agentwire-dev 1.49.0__tar.gz → 1.50.1__tar.gz

{agentwire_dev-1.49.0 → agentwire_dev-1.50.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentwire-dev
-Version: 1.49.0
+Version: 1.50.1
 Summary: Multi-session voice web interface for AI coding agents
 Project-URL: Homepage, https://agentwire.dev
 Project-URL: Repository, https://github.com/dotdevdotdev/agentwire-dev
@@ -266,11 +266,11 @@ Description-Content-Type: text/markdown
 </p>
 
 <p align="center">
-  <strong>Self-hosted, voice-native control for Claude Code — your machine, your keys, no cloud account.</strong>
+  <strong>A self-hosted, keyboard-driven cockpit for running a whole <em>fleet</em> of Claude Code agents at once.</strong>
 </p>
 
 <p align="center">
-  Orchestrate many agents by voice — neural TTS on your CPU, voice cloning via a self-hosted GPU shim.
+  From a wall of tmux panes to one cockpit. Worktrees, command palette, scheduler, voice — every layer stacks, so you ship far more than you ever could hand-juggling terminals. Your machine, your keys, no telemetry.
 </p>
 
 <p align="center">
@@ -285,39 +285,51 @@ Description-Content-Type: text/markdown
 
 ## Why AgentWire
 
-Remote access to a coding agent is now table stakes — first-party and hosted tools
-all let you reach an agent from your phone. AgentWire competes on what those
-structurally can't match:
-
-- **Self-hosted, keys-local.** Runs entirely on your hardware. No cloud account, no
-  third party in the loop, no telemetry — your code and your API keys never leave
-  the machine.
-- **Voice-native, not voice-bolted-on.** Push-to-talk in, **neural TTS out on plain
-  CPU** (Kokoro, no GPU). Want your agents to talk back in a **cloned voice**? Run a
-  self-hosted GPU shim (Chatterbox / Zonos) — same machine, still your keys.
-- **Built to orchestrate _many_ agents, not watch one.** Council (multi-soul
-  deliberation), a recurring scheduler, briefing mode, and worktree-parallel
-  workers — tmux- and Claude-Code-native from the ground up.
-
-> **Use the official app to watch one session; use AgentWire to orchestrate many,
-> by voice, on your own hardware.**
+Running several Claude Code agents by hand means a wall of tmux panes: constant
+context-switching, copy-pasting between sessions, tracking which branch each agent
+sits on, babysitting every one. It doesn't scale past one or two before **your
+attention** — not the model — becomes the bottleneck. AgentWire turns that wall into
+one cockpit, and its capabilities **compound**:
+
+- **It compounds.** Worktrees + command palette + quick-action keys + tab-switching
+  + scheduler + voice each stack on the others to multiply how much one person can
+  run at once. That's the whole point — not any single feature.
+- **Orchestrate _many_ agents, not watch one.** Every session is a window; **F3**
+  fans them into a Mission-Control collage; **Tab** cycles between them; **Cmd/Ctrl+K**
+  runs any action. Council deliberates, briefing mode fans out, worktree sessions
+  isolate parallel work — tmux- and Claude-Code-native from the ground up.
+- **Autonomous and unattended.** A scheduler runs recurring jobs (research, test runs,
+  cleanup, doc-drift); gates verify *real work happened*; repo tasks open *draft PRs,
+  not commits to main*; reliability plumbing (verified delivery, dead-letter,
+  email-on-fail, usage-limit recovery) means nothing silently fails.
+- **Your machine, your keys, no telemetry.** Runs entirely on your hardware, loopback
+  by default. No cloud account, no third party in the loop — your code and your API
+  keys never leave the machine. 300+ damage-control blocks make running unattended sane.
+
+> **Use the official app to watch one session; use AgentWire to command a whole
+> fleet — on your own hardware, by keyboard, with voice layered on top.**
 
