[feat] Run prompt and agent steps with the local driver #1077

Parent: #818

Goal

Extend the local driver to execute real AI-tool steps while leaving workflow semantics and output validation in TeamAI’s engine.

Scope

• Add prompt, prompt_file, and optional agent.
• Resolve prompt files safely inside the team repo when the workflow is pulled and cache them with the validated definition.
• Keep cached prompts revision-bound so a run at an older revision does not read a later edit or a deleted file.
• Resolve named agents from files already deployed by teamai pull.
• Add a small local tool-adapter interface for supported headless tools.
• Read per-machine workflow tool selection, with CLI overrides suitable for CI.
• Give each attempt a private result.json location outside the product checkout.
• Have the engine validate reported output against the declared shorthand schema.
• Report tool, concrete model when known, session handle, token counts, and structured failure details.
• Detect a read-only agent step that leaves repository changes and fail it while preserving those changes for inspection.
• Consume deployed agent configuration directly; do not make model aliases a runtime dependency.
• Add shared conformance scenarios that run through both the in-process engine and next/report --json, including the CLI-only parseReport path.

Out of scope

Session continuation, editable worktrees, branch delivery, graph control flow, human gates, and remote drivers.

Dependencies

• Blocked by: [feat] Add a minimal command-only workflow runtime #1076
• Unblocks: [feat] Add workflow workspaces, branches, and session continuation #1080, #10761, #10762

Acceptance criteria

• A prompt-only workflow runs through one configured local AI tool.
• A named agent is leased only when available in the selected tool environment.
• Missing tools, missing agents, invalid output, and child-process failures become stable step failures.
• The adapter cannot bypass engine output validation.
• Prompts, outputs, and secrets are not logged by default.
• A raw model or tool default works without model aliases.
• A run at an older revision uses its cached prompt_file content after the source file is changed or deleted.
• The same conformance scenarios produce equivalent results through the in-process engine and next/report --json.

Verification

Focused adapter, agent-lookup, result-parsing, diagnostic, prompt-cache, and direct/CLI conformance tests; build/typecheck/lint; one real-CLI E2E with one representative AI executable. Update the bilingual workflow guide for agent steps.