ouroboros pm ignores configured Codex backend and always uses LiteLLMAdapter

[srausser]

Summary

ouroboros pm currently constructs LiteLLMAdapter() directly, so the PM interview path ignores the configured LLM backend and requires provider/API-key setup even when Ouroboros is configured for Codex.

Repro

1. Run ouroboros setup and select Codex runtime.
2. Confirm ~/.ouroboros/config.yaml shows:

llm:
  backend: codex
orchestrator:
  runtime_backend: codex

3. Run ouroboros pm on a stock install.

Actual

The PM path imports and instantiates LiteLLMAdapter directly from ouroboros/cli/commands/pm.py.

That means:

• Codex runtime selection does not control PM interview execution.
• PM can still require provider-specific model strings and API keys.
• Backend-safe defaults like default are not valid for the LiteLLM path.

Expected

The PM interview command should create its adapter through the provider factory so it respects the configured backend, e.g. Codex should use the Codex CLI adapter.

Evidence

Current code path in ouroboros/cli/commands/pm.py:

from ouroboros.providers.litellm_adapter import LiteLLMAdapter
adapter = LiteLLMAdapter()

There is already a factory abstraction in ouroboros.providers.factory.create_llm_adapter(use_case="interview") that resolves Codex correctly.

Local workaround

I patched my local install to replace the direct LiteLLMAdapter() construction with create_llm_adapter(use_case="interview"). After that, ouroboros pm started generating interview questions through the Codex-backed path instead of failing in the LiteLLM path.

[ouroboros-agent]

Triage result: bug, feasible for a single-PR fix.

Why:

• Reproduction steps are present.
• Expected vs actual behavior is clear.
• The report points to a likely root cause in src/ouroboros/cli/commands/pm.py, where the PM path constructs LiteLLMAdapter() directly instead of using the provider factory.
• No existing linked PR was found.

Routing:

• Added the bug label.
• This appears safe to proceed through the implementation workflow as a focused adapter-selection fix.

ouroboros-agent[bot]

[srausser]

Related to #240, but not a duplicate. #242 is the backend-selection bug: ouroboros pm constructs LiteLLMAdapter() directly instead of using the provider factory / Codex adapter path. #240 is the separate hardcoded Claude model default in the CLI option definition.

[ouroboros-agent]

Follow-up triage update:

Thanks for clarifying the split from #240. Keeping #242 scoped as a standalone bug for the backend-selection path in ouroboros pm:

• ouroboros pm ignores configured Codex backend and always uses LiteLLMAdapter #242 = adapter selection bug (LiteLLMAdapter() is constructed directly instead of using the provider factory)
• ouroboros pm hardcodes Claude model instead of using configured clarification model #240 = separate CLI/default-model behavior

This still looks like a focused single-PR fix in the PM interview path, not a duplicate. No linked PR exists yet.

ouroboros-agent[bot]