Refactor: unify agent adapter capabilities with per-stage configuration

[zhangweijian97]

Current State

Three agent adapters exist (Codex, Claude Code, Cursor), all implementing the same AgentAdapter interface. However, capabilities are heavily skewed:

Codex (most mature):

• Per-stage model selection (models.plan, models.implement, etc.)
• Per-stage reasoning effort (reasoningEfforts.plan, etc.)
• Per-stage fast mode toggle
• Sandbox control, Docker support, plugins/skillsets
• Output file management, usage extraction

Claude Code / Cursor (minimal):

• All 5 adapter methods implemented, but each routes through the same code path
• No per-stage model or configuration differentiation
• Claude: global model, maxTurns, allowedTools, permissionMode
• Cursor: global model, force flag

Meanwhile, agent-registry.types.ts already defines AgentStage and AgentModelDefinition.defaultFor — suggesting per-stage config was anticipated but not yet adopted by Claude/Cursor adapters.

Proposal

Phase 1 — Per-stage model for all adapters (the quick win)

• Extend AgentAdapterRuntimeConfig.agent with models?: Partial<Record<AgentStage, string>> (mirroring codex config shape)
• Claude Code: use per-stage model in runPlan / runReview / runGithubComment / runTaskIntake
• Cursor: same
• Codex: no change (already has this)

Phase 2 — Optional capability declaration

• Add capabilities to AgentBackendDefinition (e.g., perStageModel, perStageReasoning, sandbox, docker, sessionResume, usageExtraction)
• Workflow layer can check capabilities before attempting adapter-specific features
• Adapters declare what they support; callers degrade gracefully

Phase 3 — Per-stage reasoning/effort for applicable adapters

• Extend the per-stage config to include reasoning effort where the underlying agent supports it
• Codex: already done, just needs to read from the unified config shape
• Claude: no equivalent (model choice is the tuning knob)

Benefits

• Claude Code and Cursor users get the same per-stage model flexibility as Codex
• New adapters have a clear capability contract to implement against
• Reduces the config shape divergence between adapter-specific sections (codex.* vs agent.* vs cursor.*)

Notes

• This should be non-breaking — existing single-model configs continue to work as fallback
• The AgentAdapter interface itself doesn't need to change (capabilities are metadata, not methods)