feat: agentic-loop automation + multi-agent CLI mode (codex live; 7 agents staged)#23
Merged
Merged
Conversation
Generalize the tmux-backed CLI pane beyond claude so any agent can run its interactive TUI in the main workspace, reusing the same per-agent settings as headless mode (without touching the managed/UIX executors). - Add CliLaunchSpec (program/base_args/resume/prompt/continue forms) + an interactive_cli_spec() trait method; agents opt in, None falls back to claude. - Implement it for ClaudeCode (refactor of the old hardcoded launch — no behavior change) and Codex (-m model, -c model_reasoning_effort, -s sandbox, -a approval, --no-alt-screen, 'codex resume <id>'). - Make cli_bootstrap/PtyCommand::TmuxCli agent-agnostic; resolve the agent from Session.executor in terminal.rs; honor the selected executor in start_workspace_cli instead of hardcoding claude. - Pre-seed codex local-env friction (per-folder trust + the blocking 'Update available' startup modal) so an unattended launch reaches the TUI. Codex CLI mode verified live in a tmux pane (YOLO autonomy + model + worktree).
…wake-ups, re-prompt) Keep an opt-in CLI workspace's agent going when a chat stops for a non-completion reason (usage limit / transient rate limit), so agentic loops survive provider throttling instead of stalling until a human notices. - DB: loop_automations (per-workspace opt-in policy, default OFF) + scheduled_wakeups (pending/delivered re-prompts) migration + models. - Loop supervisor (tokio interval task, spawned from LocalDeployment::new()): capture-pane to detect the limit banner, schedule a wake-up (short retry for a rate limit, parsed reset time for a usage limit), and re-prompt via send-keys. Idle-safe: limit-kind wake-ups only fire while the banner is still showing. - tmux capture_cli_pane + send_cli_keys helpers (pane output is otherwise ephemeral; this is the only server-side view of / write to a live detached pane). - REST API: GET/PUT per-workspace policy, POST manual wake-up, DELETE wake-up. - Kill-switch env DISABLE_LOOP_AUTOMATION; banner + reset-time parsing unit-tested; tmux send/capture mechanics validated live.
…ilot CLI specs Implement interactive_cli_spec for the remaining 7 agents so any of them can run in the CLI pane via the same generic path as claude/codex. Each maps its existing managed config (model / reasoning / autonomy) to interactive flags, autonomous by default (Supervised dials back): - gemini/qwen: --approval-mode yolo, -i prompt, --resume - opencode: -m provider/model, --prompt, -s/-c (autonomy via opencode.json) - cursor-agent: resolved --model, --force, positional, --resume/--continue - droid: --model, --reasoning-effort, --auto/--skip-permissions-unsafe - amp: stdin-piped prompt (new CliPromptArg::StdinPipe), 'amp threads continue' - copilot: --model, --allow-all-tools, --no-banner, -i These are NOT installed locally, so they're staged with unit tests + an exact per-agent install/auth/seeder/e2e plan (docs/.../cli-agents-install-and-e2e-plan.md) awaiting approval before any install. cli_spec_tests verifies all 9 agents wire up with the right binary + autonomous defaults.
Front-end for part 1: a per-workspace, default-OFF 'Keep going' toggle in the CLI
pane header with a settings popover (retry interval, continuation prompt, max
attempts) and a next-wakeup status line ('Retrying at 14:32' / 'Waking at 05:00
UTC') with a cancel button.
- loopAutomationApi (get status / update policy / create + delete wakeup) in api.ts
- useLoopAutomation react-query hook (status query + mutations, cache seeding)
- LoopAutomationControl wired into CliMainPane header; i18n strings added
Typechecks + lints clean (web-core/local-web/remote-web, eslint, i18n).
…ft on installed versions Installed gemini/qwen/opencode/amp/copilot (npm) + cursor-agent/droid (curl) and launched each in a real tmux pane with the exact flags interactive_cli_spec emits. All 7 binaries resolve and every coded flag is accepted; each reaches its own UI (TUI / login / provider menu) — opencode even answered on its free built-in model. Only auth keys/login remain. Flag fixes for the newer installed versions: - droid 0.161: interactive launch no longer takes --model/--reasoning-effort/ --skip-permissions-unsafe; emit only --auto low|medium|high (Skip→high). - copilot 1.0.65: no --no-banner (banner off by default); removed. - opencode: use confirmed long-form --session/--continue. Verification table + remaining per-agent friction recorded in the install plan.
