webbee 0.3.22__tar.gz → 0.3.24__tar.gz

{webbee-0.3.22 → webbee-0.3.24}/CHANGELOG.md

@@ -1,5 +1,30 @@
 # Changelog
 
+## 0.3.24
+
+The tab bar got a real design: the active tab is a solid highlighted chip,
+tabs are clearly separated, and spacing is uniform — plus a message typed
+on Home no longer shows up twice in the new tab.
+
+- Tabs now render as padded chips with uniform spacing, a clear dim
+  separator between every pair, and the active tab as a solid
+  bee-yellow block — unmistakable at a glance, no more squinting to find
+  which tab you're on.
+- Typing a task on Home and landing in the new session tab no longer
+  double-prints your first message in that tab's transcript.
+- Each tab now keeps its own input draft — switch away and back, your
+  unsent text is right where you left it.
+
+## 0.3.23
+
+If a task on your session is waiting for your terminal — sent from the panel
+or Telegram while the terminal was idle — the terminal now picks it up
+within seconds instead of the task stalling.
+
+- An idle terminal now notices when a running task on its own session is
+  waiting for a tool approval or result, and reconnects to it automatically
+  — no more silent stalls that eventually get marked "unresponsive".
+
 ## 0.3.22
 
 Open the same project in two tabs and they truly run in parallel — each tab

{webbee-0.3.22 → webbee-0.3.24}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: webbee
-Version: 0.3.22
+Version: 0.3.24
 Summary: Webbee 🐝 — the Imperal Cloud coding agent in your terminal
 Project-URL: Homepage, https://imperal.io
 Project-URL: Documentation, https://docs.imperal.io

{webbee-0.3.22 → webbee-0.3.24}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "webbee"
-version = "0.3.22"
+version = "0.3.24"
 description = "Webbee 🐝 — the Imperal Cloud coding agent in your terminal"
 readme = "README.md"
 license = "MIT"

webbee-0.3.24/src/webbee/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.3.24"

{webbee-0.3.22 → webbee-0.3.24}/src/webbee/repl.py

@@ -573,11 +573,19 @@ async def _home_input(text: str, *, slots: SlotManager, cfg, token_provider, mod
     ui_hooks.get("switch", slots.switch)(idx)
     if spawn_poller is not None:
         spawn_poller(new_slot)
-    new_slot.sink.user_echo(text)
     start_turn_in = ui_hooks.get("start_turn_in")
     if start_turn_in is not None:
+        # 0.3.24 fix: the dock's start-turn seam threads through on_line ->
+        # `_handle`, which ALREADY echoes the typed line itself for every
+        # non-command line (the SAME canonical echo every other typed line
+        # in the dock gets) -- echoing here too doubled it up in the new
+        # tab's transcript (Valentin, live on 0.3.22/0.3.23).
         start_turn_in(new_slot, text)
     else:
+        # No dock: `run_turn` (repl's `_run_turn`) never echoes on its own
+        # (it only runs the agent turn), so this IS the one echo the typed
+        # line gets.
+        new_slot.sink.user_echo(text)
         await run_turn(new_slot, text)
 
 
@@ -977,6 +985,72 @@ async def run_repl(cfg, mode: str = "default", *, once: bool = False, sink=None,
         slot.sink.foreign_turn(surface, "user", text)
         await _run_turn_on(slot, text, surface=surface, steer_iid=steer_iid)
 
+    async def _attach_turn_on(slot: SessionSlot, attach_info: dict) -> None:
+        """poll_idle_steer's `attach_turn` seam (attach-on-poll, URGENT
+        live-user-pain fix): a kernel-side marathon turn woken by a panel/
+        Telegram message dispatched a local tool_request/confirm_request
+        while THIS terminal sat idle-but-online -- nothing attached the
+        stream, so the dispatch just burned tool_wait (5-65min) before the
+        kernel parked it "unresponsive". `attach_info` is the gateway's
+        {task_id, last_id, kind} (GET /pending-steer's additive `attach`
+        field); `slot.agent.attach()` opens the SSE with NO start POST at
+        all -- doing so alone fires client_connected server-side, which is
+        what makes the kernel re-dispatch the pending request. `marathon=not
+        once` (the SAME flag `_spawn_slot_poller` already threads into
+        poll_idle_steer/derive_session_id for THIS slot's poller) so a
+        --once slot's attach derives its "coding-"-prefixed id, never the
+        "marathon-" one -- a mismatched prefix would derive a DIFFERENT,
+        wrong session id than the one whose `attach` field this actually is.
