The target-extraction regex stopped at ':' and '/', so a dialog
containing a full URL (e.g. 'http://verify.muse-dev.online/api/...')
was captured as just the scheme ('http'). Add an optional
'(?:https?://)?' prefix so the hostname is captured instead.
This is defense-in-depth: the 2026-10-11 bare-scheme dialogs render
'http' directly from the app (closed source), but if the app resumes
rendering full URLs, the extractor will now yield the hostname rather
than the scheme.
Verified against 5 cases: bare 'http' unchanged, full http/https URLs
now yield hostnames, IPs and single-label hosts unchanged.
Adds bin/verify-node-ssh.sh: checks reverse-tunnel listeners and SSH
auth for each fleet node port from the VM. Distinguishes dark nodes
(no listener) from auth failures (authorized_keys perms/keys).
Verification 2026-10-09:
- muse/2225, 646/2226, pip/2227, muse-main/2224, opm/2228: LISTEN
- def/2229, dev/2230: DARK (no reverse tunnel)
- All listening nodes reject VM super key (expected: nodes authorize
per-operator/id_frontdoor keys, not the VM super key)
Fixes#211
Read-only curses TUI (box-fleet-tui structure): per-device SSH scan of
tmux panes + process table, joined locally, runs sorted by agent type
across tailnet devices. Open harness taxonomy (muse/agy/known plus
other:<bin>), multi-socket enumeration, --once/--json dump mode.
- Add dedicated watchers/ project folder with box-stability-watcher.py supervisor
- Monitor host load, memory, swap saturation, and crash-looping services
- Implement tiered mitigations: yellow renicing, orange SIGSTOP pause with 60s grace, red shedding
- Distinguish user-launched agents (allowed on desktop default socket) from automated box workloads
- Wire first-class box stability CLI subcommand and top-line host status in fleet status
- Harden tmux.service with cgroup memory limits to prevent OS freeze and OOM avalanches
- Add 10-test unit test suite covering thresholds, safety whitelist, pause/resume, and isolation
- bin/tmux_auto_approver.py: multi-socket worker discovery across user and netns sockets
- Regex matching engine with 7 terminal prompt rules and hard security guardrails
- bin/box-onboard-tui.py: dedicated 4-tab curses TUI for fleet connects, tmux workers, rules, and audit logs
- Audit logging stream in logs/tmux/auto-approvals.jsonl and state in .state/
- Unit test suites covering engine, rules, guardrails, and curses rendering
- Support invite code discovery in main chat and redemption in settings menu
- Add Settings RPA primitives with Radix UI mouse dispatch and retry polling for async DOM
- Attribute 'has_redeemed' binary from the Additional tokens ticker / entrypoint visibility
- Unblock agent @646 by repairing warp-def/dev tunnels and redeeming REDCJ7 via dev (+1B tokens)
- Implement 'box onboard' pipeline to provision infra, authenticate, and auto-redeem queued codes
- Add 'box onboard feed-matrix' ranking all agents by work done over time, job count, and role
- Register 'InputType.SALVAGE' in loop modulation (CRITICAL priority, 600s timeout, 3 nudges to opm)
- Ingest recurring balance audits into canonical HEARTBEAT.md and TOOLS.md
- bin/muse-tui.py:
* Parse raw SGR 1006 (\033[<btn;x;yM/m) and Xterm mouse escape sequences in _handle_escape_sequence fallback.
* Multi-trigger context menus: Button 3, Button 2, Ctrl/Shift/Alt+Click, double-click, click on active item, or click [⚡] / [sid] target.
* Render permanent [⚡] action target across all sidebar thread rows.
* Separate focus partitions for FLEET AGENTS and SIDECHATS with partition-specific wheel scrolling and keyboard navigation (j/k, Enter, h/l).
* Space key support in NORMAL mode to open context menus.
* Active pane highlighting and updated footer hints.
- bin/super-cli.py:
* Add 'box tui' command dispatching directly to muse-tui.py --mode box.
- tests:
* Add unit tests in test_context_menus.py, test_focus_highlight.py, and test_main_nav.py (51/51 passing).
Close the def/dev supervision gap at the source: every node brought
up gets watched, and every supervisor enumerates the registry.
- bin/ensure-node-supervision.sh (new, idempotent): appends the
NODES.md row (netvm-names port, honors CDP_PORT_OVERRIDE so it
never fights provision's picker) and installs/enables
chromebox-watchdog-<node>.timer. --all heals drift (registry +
/etc/netvm identities). Template verified byte-identical to the
installed def unit.
- netvm-node-up.sh: calls ensure (non-fatal) at the end. Provision
and the onboarding pipeline reach it transitively.