 ---
 
 ## What It Does
 
-Push-to-talk voice control for [Claude Code](https://github.com/anthropics/claude-code) or any AI coding assistant running in tmux.
+A self-hosted, desktop-style portal where **every agent session is a window** —
+reachable from any device on your LAN, running on your own hardware with your own keys.
 
 ```
-Phone → AgentWire Portal → tmux session → Claude Code
- 🎤        (WebSocket)         📺           🤖
+You → AgentWire Portal → tmux sessions → a fleet of Claude Code agents
+ ⌨️🎤      (WebSocket)         📺                    🤖🤖🤖
 ```
 
-**From your phone, tablet, or laptop on your network:**
-- Hold to speak, release to send
-- Watch agents work in real-time
-- Hear responses via TTS
-- Manage multiple projects simultaneously
+**From your laptop, phone, or tablet on your network:**
+- Every session is a draggable window; **F3** collage, **Tab** to cycle
+- Command palette (**Cmd/Ctrl+K**) runs any action; quick-action keys throughout
+- Branch a session into an isolated git worktree for parallel work
+- Put repeat work on a scheduler; hand hard calls to a council
+- Push-to-talk in, spoken summaries back — drive sessions by ear while off-tab
+
+**The outcome:** the whole fleet feels like one machine. You go from babysitting one
+session to commanding many.
 
 ---
 
@@ -332,7 +344,7 @@ agentwire init
 
 # Run
 agentwire portal start
-# Open http://127.0.0.1:8765 in Chrome — voice works immediately
+# Open http://127.0.0.1:8765 in Chrome — your cockpit is live (voice works immediately too)
 ```
 
 **Requirements:** Python 3.10+, tmux, ffmpeg, Claude Code
@@ -382,18 +394,20 @@ pip install agentwire-dev
 
 ## Features
 
+The cockpit, then the layers that stack on it:
+
 | Feature | Description |
 |---------|-------------|
-| **Voice Control** | Push-to-talk from any device on your network |
-| **Neural Voice + Cloning** | Kokoro neural TTS on CPU out of the box; voice cloning via a self-hosted GPU shim |
+| **Desktop UI** | Every session is a draggable window; **F3** collage grid, **Tab** to cycle, **Cmd/Ctrl+K** command palette |
+| **Multi-Session** | Run multiple agents on different projects simultaneously |
+| **Git Worktrees** | Same project, multiple branches, parallel agents — each in an isolated worktree session |
+| **Worker Orchestration** | Spawn worker panes, coordinate tasks, auto-reap when idle |
 | **Council** | Fan a prompt to multiple "soul" sessions (brain, conscience, critic…), synthesize with attribution |
-| **Scheduler** | Recurring autonomous tasks with gates, priorities, and usage-limit recovery |
+| **Scheduler** | Recurring autonomous tasks with gates, priorities, and usage-limit recovery — draft PRs, not commits to main |
 | **Briefing Mode** | A terse human-facing anchor fans out verbose worktree correspondents, then briefs you on cue |
-| **Multi-Session** | Run multiple agents on different projects simultaneously |
-| **Git Worktrees** | Same project, multiple branches, parallel agents |
+| **Voice** | Push-to-talk in, neural TTS out on CPU (Kokoro); voice cloning via a self-hosted GPU shim; global macOS hotkeys dictate to a background session from any app |
 | **Remote Machines** | SSH into GPU servers and talk to agents there |
-| **Worker Orchestration** | Spawn worker panes, coordinate tasks, voice commands |
-| **Safety Hooks** | 300+ dangerous commands blocked (rm -rf, force push, etc.) |
+| **Safety Hooks** | 300+ dangerous commands blocked (rm -rf, force push, etc.), all logged |
 | **Outbound Channels** | Email (Resend) + SMS (Quo / OpenPhone) for cross-device notifications |
 | **Session Roles** | Leader/worker patterns for multi-agent workflows |
 
@@ -407,13 +421,13 @@ agentwire new -s myproject -p ~/projects/myproject
 ```
 
 **2. Open the portal:**
-Visit `http://127.0.0.1:8765` in Chrome (or your phone/tablet with LAN access configured)
+Visit `http://127.0.0.1:8765` in Chrome (or your phone/tablet with LAN access configured). Each session is a window — open as many as you want and **Tab** between them, or **F3** for the collage grid.
 
