Run agent shell commands as independent background jobs

AGENTS.md

@@ -227,10 +227,11 @@ Key invariants:
 - Progressive summaries are prepended to provider requests as a user/assistant pair; direct-provider runs compact their live transcript at a checkpointed `ReadyToCallModel` boundary before another model call, while background summarization remains independent for session-owned runtimes
 - Direct-provider progressive compaction runs in a separate request alongside foreground work. Start early against the smaller usable history budget of the current and terminal-use profiles, including each profile's prompt/tools/output reservation. Apply a result only at a safe model boundary after validating its exact snapshotted prefix, preserving all newer history and recomputing live context accounting. Wait for pending compaction only after context exhaustion; cancellation and terminal cleanup must remain immediate. Cleanly completed conversations may accept a late summary while idle; stale, cancelled, or superseded generations must never replace a newer run's history.
 - Loop prevention in `controller.rs` detects repeated tool failures (3+ identical) and injects a corrective instruction
-- Direct-provider long-running shell follow-ups retain CLI tasks as UI projections while command monitoring and completion stay owned by the same provider run
-- A direct-provider command completion is only queued when the terminal reports it; the CLI task remains active until the provider run applies that completion at a safe boundary and deactivates it
-- Provider command ownership is resolved from the active slot or its durable snapshot by block/action identity; completion arriving during restore is persisted into that snapshot and must never fall back to the legacy assessment path
-- Direct-provider completed-command assessments are hidden, tool-free root-task turns whose output and hidden input survive CLI-task deactivation and restoration
+- Agent-requested shell commands run as independent, noninteractive background jobs with closed stdin and no PTY. User-typed commands retain the normal terminal execution and rendering path. No model turn may poll commands, write process input, or take over a shell; unavoidable interactive commands are given to the user to run themselves.
+- A successful shell launch returns `BackgroundJobStarted` and yields the provider generation only after every tool result in that batch is committed. The process belongs to `BackgroundJobsModel`, independently of conversation cancellation or subsequent user input. Running or undelivered jobs keep orchestrated children pending rather than reporting premature success.
+- Background completion sends command, bounded logs, and actual exit status as hidden execution evidence to the owning conversation when it is idle. Completion is not a user request or new authorization, must preserve the user's draft and selected conversation, and is acknowledged only after acceptance. Failed, cancelled, and successful jobs all deliver terminal evidence.
+- `/background-jobs` lists persistent history scoped by git project or directory with adjustable filters, logs, and process-tree termination. Inline tool panes show status only: animated running indicator, green success, red failure, gray cancelled square. Full logs live in private temporary files; View logs appears only while a retained log is readable, and truncation in model context is explicit.
+- Restoring history never replays a command or wakes an agent from old completion records. Local interrupted jobs become cancelled; remote running jobs reconcile with their owning daemon before claiming an outcome and remain controllable after reconnect. Process cancellation is scoped to an owned process group/job object, never a reused PID or another conversation's command.
 - ACP remains a separate session-owned runtime; direct-provider cleanup must not move ACP lifecycle into `ProviderRun`
 - Orchestrated child conversations are leaf workers: nested `RunAgents` and legacy `StartAgent` calls are rejected before autonomy or permission bypasses, and child requests do not advertise delegation tools
 - Direct-provider `RunAgents` acknowledges linked local children and yields the parent generation after committing its launch results. Other tools in that launch batch are explicitly not executed. Persistent `LocalAgentGroup` membership survives parent follow-ups; only when every child reaches `Success`, `Error`, or `Cancelled` (with provider cleanup complete), or is removed/deleted, does one combined completion message wake the parent. `Blocked`, `TransientError`, and `WaitingForEvents` children remain pending. Restarts invalidate undelivered group results, and group delivery is confirmed only after the parent input is accepted.
@@ -245,7 +246,7 @@ Key invariants:
 - `app/src/ai/blocklist/action_model/risk_review.rs` assesses Galaxy Decides actions using the terminal-use model in a separate, tool-free request, falling back to the base model when unconfigured. Risk 0–3 runs; risk 4–5 and denylist/protected-path/workspace restrictions deny automatically. Galaxy Decides never requests human permission. Denials return reasons so the agent can revise its scope/justification or choose a safer approach.
 - Risk reviews carry chronological available recorded user requests, including intermediate restrictions; oversized review context denies rather than discarding revocations, separate from agent rationale/delegated scenarios. Worker reviews inherit root-conversation authorization; agent claims never establish permission. Later revocations take precedence. An explicitly requested ordinary non-force push within the stated/established remote and branch may be risk 3; external visibility alone is not risk 4. Force/deletion/out-of-scope pushes, secrets, and production changes remain stricter.
 - Invalid, unavailable, oversized, or timed-out assessments never grant execution. Assessment calls are bounded to 30 seconds. Reviews bind conversation, action contents, execution identity, profile mode and environment. Cancellation, replacement and cleanup invalidate callbacks; stale approval clicks are not authorization.
-- Approval of a running agent command covers subsequent output monitoring and PTY input for that same conversation-owned active block while the agent retains control. Validate block and launch-action identity; never extend PTY writes to a finished command, another conversation, or a fresh shell prompt. A finished owning block may return its final snapshot without additional permission.
+- Approval of an agent shell command covers launching that exact background job and terminating its owned process tree. It does not grant interactive input or another command. Cancellation must match the job and owning conversation; users can manage jobs directly in `/background-jobs`.
 - Galaxy Decides risk assessment appears only as a transient “Checking permissions…” status above the input, not a per-tool permission pane. Ask First keeps its actionable approval panes. `PermissionReviewChanged` refreshes the owning conversation’s status bar.
 - Safe inspection and output polling do not add risk-review model calls. A narrow shell fast path accepts validated `pwd`, `ls`, Git version, and hardened Git status/diff forms; shell expansion, chaining, redirects, unknown flags and unhardened Git commands still require review. Explicit restrictions are checked before fast approval. Galaxy Decides skips user-question pauses. External ACP/CLI agents retain their own protocol permission flows; unstructured ACP requests must not grant Galaxy terminal mutations.