Reopening a conversation in an editor does not establish that background commands are represented correctly. Context continuity and execution visibility are separate contracts. Treating every Claude Agent ACP 0.101.0 change as “better restoration” obscures the regression paths that matter.
The official release is dated November 5. The changelog, checked November 9, also lists 0.101.1 and 0.102.0. This article analyzes the briefing's 0.101.0 release; it is not a recommendation of the latest version. No runtime testing was performed.
Compaction needs a preservation test
PR #1522 introduces idle-session compaction before prompt-cache expiration. Compacting conversational context is different from restoring operating-system processes. Whether requirements, constraints and file-change information survive adequately needs a task-based test.
Record a small task's acceptance criteria and unresolved failures before the idle interval. After resumption, check whether the agent still recognizes prohibited scope and failing checks. Repeating a previous sentence is weaker evidence. This is a proposed regression scenario, not an executed test.
Visible output does not prove command success
PR #1512 streams background-command output in v2. Do not assume the same behavior for a v1 editor connection. Test output delivery, process termination and exit-status presentation separately.
Editor ↔ ACP adapter ↔ agent execution environment
↑ ↓
output events ← background command
This original role diagram does not specify the implementation's exact scheduling. If output stops, the process could have finished, disconnected or simply stopped emitting text. An intermediate log message is insufficient to mark the task successful.
The #1512 diff in src/acp-agent.ts awaits finishBackgroundTerminalOutput before sending a task's terminal-exit update. followBackgroundTerminalOutput tracks terminals with an output-file path. This introduces a delivery-order requirement: follow the file, then drain remaining output before exit. Added upstream tests cover files appearing late, reads splitting UTF-8 characters, shrinking files and stopping reads after completion. We reviewed these tests without executing them.
Defaults change the regression surface
PR #1533 makes Auto mode the default for new sessions. PR #1520 honors the configured fallback when a restored mode is unavailable. A mode name alone does not establish every tool permission; check the editor, adapter and runtime combination.
| Scenario | Required observation | Easy-to-miss failure |
|---|---|---|
| Start a new session | Default mode and permission display | Assuming previous defaults |
| Restore an unavailable mode | Configured fallback applies | Silent permission change |
| A v2 command emits chunks then fails | Output order and final failure | Treating the last log as success |
| Disconnect and reconnect | Work state and duplicate execution | Re-running a command to recover its log |
The single-file executable work in PR #1524 changes distribution. Packaging alone does not establish editor compatibility or complete session restoration.
Upgrade the tested combination
Record the editor, ACP protocol, adapter, SDK/CLI and operating system in regression results. A newer release on the same day is not sufficient reason to upgrade without checking it. Require correct output delivery, permission fallback and command termination states.
For a continuous-session experience, evaluate context preservation separately from work-state delivery. The same distinction appears in checkpoint and stream observation. The useful result of 0.101 is a set of changed, testable contracts—not a demonstrated productivity multiplier.