-**3. Talk:**
-Hold the mic button, speak your request, release. In instant mode the transcript appears for a quick glance — Enter sends it to Claude Code.
+**3. Drive it by keyboard:**
+**Cmd/Ctrl+K** runs any action; quick-action keys send to the focused session; branch a session into a worktree for parallel work; put repeat jobs on the scheduler.
 
-**4. Listen:**
-Agent responses are spoken back — Kokoro neural voice out of the box, or any TTS model behind a custom shim.
+**4. Add voice on top:**
+Hold the mic to speak (in instant mode the transcript appears for a quick glance — Enter sends it), and agent responses are spoken back — Kokoro neural voice out of the box, so you can drive sessions and hear summaries by ear while your eyes are elsewhere.
 
 ---
 

{agentwire_dev-1.49.0 → agentwire_dev-1.50.1}/README.md

@@ -3,11 +3,11 @@
 </p>
 
 <p align="center">
-  <strong>Self-hosted, voice-native control for Claude Code — your machine, your keys, no cloud account.</strong>
+  <strong>A self-hosted, keyboard-driven cockpit for running a whole <em>fleet</em> of Claude Code agents at once.</strong>
 </p>
 
 <p align="center">
-  Orchestrate many agents by voice — neural TTS on your CPU, voice cloning via a self-hosted GPU shim.
+  From a wall of tmux panes to one cockpit. Worktrees, command palette, scheduler, voice — every layer stacks, so you ship far more than you ever could hand-juggling terminals. Your machine, your keys, no telemetry.
 </p>
 
 <p align="center">
@@ -22,39 +22,51 @@
 
 ## Why AgentWire
 
-Remote access to a coding agent is now table stakes — first-party and hosted tools
-all let you reach an agent from your phone. AgentWire competes on what those
-structurally can't match:
-
-- **Self-hosted, keys-local.** Runs entirely on your hardware. No cloud account, no
-  third party in the loop, no telemetry — your code and your API keys never leave
-  the machine.
-- **Voice-native, not voice-bolted-on.** Push-to-talk in, **neural TTS out on plain
-  CPU** (Kokoro, no GPU). Want your agents to talk back in a **cloned voice**? Run a
-  self-hosted GPU shim (Chatterbox / Zonos) — same machine, still your keys.
-- **Built to orchestrate _many_ agents, not watch one.** Council (multi-soul
-  deliberation), a recurring scheduler, briefing mode, and worktree-parallel
-  workers — tmux- and Claude-Code-native from the ground up.
-
-> **Use the official app to watch one session; use AgentWire to orchestrate many,
-> by voice, on your own hardware.**
+Running several Claude Code agents by hand means a wall of tmux panes: constant
+context-switching, copy-pasting between sessions, tracking which branch each agent
+sits on, babysitting every one. It doesn't scale past one or two before **your
+attention** — not the model — becomes the bottleneck. AgentWire turns that wall into
+one cockpit, and its capabilities **compound**:
+
+- **It compounds.** Worktrees + command palette + quick-action keys + tab-switching
+  + scheduler + voice each stack on the others to multiply how much one person can
+  run at once. That's the whole point — not any single feature.
+- **Orchestrate _many_ agents, not watch one.** Every session is a window; **F3**
+  fans them into a Mission-Control collage; **Tab** cycles between them; **Cmd/Ctrl+K**
+  runs any action. Council deliberates, briefing mode fans out, worktree sessions
+  isolate parallel work — tmux- and Claude-Code-native from the ground up.
+- **Autonomous and unattended.** A scheduler runs recurring jobs (research, test runs,
+  cleanup, doc-drift); gates verify *real work happened*; repo tasks open *draft PRs,
+  not commits to main*; reliability plumbing (verified delivery, dead-letter,
+  email-on-fail, usage-limit recovery) means nothing silently fails.
+- **Your machine, your keys, no telemetry.** Runs entirely on your hardware, loopback
+  by default. No cloud account, no third party in the loop — your code and your API
+  keys never leave the machine. 300+ damage-control blocks make running unattended sane.
+
+> **Use the official app to watch one session; use AgentWire to command a whole
+> fleet — on your own hardware, by keyboard, with voice layered on top.**
 