+
+        Runs through the SAME slot-explicit start seam `_home_input` uses
+        (`ui_hooks["start_attach_in"]`, tui's own `_start_attach_in`) so
+        `slot.turn["task"]` is genuinely set under a dock -- Esc/Ctrl-C can
+        cancel it locally exactly like a typed turn, and begin_turn/
+        end_turn/busy-gate/queue-drain semantics are identical to any other
+        turn (mirrors `_run_turn_on`'s own shape). The fallback loop (no
+        dock, no ui_hooks entry) just awaits the coroutine directly -- no
+        Esc/Ctrl-C key handling exists there to satisfy. Either way this
+        BLOCKS for the whole turn: poll_idle_steer's `attach_turn` contract
+        requires pausing polling for its duration, same discipline as
+        `submit`/`_steer_submit_on` above."""
+        async def _drive() -> None:
+            _sink = slot.sink
+            _sink.begin_turn()
+            _note = getattr(_sink, "note", None)
+            if _note is not None:
+                _note("⚡ attaching — a running task on this session needs this terminal")
+            task_id = str(attach_info.get("task_id") or "")
+            start_id = str(attach_info.get("last_id") or "0-0")
+            try:
+                text = await slot.agent.attach(_sink, task_id=task_id, start_id=start_id,
+                                               marathon=not once)
+            except (KeyboardInterrupt, asyncio.CancelledError):
+                _sink.abort()
+                _sink.note("Interrupted.")
+                _sink.end_turn("")   # clear busy (poller starvation guard)
+                return
+            except Exception as e:  # network/auth/etc — never crash the poller
+                _mark = getattr(_sink, "mark_turn_failed", None)
+                if _mark is not None:
+                    _mark()
+                if type(e).__name__ in ("StreamAuthError", "NotLoggedInError"):
+                    _sink.note("Session expired or access revoked — run /login to sign in again.")
+                else:
+                    _sink.note(f"Error: {type(e).__name__}: {e}")
+                _sink.end_turn("")
+                if slot.pending:
+                    _sink.note(f"⏸ queue held: {len(slot.pending)} queued message(s) wait "
+                               "— ↑ pulls the next into the input, /queue clear drops them")
+                return
+            _sink.end_turn(text)
+
+        coro = _drive()
+        start_attach_in = ui_hooks.get("start_attach_in")
+        if start_attach_in is not None:
+            await start_attach_in(slot, coro)
+        else:
+            await coro
+
     async def _stop_active_turn() -> None:
         """tui.run_session's `stop_turn` leg (Esc/Ctrl-C) -- resolves the
         ACTIVE slot's agent AT CALL TIME (W4a Task 3: the injected callable
@@ -1006,7 +1080,10 @@ async def run_repl(cfg, mode: str = "default", *, once: bool = False, sink=None,
         shared `first_session_slot`: `is_busy`/`mode_getter` read its OWN
         sink/turn/mode, `live_session_id` its OWN agent, `submit` renders the
         remote line into its OWN pane and runs the turn through
-        `_run_turn_on(slot, ...)`. Staggered (item 3): each poller's FIRST
+        `_run_turn_on(slot, ...)`, and `attach_turn` (attach-on-poll) runs
+        `_attach_turn_on(slot, ...)` against this SAME slot when the
+        gateway's drain reports an unanswered request already dispatched on
+        this session. Staggered (item 3): each poller's FIRST
         tick carries a small incrementing offset (0, 1, 2, 0, 1, 2... seconds
         -- `poll_idle_steer`'s own `initial_delay`) so several tabs opened
