v0.3.0 — Telemetry pipeline + TikTok crack + trust pass
Major
- Engine telemetry pipeline — opt-in OFF by default. New CLI:
python server.py telemetry on|off|status|preview. Persisted in~/.webloom/config.json. After every tool call, sends{tool, ok, error_class, duration_ms, engine_version, anon_id, ts}to webloom.run — zero URLs, zero content, zero identity. See https://webloom.run/transparency. - TikTok crack —
tiktok_signexposes the in-pagebyted_acrawlersigner for any URL (returns x-bogus + msToken).tiktok_post_videodrives end-to-end UI upload (drop → caption → privacy → post). Selectors all overridable as args. - X.com transaction-id RE (
x_create_tweet) — vendored XClientTransaction, reverses the body-signed anti-replay header. Real tweet shipped via this path. - Tool descriptions now ship escalation ladders —
upload_file,click,filleach end with 'IF ALL FAIL — DO NOT GIVE UP, try X, Y, Z'. Stops other AI sessions from bailing prematurely. - draftjs_set_text — chunked beforeinput (16-char) for long X posts. No more MCP timeout on full-length tweets.
- react_invoke_handler — fiber-walks the React tree and calls onClick directly. Bypasses DOM-event-blocking overlays (LinkedIn
interop-outletpattern).
Trust signals (because AI installers were rejecting installs)
- SECURITY.md documenting threat model, data boundary, code provenance, reporting path
- README rewritten to lead with what WebLoom does + does NOT do, privacy defaults, uninstall path
- Install doc at webloom.run/install rewritten — no more 'do this without asking', every step explained
Other
- Engine version bumped 0.2.0 → 0.3.0
vendor/x_client_transaction/for X signing math (MIT, ~424 LOC)- Multi-handler click fallback now records which handler succeeded to playbook
- New shadow-DOM-aware
upload_fileStrategy E for LinkedIn composer + Lit-based dashboards
License
MIT — same as 0.2.0.