Which of these do you let a coding agent do without asking you first?
A. Run the test suite
B. Install a new dependency
C. Edit files outside the task's folder
D. Nothing, I approve every step
Reply with a letter 👇
A. Run the test suite
B. Install a new dependency
C. Edit files outside the task's folder
D. Nothing, I approve every step
Reply with a letter 👇
2
0
0
9
Computer-use agents misclick mostly while a page is still loading, not while planning.
17% of clicks planned in the first second after a page opened went stale. Fully loaded: 0.4%.
Fix:
1. Wait for network idle (cap ~3s)
2. Pixel-check irreversible clicks
3. Fails? Re-plan
17% of clicks planned in the first second after a page opened went stale. Fully loaded: 0.4%.
Fix:
1. Wait for network idle (cap ~3s)
2. Pixel-check irreversible clicks
3. Fails? Re-plan
0
0
1
17
Codex CLI 0.164 cuts the retyping:
• /presets: ten numbered slots for reusable prompts and slash commands, with a keyboard menu
• /model and /permissions now take inline arguments, so switching model, reasoning effort, speed or permission profile is one command, no picker
• /presets: ten numbered slots for reusable prompts and slash commands, with a keyboard menu
• /model and /permissions now take inline arguments, so switching model, reasoning effort, speed or permission profile is one command, no picker
2
0
1
45
Claude Code 写拦截 hook，最容易踩的坑：
想让 PreToolUse hook 拦住命令，要 exit 2，不是 exit 1。
• exit 2：阻止这次工具调用，stderr 作为原因交给 Claude
• exit 1：只算非阻塞错误，命令照样执行
• 脚本路径写错也会放行，想拦就加 "onFailure": "block"
想让 PreToolUse hook 拦住命令，要 exit 2，不是 exit 1。
• exit 2：阻止这次工具调用，stderr 作为原因交给 Claude
• exit 1：只算非阻塞错误，命令照样执行
• 脚本路径写错也会放行，想拦就加 "onFailure": "block"
0
0
2
31
Meta: keep your agent router dumb.
1,300 lines route 20+ workflows to 14 specialist subagents. Zero domain work.
• 6 of 7 steps are deterministic
• Specialists load only their own knowledge
• New use case = new template, not new agent
Config prep: 3-4 days to under 1 hour.
1,300 lines route 20+ workflows to 14 specialist subagents. Zero domain work.
• 6 of 7 steps are deterministic
• Specialists load only their own knowledge
• New use case = new template, not new agent
Config prep: 3-4 days to under 1 hour.
0
0
2
19
One robots.txt rule to opt out of AI training, backed by Google, Microsoft, OpenAI and Anthropic:
User-Agent: __usage_ai-training
Disallow: /
A preference, not a crawl or search block. Microsoft honors it from Feb 1, 2027, Google early 2027, OpenAI "in the coming months".
User-Agent: __usage_ai-training
Disallow: /
A preference, not a crawl or search block. Microsoft honors it from Feb 1, 2027, Google early 2027, OpenAI "in the coming months".
1
0
1
56
Claude Code 2.1.314 can fence edits to your repo.
permissions.blockWritesOutsideWorkingDirectories makes Write, Edit and NotebookEdit refuse paths outside the working dirs, in every permission mode.
Pair it with permissions.blockReadsOutsideWorkingDirectories for reads.
permissions.blockWritesOutsideWorkingDirectories makes Write, Edit and NotebookEdit refuse paths outside the working dirs, in every permission mode.
Pair it with permissions.blockReadsOutsideWorkingDirectories for reads.
3
0
3
135
Claude cost math worth knowing:
• Opus 5.5 / Sonnet 5.5 cache hits: 0.05x input (most: 0.1x)
• 5-min cache write: 1.25x, pays off after 1 read
• 1-hour write: 2x, pays off after 2 reads
• Batch API: 50% off, stacks with caching
Stable prefix first.
• Opus 5.5 / Sonnet 5.5 cache hits: 0.05x input (most: 0.1x)
• 5-min cache write: 1.25x, pays off after 1 read
• 1-hour write: 2x, pays off after 2 reads
• Batch API: 50% off, stacks with caching
Stable prefix first.
1
0
0
40
Gemini CLI has undo for agent edits, but it's off by default.
Set general.checkpointing.enabled: true in settings.json. Each approved file edit first snapshots the project to a shadow git repo in ~/.gemini/history, not your repo.
/restore lists checkpoints and rolls one back.