         back-to-back don't all hit the gateway in the same instant. Appended
@@ -1025,6 +1102,7 @@ async def run_repl(cfg, mode: str = "default", *, once: bool = False, sink=None,
                 slot, text, surface, steer_iid),
             on_mode=lambda mode, surface: _on_mode(slot, mode, surface),
             mode_getter=lambda: slot.mode,
+            attach_turn=lambda attach_info: _attach_turn_on(slot, attach_info),
             client=shared_client, slot_id=slot.slot_id, initial_delay=offset))
         slot.bg_tasks.append(task)
 

webbee-0.3.24/src/webbee/session.py

@@ -0,0 +1,501 @@
+import asyncio
+import time
+from collections import OrderedDict, deque
+
+from webbee.coding_context import build_coding_context, detect_verify_cmd
+from webbee.consent import _retire, race_consent
+from webbee.frames import (
+    _MARATHON_FACT_TYPES,
+    _first_time,
+    _is_foreign_frame,
+    _origin_tag,
+    _progress_text,
+    _summary,
+    handle_action_frame,
+    handle_step_finished,
+    handle_step_started,
+    handle_tool_request,
+    marathon_note,
+    render_foreign_frame,
+    render_todo_frame,
+)
+
+
+# W1 Task 10 (own-task watchdog): a test seam for the deduped-twin-hang clock
+# below -- monkeypatched as `webbee.session._monotonic` so a test can drive
+# elapsed-time without a real sleep.
+_monotonic = time.monotonic
+
+# Deduped-twin hang (W1 front-3b): a turn's task_id can be ring-dropped
+# kernel-side as an at-least-once twin, so the shared stream carries ONLY
+# frames tagged with OTHER task_ids while this turn's own task_id never
+# appears. Left alone the dock spins "working" forever. Past this many
+# seconds of foreign-only traffic with zero own frames, end the turn honestly.
+_FOREIGN_ONLY_DEADLINE_S = 90.0
+
+
+_SEEN_KEEP = 64   # kernel re-dispatch only ever targets the CURRENT pending
+                  # req_id -- a small recency window is enough; unbounded full
+                  # results (file bodies, bash output) were the many-hour-
+                  # marathon RAM leak (W1 front-1).
+
+
+def _remember(seen: OrderedDict, rid, out) -> None:
+    """LRU-bounded store for `seen` -- the run() dedup cache. A (re-)store
+    always counts as the most-recent touch; the least-recently-touched entry
+    is evicted once the store grows past _SEEN_KEEP."""
+    seen[rid] = out
+    seen.move_to_end(rid)
+    while len(seen) > _SEEN_KEEP:
+        seen.popitem(last=False)
+
+
+def _is_transient_status(status: int) -> bool:
+    return status >= 500 or status in (408, 429)
+
+
+def _transient_exceptions():
+    # Lazy import (module-level `import httpx` would move httpx's cost to CLI
+    # boot -- session.py is imported by repl.py at module top, and the rest of
+    # this module already keeps httpx as a run()-local, boot-time-optimization
+    # import). `except <call expression>` is valid Python -- the call just
+    # has to happen on every except test, which is cheap (import is memoized).
+    import httpx
+    return (httpx.HTTPError, OSError)
+
+
+async def _transient_retry(send, *, attempts: int = 5, base: float = 1.0, cap: float = 8.0):
+    """Bounded transient-retry for gateway WRITE calls (turn-start POST): a
+    502 during a deploy must not kill the turn. Verdict/normal statuses return
+    immediately; TRANSPORT errors (httpx.HTTPError/OSError) and 5xx/408/429
+    retry with capped backoff -- anything else (a bug in the caller) is not a
+    transient condition and must surface immediately, not be retried 5x.
+    The LAST failure is returned/raised for the caller's raise_for_status."""
