## Description
Removes the `orchestration_event_push` feature flag and the polling
fallback in the orchestration event delivery path. SSE-based event push
has been on in dogfood and staging long enough that it's now the only
path; the dual-mode `OrchestrationEventPoller` is renamed to
`OrchestrationEventStreamer` and only opens persistent SSE connections.
- Removed `FeatureFlag::OrchestrationEventPush` and the matching Cargo
feature.
- Renamed `app/src/ai/blocklist/orchestration_event_poller.rs` (and its
tests) to `orchestration_event_streamer.rs`. Renamed the public type to
`OrchestrationEventStreamer`; renamed the shared event-injection sink
`handle_poll_result` → `handle_event_batch`.
- Removed polling-only state (`poll_backoff_index`, `poll_in_flight`),
methods (`poll_and_inject`, `start_idle_poll_timer`), and constants
(`POLL_BACKOFF_STEPS`, `EVENT_POLL_BATCH_LIMIT`). Kept `event_cursor`,
`pending_delivery`, and `conversation_statuses` since they are also used
by the SSE path.
- Removed the now-dead `AIClient::poll_agent_events` trait method and
its `ServerApi` implementation in `app/src/server/server_api/ai.rs`.
The `OrchestrationV2` flag continues to gate streamer instantiation and
watched-run registration; runtime behavior under v2 is unchanged.
Pairs with the warp-server PR that drops the flag server-side:
https://github.com/warpdotdev/warp-server/pull/10736 — that PR should
land first so the polling endpoint stays available for older clients
during rollout.
## Testing
- `cargo build -p warp` clean.
- `cargo clippy -p warp -p warp_features --all-targets -- -D warnings`
clean (after `cargo fmt`).
- `cargo test -p warp --lib ai::blocklist::orchestration_event_streamer`
— all 7 tests pass.
## Agent Mode
- [x] Warp Agent Mode - This PR was created via Warp's AI Agent Mode
---
[Plan: Remove orchestration_event_push flag (server +
client)](https://staging.warp.dev/drive/notebook/yAlAxMAr4EO65A9LEy40cX)
[Conversation](https://staging.warp.dev/conversation/ff5e82cb-bba7-4cbf-ae0b-e51f2c542f4a)
---------
Co-authored-by: Oz <oz-agent@warp.dev>
I want to run an a/b test where 50% of free users get defaulted to auto
cost efficient, and other 50% to auto open weights.
## Tests
- 50% of the time when i open the app with a different WARP_DATA_DIR i
get the expected flip/flopping
- my default gets persisted throughout signup
---------
Co-authored-by: Oz <oz-agent@warp.dev>
## Description
Running `/pr-comments` (or any other slash command that targets the currently-selected conversation) while an AI response stream is in-flight crashed in debug builds. `BlocklistAIController::send_request_input` hits its in-flight invariant and fires `safe_assert!(false, ...)` (panics in debug, returns `Err` in release).
The user-query path (`send_query`) avoids this because it pre-cancels any active stream on the target conversation with `CancellationReason::FollowUpSubmitted` before calling `send_request_input`. Slash commands bypass that path and dispatch directly via `SlashCommandRequest::send_request`, so the cancel never happens.
This PR makes `send_slash_command_request` mirror that cancel-and-resend: if the target conversation has an in-flight stream, cancel it before dispatching. All `SlashCommandRequest` variants are conceptually a fresh user turn (a follow-up), so this matches the semantics users already get from typing a follow-up message.
`send_queued_slash_command_request` is unchanged — it's only invoked from `Input::submit_queued_prompt` once the conversation is idle, so no pre-cancel is needed.
## Testing
Verified locally. Demo [here](https://www.loom.com/share/a9638ac9a53b48349ce21d03eaba516a)!
- Manually reproduced the crash on a dogfood debug build by triggering `/pr-comments` mid-stream; confirmed the panic at the `safe_assert!` in `BlocklistAIController::send_request_input`.
