Skip to content

openclaw-zero-token 2026.4.5

Latest

Choose a tag to compare

@linuxhsj linuxhsj released this 04 Apr 23:57
· 20 commits to main since this release

What's New in 2026.4.5

Upgraded to OpenClaw 3.28 — all 13 web models tested and verified working.

Highlights

  • All 13 web models pass end-to-end testing: ChatGPT, Claude, DeepSeek, Doubao, Gemini, GLM, GLM International, Grok, Kimi, Perplexity, Qwen, Qwen CN, Xiaomi MiMo
  • Fixed long prompt issue: web models now receive only the user's message instead of the full system prompt + 24 tool definitions (~35,000 chars). This caused garbled responses, timeouts, and system prompt leaks
  • Fixed repeated content: Qwen CN and Doubao SSE streams send accumulated content (not incremental deltas) — now correctly deduplicates by emitting only the new portion
  • Simplified chat UI: removed the redundant session/agent selector, keeping only the model selector dropdown

Web Model Fixes (per provider)

Provider Fix
ChatGPT Playwright native keyboard API for reliable input; updated send button selectors
Gemini Playwright keyboard API; innerText instead of textContent for response extraction
Grok Playwright keyboard API for message submission
Perplexity Rewritten as DOM simulation with "New Thread" button click; only sends user message
Kimi Fixed Connect binary protocol parser (only accept append/set ops)
Qwen CN Added credentials: "include" to fix 403 signature error; fixed auth loop
Qwen Web Fixed model ID (qwen-maxqwen3.5-plus)
DeepSeek Enhanced tool_call regex, data.v array handling, XML argument fallback
Xiaomi MiMo Filter <think> blocks and dialogId events; strip null bytes
Doubao Deduplicate accumulated SSE content
GLM/Grok Deduplicate repeated SSE content

Infrastructure

  • Shared stripInboundMeta(): all web streams now strip OpenClaw metadata (Sender, Conversation info, timestamps) from user messages before sending to web models
  • E2E test suite (scripts/test-web-model-e2e.sh): two-phase testing (TUI first, then WebUI), anti-ban adaptive delays, message pool with 20 CN/EN natural prompts, HTML report output
  • onboard.sh webauth auto-exits after authorization completes (no more Ctrl+C needed)

Dependencies

  • Upgraded to upstream OpenClaw v2026.3.28
  • Bumped Swift deps (swift-log 1.11.0, Sparkle 2.9.1)
  • Bumped Android deps (Compose BOM, commonmark, kotest)

Full Changelog: v2026.3.27...v2026.4.5