+    backoff = base
+    last_exc = None
+    last_resp = None
+    n = max(1, attempts)
+    for i in range(n):
+        try:
+            resp = await send()
+            if not _is_transient_status(resp.status_code):
+                return resp
+            last_resp, last_exc = resp, None
+        except _transient_exceptions() as e:
+            last_exc, last_resp = e, None
+        if i < n - 1:               # never sleep after the FINAL attempt
+            await asyncio.sleep(backoff)
+            backoff = min(backoff * 2, cap)
+    if last_exc is not None:
+        raise last_exc
+    return last_resp
+
+
+class AgentSession:
+    """Client-side driver for one coding turn against the Imperal cloud.
+    The brain runs server-side; this is the hands — it streams kernel-
+    pre-approved tool_request frames over SSE, runs each tool locally, relays
+    confirm_request replies RAW for the brain to interpret, drives the sink
+    for live UI, and posts results back until a final frame arrives.
+
+    P1: one POST per turn (server reloads the shared webbee-terminal thread,
+    so context carries across turns). Persistent signal-based sessions are P3."""
+
+    def __init__(self, cfg, token_provider, workspace_root: str, mode: str = "default", intel=None,
+                 shadow=None, slot_id: str = "") -> None:
+        self.cfg = cfg
+        self.token_provider = token_provider
+        self.workspace_root = workspace_root
+        self.mode = mode
+        self.session_id: str = ""
+        self.steps: deque = deque(maxlen=200)
+        self._task_id: str = ""
+        self._intel = intel  # IntelService, or None (base install / boot failure)
+        self._shadow = shadow  # ShadowGit, or None (git unavailable / boot failure)
+        # W4b T5: "" (tab-1 / the fallback loop's only slot) keeps every turn's
+        # POST body byte-identical to before this feature existed -- a later
+        # tab's own short slot_id rides the body as `slot`, and the gateway
+        # mints an `-s{slot}`-suffixed session id from it.
+        self.slot_id = slot_id
+
+    async def _headers(self) -> dict:
+        token = await self.token_provider()
+        return {"Authorization": f"Bearer {token}"}
+
+    async def run(self, task: str, sink, *, marathon: bool = False, goal: str = "",
+                  surface: str = "", steer_iid: str = "") -> str:
+        import httpx
+        from uuid import uuid4
+
+        from webbee.tools import LocalToolExecutor
+        from imperal_mcp.client import ImperalClient
+
+        # FIX1 (W1 final review): every turn-start POST needs a stable dedup
+        # key. A typed turn mints its own (steer_iid arrives empty); a steer
+        # pickup keeps the queue entry's OWN id. Either way, _transient_retry
+        # below re-sends this SAME closure -- so a retry after an ambiguous
+        # failure (timeout / edge-504 whose first attempt may have already
+        # landed) carries the IDENTICAL key, and the kernel's steer-iid ring
+        # drops the twin instead of executing the instruction twice.
+        steer_iid = steer_iid or uuid4().hex
+
+        # Offload to a worker thread — build_coding_context runs sync
+        # subprocess.run(git status, timeout=10) + os.walk; inline on the dock's
+        # asyncio loop it froze the whole UI at every turn start.
+        coding_context = await asyncio.to_thread(build_coding_context, self.workspace_root, self._intel)
+        if marathon:
+            # verify_cmd is CLIENT-detected here and carried in coding_context —
+            # the trusted proof-of-done the kernel runs (never brain-authored).
+            root = coding_context.get("repo_root") or coding_context.get("cwd") or self.workspace_root
+            verify_cmd = await asyncio.to_thread(detect_verify_cmd, root)
+            coding_context = {**coding_context, "verify_cmd": verify_cmd}
+        imperal_id = await ImperalClient(self.cfg, self.token_provider).whoami()
+        executor = LocalToolExecutor(self.workspace_root, indexer=self._intel,
+                                     shadow=self._shadow)
+
+        body = {"user_id": imperal_id, "task": task, "coding_context": coding_context}
+        if surface:
+            # Liveness v2 §B: an idle-steer pickup carries the queued item's
+            # origin surface so the kernel adopts it start-path (provenance +
+            # [surface] tags). Additive-only -- a typed turn's body is
+            # byte-identical to before.
