Summary

When the harness already has provider credentials configured for the agent, the plugin offers them to akm as the LLM backend for index-time passes (memory inference, graph build).

Today

If no LLM is configured for akm, the new index passes (inference, graph) cannot run.

Proposal

• Plugin detects that akm has no akm.llm configured.
• Plugin exposes a small shim (e.g., a local script or socket) and sets a CLI-side env / config such as AKM_LLM_PROXY_CMD=<shim path>.
• akm consults the proxy command when no native LLM is configured; the shim forwards prompts through the harness's existing provider connection and returns the result.
• No provider credentials cross the plugin / CLI boundary in plaintext — the shim handles invocation.

Why plugin, not CLI

Only the harness has the user-authenticated provider connection; akm itself stays harness-agnostic.

Open questions

• Exact shim contract (stdio / unix socket / file).
• Streaming vs. one-shot.
• Per-pass model selection.

Depends on

• itlackey/akm issue: "Single configurable LLM block reused across index passes" (the proxy hooks into the same config path).

Acceptance

• With no akm.llm configured but a harness LLM available, akm index passes succeed.
• Disabling the harness LLM falls back to no-op (not error) on index passes.
• Behaviour parity-matched in opencode/ and claude/.

Sourced from upstream plan: https://github.com/itlackey/openpalm/blob/claude/simplify-openpalm-akm-wsA6t/docs/plans/upstream-issues/akm-plugins-issues.md