* perf(aggregator): async-sweep — 243 blocking route handlers async def→def (#738)
Mounted in the aggregator's single event loop, an 'async def' route handler
that runs blocking code (subprocess/journalctl/openssl/argon2) freezes the
WHOLE loop -> aggregator.sock Connection refused -> 502 board-wide.
Deterministic AST codemod (scripts/async-sweep.py) converts route handlers
that (a) are decorated with an HTTP verb, (b) contain a known blocking call,
(c) have NO await/async-with/async-for/yield, (d) are never used as a
coroutine elsewhere -> plain 'def'. Starlette then runs them in the AnyIO
threadpool, so the blocking call no longer stalls the gateway. await/stream/
websocket handlers are left untouched. Every file py_compile-checked.
243 handlers across 56 modules (system 17, qos 13, netdiag/hexo 12, hub 11...).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* perf(aggregator): raise AnyIO threadpool to 80 tokens (#738)
The async-sweep moves ~243 blocking handlers to the threadpool. With ~110
modules in one process, the default 40-token pool can queue head-of-line under
concurrent blocking load. Raise to 80 on startup (best-effort, never breaks
boot).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* perf(hub): double-buffer status cache + emojized health page (#738)
Navbar status (menu + health-batch) is now a strict double-buffer cache:
- request handlers NEVER compute on the request path — they return the current
snapshot instantly (or a 'warming' placeholder), so the sidebar's polling can
no longer serialize behind a ~3s systemctl walk and starve the loop;
- the background refresher is kicked from the request path (_ensure_bg) because
mounted sub-apps receive neither startup nor @app.middleware events under the
aggregator — the previous lazy-start middleware never fired there;
- snapshots are built complete then swapped atomically, so the dashboard never
shows partial/bad counts.
Served by the dedicated secubox-hub process (:8001, isolated loop) the navbar
stays <50ms and holds 200 under 25+ concurrent polls where the aggregator-
mounted copy wedged (000). health.js: 🟢🟡🔴 emoji status indicators.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* style(hub): render health status emoji cleanly (neutralize .led dot) (#738)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* perf(sidebar): kill per-module health storm, use batch endpoint (#738)
The navbar refreshed LEDs by firing ONE /api/v1/<module>/health request per
module — ~119 requests every 30s, in batches of 8 — straight at the aggregator's
single shared event loop. Combined with the in-process module mount this is a
prime driver of the recurring board-wide 502 wedge (user-identified).
checkAllHealth + refreshStaleHealth now call /api/v1/hub/public/health-batch
ONCE (served by the dedicated, double-buffered hub process) and populate every
module's LED from that single response. 119 reqs/cycle -> 1.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: CyberMind-FR <gandalf@Gk2.net>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
build-image.sh / build-installer-iso.sh / validate-staged-repo.sh fetched the
ASCII-armored secubox-keyring.gpg and used it directly as
`signed-by=/usr/share/keyrings/secubox.gpg`. apt's signed-by requires a
DEARMORED (binary) keyring, so the chroot apt-get update failed with
"NO_PUBKEY 44E50F0178E8BC7E / repository not signed" even though the repo
signature and key are correct. Pipe the key through `gpg --dearmor`.
Root cause of all image-build failures since the keyring became armored;
not a key desync (repo signer == published keyring == 44E50F0178E8BC7E).
ingest_site() runs git init/commit inside /srv/metablogizer/sites/<name> over
ssh as root, so a freshly created .git is owned root:root — but the
metablogizer service runs as 'secubox' and cannot write the repo (blocked
sub-E #113 webhook deploys). Add a fix_perms helper that chown -R
secubox:secubox the site dir after every .git-touching ingest path
(ingested-fresh x2 + ingested-with-history). Matches module postinst pattern.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Systemic clobber: the scaffold boilerplate (install -d -m 750 /var/lib/secubox,
/run/secubox) put restrictive modes on SHARED parents in ~56 module postinsts,
reverting them to 0750 on every install/upgrade and breaking traversal for
non-secubox daemons (kbin/toolbox 500). Empirically confirmed install -d -m only
modes the final component, so /parent/leaf forms are harmless — only bare-parent
targets were rewritten. Multi-arg lines (incl. ones making /var/lib world-writable
1777) split per-parent: /run/secubox=1777 root:root, /var/lib|cache|etc=0755
secubox:secubox; module-private leaves keep 0750. Scaffold + PATTERNS.md fixed so
new packages don't reintroduce it.
build-packages.sh omitted -d, so dpkg-checkbuilddeps silently dropped
arch:all packages (incl. secubox-core) from the repo. All 144 packages now
build + are published to apt.secubox.in, signed with apt@secubox.in.