+            body["surface"] = surface
+        # steer-iid-dedup: ALWAYS present now (FIX1 above) -- W1 turn-start
+        # retries need the ring dedup, so a typed turn's minted id rides the
+        # body exactly like a pickup's own id. No longer additive-only for
+        # this key (surface/marathon/goal above still are).
+        body["steer_iid"] = steer_iid
+        if self.slot_id:
+            # W4b T5: slot-suffixed session ids -- the gateway mints
+            # {prefix}-{imperal_id}-r{repo_key}-s{slot} when a slot rides the
+            # body. Additive-only: "" (tab-1 / fallback's only slot) keeps
+            # today's body byte-identical, no "slot" key at all.
+            body["slot"] = self.slot_id
+        if marathon:
+            body["marathon"] = True
+            body["goal"] = goal
+
+        headers = await self._headers()
+        async with httpx.AsyncClient(base_url=self.cfg.api_url, timeout=60) as client:
+            resp = await _transient_retry(lambda: client.post(
+                "/v1/agent/sessions", json=body, headers=headers))
+            resp.raise_for_status()
+            _sess = resp.json()
+            session_id = _sess["session_id"]
+            start_id = _sess.get("last_id", "0-0")
+            task_id = _sess.get("task_id", "")
+            return await self._serve_stream(sink, client, session_id, start_id, task_id,
+                                            executor, marathon=marathon)
+
+    async def attach(self, sink, *, task_id: str, start_id: str, marathon: bool = True) -> str:
+        """Attach-on-poll (idle terminal picks up a kernel-dispatched task):
+        NO start POST. A marathon turn woken by a panel/Telegram message
+        while this terminal sat idle can dispatch a local tool_request/
+        confirm_request that nothing attaches -- the gateway's
+        `GET .../pending-steer` response flags this via its additive
+        `attach` field (set only when the drain found no items AND the
+        session's stream tail holds an unanswered request) and hands the
+        client `task_id`/`last_id` to resume from. This session id is
+        DERIVED the same way `webbee.steer.derive_session_id` derives it
+        for polling in the first place (this AgentSession already knows its
+        own workspace + slot_id, so the derivation is reused verbatim, not
+        reinvented) -- simply opening the SSE stream against it fires
+        `client_connected` server-side, which makes the kernel re-dispatch
+        the pending request onto the stream. From there this runs through
+        the EXACT SAME frame-loop `run()` uses (`_serve_stream`): tool
+        execution, consent racing, dedup, and the final/marathon_paused
+        exits all behave identically to a normal turn -- including
+        begin_turn/end_turn semantics, busy gates, and Esc/stop (both
+        driven by the caller around this call, same as any other turn)."""
+        import httpx
+
+        from webbee.steer import derive_session_id
+        from webbee.tools import LocalToolExecutor
+
+        session_id = await derive_session_id(
+            self.cfg, self.token_provider, self.workspace_root,
+            marathon=marathon, slot_id=self.slot_id)
+        executor = LocalToolExecutor(self.workspace_root, indexer=self._intel,
+                                     shadow=self._shadow)
+        async with httpx.AsyncClient(base_url=self.cfg.api_url, timeout=60) as client:
+            return await self._serve_stream(sink, client, session_id, start_id, task_id,
+                                            executor, marathon=marathon)
+
+    async def _serve_stream(self, sink, client, session_id: str, start_id: str,
+                            task_id: str, executor, *, marathon: bool = False) -> str:
+        """The frame-loop phase of ONE coding turn -- factored out of `run()`
+        (behavior-preserving split) so `attach()` (no start POST at all --
+        see there) can drive the identical loop against a session id it
+        derived itself instead of one a fresh POST just minted. Sets
+        `self.session_id`/`self._task_id` itself (rather than trusting the
+        caller already did) so every frame-loop reference to `self._task_id`
+        below (the C7 foreign-frame filter, the consent race) sees the
+        RIGHT turn's id regardless of which caller reached this point."""
+        self.session_id = session_id
+        self._task_id = task_id
+        self.steps = deque(maxlen=200)
+
+        seen: OrderedDict = OrderedDict()  # req_id -> already-posted result
+                                            # (at-least-once dedup); LRU-64 via
+                                            # _remember -- see _SEEN_KEEP.
