Generated via multi-agent workflow (3 planners → adversarial review → synthesis).
Covers frp tunnel on shared :443, mTLS device-cert auth, and client-cert
implementation for iOS/Android/desktop. Trust Model A (single-owner fleet) for v1.
The message listener was attached only when the mux consumer called onMessage — which
happens inside attach(), AFTER the async mTLS-registry lookup. On loopback the agent's FIRST
heartbeat ping arrived in that gap and ws dropped it (no listener), so the agent's heartbeat
died on the first miss at 15s and the tunnel flapped forever. Attach the listener at
construction and buffer early frames (and an early close) until the consumer wires up.
Node/OpenSSL rejects an agent client leaf whose only trust anchor is a non-self-signed
intermediate (no PARTIAL_CHAIN flag) → every agent was reset at the TLS layer before attach(),
surfacing as a silent 1006. Pass the full agent-ca bundle and stop swallowing tlsClientError.
- ca/issue.ts: real X.509 v3 Ed25519 leaf with SPIFFE URI SAN, signed by the
intermediate; SAN built via relay-auth spiffeIdFor so it can't drift
- ca/csr.ts: parse real PKCS#10 (agent's PEM) + async PoP verify; decodeCsrWire
accepts PEM or base64(DER)
- main.ts/env.ts: real issuer when CA_INTERMEDIATE_KEY_PATH present, else dev fallback
- interop.test.ts: oracle proving relay-auth verifyAgentCert accepts the leaf (6 tests)
- deps: @peculiar/x509, reflect-metadata
relay-run imported bare specifiers (control-plane/relay-auth/relay-contracts/relay-e2e/
term-relay + ws/pg/ioredis) with an empty package.json, relying on a hand-populated
node_modules from local dev. Declare them (file: siblings + runtime deps + tsx) so
'npm install' resolves relay-run on a clean host.
viewport-fit=cover spans the physical screen edges; add --safe-t/--safe-b from
env(safe-area-inset-*) threaded through #tabbar/#term/#keybar/#approvalbar/#searchbox/
#settingspanel. Falls back to 0px (zero regression on non-safe-area platforms).
New brand icon (Orbit — a node circling a block-cursor core, amber-gold on
warm-dark, matching the desktop theme):
- iOS: Assets.xcassets/AppIcon.appiconset (opaque 1024, was no icon)
- desktop: resources/icon.png (rounded 1024 → .icns at package time)
- desktop tray: trayTemplate.png/@2x — a proper macOS TEMPLATE (black shapes on
transparent, system-tinted) replacing the old white blob
- web: public/icon.svg (vector favicon/PWA)
Master SVGs rendered with rsvg-convert.
The projects panel listed bare parent folders (~, ~/Documents) as projects and
a single session lit up every ancestor project's 'Active now'. dropParentFolders
filters non-git entries that merely contain other listed projects; assignSessions
attributes each live session to the DEEPEST containing project (detail page keeps
prefix matching). +3 tests; projects 29/29, tsc clean. Frontend unchanged.
Amber-gold accent is designed for the desktop's warm-dark background; on light
mode gold text/badges wash out. Match the desktop (dark theme by default):
.preferredColorScheme(.dark) at the root. DirtyBadge becomes a soft filled
capsule (amber on amber-18%) instead of amber-on-quaternary, readable in any mode.
ios/README: add a comprehensive Features section (P0 pocket cockpit, P1 walk-away
complete, iPad adaptive split, design system, verification status); note min iOS
17, both-idiom test commands, P1 server touch-points.
root README: add a Clients section covering the iOS/iPad app, desktop Electron
shell, and the pre-production relay stack (with Tailscale as the remote-access
path today).
Match the desktop/web theme (public/style.css): accent #E3A64A gold (dark) /
#C9892F (light) replacing the indigo, +onAccent #1A1305 dark ink for gold-fill
buttons, +accentSoft. Status colors realigned to the web warm palette
(#46D07F/#F5B14C/#FF6B6B). Terminal canvas fixed to the desktop warm-dark
(#100F0D bg / #ECE9E3 fg) with gold caret/selection. DSButtonStyle primary now
uses onAccent (dark) instead of white for contrast on gold. All via the single
accent token + DS palette — no scattered edits.
