Improve skill trigger routing accuracy across models

Description

Multi-model skill trigger evals reveal systematic failure patterns affecting skill routing accuracy across Claude Sonnet, Claude Opus, GPT-5.4, and Gemini 2.5 Pro. All models pass the 90% threshold but there are clear skill description issues to fix.

CI results (2026-04-10): https://github.com/systeminit/swamp/actions/runs/24239470178

Model	Pass Rate	Passed	Failed
Sonnet	99.0%	200/202	2
GPT-5.4	98.0%	198/202	4
Opus	94.1%	190/202	12
Gemini 2.5 Pro	91.6%	185/202	17

Cross-Model Failures (highest priority)

1. extension-model should NOT trigger for "Can I run this custom model as part of a scheduled workflow?" - fails on 3/4 models (sonnet, gpt-5.4, opus). Description too broad.
2. report SHOULD trigger for "What methods does UnifiedDataRepository expose in reports?" - fails on 3/4 models (sonnet, opus, gemini). Routed to troubleshooting instead.
3. extension-driver should NOT trigger for "Run this workflow step on a remote Kubernetes cluster" - fails on 3/4 models (gpt-5.4, opus, gemini). Description too broad.
4. model should NOT trigger for "How do I chain this model into an automated workflow?" - fails on 2/4 (gpt-5.4, opus).
5. workflow should NOT trigger for "The workflow is erroring on the second step" - fails on 2/4 (opus, gemini).

Pattern 1: swamp-report poorly differentiated

3/4 models affected. Gemini has 7 report failures alone. Reports consistently routed to troubleshooting. Fix: Update .claude/skills/swamp-report/SKILL.md to mention report creation, output formats, dataRepository, UnifiedDataRepository, dataHandles. Differentiate from troubleshooting.

Pattern 2: Extension descriptions too broad

3/4 models affected each. extension-model and extension-driver trigger on usage queries not about creating extensions. Fix: Update .claude/skills/swamp-extension-model/SKILL.md and swamp-extension-driver/SKILL.md to emphasize creating new TypeScript extensions. Add exclusions.

Pattern 3: Text responses instead of tool calls

2/4 models affected (Opus: 8 cases, Gemini: 10 cases). Both respond conversationally instead of routing via tool call. Fix: Strengthen system prompt in evals/promptfoo/generate_config.ts (~line 242) or adjust expectations per model.

Pattern 4: Ambiguous test cases

Some test cases may need reclassification in trigger_evals.json files. Review cases that fail on 2-3 models where routing is genuinely ambiguous.

Priority

1. Fix report SKILL.md description
2. Fix extension-model and extension-driver SKILL.md descriptions
3. Review ambiguous test cases
4. Address text response handling for Opus and Gemini

Reproduction

Set ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY then run eval-skill-triggers for each model. See scripts/analyze_eval_results.ts for cross-model analysis.