+        # Slice-5 T9: id-sets shared across BOTH vocabularies for EXT tools
+        # (the kernel reuses the SAME tc["id"] as step_id there), so a step
+        # announced by one vocabulary is never re-announced by its
+        # dual-emitted twin. step_labels carries a v2 step_started's label
+        # forward to its later step_finished (which carries no app_id/tool
+        # of its own — facts-only). local_ids tracks LOCAL-tool v2 step_ids,
+        # which use a disjoint id space from the tool_request round trip
+        # (see webbee.frames' module docstring) and are a pure no-op.
+        started: set = set()
+        finished: set = set()
+        step_labels: dict = {}
+        local_ids: set = set()
+        from webbee.stream import stream_frames
+        _fr = getattr(self.token_provider, "force_refresh", None)
+        _rc = getattr(sink, "reconnecting", None)
+        stream = stream_frames(client, session_id, self._headers, start_id=start_id,
+                               force_refresh=_fr, on_retry=_rc)
+        # Liveness A: explicit __anext__ pulls (not `async for`) so a
+        # pending local consent can RACE the stream. Between iterations at
+        # most ONE of carry_task/carry_frame is set — a consent race hands
+        # ownership of its pulled-ahead pull back to this loop: a pending
+        # task when consent won, an already-pulled frame when the stream
+        # won. Everything else is byte-identical to the old async-for.
+        carry_task = None    # a still-pending __anext__ task (consent won)
+        carry_frame = None   # an already-pulled frame (the stream won)
+        _t0 = _monotonic()   # W1 Task 10: own-task watchdog start
+        _own_frames = False  # -> True the instant ANY non-foreign frame lands
+        try:
+            while True:
+                if carry_frame is not None:
+                    frame, carry_frame = carry_frame, None
+                else:
+                    if carry_task is not None:
+                        pull, carry_task = carry_task, None
+                    else:
+                        pull = asyncio.ensure_future(stream.__anext__())
+                    try:
+                        frame = await pull
+                    except StopAsyncIteration:
+                        break
+                ftype = frame.get("type")
+
+                # A frame from a DIFFERENT turn on the shared persistent stream
+                # (task_id absent on legacy kernels -> treated as own). C7 safety:
+                # foreign actionable frames are NEVER executed/consented and NEVER
+                # end this turn -- but instead of vanishing they (and any origin-
+                # stamped cross-surface display frame, e.g. a Telegram-steered
+                # turn's progress) now render ONE tagged, display-only line.
+                if _is_foreign_frame(frame, self._task_id):
+                    render_foreign_frame(frame, sink)
+                    if (not _own_frames and self._task_id
+                            and _monotonic() - _t0 > _FOREIGN_ONLY_DEADLINE_S):
+                        # Deduped-twin hang (W1 front-3b): the kernel ring
+                        # dropped this turn's task as a twin -- its task_id
+                        # will never appear on the stream. Foreign traffic
+                        # flowing + zero own frames is the signature; end
+                        # honestly instead of spinning "working" forever.
+                        _note = getattr(sink, "note", None)
+                        if _note is not None:
+                            _note("⚠ this message produced no work of its own "
+                                  "(likely a duplicate the kernel dropped) — done waiting")
+                        return ""
+                    continue
+                _own_frames = True
+
+                # Live steer topology: a Telegram/panel-steered turn keeps THIS
+                # client's task_id (the terminal stays the sole executor) with
+                # `origin` stamped -- tag the text renders below; everything
+                # else (execution, dedup, consent, accounting) is unchanged.
+                _tag = _origin_tag(frame)
+
+                if ftype == "tool_request":
+                    rid = frame.get("req_id")
+                    sid = str(rid or "")
+                    if rid in seen:
+                        seen.move_to_end(rid)
+                        out = seen[rid]
+                    else:
+                        # UI rendering is guarded so it can never block the
+                        # result POST below (an unposted result hangs the kernel
+                        # dispatch and freezes the dock).