 ---
 
 ## What It Does
 
-Push-to-talk voice control for [Claude Code](https://github.com/anthropics/claude-code) or any AI coding assistant running in tmux.
+A self-hosted, desktop-style portal where **every agent session is a window** —
+reachable from any device on your LAN, running on your own hardware with your own keys.
 
 ```
-Phone → AgentWire Portal → tmux session → Claude Code
- 🎤        (WebSocket)         📺           🤖
+You → AgentWire Portal → tmux sessions → a fleet of Claude Code agents
+ ⌨️🎤      (WebSocket)         📺                    🤖🤖🤖
 ```
 
-**From your phone, tablet, or laptop on your network:**
-- Hold to speak, release to send
-- Watch agents work in real-time
-- Hear responses via TTS
-- Manage multiple projects simultaneously
+**From your laptop, phone, or tablet on your network:**
+- Every session is a draggable window; **F3** collage, **Tab** to cycle
+- Command palette (**Cmd/Ctrl+K**) runs any action; quick-action keys throughout
+- Branch a session into an isolated git worktree for parallel work
+- Put repeat work on a scheduler; hand hard calls to a council
+- Push-to-talk in, spoken summaries back — drive sessions by ear while off-tab
+
+**The outcome:** the whole fleet feels like one machine. You go from babysitting one
+session to commanding many.
 
 ---
 
@@ -69,7 +81,7 @@ agentwire init
 
 # Run
 agentwire portal start
-# Open http://127.0.0.1:8765 in Chrome — voice works immediately
+# Open http://127.0.0.1:8765 in Chrome — your cockpit is live (voice works immediately too)
 ```
 
 **Requirements:** Python 3.10+, tmux, ffmpeg, Claude Code
@@ -119,18 +131,20 @@ pip install agentwire-dev
 
 ## Features
 
+The cockpit, then the layers that stack on it:
+
 | Feature | Description |
 |---------|-------------|
-| **Voice Control** | Push-to-talk from any device on your network |
-| **Neural Voice + Cloning** | Kokoro neural TTS on CPU out of the box; voice cloning via a self-hosted GPU shim |
+| **Desktop UI** | Every session is a draggable window; **F3** collage grid, **Tab** to cycle, **Cmd/Ctrl+K** command palette |
+| **Multi-Session** | Run multiple agents on different projects simultaneously |
+| **Git Worktrees** | Same project, multiple branches, parallel agents — each in an isolated worktree session |
+| **Worker Orchestration** | Spawn worker panes, coordinate tasks, auto-reap when idle |
 | **Council** | Fan a prompt to multiple "soul" sessions (brain, conscience, critic…), synthesize with attribution |
-| **Scheduler** | Recurring autonomous tasks with gates, priorities, and usage-limit recovery |
+| **Scheduler** | Recurring autonomous tasks with gates, priorities, and usage-limit recovery — draft PRs, not commits to main |
 | **Briefing Mode** | A terse human-facing anchor fans out verbose worktree correspondents, then briefs you on cue |
-| **Multi-Session** | Run multiple agents on different projects simultaneously |
-| **Git Worktrees** | Same project, multiple branches, parallel agents |
+| **Voice** | Push-to-talk in, neural TTS out on CPU (Kokoro); voice cloning via a self-hosted GPU shim; global macOS hotkeys dictate to a background session from any app |
 | **Remote Machines** | SSH into GPU servers and talk to agents there |
-| **Worker Orchestration** | Spawn worker panes, coordinate tasks, voice commands |
-| **Safety Hooks** | 300+ dangerous commands blocked (rm -rf, force push, etc.) |
+| **Safety Hooks** | 300+ dangerous commands blocked (rm -rf, force push, etc.), all logged |
 | **Outbound Channels** | Email (Resend) + SMS (Quo / OpenPhone) for cross-device notifications |
 | **Session Roles** | Leader/worker patterns for multi-agent workflows |
 
@@ -144,13 +158,13 @@ agentwire new -s myproject -p ~/projects/myproject
 ```
 
