Files
hermes-agent/tui_gateway/AGENTS.md
teknium1 6406be0f97 fix(tui_gateway): subagent.list follows the conversation across UI-sid and compression rotation
`subagent.list` matched live child records only on the exact `owner_session_id`
and session-record identity frozen at spawn. A Desktop reconnect / resume remints
the UI session id and rebuilds the session record, and compression rotates the
durable key, so every still-running child vanished from the Desktop subagent
panel for good while it kept working in the background (#114909, bug 2).

The read-only roster now also admits records whose durable lineage
(`owner_agent_session_id` resolved to its compression tip via the parent's
SessionDB) is the calling session's agent — the same `_owns_subagent_record`
spine the in-process `delegate_task(action="list")` already uses, so no new
record field, no registry rewrite at attach/compress sites. Control RPCs
(steer / interrupt / tail) keep the exact generation authority: visibility is
widened, authority is not.

Slimmer redo of the visibility half of #115189 by @aydnOktay, which re-pointed
`owner_session_id` / `owner_session_record` on every list/steer call.

Co-authored-by: aydnOktay <aydnoktay@users.noreply.github.com>
2026-09-19 23:57:59 -07:00

11 KiB

tui_gateway/ + ui-tui/ — the TUI and its JSON-RPC backend

Applies on top of the root AGENTS.md. The TUI fully replaces the classic prompt_toolkit CLI; activate with hermes --tui or HERMES_TUI=1. tui_gateway is ALSO the backend the Desktop app and the dashboard /chat talk to — changes here have three consumers.

Process model

hermes --tui
  └─ Node (Ink)  ──stdio JSON-RPC──  Python (tui_gateway)
       │                                  └─ AIAgent + tools + sessions
       └─ renders transcript, composer, prompts, activity

TypeScript owns the screen. Python owns sessions, tools, model calls, and slash-command logic. Never move agent behaviour into the renderer.

Transport

Newline-delimited JSON-RPC over stdio, peer-to-peer: client→server method calls, server→client requests (the agent asking the user something: approval, clarify, sudo, secret, vault.*, connection, the desktop read/act bridges) and server→client event notifications. tui_gateway/server.py is the facade with the method/event catalog; methods live in methods_*.py siblings (methods_config, methods_complete, methods_browser, methods_bot_relay, ...), event publishing in event_publisher.py / event_replay.py, server→client requests in server_requests.py (send() blocks the agent thread until the response frame with the same srq-<n> id arrives; cancel* withdraws with a request.cancel event; open_requests(sid) is what session.resume / session.events.since replay so a reconnecting client re-renders the still-open questions). A client says once per connection that it answers them (client.capabilities {server_requests: true}, sent by the shared channel on gateway.ready); a WebSocket client that never did is an app build older than server→client requests, and send() fails fast for it instead of stalling the agent for the deadline. Desktop reaches the same server over WebSocket via apps/shared (JsonRpcGatewayClient, onRequest). New RPC = a new methods_<topic>.py or an entry in an existing topical sibling, registered in the table — no if method == ... chain (root shape rules).

The wire is declared in Python and generated for TypeScript (tui_gateway/contracts/). Every method has a Params + Result model, every server→client request a Params + Result, every event a Payload — one Pydantic class each, extra="forbid" by default (OpenModel for producer-owned dicts). register_method refuses an undeclared name at import; the dispatcher rejects unknown param keys (4000 + key path) and, under HERMES_TEST_ISOLATION=1, raises ContractViolation when a handler's result or an emitted payload does not match its model (production only logs). apps/shared/src/ gateway-contract.generated.ts (RpcMethods, ServerRequestMap, BackendGatewayEventMap + every value shape) and gateway-contract.openrpc.json are rendered by scripts/gen_gateway_contracts.py; tests/tui_gateway/contracts/test_generated.py fails when they are stale, so the loop is: change the model → regenerate → tsc shows every consumer the field moved. apps/shared/src/gateway-events.ts only adds the client-local synthetic events and the GatewayEvent envelope on top. New question for the user = _ask("<method>", sid, params, timeout) in the emitter, a handler in apps/desktop/.../gateway-event/server-requests.ts and ui-tui/src/app/createServerRequestHandler.ts, and a server_request(...) in contracts/server_requests.py. New event = event("<type>", Payload) in contracts/events.py; the emitter is checked against it.

