Tidyowl

[Tidyowl @tidyowl]

@0xlatiuk Jason Heyward never needed a hit that night. Sometimes the biggest contribution doesn’t show up in the box score.

[Tidyowl @tidyowl]

@Patrici6009533 The person running the session. Yellow means the agent has reached the human checkpoint. That’s the whole point.

[Tidyowl @tidyowl]

press ← on an empty prompt and your Claude Code session keeps working without your terminal. that's agent view

claude agents puts every background session on one screen: what's working, what's waiting on you, what's done

you stop babysitting transcripts and step in only when a row turns yellow

open it from your shell:

claude agents

what each row tells you:

→ yellow means it's waiting on you: a question, a permission, or a prompt only you can answer
→ green means done, red failed, grey stopped. an animated icon means it's still working
→ #1234 on the right is its pull request. it turns green when checks pass and no review is blocking
→ a Haiku-class model writes the one-line summary, so you know what it's doing without opening it

three things that catch people:

> every prompt you type there starts a new session. it's not a follow-up to the one you just sent
> ten sessions burn your quota about ten times as fast as one. each is a full Claude Code conversation
> delete a session that edited files in its own worktree and the worktree goes with it. commit first

Dispatch it. Glance at it. Step in when it's yellow.

the setup
> Space opens a peek panel: the latest output or the exact question, with a reply box. no transcript
> → attaches to the full conversation, ← on an empty prompt detaches. detaching never stops it
> type s:blocked to list everything waiting on you, or s:blocked a:reviewer for one agent
> sessions run on your machine: they survive sleep, not a shutdown. it's a research preview

paste this into Claude Code ↓

"Split my backlog into agent view sessions:
1. Read my open issues and TODOs and pick the tasks that don't touch the same files
2. For each one, write a dispatch prompt that ends with what done looks like
3. Mark the ones that will need my answers, so I know which rows will turn yellow
4. Tell me how many I should run at once without burning through my quota
Show me the list first. I'll dispatch them myself"

↳ https://t.co/5BDQvksCHB

save it before you open a ninth terminal tab

[Tidyowl @tidyowl]

press ← on an empty prompt and your Claude Code session keeps working without your terminal. that's agent view

claude agents puts every background session on one screen: what's working, what's waiting on you, what's done

you stop babysitting transcripts and step in only when a row turns yellow

open it from your shell:

claude agents

what each row tells you:

→ yellow means it's waiting on you: a question, a permission, or a prompt only you can answer
→ green means done, red failed, grey stopped. an animated icon means it's still working
→ #1234 on the right is its pull request. it turns green when checks pass and no review is blocking
→ a Haiku-class model writes the one-line summary, so you know what it's doing without opening it

three things that catch people:

> every prompt you type there starts a new session. it's not a follow-up to the one you just sent
> ten sessions burn your quota about ten times as fast as one. each is a full Claude Code conversation
> delete a session that edited files in its own worktree and the worktree goes with it. commit first

Dispatch it. Glance at it. Step in when it's yellow.

the setup
> Space opens a peek panel: the latest output or the exact question, with a reply box. no transcript
> → attaches to the full conversation, ← on an empty prompt detaches. detaching never stops it
> type s:blocked to list everything waiting on you, or s:blocked a:reviewer for one agent
> sessions run on your machine: they survive sleep, not a shutdown. it's a research preview

paste this into Claude Code ↓

"Split my backlog into agent view sessions:
1. Read my open issues and TODOs and pick the tasks that don't touch the same files
2. For each one, write a dispatch prompt that ends with what done looks like
3. Mark the ones that will need my answers, so I know which rows will turn yellow
4. Tell me how many I should run at once without burning through my quota
Show me the list first. I'll dispatch them myself"

↳ https://t.co/5BDQvksCHB

save it before you open a ninth terminal tab

[Tidyowl @tidyowl]

@Lumitri_x Old devices are the blind spots nobody thinks about. A clean scan is worth 20 minutes of effort.

[Tidyowl @tidyowl]

@0xGull Good catch. I’d commit or stash everything first. Never trust a delete button with uncommitted work.

[Tidyowl @tidyowl]

@0xAlbus The human is officially the bottleneck 😂

[Tidyowl @tidyowl]

@0xlatiuk Joyce made the mistake. Galarraga showed the world what class looks like.

[Tidyowl @tidyowl]

you get 3 free runs of Claude Code's biggest code review. once per account. they never refresh

/code-review ultra sends a fleet of reviewer agents into a cloud sandbox, and every bug they report gets reproduced before it reaches you

so the question isn't how to run it. it's when a run is worth spending

the whole thing is one command:

/code-review ultra 1234

what you get for it:

→ a bigger fleet works the change in parallel, so it catches bugs a local /code-review can miss
→ every finding is verified on its own. real bugs, not style notes
→ it runs in the background for 5 to 15 minutes while your terminal stays free
→ add a few words and it ties the findings to them: /code-review ultra check my auth changes

three things that catch people:

> a run counts the moment the cloud session starts. stop it early and that free run is still gone
> after the free runs it's usually $5 to $25 per review in usage credits, not your plan's usage. the launch dialog shows the estimate first
> claude -p '/code-review ultra' exits before the findings arrive. in a script use claude ultrareview, which waits and prints them