 **2. Open the portal:**
-Visit `http://127.0.0.1:8765` in Chrome (or your phone/tablet with LAN access configured)
+Visit `http://127.0.0.1:8765` in Chrome (or your phone/tablet with LAN access configured). Each session is a window — open as many as you want and **Tab** between them, or **F3** for the collage grid.
 
-**3. Talk:**
-Hold the mic button, speak your request, release. In instant mode the transcript appears for a quick glance — Enter sends it to Claude Code.
+**3. Drive it by keyboard:**
+**Cmd/Ctrl+K** runs any action; quick-action keys send to the focused session; branch a session into a worktree for parallel work; put repeat jobs on the scheduler.
 
-**4. Listen:**
-Agent responses are spoken back — Kokoro neural voice out of the box, or any TTS model behind a custom shim.
+**4. Add voice on top:**
+Hold the mic to speak (in instant mode the transcript appears for a quick glance — Enter sends it), and agent responses are spoken back — Kokoro neural voice out of the box, so you can drive sessions and hear summaries by ear while your eyes are elsewhere.
 
 ---
 

{agentwire_dev-1.49.0 → agentwire_dev-1.50.1}/agentwire/__init__.py

@@ -1,3 +1,3 @@
 """AgentWire - Multi-session voice web interface for AI coding agents."""
 
-__version__ = "1.49.0"
+__version__ = "1.50.1"

{agentwire_dev-1.49.0 → agentwire_dev-1.50.1}/agentwire/hooks/idle-handler.sh

@@ -370,10 +370,17 @@ complete | incomplete | error
           echo "[$(date -Iseconds)] TASK-ORPHAN: killing tmux session" >> "$dlog"
           tmux kill-session -t "$tmux_session" 2>/dev/null &
         ) &
-      elif [[ -n "$parent_session" ]]; then
-        # Orchestrator pane with parent: notify parent
-        message="${session_name} is idle"
-        $AGENTWIRE notify-parent -q --to "$parent_session" "$message" 2>/dev/null &
+      else
+        # Pane-0 session went idle: tell its parent. Resolution lives inside
+        # notify-parent (prompt_router.resolve_parent) so all three precedence
+        # sources are honored — worker pane → pane 0, creator recorded at
+        # `agentwire new` time (session metadata), then `.agentwire.yml parent:`.
+        # The old code passed only the .agentwire.yml value and skipped the call
+        # entirely when it was empty, so worktree / `agentwire new` children
+        # (whose parent lives in metadata, not config) never notified anyone.
+        # TMUX_PANE is inherited, so the CLI knows which session this is; a
+        # top-level session resolves to no parent and this is a harmless no-op.
+        $AGENTWIRE notify-parent --on-idle -q "is idle and done working" 2>/dev/null &
       fi
     fi
   fi

{agentwire_dev-1.49.0 → agentwire_dev-1.50.1}/agentwire/notify_cli.py

@@ -51,11 +51,30 @@ def cmd_notify_parent(args) -> int:
     current_session = pane_manager.get_current_session()
     current_pane = pane_manager.get_current_pane_index()
 