- After the fix: same repro cancels the in-flight turn and dispatches `/pr-comments` cleanly. Same behavior verified for `/skills` (InvokeSkill), `/compact` (Summarize), and `/create-environment`.
## Agent Mode
- [x] Warp Agent Mode - This PR was created via Warp's AI Agent Mode
## Changelog Entries for Stable
CHANGELOG-BUG-FIX: Fixed an issue where slash commands sent while an agent was still responding were silently dropped. Now, slash commands like `/pr-comments` run as follow-ups, just like typed messages.
---
Run: https://staging.warp.dev/conversation/aeecfa2d-1456-440b-9961-c27295d82531
Plan: https://staging.warp.dev/drive/notebook/Ryj2w6xDSqDwNJYhpwwI0Vly
## Description
Restore orchestration event delivery on the client after a Warp restart
so that a parent conversation continues to receive lifecycle events and
inbox messages from its children — including terminal events that
arrived while Warp was not running.
See the full design in `specs/replay-agent-events-on-restore/PRODUCT.md`
and `specs/replay-agent-events-on-restore/TECH.md`.
This is a re-land of warpdotdev/warp-internal#24999, which was reverted
in warpdotdev/warp-internal#25055 due to a CI race. This version fixes
the test compilation issue (missing fields in `AmbientAgentTask` struct
literal in `conversation_ended_tombstone_view_tests.rs`).
### What
- Persists the per-conversation event cursor across restarts.
- Adds `last_event_sequence: Option<i64>` to `AgentConversationData`
(SQLite) and `AIConversation`.
- New `BlocklistAIHistoryModel::update_event_sequence` helper writes the
cursor through `write_updated_conversation_state` after each event
batch.
- Also persists the cursor to the server (fire-and-forget) so driver /
cloud restarts can resume without local SQLite state.
- Restores `OrchestrationEventPoller.watched_run_ids` and re-establishes
event delivery on `BlocklistAIHistoryEvent::RestoredConversations`.
- New `on_restored_conversations` handler issues `GET
/agent/runs/{run_id}` for each restored parent and uses the response
inline `children` and `last_event_sequence` to populate watched run ids
and merge the cursor (`max(SQLite, server)`).
- Fetch failures retry with exponential backoff (1s, 2s, 5s, 10s capped)
keyed off a per-conversation `restore_fetch_failures` counter; reset on
success and on conversation removal.
- Gated on `OrchestrationV2`. Shared-session viewers and conversations
without children are skipped.
- `Success` parents resume delivery immediately; `InProgress` parents
defer to the existing `on_conversation_status_updated` path.
- Restores V1 lifecycle subscriptions on restart by extending the
existing `RestoredConversations` handler in `OrchestrationEventService`
to re-register `lifecycle_subscription_routes` for restored child
conversations whose parents are present locally.
### Why
After a Warp restart, parent conversations were silently receiving no
further events from children. The `event_cursor` in
`OrchestrationEventPoller` was always initialized to 0, so even if
delivery had resumed, every event since the start of the conversation
would have replayed and produced duplicate messages. V1 lifecycle
subscription routes were also not restored, so V1 parents missed child
status transitions.
## Testing
- Added unit tests in
`app/src/ai/blocklist/orchestration_event_poller_tests.rs` covering:
cursor merge from server vs SQLite, retry on `get_ambient_agent_task`
failure, V2 gating, shared-session-viewer exclusion, cleanup on delete,
and `last_event_sequence` round-trip through
`AIConversation::new_restored`.
- Added unit test coverage in
`app/src/ai/blocklist/orchestration_events_tests.rs` for V1 lifecycle
re-registration on restore.
- Manual verification per
`specs/replay-agent-events-on-restore/TECH.md`.
## Server API dependencies
- [x] Does this change rely on a [new server
API](https://www.notion.so/warpdev/How-to-add-a-new-full-stack-feature-8412cede405a4ec194b32bdd4b951035?pvs=4#04da1e6a493542d68b3e998c7d339640)?
