feat: add command step type execution to workflows

[ereslibre]

Add support for command step type in workflow execution, allowing workflows to run shell commands directly alongside AI agent steps. This enables mixed workflows where deterministic tasks (builds, linting, tests) execute as subprocesses while AI-driven tasks are dispatched to agent runners.

The step execution logic has been refactored from the run command into WorkflowManager.run(), centralizing the orchestration of both step types. Command steps capture stdout/stderr as step outputs and support allow_failure to continue execution on non-zero exit codes.

Changes

• Enabled WorkflowStepSchema as a discriminated union of agent and command step types, replacing the previous agent-only schema
• Added isCommandStep type guard and exported it from rover-schemas
• Added allow_failure field to WorkflowCommandStepSchema for fault-tolerant command steps
• Introduced WorkflowManager.run() method that orchestrates step execution: command steps run via launchSync, agent steps are delegated to a caller-provided AgentStepExecutor callback
• Extracted StepResult, AgentStepExecutor, OnStepComplete, and WorkflowRunResult types into rover-core as shared interfaces
• Refactored RunnerStepResult and ACPRunnerStepResult to extend the shared StepResult interface
• Simplified the run command by removing duplicated step-loop logic for ACP and standard modes, replacing it with a single agentStepExecutor callback passed to WorkflowManager.run()
• Added unit tests for command step schema validation, type guards, and WorkflowManager.run() covering command execution, agent delegation, continueOnError, allow_failure, output accumulation, and step completion callbacks
• Added E2E tests for workflow creation/inspection with command steps, allow_failure, and mixed step types
• Added command step execution feature documentation to E2E_TESTS.md
• Added process execution guidelines to AGENTS.md

Notes

Context injection (handleContextInjection) now only applies to agent steps via the isAgentStep guard, since command steps do not have prompts.

Closes: #524

[ereslibre]

To be merged when this gets released: endorhq/docs#39

[Angelmmiguel]

LGTM! Nice job and refactor 😄

[Angelmmiguel]

Nice!