Casters: a hung voice is given up in 15 s, GPT-6 casts, and the review's loose ends

[earakely-scale]

What this does

The voice timeout, GPT-6 support, and the smaller findings left over from the #14 review.

• A voice that hangs is given up in 15 s, not 45. The line is still captioned. Voices normally take about 2.5 s, at most 8 s. In two of the four smoke runs for Casters: two agents that look the match up and research it, a line at a time #14, the speech service hung, for 45 s of dead air each time.
• OpenAI's reasoning models can cast. GPT-6 on chat completions rejects max_tokens, and it rejects tools with reasoning on. The caster takes up what the 400 asks for and retries, and the model's calls carry it from then on. Checked live: GPT-6 Luna takes both adjustments, then calls say.
• A tool refusal in vLLM's words ("requires --enable-auto-tool-choice") turns tools off, as other refusals do. Before, every line failed and the casters went silent.
• Text alone is not a line while there is time to look things up. "Let me check the standings first" is sent back with a note. On the last round, text is still taken as the line, for models that won't call say.
• A tool call without an id gets one, so the call and its answer go back with the same id.
• An argument a lookup doesn't take is ignored and noted, instead of an error that costs a round.
• battles labels a fight with the turn it was fought, the turn before its entry, as messages and notes are labelled.
• Small tidy-ups:
  • the Barbarians are named as such, not "the game's own AI";
  • research's unknown-tool error lists jot, not say;
  • the unused Caster.model is gone;
  • an HTTP error's text is read once and kept, so every reader sees it.

Tests

• 516 pass; ruff is clean.
• Five new tests, each failing on main:
  • the GPT-6 adjustments;
  • the vLLM refusal;
  • text alone sent back, together with the id fix;
  • the 15 s voice timeout;
  • the battles' turns.

A real match on EC2, with the client view, follows in a comment.

🤖 Generated with Claude Code

[earakely-scale]

A real match on this branch (EC2, client view)

Setup:

• Opus 5.5 (Claude Code, Rome) against GPT-6 Sol (Codex, America) on a Tiny map with one AI civilization (Spain), at most 100 turns at a 20 s pace;
• the env image rebuilt from this branch, with the client;
• the casters on, with Haiku 4.5 for the play-by-play and Sonnet 5.5 for the analyst;
• recorded offline with --client-view.

Result: grade 1, in 22.6 minutes.

• The game: war at turn 39, Washington taken at turn 40 and New York at turn 56, and Sol eliminated at turn 59. That's a conquest.
• The ending, as Game over is the game's: a seat knocked out no longer ends it for the casters, the viewer and the play page #15 meant it: the casters signed off at GAME OVER with the victory in the data, "Opus 5.5 wins by conquest, 192 to 36". In the public sol-vs-opus broadcast they had signed off a turn early, calling the win "on score".
• The casters: 122 lines, and no voice timed out.
  • The only error logged was Sonnet 5.5 turning a forced say down, which is handled.
• Talk: 32 public messages, 18 from Sol and 14 from Opus.
• Cost: $3.26 for the agents (Opus $2.09, Sol $1.18), plus about $1.50 for the casters.

🤖 Generated with Claude Code