feat: per-call model + reasoning override with list_models discovery

[harshav167]

Summary

• Lets the orchestrator pick a model (and reasoning/variant) per delegation — on task (both subagent_type and category), call_omo_agent, and team members — gated to connected/available models, with the configured default and fallback chain preserved when omitted.
• Adds a list_models tool so the agent can discover connected provider/model ids and the valid reasoning vocabulary before overriding.
• Fixes team_create inline members to actually accept model / reasoning_effort (the tool schema previously stripped them).

Changes

• @oh-my-opencode/delegate-core — new pure, dependency-injected helper resolveRequestedModelOverride(): parses provider/model[ variant], lets reasoning_effort override the embedded variant, gates via fuzzyMatchModel, with cold-cache passthrough so the first call isn't blocked.
• task + call_omo_agent — new optional model / reasoning_effort args; the override replaces the resolved model at top precedence while preserving the agent/category fallbackChain.
• team-core MemberBaseSchema — optional model / reasoning_effort (schema stays .strict()); and team_create's inline-member tool schema now advertises both fields so the framework no longer strips them.
• list_models — new always-on tool: connected provider/model ids grouped by provider plus the reasoning vocabulary; pointer seeded into the model arg descriptions. Registered via the standard factory → tool-registry-factories → createCoreTools() → tools/index.ts path.
• Docs — docs/reference/features.md (per-call override + list_models) and docs/guide/team-mode.md (per-member override).

Testing

bun run typecheck   # 0 errors (tsgo, all packages)
bun test            # green locally; CI passes on ubuntu/macos/windows

• delegate-core helper — 8 (parse / gate / variant / reasoning precedence / cold cache)
• delegate-task tool-boundary integration — 5 (asserts the real session.prompt/launch body carries model, reasoning, gate-reject + no-spawn, default-preserved, category)
• team member schema — 4; team_create inline-member model — 4; list_models formatting — 5
• Manual QA on the real OpenCode harness: list_models returns the connected set; task(model="openai/gpt-5.5", reasoning_effort="high") and task(model="xai/grok-composer-2.5-fast") each launch on exactly that model/provider; an invalid id (openai/ghost-999) fails at the gate/provider and the preserved fallback chain recovers it. Plugin load verified in an isolated XDG sandbox (Built tool registry … totalTools 26, teamModeEnabled true, 25→26 = list_models registered).

Related Issues

None — standalone feature. Three design questions for maintainers:

1. Opt-in gating? On by default (the gate is the safety). Happy to add a delegate_task.allow_runtime_model config flag if you'd prefer it gated, given the caution in docs/guide/agent-model-matching.md.
2. Cold-cache passthrough in requested-model.ts is the one gate relaxation (first-call safety); stricter behavior is a one-liner.
3. reasoning_effort accepts any string (normalized downstream per model); could be enum-validated.