Evidence-first tooling for AI coding agents.
Don't trust the model's self-report. Make every claim checkable.
codeblast · paired-eval · AgentXRay · coding-agent-internals · blog
Every tool here exists because an LLM told me something that turned out to be false — a diagram that was the model's opinion, a judge score that could be bluffed, a "refactor" that dropped a symbol, a benchmark number with no source. So each one replaces a claim with a check:
| The claim it refuses to take on faith | What it does instead | |
|---|---|---|
codeblast npx codeblast |
"This change only affects X" | Deterministic code graph from tsc/AST. impact / change / archmap, every edge with file:line. Recall verified by mutation testing (tRPC, 950 files: 28/28), blind spots reported, never hidden. Agent skill + PR bot. |
paired-eval pip install paired-eval |
"Model A scored 0.72, B scored 0.65, so A is better" | Program gate first, rubric only for what survives; paired McNemar/permutation with Holm and bootstrap CI; verdicts say significant / bounded null / uninformative / powerless — never a bare p. Zero dependencies. |
AgentXRay npx @alloevil/agent-xray |
"The agent did what I asked" | Read the actual session log — Claude Code, Codex, OpenClaw — turn by turn, tool call by tool call. |
| coding-agent-internals | "Agent X supports feature Y" | How 12 coding agents implement search, edit, LSP, DAP, sub-agents — shell-fork vs in-process, str_replace vs hash-anchored. Ceilings, not checklists. |
| agents-with-receipts | Any practice claim about agentic coding | Cell-by-cell sourced comparison table + an AGENTS.md linter. |
| deepresearch-arms-lab | "Deep research pipeline X helps" | 14-arm ablation on a weak base model with the negative results left in and an n=15 re-validation. |
Also: weibo-chat-auto (desktop archiver with daily AI digest), llm-benchmarks-tracker (benchmark catalogue with provenance on every score), pocket-agent (archived).
How the repos are run. Ship on a registry (npm / PyPI), one command to first output, README first
screen readable in 30 seconds, demo output in the README is regenerated by a test so it cannot drift from
the code, and a CI gate on the property the tool actually promises (codeblast: 100% recall on a weekly
mutation run; paired-eval: a mutation ratchet on the statistics).
Reading lately
Coding activity
I'm an Early 🐤
🌞 Morning 1684 commits █████████████░░░░░░░░░░░░ 52.28 %
🌆 Daytime 1251 commits ██████████░░░░░░░░░░░░░░░ 38.84 %
🌃 Evening 274 commits ██░░░░░░░░░░░░░░░░░░░░░░░ 08.51 %
🌙 Night 12 commits ░░░░░░░░░░░░░░░░░░░░░░░░░ 00.37 %
📅 I'm Most Productive on Monday
Monday 1492 commits ████████████░░░░░░░░░░░░░ 46.32 %
Tuesday 182 commits █░░░░░░░░░░░░░░░░░░░░░░░░ 05.65 %
Wednesday 183 commits █░░░░░░░░░░░░░░░░░░░░░░░░ 05.68 %
Thursday 527 commits ████░░░░░░░░░░░░░░░░░░░░░ 16.36 %
Friday 188 commits █░░░░░░░░░░░░░░░░░░░░░░░░ 05.84 %
Saturday 58 commits ░░░░░░░░░░░░░░░░░░░░░░░░░ 01.80 %
Sunday 591 commits █████░░░░░░░░░░░░░░░░░░░░ 18.35 %
📊 This Week I Spent My Time On
🕑︎ Time Zone: Asia/Shanghai
💬 Programming Languages:
Python 4 hrs 29 mins ███████████░░░░░░░░░░░░░░ 42.69 %
Bash 1 hr 42 mins ████░░░░░░░░░░░░░░░░░░░░░ 16.24 %
Markdown 1 hr ██░░░░░░░░░░░░░░░░░░░░░░░ 09.63 %
Other 49 mins ██░░░░░░░░░░░░░░░░░░░░░░░ 07.87 %
TypeScript 46 mins ██░░░░░░░░░░░░░░░░░░░░░░░ 07.34 %
🔥 Editors:
OMP 5 hrs 10 mins ████████████░░░░░░░░░░░░░ 49.14 %
Cursor 2 hrs 35 mins ██████░░░░░░░░░░░░░░░░░░░ 24.59 %
Claude Code 2 hrs 25 mins ██████░░░░░░░░░░░░░░░░░░░ 22.97 %
Agent 12 mins █░░░░░░░░░░░░░░░░░░░░░░░░ 02.02 %
Unknown Editor 8 mins ░░░░░░░░░░░░░░░░░░░░░░░░░ 01.28 %
🐱💻 Projects:
isc-world-model 4 hrs 51 mins ████████████░░░░░░░░░░░░░ 46.14 %
llm-benchmarks-tracker 1 hr 22 mins ███░░░░░░░░░░░░░░░░░░░░░░ 13.03 %
codeblast 1 hr 10 mins ███░░░░░░░░░░░░░░░░░░░░░░ 11.23 %
openclaw-changelog 1 hr 10 mins ███░░░░░░░░░░░░░░░░░░░░░░ 11.17 %
ax 29 mins █░░░░░░░░░░░░░░░░░░░░░░░░ 04.70 %
💻 Operating System:
Linux 10 hrs 31 mins █████████████████████████ 100.00 %
I Mostly Code in Python
Python 13 repos █████████░░░░░░░░░░░░░░░░ 34.21 %
TypeScript 8 repos █████░░░░░░░░░░░░░░░░░░░░ 21.05 %
JavaScript 8 repos █████░░░░░░░░░░░░░░░░░░░░ 21.05 %
HTML 3 repos ██░░░░░░░░░░░░░░░░░░░░░░░ 07.89 %
Shell 1 repo █░░░░░░░░░░░░░░░░░░░░░░░░ 02.63 %
Last Updated on 06/09/2026 01:09:17 UTC