-    # If no explicit target, try parent from config
-    if not target_session:
-        parent = get_parent_from_config()
-        if parent:
-            target_session = parent
+    # --on-idle: the idle hook is reporting that a pane-0 session went idle.
+    # Infrastructure services (portal, tts/stt, scheduler, the idle-nag bridge,
+    # custom services) cycle active→idle constantly and are NOT delegated work,
+    # so they must never fire a "child finished" ping at their parent. A
+    # worktree / `agentwire new` child is not a service and is unaffected.
+    if getattr(args, 'on_idle', False) and current_session:
+        from agentwire import services
+
+        if services.is_service_session(current_session):
+            return _output_result(True, json_mode, "", skipped="service-session")
+
+    # If no explicit target, resolve the parent through the SAME precedence the
+    # prompt router uses (worker pane → pane 0; else creator recorded at
+    # `agentwire new` time; else `.agentwire.yml parent:`). The old path looked
+    # only at `.agentwire.yml`, so a worktree/`agentwire new` child — whose
+    # parent lives in session metadata, not config — resolved to nothing and its
+    # idle notification silently dropped (the parent never heard the child go
+    # idle). resolve_parent reads the creator metadata, closing that gap.
+    if not target_session and current_session is not None and current_pane is not None:
+        from agentwire import prompt_router
+
+        resolved = prompt_router.resolve_parent(current_session, current_pane)
+        if resolved:
+            target_session = resolved[0]
 
     # Build notification message (--raw sends verbatim — queued messages
     # already carry their own [WORKER SUMMARY ...] / [PROMPT ...] headers)
@@ -255,6 +274,9 @@ def register_notify_parser(subparsers) -> None:
     notify_cmd_parser.add_argument("-q", "--quiet", action="store_true", help="Suppress output")
     notify_cmd_parser.add_argument("--raw", action="store_true",
                                    help="Send the message verbatim (no [NOTIFY from ...] prefix)")
+    notify_cmd_parser.add_argument("--on-idle", dest="on_idle", action="store_true",
+                                   help="Idle-hook mode: suppress the notify if the current session "
+                                        "is an infrastructure service (they cycle idle constantly)")
     notify_cmd_parser.add_argument("--json", action="store_true", help="Output as JSON")
     notify_cmd_parser.set_defaults(func=cmd_notify_parent)
 

{agentwire_dev-1.49.0 → agentwire_dev-1.50.1}/agentwire/send_cli.py

@@ -186,13 +186,17 @@ def cmd_send(args) -> int:
 
     if verify:
         # Same delivery verification as --wait-ready, but without waiting for a
-        # fresh-boot banner — for an already-running session, just confirm the
-        # paste landed and report verified vs unconfirmed. retries=0: send ONCE
-        # and report — a re-paste on a false miss (a busy session that scrolled
-        # the echo past the capture window) would deliver the message twice.
+        # fresh-boot banner — for an already-running session, confirm the paste
+        # landed AND submitted. retries=1: a whole-send retry only fires when the
+        # paste never landed in the box (phase 1 of _deliver_once) — i.e. it
+        # genuinely vanished, so re-pasting can't double-deliver. Once text has
+        # landed, _deliver_once re-presses Enter in place (never re-pastes), so
+        # the old double-delivery worry the retries=0 choice guarded against
+        # can't happen anymore. Staying patient here is what keeps a laggy host
+        # from surfacing a false failure the parent has to clean up by hand.
         from agentwire.session_ready import send_verified
 
-        ok = send_verified(session, prompt, retries=0)
+        ok = send_verified(session, prompt, retries=1)
         if json_mode:
             print(json.dumps({"success": True, "session": session_full, "machine": None,
                               "verified": ok,

{agentwire_dev-1.49.0 → agentwire_dev-1.50.1}/agentwire/services.py

@@ -78,6 +78,42 @@ def notifications_session_name() -> str:
     )
 