…an agent isn't installed Previously, launching CLI mode for an agent whose binary isn't on PATH (the 'command -v <prog> && ...' gate failed) dropped silently into a bare shell — the user picked an agent and got an empty terminal with no reason. Now the pane prints a clear '[!] <agent> is not installed — Install: <cmd>' notice with the per-agent install command, then leaves a usable shell so they can install it there or switch agents. (Installed-but-not-authed is unchanged: the agent shows its own login screen, which is the right place for it.)
… don't hide) Proactive complement to the in-pane install notice: the create-flow agent picker now floats installed agents to the top and tags uninstalled ones '· not installed', so you see it before launching rather than after. - GET /api/agents/cli-availability returns which agents' interactive CLI binary is on PATH (a CLI-specific 'command -v' check via the which crate — distinct from the config-file-based AvailabilityInfo; managed mode runs via npx and is unaffected). - useCliAgentAvailability hook + configApi.cliAgentAvailability. - CreateChatBoxContainer sorts installed-first and annotates via the existing formatExecutorLabel prop (no @vibe/ui change). Loading state reports all installed so it never flashes a misleading tag. Picker stays honest/consistent with the managed UI (shows all agents); the in-pane notice still covers launch if a missing agent is picked anyway.
… the hot path) The TmuxCli variant carries the resolved CliLaunchSpec and is the common case (CLI mode is the default pane), so the size gap vs the unit Shell variant is expected; boxing the hot path would only add an allocation. Keeps `cargo clippy -- -D warnings` (as CI runs it) green.
…ntend-checks)
The 'Keep going' UI strings were added to en/common.json only; CI's check-i18n
requires every key in every locale (es, fr, ja, ko, zh-Hans, zh-Hant). Add
translations for all 16 keys across the 6 locales, inserted in en's key order
(after cliMode). Preserves {{time}}/{{used}}/{{max}} interpolation placeholders.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Two additive features that keep agentic loops running and bring every supported agent into CLI mode. The older managed/UIX executors are untouched — everything here is additive.
Part 1 — Loop automation (keep going when a chat stops for a non-completion reason)
Opt-in per workspace (default OFF). The supervisor watches the CLI pane, and when the agent stalls on a usage limit or a transient rate limit it schedules a wake-up and re-prompts the agent — so loops survive throttling instead of waiting for a human.
loop_automations(per-workspace policy) +scheduled_wakeupsmigration + models.local_deployment::loop_supervisor, a tokio-interval task):tmux capture-paneto detect the banner, schedule a wake-up (short retry for a rate limit, parsed reset time for a usage limit), and re-prompt viatmux send-keys. Idle-safe — a limit-kind wake-up only fires while the banner is still showing, so a manually-resumed agent is never interrupted.GET/PUTper-workspace policy,POSTmanual wake-up ("ping at 05:00 UTC"),DELETEwake-up.DISABLE_LOOP_AUTOMATION. Banner + reset-time parsing unit-tested;send-keys/capture-panemechanics validated live.Part 2 — Multi-agent CLI mode
The persistent tmux CLI pane was Claude-only. It's now generic: any agent runs its interactive TUI in the pane, reusing the same per-agent settings the managed UI already exposes (model / reasoning / sandbox / approval).
CliLaunchSpec+interactive_cli_spec()trait method; agents opt in,Nonefalls back to claude.cli_bootstrap/PtyCommand::TmuxCliare agent-agnostic; the agent is resolved fromSession.executor;start_workspace_clihonors the selected executor.Agent status
interactive_cli_specimplemented + unit-tested, wired through the same path. Not installed locally, so they're staged behind your approval: seedocs/superpowers/specs/2026-06-30-cli-agents-install-and-e2e-plan.mdfor each agent's exact install command, auth/onboarding requirements, the code already written, and the e2e checklist. No tools are installed by this PR.Verification
cargo testgreen for the changed crates: executors (73), local-deployment (42).generate-types:check✅ andprepare-db:check✅ (CI guards).tsc(web-core/local-web/remote-web), eslint, and i18n checks ✅.cargo check --workspacefails only on thetauri-appGUI crate (glib-sys, no GTK in this headless env) and theremotecrate's privatebillingdep — both pre-existing and unrelated to this change.Design docs
docs/superpowers/specs/2026-06-30-cli-agents-and-loop-automation-design.mddocs/superpowers/specs/2026-06-30-cli-agents-install-and-e2e-plan.md