fix(evals): name the model of a spec hash rather than stringifying it

[TonsOfFun]

A scenario evaluation started from the dashboard failed every scenario with HTTP 400 before reaching the model.

Root cause

ModelSpec.parse_all calls to_s on each value:

values.map { |value| value.to_s.strip }

A Hash stringifies to its inspected form, which then travels as the model ID. OpenRouter answers:

{"label" => "openrouter/openai/gpt-4o-mini", "model" => "openai/gpt-4o-mini",
 "provider" => "openrouter"} is not a valid model ID

This is reachable from the dashboard by design, not by misuse: ScenarioEvaluationRunner#selection_summary persists a run's models as specs.map(&:to_h), and re-running that selection hands the hashes straight back to parse_all. A caller passing a plain string (test_execute) was unaffected, which is why single runs worked while every batch run failed.

Fix

parse_all names the model of a hash — label first, then model — and leaves strings alone.

Verification

Against a host application whose agent calls MCP tools, the same three-scenario run before and after:

	before	after
errors	3 × HTTP 400	0
tool calls	none	list_filters, search_providers
scores	0.0	0.68 – 0.875
duration	~1s (failed early)	2.8 – 4.3s

The remaining failures are genuine quality results rather than infrastructure.

Three regression tests cover the label-bearing hash, a symbol-keyed hash, and a hash carrying only model.

[Copilot]

🟢 Approval recommended

The focused implementation resolves the reported round-trip failure and is adequately covered by regression tests.

Pull request overview

Fixes dashboard evaluation reruns by correctly parsing persisted model-spec hashes.

Changes:

• Extracts label, falling back to model, from hash-like values.
• Adds regression coverage for string-keyed, symbol-keyed, and label-less hashes.

File summaries

File	Description
lib/active_agent/evals/model_spec.rb	Coerces persisted model specs before parsing.
test/evals/model_spec_test.rb	Tests supported hash representations.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.