Singh_AP

SDE. Views my own.

[Singh_AP @singh_builds]

Didn't help: predicting the hidden tests up front (wrong guesses became requirements), or swapping in Opus 5.5.

Caveats: offline subset of 46 tasks, 1 trial each, not leaderboard-comparable.

Same loop on Reasoning Gym: 20% to 91%.

Code: https://t.co/aYGztpL6AY

[Singh_AP @singh_builds]

I built a coding-agent harness that rewrites itself from its own failures.

Same model (Claude Sonnet 5.5), same tasks, only the harness changed: on Terminal-Bench 4.0 tasks it never saw, it went from 50% to 67%.

Based on the Harness Learning paper (Zhang et al., 2026). 🧵

https://t.co/zTMbnkJctU

[Singh_AP @singh_builds]

What changed:
- Resume replies cut off by the output limit (up to 1 in 3 runs silently stopped mid-task)
- Treat "infeasible" or a near miss as a bug signal
- Tie-breakers for judgment calls
- A reviewer with a failure checklist (a generic one passed 16/18, failed runs too)

[Singh_AP @singh_builds]

Corrections and contributions are very welcome, especially from people who work in these fields.

⭐ the repo if it's useful: https://t.co/RkNBbMryfC

[Singh_AP @singh_builds]

OpenAI released 722 math papers written by an internal AI model, with claimed results on problems like the Hodge conjecture (CM case), BSD, Unique Games and Kakeya.

They're written for specialists, so I started a repo that explains them to beginners.

https://t.co/RkNBbMr0q4

[OpenAI @OpenAI]

We’re releasing a broad range of new mathematical results produced by an internal frontier model.

We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.

https://t.co/7N6TPlft1P

[Singh_AP @singh_builds]

21 explainers so far, including:

• no zeros of the Riemann zeta function with real part > 7/8
• π's irrationality exponent is exactly 2
• Erdős's conjecture on arithmetic progressions
• the plane can't be 5-colored
• matrix multiplication in n^(9/4+ε)

[Singh_AP @singh_builds]

There are also 76 hand-picked community mods (24 install from the same marketplace, pinned to a reviewed commit) and a daily index of every public mod.

Search them all: https://t.co/dbLeVVK5Cs:

Install inside Claude Code:
/plugin marketplace add Singh-AP/awesome-claude-mods
/plugin install bash-guard@awesome-claude-mods

[Singh_AP @singh_builds]

Claude Code just got mods:
TypeScript hooks that run inside Claude Code itself. They can draw panes, intercept tool calls and change what the model sees.

I built 20, all tested, and indexed 2,392 more.

https://t.co/YzeCq3qRRw

[Singh_AP @singh_builds]

📡 tool-radar: every tool call live in a side pane, with args, duration and failures

🧪 test-pulse: your suite's pass/fail in the status line

💸 cost-meter: spend, context % and rate limits, plus what each turn cost

[Singh_AP @singh_builds]

There are also 76 hand-picked community mods (24 install from the same marketplace, pinned to a reviewed commit) and a daily index of every public mod.

Search them all: https://t.co/2wLtyqaN6G

[Singh_AP @singh_builds]

💡 command-explainer puts a plain-English line and a risk level right above the permission prompt:

"🛑 Danger: permanently deletes every .tmp file under /tmp/acm-demo older than 30 days"

Approve what you actually understand.

[Singh_AP @singh_builds]

Every mod has tests (793 so far), passes Claude Code's own static analyzer in strict mode, and was run in a real session.

CI re-tests everything weekly against the newest Claude Code, and a page lists exactly what each mod can touch.

[Singh_AP @singh_builds]

And some fun ones:

🐱 buddy: an ASCII pet above your prompt that celebrates green tests and levels up

🏴‍☠️ spinner-packs: "Polishing the cannons…" instead of "Sautéing…"

🎁 wrapped: Spotify Wrapped for your Claude Code life