[Feature] Allow defining harness settings (e.g. Claude settings.json) in a custom agent

[jan21deepak]

Problem or use case

Custom agents currently have no way to specify harness-level configuration, such as a Claude Code settings.json (permissions, hooks, env vars, etc.). Anyone wanting a custom agent to run with specific harness settings has to configure that outside of the agent definition, which means the settings aren't portable with the agent and have to be reapplied per environment.

Proposed solution

Support defining harness settings (e.g. a Claude settings.json-shaped config block) directly in a custom agent's definition, so the settings travel with the agent and are applied automatically whenever that agent runs.

Alternatives considered

Manually configuring harness settings outside the agent definition works but isn't portable and has to be redone per host/environment.

---

Imported from the parent repository for autonomous engineering evaluation.

Source issue: omnigent-ai#4496

[jan21deepak]

✅ Task completed

• Cursor agent ID: bc-0d9a1fcb-ae43-413b-82ef-e796c59947a6
• Pull Request: [Feature] Allow defining harness settings (e.g. Claude settings.json) in a custom agent (#73) #84
• Runtime: 4m 19s
• Completed at: 10 Aug 2026 23:20 SGT

Summary of changes:

Implemented support for portable harness settings in custom agent definitions, resolving issue #73.

What changed

Agents can now declare Claude Code settings.json-shaped configuration under executor.config.harness_settings in their config.yaml:

executor:
  type: omnigent
  config:
    harness: claude-native
    harness_settings:
      permissions:
        defaultMode: acceptEdits
      env:
        MY_FLAG: "1"

claude-native: Settings merge into the invocation-local --settings sidecar. Omnigent framework keys (hooks, statusLine, apiKeyHelper) and launch-time overrides (model, permission mode, effort) take precedence.

claude-sdk: Settings thread through HARNESS_CLAUDE_SDK_SETTINGS_OVERLAY and merge into the SDK settings option at turn time (apiKeyHelper from auth still wins).

Files added/modified

• omnigent/harness_agent_settings.py — extraction, serialization, merge helpers
• omnigent/spec/parser.py — preserve harness_settings as structured YAML
• omnigent/claude_native_bridge.py — merge agent settings in build_hook_settings / augment_claude_args
• omnigent/runner/native/orchestration.py — pass agent settings at claude-native launch
• omnigent/runtime/workflow.py — serialize settings for claude-sdk spawn env
• omnigent/inner/claude_sdk_harness.py / claude_sdk_executor.py — consume and apply overlay
• Tests in tests/test_harness_agent_settings.py plus updates to spawn-env/harness tests

How to test

uv run pytest tests/test_harness_agent_settings.py -q
uv run pytest tests/runtime/test_claude_sdk_spawn_env.py::test_harness_settings_are_serialized_for_claude_sdk -q
uv run pytest tests/inner/test_claude_sdk_harness.py::test_settings_overlay_is_resolved_from_env -q
uv run pytest tests/test_claude_native_bridge.py -k "augment_claude_args or build_hook_settings" -q

For a live check, add harness_settings to a custom agent bundle and start a session with that agent on claude-native or claude-sdk; the declared permissions/env should appear in the effective Claude settings without manual host configuration.

Branch

cursor/harness-settings-in-agent-47a6 — pushed to https://github.com/jan21deepak/omnigent

PR creation was not auto-enabled in this environment; you can open one from the branch when ready.