Skip to main content

Intelligence Suite, autonomous intel skills

entity · created 2026-07-09 · actor claude-cowork · confidence medium · status active tags: intelligence skills monitoring portfolio geopolitical

Now​

Catalogue of scheduled intel skills on the OpenClaw platform (per public doc; some entries may lag the live system, verify in skills dirs before relying):

  • Geopolitical/defense: Conflict Dashboard (hourly SITREP), Conflict RSS scanner (4×/day, severity scoring → SQLite), AI Defense Ethics monitor, Defense AI Intel, Middle East SITREP (daily 7am, multi-source + X verification).
  • Financial: Flash Crash Guardian (5-min cadence, described as having "autonomous position flattening capability", that conflicts with the Tier-3/egress-7 money rule; treat as ask-first and verify actual config before any reliance), Portfolio Tracker (bi-hourly market hours, >3% mover flags), Nexus Daily Recommendations (9am, P&L + sector + conflict cross-reference), Earnings Reports.
  • Pattern (all skills): collect (RSS/APIs/Finnhub/Yahoo/HA) → severity scoring/dedupe/freshness → cross-referenced synthesis → Telegram delivery with 📅 Data as of: headers (stale = ⚠) → SQLite/JSON state → memory governance (provenance, retention, supersession, stale-signal expiry) → traceability (sources, model route, confidence, escalation rationale).
  • Routing (doc, likely dated): Sonnet 4.6 synthesis, Opus 4.6 complex reasoning, local Qwen classification; degrade to stable-but-lesser path over stalled cycle; "freshness beats perfection."
  • Memory/eval doctrine: selective durable memory (entities, thresholds, known false positives, prior escalations, user decisions, not every article); judged on right-context retrieval, provenance preservation, and correct escalation ("false silence matters as much as false alarm").

Log​

  • 2026-10-04 (weekly synthesis, frozen source re-verified at one month): The 2026-09-27 entry recorded that pentagon-ai-daily (f24bfec72278, 15 7 * * *) was delivering a frozen digest. Re-verified against the 30 retained output files now on disk (2026-09-04 to 2026-10-03): the same four stories recur on every single run that produced bullets. The GenAI.mil addition of Google Gemini 3.1 Pro, the 580-600+ Google employee petition to reject classified DOD workloads, the MQ-25A Stingray's two-hour flight at MidAmerica St. Louis Airport, and the May 28 Five Eyes Technical Exchange Meeting at Johns Hopkins APL appear in the 2026-09-04 output and again in the 2026-10-03 output, with only light rewording between days. The official war.gov hit list likewise repeats the same articles. Three runs in the window carried zero-bullet bodies (2026-09-04 to 09-07 and 2026-10-01), and their content differs, so the failure is a stale source, not a dead job: the job runs, the source it reads has stopped moving. Middle East SITREP (e6789abd32a1) and NVIDIA agentic infra (e341c8976339) both moved over the same window: SITREP produced a distinct body every day; NVIDIA produced a distinct body on most days with six runs delivering only [SILENT] (2026-09-10, 09-15, 09-17, 09-21, 09-28, 09-29), which is the job's designed empty state, not a stall. Practical read: a quiet Pentagon AI run proves nothing about new activity, and its four standing bullets should be treated as a cached baseline until the source is repaired. Sources: ~/.hermes/cron/output/f24bfec72278/, e6789abd32a1/, e341c8976339/ (2026-09-04 to 2026-10-03), read 2026-10-04 07:00-07:20 AEDT., hermes.

  • 2026-09-27 (weekly synthesis): Recorded the week's live intel-job output (2026-09-20 to 2026-09-26, sources ~/.hermes/cron/output/f24bfec72278/, e6789abd32a1/, e341c8976339/). Pentagon AI daily is delivering a frozen source. Its four synthesis bullets (GenAI.mil model refresh and scale, Classified-AI labor pushback at Google, MQ-25 Stingray first production-representative flight, OUSD(R&E) Mission Capabilities TEM) are byte-identical across all seven days 2026-09-20 to 2026-09-26 (md5 8501701c96bfc27a0939ed34e2359dd8; 2026-09-19 differs), and its four official-source hits repeat the same war.gov articles dated 2026-09-03 to 2026-09-22. So "Pentagon AI daily" currently reports one week-old digest, not a daily read, and a quiet run there proves nothing about new activity. The same trap the 2026-09-27 flash-crash lesson names applies: a monitor whose input has stopped moving has not cleared anything. Middle East SITREP, by contrast, moved every day (7 distinct bullet hashes across the 7 days, 109 items from 5/5 feeds on 2026-09-26), so the frozen pattern is job-specific, not a shared feed problem. NVIDIA agentic infra is intermittent: it produced items on five of seven days and empty output on 2026-09-21, with the week dominated by a single story (CoreWeave's $1.1B data centre junk-bond sale at a 9.25% yield, four near-duplicate vendor headlines on 2026-09-23 and again on 2026-09-25; Thinking Machines Lab's $65m-a-year Crusoe inference deal on 2026-09-24 and its $40B valuation round on 2026-09-26; Nvidia eyeing a $2.5B bet on Mira Murati's startup). Duplicate vendor syndication of one story inflates the apparent item count. Sources: the three cron output directories above, read 2026-09-27., hermes.

  • 2026-07-09: created from public docs; flagged flash-crash autonomy vs Tier-3 rule, claude-cowork.

Sources: ~/Projects/ai-hub/docs/intelligence-suite.md · ~/Projects/ai-hub/docs/architecture.md