fix: use sonnet for skill trigger evals instead of haiku

[stack72]

Summary

• Remove EVAL_MODEL: "claude-haiku-4-5-20251001" override from the
  skill-trigger-eval CI job. Skill descriptions were tuned against Sonnet,
  and Haiku doesn't follow routing instructions (like "use this skill INSTEAD
  OF domain-specific skills") reliably enough — swamp-workflow dropped to 63%
  and swamp-troubleshooting to 65% on Haiku vs 80%+ on Sonnet.
• Keep EVAL_RUNS=1 and 25 concurrent workers for speed. 185 Sonnet calls
  dispatched in parallel is still fast.

Test plan

• Haiku results: 2 skills failing (workflow 63%, troubleshooting 65%)
• Sonnet results: all skills passing (≥80%)

🤖 Generated with Claude Code

[github-actions]

Adversarial Review

This is a one-line CI config change removing the EVAL_MODEL: "claude-haiku-4-5-20251001" env var from the skill-trigger-eval job. When EVAL_MODEL is unset, the eval script passes no --model flag to claude -p, which defaults to Sonnet.

Critical / High

None.

Medium

None.

Low

1. .github/workflows/ci.yml:137 — Implicit model default: The default model is now whatever claude -p ships with, rather than an explicit pin. If Anthropic changes the CLI default, eval results could shift without any PR touching this repo. This is unlikely to cause a real problem and is arguably the intended behavior (track the recommended model), but worth noting.

Verdict

PASS — Trivial CI config change. The removed env var override is correctly handled as optional by the eval script (Deno.env.get("EVAL_MODEL") returns undefined, which propagates to skip the --model flag). No logic, security, or correctness concerns.

[github-actions]

LGTM — clean, minimal CI config change.

The diff removes EVAL_MODEL: "claude-haiku-4-5-20251001" from the skill-trigger-eval job, falling back to the default Sonnet model. Verified that EVAL_MODEL is optional in scripts/eval_skill_triggers.ts (line 107) — when unset, claude -p uses its default model.

No blocking issues.

The rationale is sound: skill descriptions were tuned against Sonnet, and Haiku doesn't follow routing instructions reliably enough (63-65% vs 80%+ on Sonnet). Keeping EVAL_RUNS=1 and 25 workers maintains fast CI execution.

🤖 Generated with Claude Code