[REMOTE-2208] Make retained failed cloud agent sessions attachable, interactive, and kept alive by activity

[warp-agent-staging]

Description

Makes a retained failed cloud agent environment actually usable from the Warp desktop client, the client half of REMOTE-2208 (server half: warpdotdev/warp-server#12943).

Today AmbientAgentTask treats only InProgress runs as having a live execution. So a run that has already gone FAILED/ERROR — but whose sandbox is deliberately retained for debugging and still has a joinable shared session — was surfaced as a dead run: it fell back to a read-only transcript, its live session link was not offered, and a follow-up was never routed to the still-running remote VM.

What changed:

• AmbientAgentTask now gates live-session state on a new supports_live_session() helper that accepts InProgress plus terminal Failed/Error. Liveness still additionally requires a running sandbox and a joinable session signal, so this only promotes genuinely retained runs. This flows through active_live_session_state(), active_execution_session_id(), and has_active_execution(), which in turn makes a retained failure attachable, openable as a live session from the agent list/details, copyable as a live session link, and routed as LiveRemoteVm for follow-up interaction.
• entry_for_task now derives the agent-list open capability from the resolved AmbientAgentLiveSessionState instead of the raw session-id string, so a retained run whose session id is only recoverable from its session link is correctly openable rather than falling through to a transcript.

Deliberately unchanged:

• A failed run whose sandbox has ended stays transcript-only, is not attachable, and remains eligible for a cloud follow-up (can_submit_cloud_followup).
• Succeeded/Cancelled/Blocked runs never become attachable, even with stale is_sandbox_running metadata.
• No session-sharing protocol change and no self-hosted-worker change — nothing in the code required one.

Rework changes (review round 1)

Two blocking gaps were reported by the requester on the first round. Both are addressed here; the retention spec on warpdotdev/warp-server#12943 was updated by the server-side change to make the sliding-window semantics explicit.

1. The retained session tab was not interactive (no text entry, no usable terminal).
Root cause: WorkspaceView's OpenOrAttachAmbientAgentConversation handler reuses any pane registered for the run, which includes a read-only conversation transcript viewer opened earlier for the same run (transcript panes register themselves through ActiveAgentViewsModel::register_ambient_session). PaneGroup::attach_execution_session_to_ambient_pane cannot join a session from such a pane — it is backed by a MockTerminalManager with no network — but it swallowed the failed downcast and still returned true, so the workspace focused a transcript with no input box instead of opening a joinable tab. A pane that had already been switched to the ended-run view (FinishedViewer status + ended-conversation tombstone + non-editable input) was likewise reused as-is and stayed read-only after the rejoin.

Fix, in three parts:

• attach_execution_session_to_ambient_pane now refuses a conversation transcript viewer outright and propagates a failed viewer-manager downcast as false instead of reporting success.
• New TerminalView::prepare_for_live_session_reattach clears the finished/read-only state before rejoining: it removes the ended-conversation tombstone, lifts FinishedViewer (which is what makes TerminalModel::is_read_only true, and therefore what hides the input box in is_input_box_visible), and restores InteractionState::Editable.
• The workspace now falls back to opening a fresh, writable shared-session tab (add_tab_for_joining_shared_session) whenever reuse is not possible, and only focuses the existing pane when the attach actually succeeded.

2. Activity-based keepalive — retained_until must be pushed forward on activity, not fixed at failure time.
Added AIClient::extend_run_session_retention → POST /api/v1/agent/runs/{run_id}/session-keepalive (the endpoint shipped by the server half on warpdotdev/warp-server#12943), returning {retained, extended, retained_until}. It is fired from the submission router (Input::maybe_route_ai_query_to_remote_target) whenever a prompt is forwarded to a live remote VM for an ambient run — i.e. the same condition under which the viewer prompt path runs. The decision is factored into a pure session_keepalive_run_id(&AIQueryRouting) helper so it is unit-testable.

Two deliberate choices:

• The call is fire-and-forget: a prompt submission must never block or fail on the keepalive.
• It deliberately does not consult the client's cached run state. A run the client still believes is InProgress may already have failed into retention server-side, and the endpoint is a documented no-op (retained: false) for a non-retained run, so calling unconditionally on the live-VM path is both correct and simpler than mirroring server state. Read-only viewers, cloud-to-cloud follow-ups (which start a brand new VM), and shared local sessions are excluded, because none of them is activity on a retained ambient session.

