
- **Send to Agent.** In the editor and Project view context menus. It adds `@file:10-20` references to the message you typed last.
- Landing the work is left to the existing Git features: the worktree is a normal branch.
## How
- **`crates/ide-agent`**: one ACP agent process per task, running in the task's directory.
- It folds `session/update` into a timeline and answers `session/request_permission`.
- It tracks the session's settings: `configOptions`, falling back to the older `modes` / `models` fields. The user's choices are applied again on every new or reloaded session.
- Tasks are saved as JSON per task. Sessions resume with `session/load` without duplicating history, and old task files are upgraded when read.
- **`ide-jsonrpc`**: adds newline-delimited framing (ACP) next to Content-Length (LSP).
- **`ide-vcs`**: adds `numstat_with_worktree` for the changes list.
- **Editor**:
- The Switcher (⌃⇥) lists every open editor, as IDEA does, not just files.
- `EditorKindContribution.restorable` lets a kind's tabs be saved with the editor session, so agent tabs survive reloads.
- **UI**: `Select` popups now size to fit short option details (they used to cut them off).
## Testing
- `cargo test -p ide-agent` (14 tests). They use an in-memory ACP agent and cover permissions, restart with `session/load`, settings, the queue with send-now and stop, and prompt blocks.
- There is also an ignored end-to-end test against the real Claude Code adapter: `OPENIDEA_AGENT_E2E=1 cargo test -p ide-agent real_claude -- --ignored`.
- `pnpm vitest run`: 2052 tests.
- `cargo clippy -p ide-agent -p openidea --all-targets -- -D warnings` and typecheck are clean.
- Driven in the running app with Claude Code:
- A task in a worktree, with permission answers, changes listed and the diff opened.
- A follow-up after restarting the IDE.
- `@` reference with embedded content, `/` commands, Shift+Tab mode changes.
- A queued message and send-now, an image prompt, Send to Agent from a selection.
- The Switcher and restored tabs after a reload.
Not covered yet: only Claude Code was driven end to end. The other agents go through the same ACP path but weren't run here.
Merged on the maintainer's go-ahead. Merge with main: kept both sides in Cargo.toml (ide-agent + ide-clipboard), the feature registry (agents + vim) and zh.json (no duplicate keys). After the merge: typecheck clean, full vitest 268 files / 2247 tests passed, ide-agent/ide-jsonrpc/ide-vcs 144 passed, bindings unchanged after regeneration.
l0ng-ai