Skip to content

capy-flash v0.13.2 — Coding agents without error answers, same tokens

Choose a tag to compare

@ValerioDolci ValerioDolci released this 15 Nov 18:56
· 5 commits to main since this release

Fixes for what real coding agents hit, found by running Claude Code 2.1.321, OpenCode 1.19.3 and Codex CLI 0.169.0 against both models through a proxy that recorded every request and answer. The tokens are the same as v0.13.1.

Changed

  • Images and documents a model cannot read no longer fail the request. Without --vision (Flash-Next has no Vision tower), each image or video reaches the model as a line of text saying that one was attached and that the model cannot see it; an Anthropic document block (a PDF Claude Code's Read tool attaches) becomes the same kind of line, and a search_result block its title, source and text. Before, the 400 (vision_disabled, documents_not_supported) stayed in the client's history and every later turn of the conversation failed. The server's log line for the request reads media N omitted (no Vision).
  • Default output limit: the rest of the context. A request that sets no output limit (Codex CLI sends no max_output_tokens) now gets the context left after the prompt, as vLLM does; it was 8,192 tokens, reasoning included, and long file writes ended as incomplete responses. --default-max-tokens N keeps its meaning.
  • Image detail low and high are accepted on Chat Completions and Responses (Codex's view_image sends high, so Codex could not look at an image even on the 27B); the Vision frontend has one preprocessing profile for all. An unknown value is still a 400.
  • thinking.display: "updates" (Claude Code 2.1, on the first request of every session) is accepted and shows the reasoning as summarized does; before, it was refused twice per session until Claude Code retried without it.
  • Context overflow reads prompt is too long: N tokens > M maximum, the form Anthropic uses and Claude Code reads to compact the conversation, on every protocol; before, the main path gave only the limit.
  • Every error answer is logged. Client errors (an unknown model, a refused field, an unknown route) and a wrong API key now write one server log line with the request id the client receives, the status, the code and the top-level field; never the message or the request's content.

Measured (Ryzen 9 9900X, 96 GB DDR5, 2 × RTX 5070 Ti without P2P)

  • Same tokens: the 27B gate (perplexity 4.133481 / 4.385905, greedy 60/60, golden 3/3) and frozen-expert greedy runs on both Flash-Next artifacts identical to v0.13.1; the full test suite passes (223/223).
  • Real clients against the new build with the production flags: the Claude Code and Codex tasks that failed on v0.13.1 (an image on Flash-Next, an image or a long file write in Codex) now complete with no error answer.