Verified: iPhone 16 290 + iPad Pro 11 290 tests green (DesignSystemTests assert
the exact gold + web-status RGB in both schemes).
Opens WebTerm to iPhone + iPad: TARGETED_DEVICE_FAMILY 1→1,2 (project + all
targets), iPad gets all four orientations (~ipad key), ios.yml gains an
ipad-tests leg (iPad Pro 11 sim). Built plist verified UIDeviceFamily [1,2];
app launches natively full-screen on iPad Pro 11 sim (no letterbox). Adaptive
layout (NavigationSplitView) lands in T-iPad-2. Plan: docs/PLAN_IOS_IPAD.md.
T-iOS-18 (F-iOS-1..13 walkthrough) + T-iOS-19 (security audit vs built artifacts): both
PASS_WITH_FINDINGS, 0 CRITICAL/HIGH; real-device items DEFERRED with manual checklists.
Fixes (1 MED + 3 LOW, all closed):
- WebTermUITests: the single scripted happy path (manual-entry pair → list → attach →
KeyBar ^L → injected held gate via POST /hook/permission → tap Approve → server-side
behavior==allow assertion). Green twice locally (fresh + already-paired branches).
- ios.yml: ui-test leg (server boot + TEST_RUNNER_ env as PROCESS env) + iOS17-floor leg
(loud-skip if runtime absent, hard-fail if present and red); dropped CODE_SIGNING_ALLOWED=NO
from app-tests leg (unsigned hosts get -34018 from the real keychain — measured)
- URLSessionHTTPTransport defaults to .ephemeral (no disk cache of preview terminal bytes)
- TARGETED_DEVICE_FAMILY moved to target level (built plist now UIDeviceFamily [1] only)
Final: 306 automated checks green (214 pkg + 76 app + 10 integration + XCUITest); server src/
untouched; P0 complete (19/19 tasks)
project.yml gains a bundle.unit-test target hosted by the app, wired to all
5 local packages (TestSupport test-doubles included) so T-iOS-11..14 VM tests
run via xcodebuild test. Coverage gate still counts only the 4 packages (§9).
T-iOS-1: ios/ XcodeGen project (iOS 17, Swift 6 strict concurrency, ATS per PLAN §5.2), 5 SPM package shells, CI skeleton
T-iOS-2: Origin spike vs real server — URLSessionWebSocketTask custom Origin CONFIRMED (no Starscream); 16MiB replay + EMSGSIZE(40) errno correction written back to plan
T-iOS-3: WireProtocol frozen contract, 59 tests, 100% line coverage, cross-impl vectors vs src/protocol.ts via tsx
T-iOS-4: FakeTransport/FakeClock/FakeHTTPTransport doubles
Verify: independent agent re-ran all acceptance — 6/6 PASS
New relay-run/ package boots the REAL term-relay relay-node between a browser WSS
listener and an agent mTLS listener, delegating every authz verdict to the real
relay-auth onUpgrade (Origin/CSWSH + capability verify + DPoP PoP + single-use jti),
with the real P4 E2E crypto. No audited package modified.
- P1 (done): in-process integration test round-trips a sealed payload both ways
through the real createRelayNode splice + real createMuxSession; INV2 asserted
(relay sees only ciphertext); negative controls (foreign Origin / missing DPoP -> 401).
- P2 (boots): main.ts brings up self-signed TLS + 3 listeners + in-memory control-plane.
- P3 (not landed): no agent dial / node-pty / relay-web serve yet.
- 4 tests green, tsc clean. node_modules via symlinks (gitignored).
Two real seam drifts surfaced (documented in PLAN_RELAY_RUN_PHASE0, not hacked):
(1) term-relay authz-port DpopContext shape vs relay-auth — reconciled (server-derived htu/htm);
(2) control-plane CapabilityVerifier.verify is sync but relay-auth verify is async — blocks
HTTP account/pairing seeding until a 1-line async widen.
File-level plan for a single-host `npm start` relay: new relay-run/ package wiring
the real P4/P5 + in-memory control-plane + self-signed TLS + agent + relay-web.
Seam files to read, 9-step build order, verify steps. Executable spec for a fresh
build session; reuses the e2e/harness wiring.
