Skip to content

Releases: BlockRunAI/runcode

v1.3.9

04 Apr 06:15
1613d91

Choose a tag to compare

Abort (Esc): now emits clean 'aborted' event instead of error. Models command: 15s fetch timeout, handles empty API response.

v1.3.8

04 Apr 06:14
d744d62

Choose a tag to compare

SSE parser: 1MB buffer cap prevents memory growth on malformed streams. Terminal /model (no args) now shows current model and switch syntax.

v1.3.7

04 Apr 06:13
ca59a8f

Choose a tag to compare

Fallback chain: 60s timeout per attempt prevents infinite hangs. Proxy direct fetch: 2min timeout. Both prevent backend stalls from blocking the entire proxy.

v1.3.6

04 Apr 06:13
d17ded1

Choose a tag to compare

Security: Write tool now resolves symlinks before checking sensitive paths (prevents symlink traversal). Bash tool description documents output cap, timeouts, and working directory.

v1.3.5

04 Apr 06:12
703a7ef

Choose a tag to compare

Terminal UI: /retry and /compact commands. Write tool shows KB for larger files.

v1.3.4

04 Apr 06:11
7843816

Choose a tag to compare

Up/Down arrows recall previous prompts (last 50 stored). Permission prompts now styled with box drawing for better visibility in Ink UI.

v1.3.3

04 Apr 06:09
f92e4b3

Choose a tag to compare

Edit tool now shows diff preview (- old/+ new lines), error messages include actionable recovery suggestions (try /retry, /model, /compact)

v1.3.2

04 Apr 06:08
b52685f

Choose a tag to compare

/retry command resends last prompt, tool preview increased to 200 chars, running tools show elapsed seconds

v1.3.1

04 Apr 06:07
3bf9950

Choose a tag to compare

Fix: thinkingText persists after turns, /compact now forces compaction below threshold, proxy token tracking map bounded to 50 entries

v1.3.0 — /compact, Thinking Display, Auto-Retry, 20+ Context Windows

04 Apr 06:01
3bf3520

Choose a tag to compare

What's New

/compact Command

Manually compress conversation history to save tokens:

> /compact
Compacted: ~45,200 → ~12,800 tokens

Thinking Content Display

Ink UI now shows a preview of the model's thinking process instead of just a spinner.

Transient Error Auto-Retry

Network timeouts, 429 rate limits, and 500/502/503 server errors now automatically retry with exponential backoff (1s → 2s → 4s, up to 3 attempts). Sessions no longer die on momentary hiccups.

Expanded Context Window Registry

Added 20+ model context windows (xAI Grok, GLM-5, Minimax, O3/O4, GPT-4.1, Gemini 3.1, etc.) plus pattern-based inference for unknown models. Compaction thresholds are now accurate across all providers.

Other Improvements

  • System prompt: All 11 tools documented with constraints
  • 16 model shortcuts synced to terminal picker
  • Token estimation: +16 token overhead for tool_use framing
  • Router: Code block detection + byte-length estimation
  • Glob: Only ** triggers recursion (was over-recursing)
  • WebSearch: Fallback parser when DuckDuckGo changes HTML
  • First-run tips for new users
  • GLM promo auto-expiry fallback to Gemini Flash