Agent-Ecosystem Watch

A standing daily radar on the thing that compounds OmegaOS's edge: how fast you absorb new Claude and agent techniques. It watches X for concrete improvements, tells you what is worth wiring into OmegaOS, and shares the best of it on @Agentik_os on autopilot, behind a gate so your public account never ships a weak or wrong tweet.

Same input feeds three outputs: (1) a decision report for you, (2) a curated best-practice tweet feed, (3) a growing memory of what the ecosystem shipped. Marginal cost is near zero once built; the read is a few dollars a month of ScrapeCreators.

Pipeline

cron (daily) → [1] READ  watchlist topics → last30days --search x (ScrapeCreators)
             → [2] ANALYZE (claude, sonnet)  extract + score integratable_to_omega 0-10
             → [3] GATE (claude, opus)  adversarial: try to REJECT each candidate tweet
             → [4] PUBLISH (if ARMED, not mock, under cap, unseen)  omega-zernio → @Agentik_os
             → [5] REPORT  self-contained HTML in ~/.omega/artifacts/
             → [6] ALERT   Telegram (Alerts topic) via omega-alert-send.sh

Deterministic control (daily cap, dedup persistence, publish side-effects) lives in shell; judgment (extract, score, draft, refute) lives in the model. The gate is an INDEPENDENT adversarial pass (R-VERIFY): its only job is to reject, default-reject on doubt.

Publishing safety rails (non-negotiable)

Publishing to the public @Agentik_os account requires ALL of:

• armed flag present (~/.omega/state/ecosystem-watch/armed) — the kill-switch. Absent = tweets are drafts in the report only. rm the flag to stop all publishing instantly.
• a real ScrapeCreators key (mock runs never publish).
• the adversarial gate returned keep=true for that tweet.
• the tweet fingerprint is not already in the seen-store (never tweet the same idea twice).
• the daily cap (EW_DAILY_TWEET_CAP, default 5) is not exhausted.
• no em/en dash, <= 280 chars, carries its source URL (enforced in shell too, R-NODASH).

Integration of a finding into OmegaOS itself is always a human call: the report proposes with a 0-10 integratability score, the operator decides.

Run

bash scripts/watch-run.sh                    # full run (reads config/topics.txt)
EW_MOCK=1 bash scripts/watch-run.sh          # fixtures, never publishes (smoke test)
EW_ZERNIO_DRYRUN=1 bash scripts/watch-run.sh # validate zernio post without publishing

Env knobs: EW_DAILY_TWEET_CAP (5), EW_DAYS (2), EW_ANALYZE_MODEL (claude-sonnet-5), EW_GATE_MODEL (claude-opus-4-8), EW_ZERNIO_PROJECT (agentik-os), EW_TOPICS_FILE.

Arm / disarm

touch ~/.omega/state/ecosystem-watch/armed    # go fully auto (auto-publish)
rm    ~/.omega/state/ecosystem-watch/armed     # kill-switch: drafts only

Config

• config/topics.txt — one X query per line (broad AI-agent ecosystem, Claude-first).
• config/accounts.txt — reference builders, an authority hint for scoring.

State (outside repo, R-ENV)

• ~/.omega/state/ecosystem-watch/seen.jsonl — published fingerprints (dedup memory).
• ~/.omega/state/ecosystem-watch/runs/<date>/ — per-run raw pulls, analysis, report.
• ~/.omega/artifacts/ecosystem-watch-<date>.html — the daily report.
• ~/.omega/logs/ecosystem-watch.log — run log.

Cron

# 08:00 daily — disarmed by default until the operator arms it
0 8 * * * OMEGA_DIR=$HOME/.omega bash $HOME/.omega/skills/agent-ecosystem-watch/scripts/watch-run.sh

Requires

• ~/.omega/secrets/integrations.env with SCRAPECREATORS_API_KEY (X read) and ZERNIO_API_KEY (publish). Both already present in this deployment.
• omega-zernio with an agentik-os profile whose twitter account is connected.
• last30days skill checkout at ~/.omega/repos/last30days-skill/.
• claude CLI on PATH.