Rework testing

New regression tests, each verified failing before the fix and passing after:

• app/src/pane_group/mod_tests.rs — attach_execution_session_refuses_read_only_transcript_viewer_pane
• app/src/terminal/view/shared_session/view_impl_tests.rs — test_prepare_for_live_session_reattach_restores_interactive_input
• app/src/terminal/view/shared_session/cloud_conversation_continuation_tests.rs — session_keepalive_targets_ambient_run_for_executor_viewer, session_keepalive_targets_retained_failed_run, session_keepalive_targets_run_regardless_of_client_side_run_state, session_keepalive_skips_non_live_and_read_only_targets

See Rework changes (review round 2) below for the current validation run.

Rework changes (review round 2)

Round 2 raised one blocking finding: the visual proof in this PR showed the pre-rework broken fixture ("Shared session not found"), and end-to-end proof of text entry and of the keepalive was still outstanding. A genuinely retained cloud session cannot be joined from here until warp-server#12943 is deployed, so this round adds deterministic, in-process proof of both behaviors instead, and drops the misleading fixture media.

Proof that the reattached pane accepts text — test_prepare_for_live_session_reattach_accepts_typed_text (app/src/terminal/view/shared_session/view_impl_tests.rs) asserts the reported symptom directly rather than the flags behind it: text typed into the ended-run pane is dropped (buffer stays empty — exactly what the reporter saw), and the same text lands in the input buffer after prepare_for_live_session_reattach.

Proof that submitting into a live ambient session fires the keepalive — session_keepalive_is_dispatched_for_an_attached_ambient_executor_pane (app/src/terminal/input_tests.rs) builds an attached ambient executor pane, runs the production routing decision over it (resolve_ai_query_routing → session_keepalive_run_id), then runs the production dispatch (Input::extend_retained_session_window) against a recorded MockAIClient, awaits the spawned future, and asserts extend_run_session_retention was called exactly once for that run id. To make the dispatch observable, extend_retained_session_window now takes the AIClient as a parameter (resolved by the caller) and returns its SpawnedFutureHandle.

Proof of the wire contract — build_run_session_keepalive_url extracts the endpoint path (matching the existing build_run_followup_url convention) and is pinned by build_run_session_keepalive_url_routes_to_run_session_keepalive, with deserialize_run_session_keepalive_response_for_a_retained_run / ..._for_a_run_that_is_not_retained covering both response shapes (the no-op form is what makes the unconditional call safe).

Both behavior tests were verified to fail before the fix: with prepare_for_live_session_reattach neutered to a no-op the typed text stays empty (left: "" / right: "echo hello-from-retained"), and with session_keepalive_run_id returning None for an executor the dispatch test fails (left: None / right: Some(AmbientAgentTaskId(...))).

Round 2 validation

• ./script/format --check — clean.
• ./script/check_no_inline_test_modules — clean.
• cargo clippy -p warp --all-targets --tests -- -D warnings — clean.
• cargo nextest run -p warp --lib --no-fail-fast — 6077 / 6080 pass. The only failures are the same 3 pre-existing sandbox-environment ones (unchanged by this diff): ambient_agent_headers_for_task_overrides_existing_cloud_agent_header (the nsc workload-token command is denied in this container), test_decorations_with_multibyte_chars, and test_histignorespace_support_in_zsh.
• Focused run of the REMOTE-2208 tests (screenshot below): 10 passed, 0 failed.

PASS [ 0.018s] ( 1/10) warp server::server_api::ai::tests::deserialize_run_session_keepalive_response_for_a_run_that_is_not_retained
PASS [ 0.019s] ( 2/10) warp terminal::view::shared_session::cloud_conversation_continuation::tests::session_keepalive_skips_non_live_and_read_only_targets
PASS [ 0.019s] ( 3/10) warp terminal::view::shared_session::cloud_conversation_continuation::tests::session_keepalive_targets_ambient_run_for_executor_viewer
PASS [ 0.019s] ( 4/10) warp server::server_api::ai::tests::deserialize_run_session_keepalive_response_for_a_retained_run
PASS [ 0.019s] ( 5/10) warp server::server_api::ai::tests::build_run_session_keepalive_url_routes_to_run_session_keepalive
PASS [ 0.021s] ( 6/10) warp terminal::view::shared_session::cloud_conversation_continuation::tests::session_keepalive_targets_retained_failed_run
PASS [ 0.021s] ( 7/10) warp terminal::view::shared_session::cloud_conversation_continuation::tests::session_keepalive_targets_run_regardless_of_client_side_run_state
PASS [ 0.137s] ( 8/10) warp terminal::view::shared_session::view_impl::tests::test_prepare_for_live_session_reattach_restores_interactive_input
PASS [ 0.173s] ( 9/10) warp terminal::input::tests::session_keepalive_is_dispatched_for_an_attached_ambient_executor_pane
PASS [ 0.154s] (10/10) warp terminal::view::shared_session::view_impl::tests::test_prepare_for_live_session_reattach_accepts_typed_text
     Summary [ 0.185s] 10 tests run: 10 passed, 6077 skipped