+                        if _first_time(sid, started):
+                            try:
+                                sink.tool_start(_tag + frame.get("tool", ""), frame.get("args", {}))
+                            except Exception:
+                                pass
+                        out = await asyncio.to_thread(handle_tool_request, frame, executor)
+                        res = out["result"]
+                        if _first_time(sid, finished):
+                            try:
+                                sink.tool_result(_tag + frame.get("tool", ""), bool(res.get("ok")), _summary(res))
+                                self.steps.append({"step_id": sid,
+                                                   "label": frame.get("tool", ""),
+                                                   "ok": bool(res.get("ok"))})
+                            except Exception:
+                                pass
+                        _remember(seen, rid, out)
+                    await self._post_result(client, session_id, out)
+
+                elif ftype == "confirm_request":
+                    rid = frame.get("req_id")
+                    if rid in seen:
+                        seen.move_to_end(rid)
+                        await self._post_result(client, session_id, seen[rid])
+                    else:
+                        if self.mode == "plan":
+                            sink.plan_blocked(frame.get("tool", ""))
+                        # Liveness A: the local prompt races the stream so a
+                        # consent answered from ANOTHER surface (Telegram
+                        # relay) unfreezes this terminal instead of leaving
+                        # the dock stuck on `approve? y/n`.
+                        race = await race_consent(frame, sink, stream,
+                                                  mode=self.mode, task_id=self._task_id)
+                        carry_task, carry_frame = race.carry_task, race.carry_frame
+                        if race.stream_ended:
+                            break
+                        if race.out is not None:
+                            _remember(seen, rid, race.out)
+                            await self._post_result(client, session_id, race.out)
+
+                elif ftype == "panel_release_required":
+                    sink.panel_release(frame.get("panel_url", ""), frame.get("summary", ""))
+
+                elif ftype == "action":  # R2 — ext-tool call (server-side) surfaced in the feed
+                    handle_action_frame(frame, sink, started, finished, self.steps)
+
+                elif ftype == "step_started":  # v2 (Slice-5 T8 dual-emit)
+                    handle_step_started(frame, sink, started, step_labels, local_ids)
+
+                elif ftype == "step_finished":  # v2 (Slice-5 T8 dual-emit)
+                    handle_step_finished(frame, sink, finished, step_labels, self.steps, local_ids)
+
+                elif ftype == "thinking":  # system-driven reasoning -> the 💭 block
+                    _text = _progress_text(frame)
+                    (getattr(sink, "thinking", None) or sink.progress)(_tag + _text if _text else "")
+
+                elif ftype == "progress":  # P2 — dual-reads llm_text (v2) / text (legacy)
+                    _text = _progress_text(frame)
+                    sink.progress(_tag + _text if _text else "")
+
+                elif ftype == "usage":  # P2 — cumulative tokens + credits (Slice C; raw $ stays server-side)
+                    sink.usage(
+                        int(frame.get("tokens", 0) or 0),
+                        int(frame.get("credits", 0) or 0),
+                    )
+
+                elif ftype in ("task_queued", "task_dequeued"):
+                    # Full-queue-layer K1: a follow-up queued into the RUNNING
+                    # kernel session shows in the live queue panel the instant
+                    # it queues (tagged by origin) and leaves the panel when
+                    # the kernel drains it. These frames carry NO task_id
+                    # (they belong to the session, not a turn), so the C7
+                    # filter never eats them; getattr-guarded like todos/
+                    # queued_run — a minimal sink drops them, a render error
+                    # never breaks the loop. Terminal-origin rows render TOO
+                    # (mid-turn inject, 0.3.15): an injected line never sits
+                    # in the LOCAL panel — the kernel's echo is its only row,
+                    # and task_dequeued clears it when the turn absorbs it.
