[Bug]: Workflow engine hardcodes "copilot" default, ignoring project's initialized integration (e.g., opencode)

[markuswondrak]

Bug Description

When running a workflow without explicitly passing the integration via --input integration=<name>, the workflow engine defaults to copilot (defined in workflow.yml). This overrides the project's actual initialized integration stored in .specify/init-options.json and .specify/integration.json.

If the copilot CLI happens to be installed on the system, the workflow incorrectly dispatches the command to Copilot instead of the configured project AI (e.g., opencode). This results in a confusing No such agent: speckit.specify error, because the project only contains agent files for the initialized AI, not Copilot.

Steps to Reproduce

1. Initialize a project exclusively with Opencode: specify init --ai opencode
2. Ensure the copilot CLI is coincidentally installed on your system (/usr/bin/copilot).
3. Run a workflow without an explicit integration input:
   specify workflow run speckit --input spec="..."
4. The workflow fails at the first step (speckit.specify).

Expected Behavior

The workflow engine should auto-detect the project's integration from .specify/integration.json (or .specify/init-options.json) and use it as the default. If I initialized the project with opencode, the workflow should dispatch to opencode.

Actual Behavior

The run fails with: "No such agent: speckit.specify".

The Failure Path:

1. The workflow.yml (lines 22-25) sets the default input for integration to "copilot".
2. The CommandStep resolves the integration input, choosing "copilot" over the project default.
3. CopilotIntegration.dispatch_command() attempts to run: copilot -p "<spec>" --agent speckit.specify.
4. Because /usr/bin/copilot exists, shutil.which("copilot") succeeds and dispatch proceeds.
5. Copilot fails because the commands live at .opencode/command/speckit.*.md, not .github/agents/speckit.*.agent.md.

Specify CLI Version

0.7.4.dev0

AI Agent

Claude Code

Operating System

Ubuntu 24.04.4 LTS

Python Version

3.12.3

Error Logs

Additional Context

Diagnostic Information & Open Questions

If we fix the workflow to correctly dispatch to opencode (or pass --input integration=opencode manually), we might hit a secondary bug.

Currently, OpencodeIntegration delegates to MarkdownIntegration.build_exec_args(), which generates the CLI command: ["opencode", "-p", prompt].
It needs to be verified if the opencode CLI actually supports the -p shorthand in non-interactive mode. Standard opencode usage typically expects opencode run <message> or opencode --prompt <message>.
(Diagnostic commands to verify this internally: opencode --prompt "test" vs opencode run "test").

Suggested Fixes & Quality Improvements

1. Dynamic Default Resolution: The workflow YAML schema should support integration: auto, telling the engine to read .specify/integration.json instead of hardcoding "copilot".
2. Workflow Pre-flight Check: When specify workflow run is invoked, it should warn or abort if the workflow's resolved integration mismatches the project's initialization state.
3. Opencode Integration Test: Add a contract test specifically for OpencodeIntegration.build_exec_args() to ensure the generated CLI flags (like -p) are actually supported by the opencode CLI in headless mode.
4. Better CLI Error Bubbling: When CommandStep._try_dispatch() fails, surface the actual stderr from the CLI tool rather than dropping it.