Profile scope in RPC methods

One serve process may host sessions from several profile homes (Desktop pooled backends launch under a profile; the dashboard serves several). The launch profile is a profile: "default" means the launch home, never ~/.hermes. The first non-launch home hosted flips launch_profile_policy.py → set_multiplex_active(True); without it every fail-closed guard is silently off. Every method that reads or writes home-, config- or .env-derived state runs under server.py::@_profile_scoped (resolved from the live session's profile_home, or the explicit profile argument for sessionless calls) and, for tool/agent construction, methods_tools.py::_profile_scoped_rpc; the tokens come from model_switch.py:: _profile_runtime_scope_tokens(profile_home) — home + secret scope + terminal scope together. A method that sets only get_hermes_home_override() is half-bound: config paths resolve to the right profile while credentials and TERMINAL_* policy still come from the launch profile. Off-turn paths bind the same way: session_lifecycle.py::_finalize_session / _teardown_session enter _session_profile_runtime_scope(session) around on_session_end, the memory commit and agent.close() (their callers are unscoped reapers, Timers, atexit and pool threads); background threads start via agent.memory_provider.spawn_context_thread, never bare threading.Thread; children act for the served profile through tools/environments/local.py::served_profile_child_env (hermes -p X workers, key_cmd helpers, browser drivers), never dict(os.environ). Grep for unscoped handlers before adding one: rg -n "^(async )?def " tui_gateway/methods_*.py | rg -v _profile_scoped. Probe with two on-disk homes and a .env name present only in the secondary: call the method for that session and assert the secondary's value resolves and the launch profile's does not, and that os.environ is unchanged afterwards.

Key surfaces

Surface Ink component Gateway method / event
Chat streaming app.tsx + messageLine.tsx prompt.submit → message.delta / message.complete
Tool activity thinking.tsx tool.start / tool.generating / tool.complete
Approvals prompts.tsx server→client request approval → response {choice}
Clarify / sudo / secret prompts.tsx, maskedPrompt.tsx server→client requests clarify / sudo / secret (server_requests.py)
Session picker sessionPicker.tsx session.list / session.resume
Slash commands local handler + fallthrough slash.exec → _SlashWorker; command.dispatch
Completions useCompletion hook complete.slash, complete.path (under a non-local terminal.backend, complete.path lists the directory through the session's terminal backend — never the gateway host, whose same-named tree would look right and be wrong)
Theming theme.ts + branding.tsx gateway.ready carries skin data
Plugin compat notice — plugins.compat_report (see plugins/AGENTS.md)
Connection operations (desktop card) desktop store/connection-request.ts connection.request → connection.update* → connection.respond {op_id}; connectors.operation.status. The op lives in tools/connectors/live.py; the card never parks the tool thread (methods_connectors.py).

Shared subagent snapshots

subagent.list({session_id}) returns {subagents, delegations} for the calling transport's live session. The roster is read-only and follows the CONVERSATION: exact-owner records plus children whose durable lineage (owner_agent_session_id → compression tip, the same spine as in-process delegate_task(action="list")) is the session's agent, because a Desktop reconnect / resume remints the UI session id and compression rotates the key while the children keep running (#114909). Control (steer / interrupt / tail) stays pinned to the exact session record and transport. Child authority is resolved at RPC time against the owning session's LIVE transport slot, so every authenticated reattach path (prompt.submit, queued drain, resume, activate, viewer failover) carries it with no registry bookkeeping — never add a per-record transport sync at an attach site; foreign or retired generations remain uncontrollable. last_tool is the last started tool, not an in-flight indicator. Async completion units are not agents and lack exact generation authority; delegations remains an empty array for wire compatibility. No dispatch context, results, callbacks, or routing keys are sent. Clients hydrate from this snapshot on their existing poll and avoid updates when unchanged.

subagent.tail({session_id, subagent_id}) returns {subagent_id, available, text, truncated}: the last 16 KiB of the live child's existing transcript. Poll only the selected detail. Missing/finished/foreign children return an unavailable empty snapshot; no client-supplied path is opened. This is live-only, not persisted completion history. Invalid session/transport returns error 4001. subagent.steer({session_id, subagent_id, text}) remains the shared control: status: queued acknowledges acceptance, not delivery; final boundary races are reported by the existing runtime as missed_steer. subagent.interrupt({session_id, subagent_id}) requires the same exact live session/transport/generation ownership, including for subtree members. Missing RPC session authority is rejected; direct in-process interrupt_subagent(id) retains its legacy unscoped contract.

Slash command flow

  1. Built-in client commands (/help, /quit, /clear, /resume, /copy, /paste, ...) are handled locally in app.tsx.
  2. Everything else → slash.exec, which runs in the persistent _SlashWorker subprocess → command.dispatch fallback, which the gateway resolves into a skill / alias / exec directive (a skill command resolves to {type: "skill", message} and is submitted as a normal prompt).

commands.catalog (empty-query list) and complete.slash (typed-query completions) already include built-ins, user quick_commands, AND skill-derived commands (scan_skill_commands() / get_skill_commands()) — clients do not need a new RPC to see skills. The command definitions themselves come from hermes_cli/commands.py (hermes_cli/AGENTS.md).

Dev commands

cd ui-tui
npm install       # first time
npm run dev       # watch mode (rebuilds hermes-ink + tsx --watch)
npm start         # production
npm run build     # full build (hermes-ink + tsc)
npm run typecheck # tsc --noEmit
npm run lint      # eslint
npm run fmt       # prettier
npm test          # vitest

Python tests: tests/tui_gateway/ via scripts/run_tests.sh. TS tests: vitest in ui-tui. A Python test that asserts about package.json / .ts sources will not run on a JS-only PR — keep JS-side assertions in vitest (root testing rules). Root TypeScript style rules apply.

Related: web/AGENTS.md (dashboard embeds this TUI over a PTY), apps/desktop/AGENTS.md (own renderer on the same backend).