Grounded assessment that the relay service is pre-production (every cross-package
seam is a fail-closed stub, no server/TLS/DB/orchestration). Documents the target
topology + trust boundaries, the 7 wiring gaps to close (P1 socket server, P5
injection into P1/P3, pg/redis stores, real KMS CA custody, agent/relay-web build,
F2 principal + F6 replay-epoch residuals), a phased build order (dev-run → staging
→ prod), an accurate per-service env reference, infra checklist, and an ops runbook.
Two-round multi-agent security audit of the rendezvous-relay E2E/auth core.
Fixes 6 confirmed findings (F1–F6) with regression tests, adds a cross-package
e2e adversarial security harness (dynamic re-validation), and the audit report.
534 tests green across the affected packages; tsc clean. See docs/REVIEW_RELAY_SECURITY.md.
REVIEW_RELAY_SECURITY.md: two-round multi-agent audit — F1–F6 with fix plan/
resolution, refuted-but-recorded items, and the e2e harness coverage table.
PROGRESS_LOG.md: cross-session record of the audit + fixes + harness.
New e2e/ package wires the REAL P5(auth)+P4(crypto)+P2(agent) exports through
in-memory seams and an untrusted-relay attacker vantage (RelaySpy). Dynamically
re-validates F1–F4/F6 plus MITM/reflection/replay/reorder/INV2/cross-tenant/
single-use — every assertion runs the production code path. 21 tests green, tsc
clean. node_modules via symlinks (gitignored); no npm install needed.
The recoverable replay key was stable per (hostContentSecret, sessionId) while the
agent-side sealer resets its deterministic-nonce seq to 0 on every restart/re-attach
→ two sealer generations sealed distinct plaintext under the SAME (key, nonce).
Add a required 'epoch' to ReplayKeyParams, fold it into the K_content HKDF salt
(sessionId U+001F epoch), mint a fresh epoch per createReplaySealer generation and
expose it, and thread it through ReplaySource so the browser re-derives the matching
key. Fresh epoch per generation ⇒ fresh key ⇒ seq=0 can never collide; recoverability
within a generation is preserved.
Touches relay-contracts/relay-e2e/agent/relay-web. Green: contracts 81, e2e 78,
agent 133, web 99; tsc clean. Regression proves same seq-0 nonce, different key.
F1 (HIGH): bind step-up freshness to the required method (stepUpMethod) so a
login-time or weaker (TOTP) factor can no longer satisfy a passkey requirement.
F2 (HIGH): make the step-up gate host-driven and fail-closed on connect+reattach
(deny when policy.required and the principal is missing/stale/wrong-method).
F3 (MED): bind the WebAuthn assertion to the stored credential (credentialId
cross-check + forward publicKey/credentialId to the verifier).
F4 (LOW): make verifyDpopProof totally fail-safe (no unhandled throw), wrap the
authz boundary into a clean audited 401, and validate cnfJkt format at issue.
F5 (LOW): return newSignCount from finishAuthentication so the caller can persist
the advanced counter (WebAuthn clone detection).
All fixes ship regression tests. relay-auth: 122 tests green, tsc clean.
Add a `desktop/` Electron app that embeds the existing Node server + node-pty
(all-in-one): the window loads http://127.0.0.1:<port>/ and reuses the frontend
unchanged; other LAN devices can still connect. The server needs zero changes —
startServer(cfg)/loadConfig already support programmatic embedding.
- Pure, unit-tested modules: port, shell, server-config, deep-link, notify-policy,
notifications, live-poll, prefs, settings-store (94 tests; desktop/src ~97% cov)
- Electron glue: main/window/tray/menu/preload/embedded-server/logger
(hardened: contextIsolation, sandbox, deny foreign-origin navigation)
- Native value: OS notifications driven by /live-sessions status, tray, deep links
- Packaging (electron-builder -> arm64 .dmg): ships dist/ + public/ + node_modules
on-disk under Resources so the server resolves its deps from /Applications;
node-pty rebuilt for the Electron ABI
- Docs: docs/DESKTOP_PLAN.md; PROGRESS_LOG updated
Verified: desktop tsc clean; 1401 tests pass; coverage >=80% (desktop/src 97/95/100/97);
.dmg built and launch-tested on arm64 (server boots, UI serves 200, node-pty loads).
