AI Native Lang

OpenClaw requests 2–6 → AINL v1.2.8 mapping (docs-only, honest)

> Current ainl release: v1.3.3 (see docs/RELEASENOTES.md). The body below describes the v1.2.8 request-mapping baseline; later releases add features (e.g. Hermes, OpenClaw CLI polish v1.3.0, native Solana v1.3.1) without

OpenClaw requests 2–6 → AINL v1.2.8 mapping (docs-only, honest)

Current ainl release: v1.3.3 (see docs/RELEASE_NOTES.md). The body below describes the v1.2.8 request-mapping baseline; later releases add features (e.g. Hermes, OpenClaw CLI polish v1.3.0, native Solana v1.3.1) without changing this lane analysis.

Purpose: This note maps OpenClaw’s “2–6” requests (summarizer, WASM compute, vector retrieval, per-feature caps, sparse attention) to what AINL v1.2.8 already implements, what is operator/host wiring, and what is out of scope for AINL’s lane (compile-once deterministic graph runtime).

Key principle (keep AINL in its lane): AINL does not rewrite workflow logic at runtime. It does not generate new .ainl graphs during execution. It does not do dynamic prompt optimization as a first-class runtime feature. What can adapt is the self-tuning resource & budget layer around execution: caps, hydration, pruning, embedding selection, and observability feedback loops driven by OpenClaw scheduling + configuration.

2) Session Summarizer (prevent unbounded growth)

What v1.2.8 ships

  • Summarizer program: intelligence/proactive_session_summarizer.lang
  • Runner wiring: python3 scripts/run_intelligence.py summarizer
  • Scheduling guidance: docs/operations/OPENCLAW_AINL_GOLD_STANDARD.md (cron)
  • Embedding-friendly output: session summaries store the actual summary text in payload.summary so the embedding pilot can index meaningful content (see docs/operations/EMBEDDING_RETRIEVAL_PILOT.md).

What it does (and does not do)

  • Does: produce compact, durable summary artifacts/records that OpenClaw can use for bootstrap and retrieval.
  • Does not: “replace old turns inside OpenClaw’s internal session DB” unless the OpenClaw host explicitly implements that behavior. The v1.2.8 posture is write summaries + prefer curated bootstrap.

Where it plugs into OpenClaw

  • Cron: OpenClaw triggers the intelligence run on a cadence (daily is the recommended default in the gold standard; higher frequency is a host choice).
  • Bootstrap: OpenClaw should prefer injecting curated context (session_context.md and/or summary artifacts) over dumping raw history (see docs/operations/OPENCLAW_HOST_AINL_1_2_8.md).

Why it saves tokens

  • Keeps bootstrap context and retrieval payloads bounded; avoids “zombie context” growth.
  • Complements rolling budget hydration (budget_hydrate) and caps so summarization remains cost-aware.

3) WASM for compute-heavy operations (deterministic transforms)

What v1.2.8 ships

  • WASM adapter support: AINL can call deterministic compute via the wasm adapter.
  • Operator notes: docs/operations/WASM_OPERATOR_NOTES.md

What it does (and does not do)

  • Does: move parsing/aggregation/scoring into deterministic compute so LLM prompts contain only the minimal derived summary when needed.
  • Does not: ship a “standard library” of Rust/C WASM modules for every OpenClaw workload unless the operator provides them. v1.2.8 is the call surface, not a full compute pack.

Where it plugs into OpenClaw

  • In scheduled workflows / bridges: load input via fs/memory, call WASM, then store a compact result (or send a small digest to an LLM adapter).

Why it saves tokens

  • Removes prompt-based data parsing and transformation (often thousands of tokens) from recurring jobs.

4) Vector search for memory retrieval (top-k relevance)

What v1.2.8 ships

  • Embedding storage + search: runtime/adapters/embedding_memory.py
  • OpenClaw bridge verbs: embedding_workflow_index / embedding_workflow_search
  • Pilot doc: docs/operations/EMBEDDING_RETRIEVAL_PILOT.md
  • Startup context integration (optional): intelligence/token_aware_startup_context.lang supports an embedding-first candidate path when enabled.

What it does (and does not do)

  • Does: retrieve top-k relevant snapshots/snippets instead of listing many memory hits into a prompt.
  • Does not: guarantee embeddings are always enabled. It is explicitly optional and gated by:
    • AINL_STARTUP_USE_EMBEDDINGS=1 (selection logic enabled)
    • AINL_EMBEDDING_MODE != stub (real embeddings configured)

Where it plugs into OpenClaw

  • Startup context: embedding-first candidate selection can reduce session_context.md size versus filesystem-only selection.
  • Retrieval: OpenClaw bridge tools can query embeddings to fetch only the highest-signal records.

Why it saves tokens

  • Reduces retrieval payload size and improves relevance, which reduces follow-up turns and “context stuffing.”

5) Strict per-feature token caps (budget discipline)

What v1.2.8 ships

  • Caps staging + operator discipline: docs/operations/TOKEN_CAPS_STAGING.md
  • Usage observability + alerts + trends: docs/operations/TOKEN_AND_USAGE_OBSERVABILITY.md
  • Sizing probe: ainl bridge-sizing-probe (suggests a sane AINL_BRIDGE_REPORT_MAX_CHARS)
  • Startup token clamps: AINL_STARTUP_CONTEXT_TOKEN_MIN / AINL_STARTUP_CONTEXT_TOKEN_MAX (intelligence-side budget)
  • Rolling budget hydration: scripts/run_intelligence.py merges rolling budget into MONITOR_CACHE_JSON (budget_hydrate)
  • Cap auto-tuner (resource/budget layer): scripts/auto_tune_ainl_caps.py (invoked via run_intelligence.py auto_tune_ainl_caps)

What it does (and does not do)

  • Does: enforce and tune caps at the resource and interface boundaries (bridge report size, startup context allocation, gateway caps, rolling budgets, prune/cleanup flows documented in ops).
  • Does not: assume the runtime has a universal core.token_count primitive or that every program dynamically measures LLM tokens mid-run. In v1.2.8, caps are enforced via configuration + disciplined surfaces + observability.

Where it plugs into OpenClaw

  • Host config (openclaw.json env vars), cron schedules, and gateway process env for promoter caps.
  • Observability loops feed back into tightening/loosening caps via the auto-tuner and operator review.

Why it saves tokens

  • Prevents one monitor/digest from ballooning and starving the rest; makes “budget posture” explicit and enforceable.

6) Sparse attention in model calls (provider-dependent)

What v1.2.8 ships

  • No provider-specific sparse-attention switch is claimed as shipped in v1.2.8.

Honest guidance

  • Even when providers offer compute-side optimizations, billing is usually per token; the strongest, reproducible savings come from sending fewer low-value tokens.
  • The v1.2.8 path to “sparse-effective” behavior is: tight caps + hydration + summarization + embedding top-k, not a magic flag.

Bottom line

AINL v1.2.8 already provides a self-managing with adaptive intelligence posture via a self-tuning resource & budget layer (caps, hydration, pruning, embedding retrieval, observability, and cap auto-tuning) layered around a compile-once deterministic graph runtime. This is the correct technical foundation for sustaining ~84–95%+ (often 90–95%) savings on OpenClaw-style scheduled monitoring/digest workflows without claiming runtime self-rewriting of workflow logic.