After the wg-admin tunnel is confirmed working, 'harden' closes the
public SSH brute-force surface:
- sshd drop-in: PasswordAuthentication no + PermitRootLogin
prohibit-password (refuses if no root authorized_keys — no lockout).
- nft SSH-guard: drop tcp/22 from sources outside the LAN (192.168.1.0/24)
+ 10.x tunnels, inserted BEFORE the blanket 'iif eth1 accept' so the
router's port-forwarded public SSH (real public src IP, DNAT no-SNAT)
is dropped while LAN + wg-admin stay reachable. Persisted in
/etc/nftables.conf + applied live without flushing other tables.
Applied + verified on gk2: tunnel SSH (10.98.0.1) key-only works; public
admin.gk2.secubox.in:22 now times out. Tables (blacklist/wg) intact.
The while-read in write_server_conf reused the global pip/ppub names, so
after it ran the caller's $pip (the new peer IP for the client config)
was wiped to empty — the emitted client config showed 'Address = /32'.
The provisioned server state was always correct; only the printed config
was wrong. Loop now uses local _pn/_pip/_pub.
Idempotent provisioner for a dedicated out-of-band admin tunnel,
distinct from wg-toolbox:
- wg-admin UDP 51821, server 10.98.0.1/24, peers 10.98.0.2+/32.
- generates server+peer keys once (0600, never in repo), writes
/etc/wireguard/wg-admin.conf from a peers state file.
- installs /etc/nftables.d/secubox-admin-wg.nft (allow udp/51821,
persistent) — ADDITIVE, does not touch sshd or existing rules.
- 'add <name>' mints a client .conf + QR (split-tunnel, endpoint
admin.gk2.secubox.in:51821).
sshd hardening (key-only + restrict 22) is a deliberate separate step.
The PACKAGES array was hardcoded to 30 entries, last updated months
ago. As of master at a28984c1 there are 140 secubox-* package
directories with debian/control; the script silently skipped the 110
that weren't in the list (operators got 0 build output for those).
Surfaced today during the consolidation-deploy: the 9 PRs merged into
master touched 31 packages, but the build script only built 13
because the other 18 weren't in the hardcoded list. The 18 had to be
built manually with direct dpkg-buildpackage calls.
Replace the static array with a glob over packages/secubox-*/ that
keeps only directories with debian/control. Build order: secubox-core
first (universal Depends:), metapackages secubox-full / secubox-lite
last, everything else alphabetical between them. The exact order
doesn't matter for dpkg-buildpackage (each builds independently); the
sort keeps the build log readable and surfaces core/meta failures
early/late respectively.
Verified: bash scripts/build-packages.sh bookworm all --dry-run now
plans 140 builds (vs 30 before) with the expected first/last order.
Adds scripts/audit-packages.py to scan packages/secubox-*/ and emit
structured artifacts under out/ (gitignored): packages.json,
packages.dot dependency graph, empty-packages.txt, prefix-based
clusters, and a hand-curated fuzzy-cluster report with per-member
evidence (deps, routes, services).
Findings doc 2026-05-27-secubox-consolidation-audit.md synthesises the
data into Tier 0-4 merge proposals grounded in real package contents,
not the plan stub's illustrative guesses. Key corrections vs the stub:
- 141 real packages (not "100+"); secubox-core is the substrate for 131
- streamlit-idle is already an in-package service variant, not a
separate package — that consolidation is half-done
- meta{blogizer,bolizer,catalog,oblizer} are real distinct services,
not typo proliferation (recommend rename for clarity, not merge)
- Two genuine packaging bugs surfaced: secubox-c3box binary name
collision between Python dashboard and Go daemon (issue #378),
secubox-smart-strip has no debian/ directory at all
Also imports two pre-existing 2026-05-26 plan stubs (consolidation,
build-scripts-refactor) that had been sitting untracked.
Re-run after any packaging change: python3 scripts/audit-packages.py
The IS31FL319X chip on the MOCHAbin's i2c-1 bus errors out with -EIO
above brightness ~50. Pushing 100 or 255 saturates the bus, the chip
starts NACK'ing the next register writes, the LED driver logs
'Setting an LED's brightness failed (-5)' on every retry, and the
LEDs freeze at whatever value last got through.
