Intelligence programs (intelligence/*.lang)

AINL sources under intelligence/ support OpenClaw-style automation: memory compaction, session signals, token-aware bootstrap context, and scheduled digests. They are examples / operator programs — not part of the core language spec. ZeroClaw users usually integrate via docs/ZEROCLAW_INTEGRATION.md (skill + ainl-mcp), not this monitor registry layout.

Authoring note: Before Call genmem/WRITE, Call accmem/LACCESS_READ, or R memory put, bind contract fields with Set (Set memory_namespace "workflow", Set memory_kind "…", …), not X. The X op requires a real function name (get, core.substr, …); see the callout under X in docs/AINL_SPEC.md §2.3.

Graph runtime: Intelligence programs run on the compiler-emitted graph by default. Do not use X dst {…} JSON object literals (use core.parse, obj/put, or arr — see intelligence/token_aware_startup_context.lang and modules/common/generic_memory.ainl). Do not use J OtherLabel to chain labels (J returns a value; use Call or same-label fall-through). For memory.list, use null for an omitted record_id_prefix, not "". metadata.valid_at must be an RFC3339 string (e.g. R core iso) when using the memory contract. Run tests with ./.venv-py310/bin/python -m pytest per docs/INSTALL.md.

Programs

| File | Role | |------|------| | intelligence_digest.lang | Scheduled web + TikTok snapshot, cache + memory + notify (see openclaw/INTELLIGENCE_DIGEST.md) | | memory_consolidation.lang | Keyword-based promotion from memory/*.md into MEMORY.md (no LLM) | | proactive_session_summarizer.lang | Summarize prior-day logs via OpenRouter; writes MEMORY.md; needs OPENROUTER_API_KEY + http allowlist | | token_aware_startup_context.lang | Builds compact .openclaw/bootstrap/session_context.md from MEMORY.md under a token budget | | token_aware_startup_context2.lang | Variant bootstrap writer (same general pattern) | | session_continuity_enhanced.lang | Lists session memory keys, heartbeat + notify (monitoring posture) | | store_baseline.lang | One-shot seed into memory (intel / baseline) | | test_split.lang | Small harness for split/len-style checks | | infrastructure_watchdog.lang | Service health checks + optional notify path (operator-tuned; pair with your gateway / bridge allowlists) | | signature_enforcer.py | Optional signature metadata checks (# signature: ...) + bounded retry helper for ptc_runner | | trace_export_ptc_jsonl.py | Exports AINL trajectory JSONL to PTC-compatible JSONL shape |

Local runner

scripts/run_intelligence.py compiles and runs selected programs with the OpenClaw monitor adapter registry (for hosts that mirror that layout):

# from repo root, with project on PYTHONPATH / editable install
python3 scripts/run_intelligence.py context
python3 scripts/run_intelligence.py summarizer --trace
python3 scripts/run_intelligence.py consolidation
python3 scripts/run_intelligence.py continuity
python3 scripts/run_intelligence.py signature_enforcer --dry-run
python3 scripts/run_intelligence.py trace_export_ptc_jsonl --dry-run
python3 scripts/run_intelligence.py all

Rolling budget → cache: On each non–dry-run start, the runner merges SQLite workflow / budget.aggregate / weekly_remaining_v1 (from bridge rolling_budget_publish) into workflow/token_budget in MONITOR_CACHE_JSON, so startup and summarizer gates align with weekly trends without scanning days of markdown. See docs/operations/TOKEN_AND_USAGE_OBSERVABILITY.md and docs/operations/TOKEN_CAPS_STAGING.md. JSON output includes budget_hydrate.

Enable the same adapters and paths your production gateway uses (fs, cache, http, memory, queue, etc.); see docs/INSTALL.md and docs/reference/ADAPTER_REGISTRY.md.

PTC Hybrid Patterns (Reliability + Emission)

For PTC-style reliability and hybrid overlays (all opt-in), use:

• intelligence/signature_enforcer.py
  • parses optional # signature: ... metadata
  • validates result shape/types
  • supports bounded retry helper for ptc_runner
• intelligence/trace_export_ptc_jsonl.py
  • converts AINL trajectory JSONL into PTC-compatible JSONL
  • strips _-prefixed keys during export (context firewall)
• intelligence/context_firewall_audit.py
  • audits trajectories and sources for accidental non-_ keys before ptc_runner / llm_query
  • surfaces warnings without changing graph semantics
• modules/common/ptc_parallel.ainl
  • orchestrates a list of ptc_runner calls in a pcall-style pattern
  • supports an optional inline max_concurrent cap
• modules/common/recovery_loop.ainl
  • wraps ptc_runner calls with bounded retries using the runner envelope (ok:false)
  • treats signature as runner-side validation input (or mock simulation)
• intelligence/ptc_to_langgraph_bridge.py
  • detects ptc_runner steps from source or trajectory
  • emits a copy-paste LangGraph bridge snippet with thin ptc tool nodes

Reference end-to-end flow:

• docs/adapters/PTC_RUNNER.md → Canonical End-to-End Example

Host responsibilities

• Cron / scheduler: programs declare S schedules; the host must trigger runs.
• Prompt injection: token-aware output is only useful if the host loads session_context.md / MEMORY.md into the agent context. Prefer curated bootstrap over full MEMORY.md when startup context has run — see docs/operations/AGENT_AINL_OPERATING_MODEL.md (host contract).
• Secrets: summarizer and digest flows need env vars (e.g. OPENROUTER_API_KEY) where applicable.

See also

• operations/OPENCLAW_AINL_GOLD_STANDARD.md — recommended cron, paths, and budget_hydrate checks for OpenClaw-style hosts
• agent_reports/README.md — operator field reports
• docs/RUNTIME_COMPILER_CONTRACT.md
• docs/OPENCLAW_ADAPTERS.md