SeekerClaw v2.2.0

[2.2.0] - 2026-07-20

Sign in with your SuperGrok or X Premium subscription, the newest OpenAI
models, and a security & logging hardening pass. Adds xAI Grok as a full
provider with subscription sign-in, brings in the GPT-5.6 family, and tightens
credential handling in logs and web_fetch. Fully backward-compatible —
existing provider setups, agent memory, and settings are unchanged.

Added

• Sign in with Grok (xAI) — a new AI provider. Sign in with a SuperGrok or
  X Premium subscription instead of an API key, or use an xAI API key instead.
  Browser-based OAuth, Keystore-encrypted token storage, automatic refresh, and
  the Grok model family (Grok 4.5 by default) in the model picker. Available
  from first-run Setup and Settings → AI Provider. Built to survive
  interruption: token refresh is fenced across the app and agent processes, so
  a rotation cut short by a restart, force-stop, or reboot can't invalidate the
  session and leave the agent unable to reconnect. (BAT-1124, BAT-1155, #434)
• GPT-5.6 model family — Sol (the new OpenAI default), Terra, and Luna. All
  three work on both an OpenAI API key and a ChatGPT sign-in, alongside
  GPT-5.5, GPT-5.4, and GPT-5.4 Mini. The retired GPT-5.3 Codex entry was
  removed from the picker — superseded, and never usable on a ChatGPT sign-in
  even while it was listed. A model you chose explicitly is preserved when you
  update. One exception: if you picked GPT-5.3 Codex from Telegram while the
  app was closed, the picker may show GPT-5.6 Sol after updating while the
  agent keeps using Codex — re-pick in Settings, or send /model, to resync.
  (BAT-1151, #442)

Fixed

• Agent no longer stalls on a false "out of extra usage" error — on a
  Pro/Max sign-in with usage remaining, an auto-generated note summarizing
  routine background check-ins could contain a phrase the AI service rejected.
  Because that note rode along in every request, it blocked all replies until
  it aged out. Those trivial notes are no longer generated, leftovers from
  earlier versions are filtered out, and the agent now retries once without
  recent-activity context if a request is unexpectedly rejected. The background
  check-in (heartbeat) mechanism itself is unchanged. (BAT-1130)
• Clearer authentication errors — no longer suggests "API key might be
  wrong" to Pro/Max sign-in users, who have no API key. (BAT-1130)
• Real Grok (xAI) error messages — shows the provider's actual reason
  instead of a fabricated "your subscription tier doesn't include API access —
  add an xAI API key" notice. Applies to both chat replies and the Settings
  "Test connection" check, so subscription users are no longer wrongly told to
  add a key. Debug logs now record the real error code and message. (BAT-1172)

Changed

• Lower token usage per turn — the system prompt now caches more
  effectively across background check-ins by keeping frequently-changing
  recent-activity context out of the cached section. (BAT-1130)
• Refreshed Claude Code client version reported on Pro/Max sign-in
  requests, keeping the integration current. (BAT-1123)
• Timestamped, size-bounded diagnostic logs — every line now carries the
  time the event actually happened, not the time it was received (which drifted
  under load), and the debug log rotates continuously at roughly 5 MB instead
  of only being trimmed at startup, so a long session can't grow it without
  limit. The previous generation is kept alongside the current one; if that
  rotation can't rename the file, the current log is reset rather than archived
  so it always stays bounded, and the failure is recorded in the log itself.
  (BAT-1161)
• "Clear logs" is now "Clear console" — which is what it actually does: it
  clears the on-screen console, not the agent's debug log on disk. (BAT-1161)

Security

• Secrets stripped from the Logs screen and anything shared out of it — API
  keys and tokens are masked as lines are collected, as they're forwarded from
  the agent, as older entries are reloaded on startup, and in the one-tap
  Share/export output. Previously redaction ran only on the Node side, leaving
  lines written directly by the app unmasked on screen and in Share. Note this
  does not rewrite log files already on disk from an earlier version — those
  keep their original contents until rotation ages them out. (BAT-1161)
• web_fetch cross-origin redirect hardening — caller-supplied headers
  (API keys, bearer tokens) are dropped when a request redirects to a different
  origin, and a cross-origin 307/308 can no longer forward the request body
  to the new host. Same-origin flows are unchanged. (BAT-1086)
• API keys masked before reaching the AI model — the agent can still read
  its settings and save keys, but raw values in agent_settings.json are never
  exposed to the model or the chat: the read tool returns them masked, and the
  js_eval / shell paths are blocked for that file. Protects provider billing
  keys even against prompt injection. This is model-facing output masking; the
  file on disk and the save flow are unchanged. (BAT-1087)
• Stronger SSRF guard on web_fetch — blocks private, loopback, and
  link-local addresses across all IPv4 and IPv6 encodings, including IPv6
  loopback, IPv4-mapped, ULA, link-local, zone-identifier, and decimal / hex /
  octal literal forms, with the same guard applied to the file-download path.
  Only requests to internal addresses are affected. (BAT-1088)