- relay-health-check.sh, cdp-latency-check.sh: registry-driven
watched_nodes() + LIB_ONLY guards (were hardcoded 4 nodes).
- tests/test_node_supervision.py (6): row add/idempotent/override,
timer render, node-up wiring, both watched_nodes().
- CHROMEBOX-RUNBOOK.md: setup-fed supervision section.
Pairs with the registry-driven relay/chromebox watchdogs: new rows
are picked up on the next run with no per-node code edits.
box fleet status / approvals check misreported every node as STOPPED /
CDP-unreachable from sandboxed shells (own PID+net namespaces: pgrep
blind, no route to 10.201.x.x, no sudo). Fleet was healthy throughout.
- bin/host_evidence.py (new): host watchdog evidence fallback. Recent
timer runs (journal -o json, exact UNIT match) with no newer failure
line in cdp-relay-watchdog.log / chromebox-watchdog.log (both
silent-when-healthy) prove a node is up. def/dev have no watchdog
coverage: browser verdict via chromebox-<node>.log freshness
(alive-only), CDP verdict unknown.
- super-cli.py: effective status/source/evidence per node. Host
evidence decides ONLY the fully-blind pattern (both local probes
negative); live local signals always win. New UNKNOWN badge, [*]
footnote; approvals UNREACHABLE splits into BLIND / OFFLINE(host
agrees) / unreachable-evidence-inconclusive, with honest footer.
proc_alive/cdp_ok keep local-probe meaning; status/source/evidence
are new JSON fields.
- approvals.py: host_cdp_ok flag on the unreachable path.
- agent-health.sh: restart circuit breaker. 3 consecutive futile
restarts (restart leaves agent still failing) opens the circuit:
no more kills for 1800s, ALERT to log+journal, half-open probe
after cooldown, reset on any success. Stops the def murder loop
(57 restarts / 155 API FAILs for an account-layer failure).
- tests/test_fleet_status.py (25), tests/test_agent_health.py (6).
- CHROMEBOX-RUNBOOK.md: blind-shell status + futile-restart sections.
Tests: 98/98 focused green (agent_health + fleet_status +
completion + tool_calls). Live-verified: 4 ACTIVE [*] + 2 UNKNOWN.
echo "$list" | grep -q under set -o pipefail exits 141 whenever grep
matches before echo finishes writing, so healthy browsers were reported
'CDP up but no page target' and killed every 2 min fleet-wide (load 19+).
Use [[ == *glob* ]] (no pipe, no race) for the page/muse.ai stages, and
grep -c (reads to EOF, never early-exits) for the netns check.
The 2026-10-05 fix note claimed dev/def were removed from the pool but the
code still listed [dev, def, muse]. The harvester's every-minute systemd
dispatch raced the pool daemon claiming slots as dev/def, which refuse on
attribution grounds (dev FAILs fast, def freezes until the 60-min reaper) -
the fleet-wide dev-FAIL-fast + def-frozen pattern on every swarm.
Pool now matches swarm_worker/daemon.py WORKER_POOL exactly: muse only,
the single fully authenticated auxiliary worker. Reporting/harvest paths
untouched.
- daemon: route non-shell slot tasks to a fresh subagent session on dev/def/muse
via muse_hybrid; add bin/ to sys.path so muse_hybrid/prompt_envelope import
under the tmux supervisor
- prompt_envelope: add wrap_subagent_task() - plain task assignment without
[TOOL tmux/swarm/cron] meta tags (subagents refused those as relayed test traffic)
- box-ctl/reporter: swarm-attach accepts optional session-id, stored as
slot.subagent_session_id
- executor: _looks_like_shell requires an existing executable (prose like
'verify ...' no longer misread as shell)
- poller: pick up pending swarms as well as running
- response-harvester: monitor running slot subagent sessions from swarms.json,
allow '/' in RESULT/VERB job ids, archive ephemeral threads on any verdict
(OK or FAIL), reap slot sessions for terminal swarms
Follow-up to 8c39dc3: four scripts hardcoded dev's CDP port instead of
reading netvm-registry.py. Aligned with the corrected registry (9455).
Session: sidechat/dev-cdp-partition
NODES.md advertised dev's CDP on 9460, but the fleet-facing path is the
relay netvm-cdp-relay.py 10.201.36.2:9455 -> 127.0.0.1:9455, and
netvm-names.sh never pinned dev (hash default = 9455). The watchdog kept
relaunching dev's browser on 9460 while the relay sat orphaned on 9455.
- NODES.md: dev cdp_port 9460 -> 9455 (registry is the watchdog's source)
- bin/netvm-names.sh: pin def=9450, dev=9455 (was hash-default only)
Reported by 646 as xop-e09 partition; verified hop-by-hop before fix.
Session: sidechat/dev-cdp-partition
opm
operator