[BUG]: calculate_max_tokens doesn't recognize claude-haiku-4-5 (same class as #312) #317

Summary

calculate_max_tokens in rig's Anthropic provider doesn't recognize claude-haiku-4-5 (the latest Haiku model). When max_tokens is omitted from [agent.llm] config, the safety-net default returns None → the provider errors with `max_tokens` must be set for Anthropic. Same class of bug as #312, different model.

Reproduction

Use claude-haiku-4-5 in agent.llm with orchestration enabled, omitting max_tokens:

[agent]
name = "Aura Orchestrator"
system_prompt = "..."
turn_depth = 5

[agent.llm]
provider = "anthropic"
api_key = "{{ env.ANTHROPIC_API_KEY }}"
model = "claude-haiku-4-5"

[orchestration]
enabled = true

If max_tokens is set explicitly in config, the bug is not triggered (the value is propagated directly). The bug only fires when max_tokens is omitted and rig's calculate_max_tokens fallback is consulted.

Relevant log output

WARN  aura::orchestration::orchestrator: Planning attempt 1 failed after 0.0s: CompletionError: RequestError: `max_tokens` must be set for Anthropic
ERROR aura_web_server::streaming::handlers: Stream error: Planning failed: CompletionError: RequestError: `max_tokens` must be set for Anthropic

Root cause

Anthropic changed their naming convention at the 4-series boundary:

• Old: claude-3-5-haiku (generation before tier) — matched by rig's starts_with("claude-3-5-haiku") → 8192
• New: claude-haiku-4-5 (tier before generation) — matches no existing prefix in calculate_max_tokens

Verified with a standalone test:

claude-haiku-4-5               => None
claude-haiku-4-5-20251001      => None
claude-3-5-haiku               => Some(8192)
claude-sonnet-5                => Some(64000)  # fixed in #312

Fix

Add a new branch to calculate_max_tokens and calculate_max_tokens_custom in rig-core/src/providers/anthropic/completion.rs:

} else if model.starts_with("claude-haiku-4") {
    Some(64000)
}

Uses claude-haiku-4 as the prefix (covers claude-haiku-4-5 and future claude-haiku-4-x variants). Default of 64000 matches Haiku 4.5's 64k max output (per Anthropic's model docs). This is a new branch, not an || addition to an existing one — Haiku 4.5's 64k max output is 8x higher than the existing claude-3-5-haiku default of 8192, so it can't share that branch.

Same PR pattern as #312: rig fork PR to mezmo/rig targeting mezmo, then bump the rev pin in aura's Cargo.toml.

Additional Context

• This is a latent bug — no configs in the aura ecosystem currently use claude-haiku-4-5, so it hasn't been hit in production yet.
• The aura-side max_tokens propagation fix from [BUG]: max_tokens from config.toml is not passed to ochestration coordinator #312 is already merged (PR fix(orchestration): propagate max_tokens to coordinator agent #316), so users who set max_tokens explicitly in config are not affected. This only affects users who omit it.
• Related: claude-opus-4-5 through claude-opus-4-8 are matched by the existing starts_with("claude-opus-4") prefix but have an outdated default of 32000 (their actual max output is 128k). They don't error, just cap low. Separate concern.

Searched Issues

• No similar issues found

Code of Conduct

• I agree to follow this project's Code of Conduct