Multi-tenant reverse-tunnel service ("ngrok for Claude Code" with E2E): a
host-agent dials OUT to an operator-run relay; external devices reach the host
THROUGH the relay, routed by per-tenant subdomain, forwarding ciphertext only
(the relay never sees plaintext). Lets a customer reach their own self-hosted
web-terminal from anywhere with zero networking setup.
Packages — all tsc-strict + vitest green (656 tests), cross-package integration verified:
- relay-contracts: frozen shared contracts (mux frame codec, data model,
capability token, E2E envelope, pairing) — the src/types.ts analog
- term-relay: native WS mux + stateless data plane (subdomain routing, ciphertext forward)
- agent: host-agent (pairing, per-host Ed25519 + mTLS dial-out, forwards to 127.0.0.1:3000)
- control-plane: accounts/hosts registry, pairing-code flow, routing table, provisioning
- relay-e2e: browser<->agent E2E (X25519 ECDH through relay, AEAD, anti-replay, recoverable replay key)
- relay-auth: Passkey/WebAuthn, capability tokens, per-host certs, deny-by-default tenant isolation
- relay-web: browser login + Web Crypto E2E + client-side preview rendering
Security invariants INV1-15 enforced; cross-tenant isolation CI tripwire live
(.github/workflows/relay-tripwire.yml). Design + implementation-level plans in
docs/PLAN_RELAY_*.md and docs/EXPLORE_RELAY_SERVICE.md.
NOTE: generated autonomously per the reviewed plans. The security-critical
packages (relay-e2e, relay-auth) REQUIRE expert security audit before any real
deployment — passing tests prove self-consistency, not resistance to attackers.
Base app (src/, public/) unchanged; concurrent desktop work left uncommitted.
While a tool-permission gate is held on the active tab, spoken confirm
phrases resolve it via the existing approve/reject WS channel — no server
change (byte-shuttle intact). Everything else stays ordinary dictation.
Safety: whole-utterance-exact matching + leading-negation guard + reject
precedence + confidence gate (approve only) + a cancellable 1.5s confirm
window whose commit re-validates gate identity/epoch/connectivity (TOCTOU
guard) + stale-gate epoch bound at PTT-start.
New pure modules public/voice-commands.ts and public/voice-confirm.ts at
100% coverage; wiring in voice/terminal-session/tabs/main. 1307 tests pass,
typecheck clean, build:web OK. Independent code + security reviews flagged a
confirm-window TOCTOU and an unwired cancel path — both fixed and covered.
Plan: docs/PLAN_VOICE_COMMANDS.md.
The flat Projects grid was overwhelming at 50-100 repos. Explored the UX with
4 parallel agents (auto-grouping / manual folders / density-segmentation /
competitive research); all converged on namespace auto-grouping as the primary
fix — the dotted repo names (Billo.Platform.*, Billo.Infrastructure.*) already
encode the hierarchy, so it's a string-split, not a taxonomy the user builds.
Frontend:
- groupProjects(): depth-2 namespace grouping, >=2-member threshold (singletons
fall to "Other"), degrade-to-flat when repos share no prefix, running projects
duplicated into a pinned "Active now" section.
- displayLabel(): cards show the namespace tail (Payment, not
Billo.Platform.Payment); full name in Active/Other.
- Collapsible sticky group headers with an "N active" badge so a collapsed
section never hides running work; search force-expands matching groups.
- public/prefs.ts: server-first load with a localStorage offline mirror, migrates
the legacy per-device proj-favs.
Backend (server-side prefs so favourites + collapse-state follow the user across
devices — localStorage was per-device, which broke the multi-device premise):
- src/http/prefs-store.ts: ~/.web-terminal-prefs.json store (mirrors push-store),
every field bounded + sanitized.
- GET /prefs (read-only) + PUT /prefs (Origin-guarded); UiPrefs type +
PREFS_STORE_PATH config.
Tests: +12 prefs-store, +21 groupProjects/displayLabel. Full suite 1227 passed.
Verified live: grouping renders, collapse persists to the server and survives a
full page reload (cross-device sync proven end-to-end).