Observed live 2026-05-24 on gk2: all three RGBs stuck plain green at
218/255/255 with the bash bumper and the python heartbeat both pushing
high values into a saturated chip. After dropping to brightness 10
and adding a 30 ms i2c-write delay in the python set_rgb() loop,
writes succeed and the LEDs reflect the actual health state instead
of a static color.
Files:
- usr/sbin/secubox-led BRIGHTNESS 100 -> 10
- usr/sbin/secubox-healthbump ACTIVE 100 -> 30, SLEEP 20 -> 5
- usr/sbin/secubox-led-pulse BRIGHTNESS 100 -> 10
- usr/sbin/secubox-led-heartbeat python set_rgb scales any 0-255
caller down to 0-10 in write_text, then sleeps 30 ms before the
next channel — matches the bash scripts' I2C_DELAY and lets the
mv64xxx_i2c controller drain between calls.
- scripts/secubox-healthbump (the duplicate script kept in repo root
for ad-hoc dev use) gets the same BRIGHTNESS=10 cap.
The misleading 'BRIGHTNESS=100 works reliably' comment that the code
contradicted is replaced with an incident note pointing at the EIO
ceiling.
* docs(mochabin): add EP06 mPCIe GPIO investigation runbook (ref #345)
scripts/probe-mpcie-gpios.sh: empirical CP0 GPIO sweep that drives
each unrequested line HIGH for a few seconds and watches dmesg/lsusb
for new USB device. Skips [used] lines, uses libgpiod gpioset
--mode=time so the line auto-restores to input on release. Modes:
--baseline (snapshot only, safe), --line gpiochipN K (single line),
no arg (full sweep gpiochip1+gpiochip2).
docs/hardware/mochabin-mpcie-ep06-runbook.md: hypothesis, procedure,
and DTS rfkill-gpio template patterned after cn9132-clearfog.dts
lines 69-122. Apply once a candidate GPIO is identified.
--baseline smoke-tested on gk2 — sweep deferred until spare EP06
arrives (no mPCIe device in J5 = no signal regardless of which line
drives W_DISABLE#).
* docs(mochabin): drop EP06 spare assumption from runbook (ref #345)
No spare EP06 on order; the existing EP06-E will be used for the
sweep when re-seated in J5. Replaces "deferred until spare" framing
with a multimeter-scoping fallback for the inconclusive case.
* fix(mpcie): v0.2.0 probe-mpcie-gpios.sh — drop blanket sweep, danger-list, dry-run default (ref #345)
The v0.1.0 "drive every unused-input CP0 GPIO HIGH for 3s one-at-a-time"
sweep took gk2 hard-down within seconds (host became unreachable; ARP
stopped responding). gpioinfo's [used] tag only reflects lines the
kernel *requested* — several unrequested CP0 GPIOs are physically
wired to critical board functions (eth switch reset, PCIe2 PERST#,
pca9554 IRQ) and driving them HIGH wedged the board.
v0.2.0:
- No blanket-sweep mode. Removed.
- Default mode = dry-run candidate listing only. No GPIO is driven.
- --commit --line CHIP N required to actually drive a single line.
- Hardcoded DANGER_LINES (gpiochip1:1, gpiochip1:9, gpiochip1:27)
skipped unconditionally, even with --commit.
- gpiochip2 entirely held off without --allow-chip2 (no per-line
schematic mapping for that bank).
- Hard refusals are vocal (REFUSED: ... reason) before exit.
Runbook updated to walk through the new dry-run → --commit flow.
HISTORY documents the incident + the v0.2.0 mitigations.
Lesson saved in memory feedback_gpio_blanket_sweep_crashed_board.md.
---------
Co-authored-by: CyberMind-FR <gandalf@Gk2.net>
Symptom: pre-login banner on ttyS0 (Minicom 2.9 VT102) rendered as
trailing `�` after every emoji line. The console driver couldn't
decode the UTF-8 multi-byte sequences for ⚡🔐🌐📡 and emitted the
replacement character instead.
The `█` box-block characters (U+2588) DO render correctly in VT102+
UTF-8 — they're a single-byte-class glyph in most modern fonts. We
keep those so the SECUBOX wordmart still looks like SecuBox.
Replacements (compatible with VT102 minicom AND SSH UTF-8 alike):
⚡ -> >>
🔐 -> [user]
🌐 -> [web]
📡 -> [ssh]
The /etc/motd block (shown after SSH login) is left untouched — SSH
sessions have proper UTF-8 terminals so the rich emojis are fine
there.
