Skip to content

v1.4.2 — Agentic loop polish + Anthropic prompt caching

Choose a tag to compare

View all tags
@DoyleDev DoyleDev released this 21 May 19:13
· 31 commits to main since this release
v1.4.2
61e70ec
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's new

A pile of agent-loop UX and cost fixes that came out of multi-hour SQL/data sessions with Opus 4.7. If you're using Mason for anything tool-heavy, this is worth updating to.

Prompt caching for Claude models — big cost win

Tool schemas are the heaviest static portion of every turn (~16K tokens for ~80 tools). Mason now marks cache_control: ephemeral on the last tool and last system message, so Anthropic caches the entire static prefix for 5 minutes. Subsequent turns within the cache window read at ~10% of input cost — typically 80–90% savings on the static prefix every turn.

Verify it's working in DevTools console after a few turns:

[CHAT] Usage (streamed) — input: 31000, output: 230, cache_created: 0, cache_read: 26017

cache_read climbing turn-over-turn = caching engaged.

OpenAI/GPT models automatic-cache prefixes >1024 tokens, no opt-in needed. Gemini/Meta/Qwen have no standard caching path yet.

Agent loop UX

  • Agent iteration budget raised 10 → 40. Polling patterns (waiting on a job run, multi-step SQL builds) used to silently hit the cap and return to idle with no explanation. Now the budget fits real workflows.
  • Clear stop message when the loop exhausts. No more "looks like a hang" mystery — the chat surfaces Agent loop hit the 40-step budget — send "continue" to keep going.
  • Building bricks stay visible during tool execution and below streaming text. Previously the indicator vanished when "Calling tool: X" appeared, and again once content started streaming — users assumed Mason had frozen during slow MCP calls or model pauses.
  • Empty-response detection. If Opus 4.7 burns its token budget on extended thinking and returns no visible content, Mason now shows a clear hint instead of silently rendering an empty bubble.

Chat reliability

  • max_tokens 4096 → 16384. Gives Opus 4.7 extended-thinking headroom before output, and reduces tool-call argument truncation.
  • trimHistory drops orphan tool results. Anthropic was rejecting whole requests when our message trimming cut between an assistant-with-tool_calls and its result. Now any tool-result message whose tool_call_id isn't matched is pruned.
  • Tool-call argument sanitization. Empty or malformed JSON in function.arguments (from truncation or providers that emit no args for no-param tools) now normalizes to {} so the conversation doesn't wedge.

Full changelog: v1.4.1...v1.4.2

Assets 5
Loading