Efficient mode bridge: AINL CLI ↔ ArmaraOS

This note ties together AI Native Lang (AINL) tooling and ArmaraOS Ultra Cost-Efficient Mode so operators do not confuse three different layers.

Three different things

| Layer | Where it runs | What it does | |-------|----------------|--------------| | Input compression | Rust (openfang-runtime / prompt_compressor) | Shortens the user message sent to the LLM before each agent turn. | | --efficient-mode / AINL_EFFICIENT_MODE | Python CLI sets env only | Signal to the host (e.g. ArmaraOS kernel injecting manifest metadata). No compression in Python. | | modules/efficient_styles.ainl | AINL graphs (optional) | Shapes model output (dense prose / JSON) via llm_query — complementary to input compression. |

CLI: ainl run --efficient-mode

• Sets AINL_EFFICIENT_MODE when not already set (values such as balanced, aggressive, off — host-defined).
• Does not implement compression in the Python runtime.

Module: modules/efficient_styles.ainl

• human_dense_response — natural, concise technical prose from prior node output.
• terse_structured — JSON-only internal steps.

Include with the usual AINL include prelude pattern (see file header comments). Replace llm_query with your deployment’s LLM adapter when not under ArmaraOS.

ArmaraOS authoritative behavior

Ground truth for retention rules, UI locations, API fields (compression_savings_pct, compressed_input), and logs: ArmaraOS repository:

• docs/prompt-compression-efficient-mode.md
• Root README.md — Ultra Cost-Efficient Mode section

Related

• AGENTS.md (AINL repo) — HTTP adapter, MCP ainl_run, and environment conventions.