Recover cloud agent shells after process exit

[warp-agent-staging]

Description

Recover cloud Oz sessions when an attributable agent command terminates the persistent Bash shell. Warp completes the interrupted tool call with the observed status and recovery guidance, then creates a fresh PTY/Bash in the existing sandbox without replaying the command.

Recovery now distinguishes a temporarily suspended event loop from a terminal disconnect. Resize and input intents queue while the replacement PTY bootstraps, then drain in order after reconnection. Recovery completion no longer depends on a live output waiter: a result persists until a later read consumes it exactly once.

The replacement follows the production ambient-agent shared-session path and preserves external command_finished delivery. Signal and unavailable outcomes remain explicit in stable output but do not fabricate an exit code; an observed zero remains distinguishable from an unknown status throughout the terminal model and API conversion.

Approved design and review requirements: #16144

Review guide

Start with app/src/terminal/local_tty/terminal_manager.rs and app/src/terminal/writeable_pty/pty_controller.rs for PTY suspension and queued intent delivery. Then review app/src/ai/blocklist/action_model/execute/shell_command.rs for durable, exactly-once recovery results and app/src/terminal/model/terminal_model.rs for observed-status handling. The production-wired shared-session integration is in crates/integration/src/test/cloud_shell_recovery.rs; it now uses a hermetic local Bash session rather than requiring Docker or sbx.

Plans:

• Repair PR #16148 cross-target CI failures

Linked Issue

https://linear.app/warpdotdev/issue/REMOTE-2243/recover-oz-cloud-sessions-when-an-agent-command-exits-the-shell

• The linked issue is labeled ready-to-spec or ready-to-implement.
• Where appropriate, screenshots or a short video of the implementation are included below (this change has no new visual state).

Testing

• cargo nextest run -p warp -E 'test(replacement_sender_queues_resize_and_input_while_suspended) | test(shell_recovery)' — 15 passed.
• cargo nextest run -p ai -E 'test(recovered_signal_and_unavailable_omit_wire_exit_code) | test(recovered_observed_zero_is_explicit_in_stable_output) | test(recovered_nonzero_exit_code_is_serialized)' — 3 passed.
• cargo nextest run --workspace --locked --no-run with the CI RUSTFLAGS — all workspace test targets compiled.
• Focused warp_tui recovery-call-site tests — 8 passed.
• cargo check -p warp --tests — passed.
• cargo check -p ai --tests — passed.
• cargo check -p integration — passed.
• cargo clippy --locked --workspace --all-targets --tests -- -D warnings — passed.
• cargo clippy --locked --target wasm32-unknown-unknown --profile release-wasm-debug_assertions -- -D warnings — passed.
• cargo clippy -p warp --all-targets --tests --no-deps -- -D warnings — passed.
• cargo clippy -p ai --all-targets --tests --no-deps -- -D warnings — passed.
• cargo clippy -p integration --all-targets --tests --no-deps -- -D warnings — passed.
• Linux CI-equivalent cargo nextest run --workspace --locked -E 'package(integration) and test(test_cloud_agent_shell_respawn)' under xvfb-run after the macOS timing fix — 1 passed, 336 skipped.
• cargo clippy -p integration --all-targets --tests --locked -- -D warnings with the CI RUSTFLAGS after the macOS timing fix — passed.
• Warp CI run 36637389996 passed Linux, Windows, all compile targets, and all Clippy targets; macOS exposed Bash 3.2 test races in transient session lookup and post-recovery prompt readiness. Commit 013d4b371 removes both timing assumptions.
• macOS-specific execution is pending replacement CI because this environment has no macOS runner.
• ./script/format — completed after the final source changes.
• git diff --check — passed.

test_cloud_agent_shell_respawn now drives the normal ambient-agent shared-session path through local /bin/bash, without Docker or sbx. It resolves recovery state from durable action metadata and waits for the replacement shell's post-bootstrap prompt before submitting another command. It verifies setup through the agent command executor, working-directory/environment/filesystem preservation, no command replay, exactly-once external command_finished delivery with the observed exit code, continued sharing across three recoveries, and tab closure when the fourth shell death exhausts the recovery budget.

• I have manually tested my changes locally with ./script/run — the CI-equivalent real-display integration test covers the end-to-end recovery path.

Agent Mode

• Warp Agent Mode - This PR was created via Warp's AI Agent Mode

CHANGELOG-BUG-FIX: Recover cloud Oz sessions when an agent command exits the persistent shell.

[warp-agent-staging]

This PR was generated with Warp.

Comment @warp-staging-factory on this PR to send it follow-up work.

  

[harryalbert]

• Recover cloud agent shells after process exit #16148 👈 (View in Graphite)
• master

This stack of pull requests is managed by Graphite. Learn more about stacking.