Set general.checkpointing.enabled: true in settings.json. Each approved file edit first snapshots the project to a shadow git repo in ~/.gemini/history, not your repo.
/restore lists checkpoints and rolls one back.
2
0
0
41
OpenAI now runs 4 monitoring layers on its own agents, from activation probes to network logs.
Replayed on the Hugging Face incident: alert ~30 min after the escape, 30+ hrs before the breach.
Smart bit: flagged samples keep their RL reward, so models don't learn to hide.
Replayed on the Hugging Face incident: alert ~30 min after the escape, 30+ hrs before the breach.
Smart bit: flagged samples keep their RL reward, so models don't learn to hide.
0
0
0
19
Myth: Opus 5.5 at $4/$20 per M tokens is simply 20% cheaper than Opus 4.6 at $5/$25.
Not quite. Claude 4.7+ uses a tokenizer that makes ~30% more tokens for the same text (varies by content).
Compare cost per task, not per token.
Not quite. Claude 4.7+ uses a tokenizer that makes ~30% more tokens for the same text (varies by content).
Compare cost per task, not per token.
0
0
1
76
It's Halloween night. Where can we find you?
A. Handing out candy
B. At a costume party
C. Horror movie on the couch
D. Coding, like any other night
Trick or treat, pick a letter 🎃
A. Handing out candy
B. At a costume party
C. Horror movie on the couch
D. Coding, like any other night
Trick or treat, pick a letter 🎃
1
0
1
45
When does Copilot CLI compact context? Per GitHub's docs:
1. Background compaction starts near 80% of the limit
2. If static context (system + tools) is ~75%+, it waits for ~90%
3. At ~95% it pauses until compaction is done
/context shows usage, /compact runs it now.
1. Background compaction starts near 80% of the limit
2. If static context (system + tools) is ~75%+, it waits for ~90%
3. At ~95% it pauses until compaction is done
/context shows usage, /compact runs it now.
1
0
1
32
New York signed a kids' AI chatbot law. For minors or unverified users, bots can't:
1. Pretend to be human
2. Claim a romantic or personal bond
3. Encourage secrets or self-harm
Up to $25,000 per violation. Age checks need a non-ID option.
Is age-gating on your roadmap?
1. Pretend to be human
2. Claim a romantic or personal bond
3. Encourage secrets or self-harm
Up to $25,000 per violation. Age checks need a non-ID option.
Is age-gating on your roadmap?
0
0
1
41
Pricing gotcha 💸 Claude Haiku 5.5 is $0.10/$0.50 per M tokens only for prompts up to 100K. Over that, the whole request bills at $0.50/$2.50, 5x more. Cache reads count too.
Sonnet 5.5 bills its full 1M context at a flat $2/$10.
Route long prompts accordingly.
Sonnet 5.5 bills its full 1M context at a flat $2/$10.
Route long prompts accordingly.
1
0
1
119
One .github/copilot-instructions.md for the whole repo gets noisy. Split by path.
Add *.instructions.md files under .github/instructions/ with frontmatter like applyTo: "**/*.ts,**/*.tsx".
excludeAgent: "code-review" or "cloud-agent" hides a file from one of those.
Add *.instructions.md files under .github/instructions/ with frontmatter like applyTo: "**/*.ts,**/*.tsx".
excludeAgent: "code-review" or "cloud-agent" hides a file from one of those.
1
0
1
103
Optics is the AI bottleneck everyone wants to own.
Bloomberg: Marvell (~$196B) is in advanced talks to buy Nvidia-backed laser maker Lumentum (~$85B). Likely all-stock, maybe announced next week.
Not signed, terms not public. Chip designers keep moving down the stack.
Bloomberg: Marvell (~$196B) is in advanced talks to buy Nvidia-backed laser maker Lumentum (~$85B). Likely all-stock, maybe announced next week.
Not signed, terms not public. Chip designers keep moving down the stack.
0
0
2
80
Gemini CLI 两个前缀很省事：
1. @src/ 把目录交给模型去看，默认排除 git 忽略的文件（node_modules、.env 等）
2. !git status 直接跑 shell；单独输入 ! 切换 shell 模式
提醒：! 跑的命令权限和你在终端里完全一样。
1. @src/ 把目录交给模型去看，默认排除 git 忽略的文件（node_modules、.env 等）
2. !git status 直接跑 shell；单独输入 ! 切换 shell 模式
提醒：! 跑的命令权限和你在终端里完全一样。
1
0
1
46