Persistent agent memory, deterministic verification and operations across machines.

The work here asks a practical question: what evidence shows that an AI agent did what it claimed? The repositories include small tools, fault-injection experiments, operating patterns and experience reports. Each project documents its own tests and limits.

Portfolio and resume · All upstream contributions · Research and publications · ORCID

Lost among the repositories? Of the 200+ public ones, about 150 are working forks used to send fixes to other projects. Our own work is a short list. Read in this order:

1. second-brain-starter-kit: the entry point, a second brain for a coding agent that keeps receipts.
2. sqlite-graph-memory: the memory it runs on.
3. claude-bible: the rules that keep the agent consistent.
4. claw-consensus and charm-os: several agents and people working together.
5. verbatim-citation-gate: checking that an agent's quotes are real.
6. Connectors: WhatsApp and Telegram.

Everything else is a smaller tool or an experiment, listed in the catalog.

Practical connectors: WhatsApp MCP kit · Telegram MCP kit.

Governance and operations: claude-bible · agent-leash · fleet-deploy · secondop-panel.

Complete public repository catalog · How the pieces fit: SYSTEM.md

• Microsoft Semantic Kernel #14371: pin the validated address for Python OpenAPI-plugin requests, closing a check/use gap. Merged 5 October 2026.
• Google ADK for Go #1543: images and PDFs (InlineData/FileData parts) now reach both OpenAI endpoints instead of failing as an unsupported payload; closes #1333, design agreed in the review thread. Merged 16 October 2026.
• Google Chrome DevTools MCP #2852: documentation fix for which extension tools are available from the CLI. Merged 20 October 2026.
• NVIDIA SkillSpector #807: documentation fix for broken source links in the LLM analyzer docs. Merged 14 October 2026.
• UK AISI Inspect #4769: decide a grader panel by strict majority rather than mode. Merged 1 September 2026.
• Obsidian Second Brain #377: flag a source that calls itself unverified but has no source-quality field. Merged 28 October 2026.

These are contributions to other owners' public repositories. Merged work, open proposals and closed submissions remain distinct in the full contribution ledger.

Live from the GitHub API, verified 2026-11-04 UTC. Own repositories and awesome-/registry-list entries are not counted.

70 pull requests merged into 42 upstream projects (8 of them at large tech companies and labs) · 50 more open in review across 47 projects · full list with live status

Merged, large vendors first:

• microsoft/semantic-kernel: 1 merged, last 2026-10-05.
• ChromeDevTools/chrome-devtools-mcp: 1 merged, last 2026-10-20.
• google/adk-go: 1 merged, last 2026-10-16.
• google-gemini/cookbook: 1 merged, last 2026-08-17.
• NVIDIA/SkillSpector: 1 merged, last 2026-10-14.
• QwenLM/qwen-code: 2 merged, last 2026-10-08.
• bytedance/deer-flow: 1 merged, last 2026-09-30.
• UKGovernmentBEIS/inspect_ai: 1 merged, last 2026-09-01.
• headroomlabs-ai/headroom: 3 merged, last 2026-11-04.
• ctxrs/ctx: 1 merged, last 2026-11-03.
• gwern/gwern.net: 1 merged, last 2026-10-28.
• eugeniughelbur/obsidian-second-brain: 4 merged, last 2026-10-28.

In review now:

• microsoft/semantic-kernel: Python: document connector extras and optional dependency compatibility, opened 2026-09-01.
• google/adk-go: fix(configurable): apply config_path containment to workflow node refs, opened 2026-08-10.
• anthropics/claude-cookbooks: fix: convert remaining broken sre_bot_slack.py links in SRE cookbook, opened 2026-07-25.
• openai/openai-cookbook: Add cookbook: check citation faithfulness in RAG with a zero-token gate, opened 2026-07-24.
• QwenLM/qwen-code: test(core): characterize leader delivery and terminal state per host class, opened 2026-10-20.
• QwenLM/qwen-code: fix(mcp): name the capped server timeout in App read timeout warnings, opened 2026-10-03.

• Persona portability benchmark: a frozen persona and memory, tested across models with blind judging and contamination checks. Read the dated results and limitations before comparing scores.
• Python vulnerability-patching probe: an audit of advisory leakage, fix dates, hint ablations and self-reported success in patching benchmarks.
• Brain Lite: a small receive-only library for another person's coding agent, with checksummed updates and rollback.
• Control-plane casebook: reproducible incidents involving schedulers, registries and watchdogs.
• Deep research collection: published research summaries with their questions, findings and source links.

Preprints, PDFs and DOIs · Full publication list

Clawrush records what was built, what broke and what was learned. The feed below updates automatically. The Journey is the longer build-in-public book, with human-readable chapters and a machine-readable edition.

• 2 November 2026: The Answer Was Already on the Desk  _{Nov 3, 2026}
• 1 November 2026: The Experiment We Already Ran and Stopped Measuring  _{Nov 2, 2026}
• 30 October 2026: Claims That Nothing Verifies  _{Nov 2, 2026}
• 29 October 2026: Five Things to Build, Five Things Already Standing  _{Oct 29, 2026}
• 28 October 2026: Somebody Else's Database, and Somebody Else's Gate  _{Oct 29, 2026}

Small tools with inspectable outputs. SQL and files for durable facts; models for judgment and language; tests and independent review for checking the result. A successful command, a delivered file and a verified outcome are different things.

This profile and the repositories are maintained with AI collaboration, including Claude, Codex, Grok and Gemini. Anton owns the work's direction and decisions. Tool roles vary by task; commit credits reflect actual contributions. AI collaboration policy.

Engineers are invited to try a project on a real task and report what broke, with a reproducer or a test result.

• WhatsApp: +1 341 222 9178
• Telegram: @tonydzi · English updates: @ClawEng · Russian updates: @ClawRus
• X: @Tony_Stef_

github.com/tonydzi