feat: OpenCode runner support

[lukeinglis]

Summary

Closes #394

• Add factory/runners/opencode.py implementing the Runner protocol
• Uses 'opencode run' subcommand pattern with prompt prepended to task message
• --dangerously-skip-permissions gated on parameter (not unconditional)
• Dry-run mode via FACTORY_OPENCODE_DRY_RUN=1
• Register 'opencode' in RunnerName and _RUNNERS registry
• Update CLAUDE.md with OpenCode runner docs and config profile example
• Tests in tests/test_opencode_runner.py

Usage

factory ceo /path --runner opencode
factory agent researcher --task '...' --runner opencode
FACTORY_RUNNER=opencode factory ceo /path

Config profile:

[credentials.opencode]
FACTORY_RUNNER = "opencode"

Test plan

• uv run pytest tests/test_opencode_runner.py -v passes
• factory ceo /path --runner opencode resolves to OpenCodeRunner
• Dry-run mode returns stub without spending tokens

[codecov]

Codecov Report

❌ Patch coverage is 92.85714% with 5 lines in your changes missing coverage. Please review.
✅ Project coverage is 87.43%. Comparing base (8097461) to head (641dc58).
⚠️ Report is 17 commits behind head on main.

Files with missing lines	Patch %	Lines
factory/runners/opencode.py	92.64%	5 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #396      +/-   ##
==========================================
+ Coverage   87.31%   87.43%   +0.12%     
==========================================
  Files          61       62       +1     
  Lines        9339     9472     +133     
==========================================
+ Hits         8154     8282     +128     
- Misses       1185     1190       +5     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[xukai92]

Clean addition that follows the existing Claude/Codex runner pattern with dry-run + auth-warn parity and good test coverage. One small contract bug noted below — happy to approve once gated (or as a quick follow-up PR).

[P2] dangerously_skip_permissions is silently ignored in QwenRunner.headless

factory/runners/qwen.py around the headless method:

        dangerously_skip_permissions: bool = True,
        ...
        cmd = [
            "qwen",
            "--append-system-prompt", prompt,
            "-p", task,
            "--yolo",
            "--output-format", "text",
        ]

The signature accepts dangerously_skip_permissions: bool = True to match the Runner protocol used by Claude/Codex/Bob, but --yolo is appended unconditionally. interactive_run in the same file gates this flag correctly, and the sibling runners all gate their equivalents. A caller that explicitly passes dangerously_skip_permissions=False to disable auto-approval will still get --yolo, bypassing all permission prompts.

The default is True so current call sites are unaffected, but the contract is broken for any future caller.

Suggested fix:

        cmd = [
            "qwen",
            "--append-system-prompt", prompt,
            "-p", task,
            "--output-format", "text",
        ]
        if dangerously_skip_permissions:
            cmd.append("--yolo")

Worth adding a dangerously_skip_permissions=False test in the headless suite — there's already an equivalent for interactive_run.

[lukeinglis]

Thanks for the review @xukai92. Heads up: this PR has been significantly reworked since your review.

Scope change: Replaced the Qwen Code runner with an OpenCode runner (opencode.ai). The runner pattern and test structure remain the same, but the underlying CLI mapping is different.

Your P2 is addressed: --dangerously-skip-permissions is now gated on the parameter in both headless() and interactive_run(), matching how ClaudeRunner handles it. A dangerously_skip_permissions=False test is included for both methods.

Key implementation details:

• OpenCode has no --append-system-prompt, so the role prompt is prepended to the task message
• Headless uses opencode run --dir --format default
• Interactive uses opencode run --interactive --dir
• No auth warning needed since OpenCode handles provider auth internally via opencode auth login

Files changed: factory/runners/opencode.py (new), tests/test_opencode_runner.py (new, 20 tests passing), factory/runners/init.py, CLAUDE.md. Old qwen.py and test_qwen_runner.py are deleted.

Ready for re-review when you get a chance.