+                    _origin = str(frame.get("origin", "") or "")
+                    _hook = getattr(sink, "remote_queued" if ftype == "task_queued"
+                                    else "remote_dequeued", None)
+                    if _hook is not None:
+                        _iid = str(frame.get("steer_iid", "") or "")
+                        try:
+                            if ftype == "task_queued":
+                                _hook(_origin, str(frame.get("text", "") or ""), _iid)
+                            else:
+                                _hook(_origin, _iid)
+                        except Exception:
+                            pass
+
+                elif ftype == "marathon_complete":  # U4 — the whole GOAL is done: terminal
+                    return frame.get("text", "")
+
+                elif ftype in _MARATHON_FACT_TYPES:  # U4 — marathon plan/milestone/pause/todo
+                    # Facts-only. A `todo` fact carries the FULL list -> the
+                    # dedicated full-checklist render (falls back to the old
+                    # one-line note on a minimal sink). The other facts render
+                    # ONE human-readable line. Guarded: a minimal sink (no
+                    # `note`) simply drops the fact rather than crashing the
+                    # turn (the stream reader already tolerates unknown frame
+                    # types by ignoring them).
+                    if ftype == "todo":
+                        render_todo_frame(frame, sink)
+                    else:
+                        _note = getattr(sink, "note", None)
+                        if _note is not None:
+                            _note(marathon_note(frame))
+                    if ftype == "marathon_paused":
+                        # Parked (out-of-credits / consent / runaway) -> end the
+                        # turn so the dock leaves "working"; the note shows why, and
+                        # the run resumes on the user's next reply. The kernel's task
+                        # queue stays alive server-side while parked -- tell the sink
+                        # via marathon_parked() so end_turn keeps the queue-panel's
+                        # remote rows (tagged parked) instead of wiping them (W1
+                        # front-3b: an empty panel over a non-empty kernel queue lies).
+                        # getattr-guarded like every other sink hook here: a minimal
+                        # sink drops it, a crashing hook never breaks the turn.
+                        _parked = getattr(sink, "marathon_parked", None)
+                        if _parked is not None:
+                            try:
+                                _parked(str(frame.get("reason", "") or ""))
+                            except Exception:
+                                pass
+                        return ""
+
+                elif ftype == "final":
+                    # In a MARATHON a `final` is a PER-MILESTONE result, NOT the end
+                    # of the run -> keep streaming (the goal ends on marathon_complete
+                    # / marathon_paused, or a user-stop `final` with stopped=true). A
+                    # non-marathon coding turn's `final` is terminal (unchanged).
+                    if marathon and not frame.get("stopped"):
+                        continue
+                    _text = frame.get("text", "")
+                    return _tag + _text if _text else ""
+        finally:
+            # Never leak a pulled-ahead __anext__ past the loop (e.g. an
+            # exit while a consent race's carry is still pending).
+            if carry_task is not None:
+                await _retire(carry_task)
+
+        return ""
+
+    _result_delays = (0.5, 2.0, 5.0)   # class attr — tests override per instance
+
+    async def _post_result(self, client, session_id: str, out: dict) -> None:
+        # Never raises into the stream loop. W1: bounded transient retries —
+        # the kernel dedups by req_id, so a duplicate post is safe, and this
+        # cuts outage recovery from the kernel's tool-wait expiry (up to
+        # 65min for bash) to seconds. Final fallback unchanged: give up
+        # silently, kernel re-dispatches the same req_id.
+        for delay in (0.0,) + tuple(self._result_delays):
+            if delay:
+                await asyncio.sleep(delay)
+            try:
+                headers = await self._headers()
+                r = await client.post(f"/v1/agent/sessions/{session_id}/result",
+                                      json=out, headers=headers)
+                if r.status_code < 500:
+                    return
+            except Exception:
+                continue
+
+    async def stop(self) -> None:
+        """P5g: server-side stop for Esc/Ctrl-C. Posts a cancel for the
+        in-flight turn so the kernel tears down server-side (not just the
+        local asyncio task). Fail-soft — the local task.cancel() the dock
+        already does is what actually tears the UI down, so a network error
+        here must never raise into the key-binding handler."""
+        if not self.session_id:
+            return
+        try:
+            import httpx
+
+            headers = await self._headers()
+            async with httpx.AsyncClient(base_url=self.cfg.api_url, timeout=10) as client:
+                await client.post(
+                    f"/v1/agent/sessions/{self.session_id}/cancel", headers=headers,
+                )
+        except Exception:
+            pass