Also add scripts/build-kernel-local.sh (new): the local kernel build
script that I authored today for #255. Documented; runs the same
defconfig + merge_config + scripts/config + olddefconfig + sed ZRAM
override + bindeb-pkg flow as .github/workflows/build-kernel.yml,
without the 20-min GitHub Actions ping-pong.
Co-authored-by: CyberMind-FR <gandalf@Gk2.net>
Anti-regression for the double-cache pattern documented in CLAUDE.md
(« Performance Patterns — Double Caching »). Catches the bug class that
caused PR #146 (secubox-waf 17 % sustained CPU under SOC polling)
before it lands.
What's added
============
- scripts/check-dashboard-cache.py — AST-based audit. Walks every
packages/secubox-*/api/main.py, flags hot dashboard routes
(/stats, /alerts, /summary, /dashboard, /metrics, /overview,
/events, /history, /recent, /bans, /sessions, /connections)
whose handler body does per-request I/O (open, read_text,
read_bytes, subprocess.run, Popen) AND doesn't go through a cache
(module-level _cache instance + .get() call, OR
asyncio.create_task(refresh_*) at startup).
* --check : exit 1 on unjustified violations (CI mode)
* --report : human-readable table (default)
* --json : machine-readable
- .cache-lint-allowlist.toml — TOML allowlist for justified
exceptions. Each entry needs a daemon + route + justification
(issue link expected). Seeded with the 4 current legacy cases:
secubox-waf (covered by PR #146 — will drop the entries when
merged), secubox-config-advisor /history (admin drill-in, not
polled), secubox-netdiag /connections (real-time TCP state),
secubox-tor /summary (small file read, follow-up cleanup).
- .github/workflows/dashboard-cache-check.yml — mirror of
license-check.yml. Runs the lint + the self-tests on PRs that
touch packages/**/api/main.py, the lint script, the allowlist,
or the workflow itself.
- docs/CACHE-PATTERN.md — short remediation guide. Three accepted
shapes (read-through, background-refresh, pure in-memory), code
example, allowlist syntax, references to the canonical
implementations (secubox-crowdsec for read-through, secubox-system
for background-refresh).
Tests (14/14 pass)
==================
- StatsCache pattern detection (compliant cases)
- Non-compliant detection: open, subprocess.run, multiple I/O ops
- Pure in-memory handler = compliant
- Background-refresh task = compliant
- Non-hot route (/health) not audited
- @router.get treated like @app.get
- Allowlist suppresses by exact (daemon, route) key
- Allowlist doesn't cross-pollinate across routes
- Missing allowlist file = no error
- CLI --check exits 1 on unjustified
- CLI --check exits 0 when clean
- CLI --json output well-formed
- **Real repo audit with seeded allowlist passes --check**
(sanity gate: catches regression if anyone removes an allowlist
entry without fixing the daemon)
Baseline against the real packages/ tree
========================================
122 daemons audited, 118 compliant, 4 with findings (all
allowlisted). When PR #146 merges, the secubox-waf entries can be
removed from the allowlist — the next lint run will then prove the
fix is wired correctly.
Follow-up
=========
- Drop secubox-waf allowlist entries when #146 merges.
- Per-daemon follow-up issues for the remaining 3 (config-advisor,
netdiag, tor) — not blocking, allowlist documents the rationale.
- Consider extending HOT_ROUTES to /health / /status once the small
daemons that call `systemctl is-active` adopt a 10 s cache.
Co-authored-by: CyberMind-FR <gandalf@Gk2.net>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The mail Phase 1 LXC went live at 10.100.0.10 (see #136, PR #141,
35ba4c2c). The 'dead container fix' was still rewriting routes that
target 10.100.0.10 to 10.100.0.1:9080, which would have clobbered
mail.gk2 / webmail.gk2 / rspamd.gk2 / autoconfig.gk2 routes. The
production gk2 board already had this correction in its local copy
(diffed 2026-05-17); bringing the repo back in sync so deploying
the script doesn't regress.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The previous regex `host_\K[a-z0-9_]+(?= )` required a trailing SPACE
after the hostname. HAProxy `use_backend mitmproxy_inspector if
host_<dotted_domain>` lines end at the hostname with no trailing space,
so the regex matched ZERO domains. Sync silently never auto-populated
mitmproxy routes for any haproxyctl-added vhost.