- [x] If so, is the use of this API restricted to client channels that
rely on the staging server (e.g. WarpDev)?
The companion server change adds:
- `last_event_sequence` on `Task` (`ai_tasks` column), surfaced inline
on `GET /agent/runs/:run_id`.
- `PATCH /agent/runs/:run_id/event-sequence` for client cursor writes.
- `children` inline on the `GET /agent/runs/:run_id` response.
## Agent Mode
- [x] Warp Agent Mode - This PR was created via Warp's AI Agent Mode
Co-Authored-By: Oz <oz-agent@warp.dev>
---------
Co-authored-by: Oz <oz-agent@warp.dev>
## Description
### Motivation
File-based stdio MCP servers spawned from a project's `.mcp.json` (e.g. `pnpm run mcp:foo`, or anything using `./tooling_scripts/...`) fail with:
```
Transport creation error: No such file or directory (os error 2)
```
Root cause: when the user's `.mcp.json` doesn't include a `working_directory`, Warp's spawner inherits whatever cwd Warp was launched from rather than the directory the config was discovered in. Two failure modes share this same root cause:
- **Repo-relative commands/args** (e.g. `./tooling_scripts/foo`, `node ./src/server.js`) can't be resolved from outside the repo, so `execvp` returns `ENOENT`.
- **Workspace-aware launchers** like `pnpm`/`npm`/`yarn` walk up from cwd looking for `package.json` (and workspace manifests). When cwd isn't inside the repo, they bail out before launching the requested script — surfacing as `ENOENT` from the spawner's perspective.
`working_directory` is a Warp/VS-Code-style extension; the canonical Anthropic/Cursor MCP schemas don't include it, so most `.mcp.json` files in the wild don't set it.
### Implementation
- New helper `FileBasedMCPManager::spawn_root_for_installation(uuid)` returning the discovery root for any file-based install: the repo root for project-scoped configs, the home directory for global configs (Warp and third-party). Returns `None` for non-file-based installs (e.g. cloud-templated ones), leaving them unaffected. Global Warp installs are remapped from `~/.warp/` (Warp internal state) to `~` so all global installs share a consistent cwd.
- In `TemplatableMCPServerManager::spawn_server_impl`, default `cli_server.cwd_parameter` from this helper when unset. Single funnel — covers both auto-spawn and the manual UI opt-in path. User-supplied `working_directory` always wins.
- ENOENT-aware logging at the spawn site: when `TokioChildProcess::spawn()` fails with `NotFound`, the MCP log file now spells out the server name, the missing executable, the cwd we used, and a hint pointing the user at `working_directory`. The user-surfaced error string is unchanged.
- Documented the new defaulting behavior and `working_directory` override in `resources/bundled/skills/add-mcp-server/SKILL.md`.
## Testing
Verified spawn-site behavior manually. Demo [here](https://www.loom.com/share/ec01d98ae9114433b8ae87f5f17adfd4)!
- Added `test_parse_cli_server_preserves_explicit_working_directory` in `mod_test.rs` to lock in that an explicitly-set `working_directory` round-trips through parsing and won't be clobbered by the new defaulting logic.
- Skipped a unit test for the new helper itself — it's a thin lookup over `file_based_servers_by_root` and the sorted-pick policy is straightforward enough to validate manually.
- Spawn-site behavior (cwd actually applied, ENOENT log emission) verified manually against a `pnpm`\-based repro.
## Agent Mode
- [x] Warp Agent Mode - This PR was created via Warp's AI Agent Mode
## Changelog Entries for Stable
CHANGELOG-BUG-FIX: Project-scoped file-based MCP servers now spawn from the repo root by default (and global ones from `~`), so configs with relative commands/args (and workspace launchers like `pnpm`/`npm`) work without an explicit `working_directory`.
Ryan Ward