Skip to content

TODO.md

Latest commit

 

History

History
76 lines (62 loc) · 4.87 KB

TODO.md

File metadata and controls

76 lines (62 loc) · 4.87 KB

Kai CMO Harness — What's Next

Done (Mar 16, 2026)

  • Built Content Quality Scorer (28 rules, 4 categories, 3 output formats)
  • Built Local Gate (YAML policies, SQLite audit trail, approve/reject/hold)
  • Deployed quality scorer to Kai-CMO server (/opt/cmo-analytics/quality/)
  • Deployed 14 marketing skills to OpenClaw (129 → 143 total)
  • Synced 7 missing framework files to server knowledge base
  • Gateway integration (POST /webhooks/quality/score)
  • 4 gate policies: blog-publish, linkedin-article, cold-email, default

Phase 1: Wire Harness Through Quality Scorer ✅

Replace the 3 separate gate scripts with the unified quality scorer.

  • Refactor run_gate() in scripts/harness_cli.py to call scripts.quality.gate.propose() instead of four_us_score.py + banned_word_check.py + seo_lint.py
  • Map harness format types to gate policies (blog→blog-publish, linkedin→linkedin-article, email→cold-email, press→press-release, tiktok→tiktok-script, meta-ads→meta-ad, google-ads→google-ad)
  • Wire revise_draft() to use quality scorer violations as fix instructions (rule IDs, line numbers, exact suggestions)
  • Update post_for_approval() to include proposal_id + score + grade + approve command
  • Add policies: press-release.yaml, tiktok-script.yaml, meta-ad.yaml, google-ad.yaml
  • Test end-to-end: kai-harness run --task blog --site kaicalls --keyword "test"

Phase 2: Gate → Remote Zehrava Handoff

Approved content auto-proposes to Zehrava Gate for delivery.

  • On local gate approve (or human approves pending): call gate_client.propose(policy="blog-publish", destination="blog.publish", payload=draft)
  • Two gates in series: quality gate (instant, local) → action gate (remote, with delivery)
  • Add --deliver flag: python -m scripts.quality gate article.md --policy blog --deliver
  • Wire Discord notification for pending proposals (post to channel with approve/reject reactions)

Phase 3: Self-Improvement Loop ✅

Quality scores feed back into the system automatically.

  • Wire performance_check.py to run quality scorer on published content retroactively (retro_score_content())
  • Correlate quality scores with GSC/GA4 performance (position, CTR, session duration)
  • pattern_extract.py extracts which quality rules correlate with winners (--correlate flag)
  • Auto-update policy YAML thresholds when n>=5 winner patterns emerge (harness_defaults_update.py)
  • Cron: nightly batch score of all published content → --batch-score flag
  • Weekly: surface quality summary to Discord → --weekly-summary flag

Phase 4: Architecture Page Rewrite ✅

meetkai.xyz/architecture should reflect harness engineering approach.

  • Reframe as "harness engineering for marketing" (parallel to OpenAI Codex approach)
  • Show 4-layer stack: Runtime → Tooling → Domain Agents → Content Harness
  • Feature quality scorer as the core moat (research-as-automated-tooling)
  • Show feedback loop diagram: write → score → gate → publish → 30d check → pattern extract → update defaults
  • 14 marketing skills grid
  • Live quality scorer demo wired to POST /webhooks/quality/score

Phase 4.5: Code Hardening ✅

Address competitive review findings — production-grade code quality.

  • Centralized config: scripts/harness_config.py + config.yaml — all hardcoded paths, IDs, thresholds extracted
  • API timeouts: 30s timeout on all Gemini API calls (configurable via api_timeout)
  • Circuit breaker: fail after 3 consecutive API errors (configurable via api_max_retries)
  • Input sanitization: sanitize_input() strips control chars, truncates, escapes prompt-injection patterns
  • Replace os.system() with subprocess.run() — proper error handling, timeouts, return code checks
  • MARKETING.md backup: .bak created before every auto-update (policy YAML files too)
  • Fix GSC pagination: rowLimit: 1 → 10 with weighted aggregation across rows
  • Fix duplicate SHORT_FORM definition in harness_cli.py
  • Remove hardcoded Discord channel fallback — now from config.yaml
  • LRU cache for quality scoring — identical content not scored twice
  • Structured JSON logging across all scripts (logging module, not bare print())
  • Test suite: gate module tests (test_gate.py) added to existing AA/GEO/CS/engine/parser tests

Phase 5: Open Source Prep

Quality scorer as standalone pip package.

  • Extract scripts/quality/ into standalone package (no dependency on knowledge_cloner)
  • pyproject.toml with optional [llm] extra for Four U's
  • PyPI: pip install kai-quality or pip install content-quality-scorer
  • GitHub Actions CI: run tests on PR
  • Contributing guide: how to add new rules
  • Pre-built rule packs: --pack aeo, --pack conversion, --pack technical