Caught 2026-05-17 when ckwa.gk2.secubox.in returned 502: the vhost was
added via `haproxyctl vhost add ckwa.gk2.secubox.in mitmproxy_inspector
ssl` (HAProxy ACL created correctly), but sync didn't write the
mitmproxy route entry, so the request landed at mitmproxy's default
(127.0.0.1:9080 placeholder) which doesn't exist inside the mitmproxy
LXC.
Fix: change lookahead to `(?=\s|$)` to also match end-of-line. Verified:
$ printf ' use_backend ... if host_x_y_z\n' | \
grep -oP 'host_\K[a-z0-9_]+(?=\s|$)' | sed 's/_/./g'
x.y.z
Live patch applied on gk2 (mitmproxy LXC) to restore ckwa to 200.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- metablog-webhook-install.sh: pages through GITEA_URL/users/<owner>/repos,
filters to metablog-*, checks existing hooks for the target URL,
POSTs a new hook with content_type=json + secret + events=[push].
Idempotent.
- metablog-webhook-uninstall.sh: reverse, DELETE on matching hook IDs.
- Both support --dry-run.
Exit non-zero if any operation failed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Same bug pattern as 34a4760e but on the source-helpers line.
Both empty-array restores now use "${var[@]}" without :-.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
set -- "${saved_args[@]:-}" produces a single empty positional
argument when the array is empty, which the case-statement then
sees as Unknown flag "". Use "${saved_args[@]}" (no :-) which
gives zero positional args correctly on empty.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
After excluding dot-dirs, __pycache__ still appeared as an app
candidate. Add explicit !-name '__pycache__' to the find filter.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
/srv/streamlit/apps/.claude exists (Claude tracking artifacts:
HISTORY.md, TODO.md, WIP.md) and was being picked up as an app
candidate. Filter out dot-prefixed directory names so only real
Streamlit app directories are considered.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
gitea CLI in 1.22 lacks 'user keys add', so enrolment goes:
generate-access-token (--config app.ini) -> POST /api/v1/user/keys via
LXC loopback -> delete token via sqlite3 (token list/delete API requires
basic auth in 1.22, not token auth).
Key finding: Gitea builtin SSH server (START_SSH_SERVER=true) rejects
the conventional "git@" username — the connecting username must match
the OS user Gitea runs as ("gitea" here). GITEA_SSH_USER var controls
this, defaulting to "gitea".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
URL uses gitea@ (Gitea built-in SSH server validates against the
OS user 'gitea', not 'git'). Mirrors scripts/lib/metablog-ingest-site.sh
from sub-project B (PR #97) with the streamlit- prefix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
URL uses gitea@ (Gitea built-in SSH server validates against the
OS user it runs as, which is 'gitea' in this LXC, not 'git').
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
gitea CLI in 1.22 lacks 'user keys add', so enrolment goes:
generate-access-token (--config app.ini) -> POST /api/v1/user/keys via
LXC loopback -> delete token via sqlite3 (token list/delete API requires
basic auth in 1.22, not token auth).
Key finding: Gitea builtin SSH server (START_SSH_SERVER=true) rejects
the conventional "git@" username — the connecting username must match
the OS user Gitea runs as ("gitea" here). GITEA_SSH_USER var controls
this, defaulting to "gitea".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The section-header rule flushed missing keys but did not set
saw_push/saw_branch to 1, so the END block then appended a second
[repository] block at EOF. Set the flags inline with the flush.
Also cleaned the live MOCHAbin app.ini which had been doubly
patched by the original buggy version.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- app.ini at /var/lib/gitea/custom/conf/app.ini (not /etc/gitea/)
- awk-based INI patcher (python3 not installed in gitea LXC)
- ENABLE_PUSH_CREATE_USER=true, DEFAULT_BRANCH=main applied
- Gitea restarted and confirmed active
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two bug fixes to the CMSD-1.0 license-header tool:
1. detect_existing() previously matched the SPDX token anywhere in the
first 10 lines, including inside docstrings and prose comments.
Tightened the matcher to require comment markers (#, //, *, <!--)
and whitespace before the SPDX token.
Regression tests:
- test_detect_existing_no_false_match_in_docstring
- test_detect_existing_no_false_match_inline_comment_prose
2. _read_enrollment() now returns ["**"] when the allowlist file is
absent, per spec §5.2 "missing file → repo-wide enforcement".
Previously it returned [] (nothing enforced), making Phase C closure
impossible without an additional file.
