feat(skill-trigger): add Codex provider tool mapping for skill detection

[christso]

Objective

Add proper skill-trigger evaluator support for the Codex provider. Currently, the evaluator maps provider kinds to tool-name semantics (e.g., Claude uses Skill/Read, Copilot uses Read File/readFile), but Codex has no mapping.

Context

Codex supports skills via .agents/ or .codex/ folders. However, its tool call format is different from Claude/Copilot:

• command_execution — shell commands (how Codex reads files, including skill files)
• file_change — file modifications
• mcp:server/tool — MCP tool calls

When Codex triggers a skill, it likely reads the skill file via a command_execution (e.g., cat .codex/skills/my-skill.md). Detecting skill triggering requires matching the skill name inside the command_execution input string rather than checking a dedicated Skill tool.

Design latitude

• Investigate what Codex's actual tool call looks like when it reads a skill file — is it always command_execution with cat? Does it use a different mechanism?
• Decide whether to add Codex to PROVIDER_TOOL_SEMANTICS with a custom matcher, or whether the skill_tools/read_tools config override (added in feat: make agentv-bench work across all agent harnesses #641) is sufficient for users to configure per-eval
• Consider whether command_execution input is a string (not JSON with fields) — the current ToolMatcher expects named fields like input.skill or input.file_path, which may not fit

Acceptance signals

• Codex has an entry in PROVIDER_TOOL_SEMANTICS in skill-trigger.ts
• Unit test showing skill-trigger detection with a realistic Codex tool call
• Or: documented recommendation for using config overrides if a static mapping isn't feasible

Non-goals

• Changes to the Codex provider itself
• Adding .agents/ or .codex/ folder discovery to agentv

Related

• feat: make agentv-bench work across all agent harnesses #641 — feat: make agentv-bench work across all agent harnesses (added multi-provider support)
• feat(evaluator): make skill-trigger target-agnostic #613 — make skill-trigger evaluator target-agnostic

[christso]

Closing as covered by #641. The config overrides (skill_tools/read_tools) added in #641 allow users to configure skill-trigger detection per-eval for Codex and any other provider whose tool-call format doesn't fit the static ToolMatcher model. A static mapping isn't feasible since Codex uses command_execution with free-form string input rather than structured fields.