 
+# Built-in infrastructure sessions — the Python mirror of the frontend's
+# SERVICE_SESSIONS allowlist (static/js/service-classification.js). These are
+# long-lived services, not delegated work, so they must never fire a
+# "child finished" idle notification at a parent (the idle-nag bridge alone
+# cycles active→idle ~1440×/day). Kept in lockstep with the JS set.
+_BUILTIN_SERVICE_SESSIONS = frozenset({
+    "agentwire-portal",
+    "agentwire-tts",
+    "agentwire-stt",
+    "agentwire-kokoro",
+    "agentwire-scheduler",
+})
+
+
+def is_service_session(name: str) -> bool:
+    """Is *name* an infrastructure service session (vs. a working session)?
+
+    True for the built-in infra sessions, the configurable idle-nag bridge, and
+    any user-defined custom service. Used to suppress idle parent-notifications
+    for services — a worktree / `agentwire new` child is NOT a service and still
+    notifies. Name-only and dependency-light so the idle hook can call it cheaply
+    on every idle event.
+    """
+    bare = (name or "").split("@")[0]
+    if not bare:
+        return False
+    if bare in _BUILTIN_SERVICE_SESSIONS or bare == notifications_session_name():
+        return True
+    try:
+        from .config import load_config
+
+        return any(s.name == bare for s in load_config().services.custom)
+    except Exception:
+        return False
+
+
 def _source_dir() -> str:
     """agentwire source dir (dev.source_dir), project for built-in services."""
     try:

{agentwire_dev-1.49.0 → agentwire_dev-1.50.1}/agentwire/session_ready.py

@@ -25,11 +25,21 @@ VERIFY_SCROLLBACK_LINES = 200
 # Enter to dismiss its ``[Pasted text]`` banner before submitting).
 POLL_INTERVAL = 0.15
 # Max wait for a paste to appear in the input box before giving up on this try.
-LAND_TIMEOUT = 5.0
-# Max wait, after an Enter, for the box to clear (the keystroke to register).
-SUBMIT_TIMEOUT = 3.0
-# Bounded Enter re-presses before declaring a hard failure.
-MAX_ENTER_ATTEMPTS = 4
+# Generous because under host load tmux ``capture-pane`` itself lags and a large
+# paste renders slowly — a tight budget here is the #1 cause of "the text landed
+# but delivery reported failure" on a bogged-down machine.
+LAND_TIMEOUT = 8.0
+# Max wait, after a single Enter, for the box to clear (the keystroke to
+# register). Per-press, not the whole submit phase — see SUBMIT_BUDGET.
+SUBMIT_TIMEOUT = 4.0
+# Wall-clock ceiling on the whole press-Enter-and-confirm phase. Re-pressing is
+# driven by this deadline rather than a fixed count so a laggy box is waited out
+# (each idle Enter on an already-submitted/empty box is a harmless no-op), but a
+# genuinely wedged session still fails in bounded time instead of hanging.
+SUBMIT_BUDGET = 20.0
+# Floor on re-presses regardless of how fast the budget burns (slow snapshots
+# can eat the wall-clock before we've pressed Enter enough times).
+MIN_ENTER_ATTEMPTS = 4
 
 # Substrings that mean "Claude is actively working" — a spinner footer, the
 # token counter, the esc-to-interrupt hint, or tool-output glyphs. A
@@ -259,19 +269,27 @@ def _deliver_once(
     if not _poll(landed_or_done, LAND_TIMEOUT):
         return False
 
-    # Phase 2 — press Enter and confirm the box cleared. Bounded re-press: a
-    # single Enter can be swallowed under load, and a large paste needs a second
-    # Enter (dismiss the ``[Pasted text]`` banner, then submit).
-    for _ in range(MAX_ENTER_ATTEMPTS):
+    # Phase 2 — press Enter and confirm the box cleared. Re-press is driven by a
+    # wall-clock budget, not a fixed count: a single Enter can be swallowed under
+    # load, a large paste needs a second Enter (dismiss the ``[Pasted text]``
+    # banner, then submit), and on a bogged-down host the box renders slowly. We
+    # keep pressing until the box clears or SUBMIT_BUDGET elapses — an idle Enter
+    # on an already-submitted/empty box is a harmless no-op, so over-pressing is
+    # safe, while a tight count would give up before a laggy box ever caught up.
+    deadline = time.time() + SUBMIT_BUDGET
+    attempts = 0
+    while True:
         if submitted(_snapshot(session, pane_index), message, marker):
             return True
         press_enter(session, pane_index=pane_index)
+        attempts += 1
         if _poll(
             lambda: submitted(_snapshot(session, pane_index), message, marker),
             SUBMIT_TIMEOUT,
         ):
             return True
-    return False
+        if attempts >= MIN_ENTER_ATTEMPTS and time.time() >= deadline:
+            return False
 
 
 def send_verified(

{agentwire_dev-1.49.0 → agentwire_dev-1.50.1}/agentwire/static/announcements.json