Empty allowlist file still returns [] (Phase A initial), so
test_main_empty_allowlist_passes_check is unaffected.
New tests:
- test_read_enrollment_missing_file_means_repo_wide
- test_main_check_missing_allowlist_enforces_repo_wide
Suite: 49 → 53 passing. --check still exits 0 repo-wide.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A merge from feature/83 reverted the fix from commit 22014865, which had
already replaced `haproxyctl regen` (non-existent sub-command, printed
help text and exited 0) with the correct `haproxyctl generate`.
This commit re-applies the same fix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Discovered during Phase B pilot: running `--fix packages/secubox-hub`
yielded zero files because walk() computed `rel` relative to the walk
root (the subdir) while allowlist patterns are repo-relative.
Add optional `repo_root` parameter to walk(); when present, allowlist
matching uses paths relative to it instead of the walk root.
Backwards-compatible (defaults to None, preserving existing tests).
main() passes the discovered repo_root through.
Regression test added: test_walk_subdir_with_repo_root_arg.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
After Session 156 fix, secubox-haproxy.service was safely restarted but
cpf.gk2.secubox.in (and other streamlit instances) immediately fell back
to "Wrong Domain" because sync-mitmproxy-routes.sh was clobbering their
container routes.
Two compounding bugs:
1. DEAD_CONTAINER_IPS included 10.100.0.50 — but that's the streamlit
LXC's actual IP (RUNNING). fix_dead_container_routes() rewrote every
route pointing to 10.100.0.50 (cpf:8523, etc.) to 10.100.0.1:9080.
2. The update conditional reapplied routes when the existing port didn't
match the computed port. The computed port logic only knows metablog
vs webui — it has no streamlit awareness, so it always tried to force
streamlit domains to port 9080 (webui default).
Fixes:
- Drop 10.100.0.50 from DEAD_CONTAINER_IPS. The streamlit LXC at that
address is alive; routes pointing to it are valid.
- Change "update if existing missing OR port mismatch" to "update only
if existing is missing." The script now only ADDS routes for ACL/
metablog domains; it never overwrites existing routes set by other
sources (sync-all-routes.sh, manual edits, etc.). fix_dead_container_
routes() still handles truly dead IPs separately.
Verified: cpf.gk2.secubox.in → HTTP 200 "Streamlit" persists across
sync-mitmproxy-routes.service runs. arm/admin still 200 with correct
content.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Triggered by cpf.gk2.secubox.in returning "Wrong Domain". Root cause: the
secubox-haproxy generators (both bash haproxyctl and Python FastAPI) emit
config rules that point to dead/wrong backends, leading to HTTPS-wide
metablog/streamlit outages when a regen runs.
Generator fixes:
- packages/secubox-haproxy/api/main.py: use_backend waf_inspector
→ use_backend mitmproxy_inspector (2 occurrences). waf_inspector
pointed to 127.0.0.1:8890 which is not listening; mitmproxy_inspector
is the actual healthy WAF backend at 10.100.0.60:8080.
- packages/secubox-haproxy/sbin/haproxyctl: default_backend fallback
→ default_backend mitmproxy_inspector (2 occurrences). The fallback
backend deny_status 503's unknown hosts; switching to mitmproxy_inspector
lets mitmproxy dispatch all hosts via its routes JSON (245 routes vs
93 declared in haproxy.toml).
New safe regen wrapper:
- scripts/secubox-haproxy-regen-safe: snapshot → regen via haproxyctl →
validate via haproxy -c -f → atomic swap → reload. On validation
failure, current cfg is preserved and the broken candidate is saved
for forensics. Prevents another silent breakage like Session 156.
Board-side state (not in repo): /etc/haproxy/haproxy.cfg patched in
place to mirror these fixes, secubox-haproxy.service stopped until
user confirms safe to re-enable. All 245 routes verified live (cpf,
arm, lldh, admin, pub, werdl, 3d, 42, zkp all HTTP 200 with correct
content).
See .claude/HISTORY.md Session 156 for full symptom chain and the
6-layer root cause analysis.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Addresses code review:
- Important #4: finish now fetches issue title via `gh issue view --json title`
and uses it as PR title (was using the branch name). Falls back to branch
name if the fetch fails or returns empty.
- Minor #8: explicit `return 4` on `git push` failure for consistency with
the documented exit-code contract.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Also patches gh-mock to sanitize dashes in branch-derived env var names,
and fixes --head two-arg parsing in the pr view mock handler.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