Still needs a deployed server to confirm live

Post-failure retention is not deployed anywhere reachable from this environment (warp-server#12943 is unmerged), so no genuinely retained cloud session exists to join and a live end-to-end capture remains impossible here. The round-1 fixture screenshots and video, which only ever showed the "Shared session not found / Starting …" state, have been removed because they documented the broken pre-rework state rather than the fix.

Once warp-server#12943 is deployed, a reviewer should confirm:

1. Fail a run in an environment with Keep session after failure enabled, then open the run from the agent list — the tab must be a writable Cloud agent run terminal, not a transcript.
2. Open the run's read-only transcript first, then open the run again from the agent list — it must open a separate, writable session tab rather than re-focusing the transcript. This is the exact regression fixed here.
3. Send a query in that tab and confirm POST /api/v1/agent/runs/{run_id}/session-keepalive fires and the run's retained_until moves forward.

Linked Issue

REMOTE-2208

Testing

Regression coverage added for both branches of the new behavior (each fails on master and passes here):

• app/src/ai/ambient_agents/task_tests.rs
  • retained_failed_and_error_tasks_have_attachable_live_sessions
  • ended_failed_task_with_stale_session_metadata_is_inactive
  • succeeded_task_does_not_become_attachable_from_stale_running_metadata
• app/src/ai/agent_conversations_model_tests.rs
  • test_resolve_open_action_opens_retained_failed_ambient_session_from_link
  • test_resolve_open_action_uses_transcript_for_ended_failed_task
  • test_resolve_copy_link_prefers_session_link_for_retained_failed_task
  • test_resolve_copy_link_uses_conversation_link_for_ended_failed_task
• app/src/terminal/view/shared_session/cloud_conversation_continuation_tests.rs
  • routing_is_live_remote_vm_for_retained_failed_execution
  • routing_starts_new_cloud_vm_for_ended_failed_execution
  • retained_failed_task_execution_returns_active_execution_error
  • ended_failed_task_uses_ordinary_cloud_continuation

Commands run:

• ./script/format — clean.
• ./script/presubmit — cargo fmt --check, the inline-test-module check, cargo clippy … -D warnings, clang-format, and wgslfmt all pass.
• cargo nextest run -p warp --lib --no-fail-fast — 6052 / 6055 pass. The 3 failures are pre-existing sandbox-environment failures unrelated to this diff and untouched by it: server::server_api::ai::tests::ambient_agent_headers_for_task_overrides_existing_cloud_agent_header (the nsc workload-token command exits non-zero in this container), terminal::input::decorations::tests::test_decorations_with_multibyte_chars, and terminal::input::tests::test_histignorespace_support_in_zsh.
• Full-workspace presubmit test phase additionally fails in build_cache (needs non-interactive sudo) and the integration suite (needs SSH password prompts, a real display, and free ports) — all environmental in this container, none touching this code.

Screenshots / Videos

The screenshot below is the focused run of the REMOTE-2208 regression tests above, captured with computer use in a terminal on this machine: 10 passed, 0 failed. It is the verifiable proof of the reattach text-entry behavior and the keepalive dispatch that round 2 asked for.

The round-1 media (a synthetic fixture pointing at a non-existent session id, stuck on "Shared session not found / Starting …") has been removed: it documented the broken pre-rework state, not the fix.

Agent Mode

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

CHANGELOG-IMPROVEMENT: Failed cloud agent runs whose environment is retained for debugging can now be opened and interacted with live from the agent list, instead of only showing a read-only transcript.

Computer-use screenshots

Originating thread: https://linear.app/warpdotdev/issue/REMOTE-2208/configurable-extended-session-lifetime-for-failed-cloud-agents

Conversation: https://staging.warp.dev/conversation/8b9817fe-8f49-4f16-9931-6377351cfe7b
Run: https://oz.staging.warp.dev/runs/019fb729-6a05-7f1f-a60b-986b8907764e

This PR was generated with Oz.