@@ -1,5 +1,23 @@
 {
   "announcements": [
+    {
+      "id": "2026-06-cockpit-positioning",
+      "placement": "modal",
+      "emoji": "🛩️",
+      "title": "Stop babysitting one agent. Command a fleet.",
+      "date": "2026-06-29",
+      "body": "Here's how to think about AgentWire: it's a self-hosted, keyboard-driven cockpit for running a whole fleet of Claude Code agents at once — not a voice toy with sessions bolted on. The capabilities compound. Every session is a window (F3 collage, Tab to cycle); the command palette (Cmd/Ctrl+K) runs any action; worktree sessions isolate parallel work; the scheduler runs recurring jobs unattended as draft PRs. Voice is a layer that stacks on all of it — push-to-talk in, spoken summaries by ear while you're off-tab, and global macOS hotkeys to dictate to a background session from any app. One person commands many agents instead of babysitting one.",
+      "highlights": [
+        "From a wall of tmux panes to one cockpit — the pieces compound",
+        "Every session is a window: F3 collage, Tab to cycle, Cmd/Ctrl+K for any action",
+        "Worktrees + scheduler + council run parallel and unattended work",
+        "Voice is a compounding layer, not the headline"
+      ],
+      "cta": {
+        "label": "Press F1 for the full tour",
+        "url": "https://github.com/dotdevdotdev/agentwire-dev"
+      }
+    },
     {
       "id": "2026-06-help-modal",
       "placement": "modal",

{agentwire_dev-1.49.0 → agentwire_dev-1.50.1}/agentwire/templates/base.html

@@ -6,6 +6,9 @@
     <meta http-equiv="Cache-Control" content="no-cache, no-store, must-revalidate">
     <meta http-equiv="Pragma" content="no-cache">
     <meta http-equiv="Expires" content="0">
+    <meta name="description" content="A self-hosted, keyboard-driven cockpit for running a whole fleet of Claude Code agents at once. Your machine, your keys, no telemetry.">
+    <meta property="og:title" content="AgentWire">
+    <meta property="og:description" content="A self-hosted, keyboard-driven cockpit for running a whole fleet of Claude Code agents at once. Worktrees, command palette, scheduler, voice — every layer stacks.">
     <link rel="icon" type="image/png" href="/static/favicon.png">
     <link rel="manifest" href="/manifest.webmanifest">
     <link rel="apple-touch-icon" href="/static/img/icon-192.png">

{agentwire_dev-1.49.0 → agentwire_dev-1.50.1}/docs/wiki/sessions/prompt-routing.md

@@ -37,6 +37,18 @@ Depth-1 and local-machine only. Remote (`@machine`) parents are out of scope:
 each machine's own watchdog sweeps its panes; cross-machine delivery falls
 back to human-only.
 
+**Idle notifications use the same resolution.** When a pane-0 session goes idle,
+`idle-handler.sh` calls `agentwire notify-parent --on-idle`, which resolves the
+parent through the precedence above — so a worktree / `agentwire new` child whose
+parent lives in **creator metadata** (not `.agentwire.yml`) now correctly pings
+its spawner on completion. Earlier the idle hook read only the `.agentwire.yml`
+`parent:` field and silently dropped the notification when it was empty.
+`--on-idle` additionally suppresses the ping when the source is an infrastructure
+**service** (`services.is_service_session` — portal/tts/stt/kokoro/scheduler, the
+idle-nag bridge, custom services); those cycle active→idle constantly and aren't
+delegated work. Worker-pane (`notify-parent` from pane > 0) and explicit `--to`
+callers skip the service check.
+
 ## Delivery safety
 
 Every delivery goes through `safe_deliver()` (also used by