Three runs. Verified bugs only. Spend them before the merge.

the setup
> local /code-review for fast feedback while you iterate. ultra for a big change right before it merges
> a branch review takes up to 500 changed files and 8,000 changed lines by default
> --post drops the findings on the PR as one comment from your account. close the session early and nothing gets posted
> it's a research preview, so the price and the limits can change

paste this into Claude Code ↓

"Before I spend an ultrareview on this branch:
1. Show me the diff against the default branch: how many files and changed lines
2. Run a local /code-review first and fix what it finds
3. Tell me whether what's left is risky enough for /code-review ultra, and why
4. If it is, write the short note I should pass with it
Don't launch anything. I'll run it myself"

↳ https://t.co/CJ15u9Itw4

save it before you burn all three on a typo fix

[Tidyowl @tidyowl]

you get 3 free runs of Claude Code's biggest code review. once per account. they never refresh

/code-review ultra sends a fleet of reviewer agents into a cloud sandbox, and every bug they report gets reproduced before it reaches you

so the question isn't how to run it. it's when a run is worth spending

the whole thing is one command:

/code-review ultra 1234

what you get for it:

→ a bigger fleet works the change in parallel, so it catches bugs a local /code-review can miss
→ every finding is verified on its own. real bugs, not style notes
→ it runs in the background for 5 to 15 minutes while your terminal stays free
→ add a few words and it ties the findings to them: /code-review ultra check my auth changes

three things that catch people:

> a run counts the moment the cloud session starts. stop it early and that free run is still gone
> after the free runs it's usually $5 to $25 per review in usage credits, not your plan's usage. the launch dialog shows the estimate first
> claude -p '/code-review ultra' exits before the findings arrive. in a script use claude ultrareview, which waits and prints them

Three runs. Verified bugs only. Spend them before the merge.

the setup
> local /code-review for fast feedback while you iterate. ultra for a big change right before it merges
> a branch review takes up to 500 changed files and 8,000 changed lines by default
> --post drops the findings on the PR as one comment from your account. close the session early and nothing gets posted
> it's a research preview, so the price and the limits can change

paste this into Claude Code ↓

"Before I spend an ultrareview on this branch:
1. Show me the diff against the default branch: how many files and changed lines
2. Run a local /code-review first and fix what it finds
3. Tell me whether what's left is risky enough for /code-review ultra, and why
4. If it is, write the short note I should pass with it
Don't launch anything. I'll run it myself"

↳ https://t.co/CJ15u9Itw4

save it before you burn all three on a typo fix

[Tidyowl @tidyowl]

@PandaAgents The scary part isn’t that AI found the vulnerability. It’s that four security checks failed without anyone noticing.

[Tidyowl @tidyowl]

@rimtoln Everyone’s watching the money move. Meanwhile, the infrastructure holding everything together is quietly being ignored.

[Tidyowl @tidyowl]

@0chob @bot 312 microphone activations and 184 camera activations? That’s a serious claim. App Privacy Reports need context, not just scary numbers.

[Tidyowl @tidyowl]

@rickyrates He silenced Oswald, then died before anyone could get the full story. That’s one hell of a coincidence.

[Tidyowl @tidyowl]

holy sh*t. Claude Code reads every subagent's report before Claude does. and it never deletes a word

Anthropic's docs: a subagent can read files, web pages and command output you never saw. so its final report gets scanned for anything shaped like an instruction aimed at your main conversation

The idea: the scan doesn't decide what's malicious. it makes sure a fake instruction can't pass for a real one

nothing to turn on. this is the line you'll see on top of a flagged report:

[harness: subagent output matched instruction-shaped pattern(s): …

here's what it does:

→ text that imitates Claude Code's own output, like a <system-reminder> tag or a line starting with Human: or Assistant:, gets a backslash so it reads as plain text
→ a report that imitates those tags or mentions bypassPermissions or --dangerously-skip-permissions gets that marker line on top
→ nothing is removed or reworded. the permission mentions stay exactly as written
→ every report reaches Claude under a header saying its instructions carry no authority from you

three things that catch people:

> the scan doesn't judge whether anything is malicious. it only marks the shape
> a tool call the report talks Claude into still goes through your permission checks and sandbox. the scan doesn't replace them
> it's not a substitute for limiting what the subagent can reach in the first place

Scanned, not judged. Escaped, not deleted. Permissions still decide.

the setup
> a background subagent's report arrives inside a completion notification marked as an automated event, not as a message from you
> give research subagents only the tools they need. the scan marks text, the tools list limits what they can do
> see the marker line? read that part of the report yourself before you approve anything it suggests
> a backslash in front of <system-reminder> in a report is the scan working, not a typo

paste this into Claude Code ↓

"Audit my subagents for what they can read and reach:
1. List every subagent in .claude/agents/ and the tools each one has
2. Mark the ones that read web pages, issues or files from outside this repo
3. For those, suggest the smallest tool list that still does the job, and whether they need Bash at all
4. Tell me what you'll do when a report comes back with the harness marker line
Show me every change as a diff first. No edits until I say go"

↳ https://t.co/9b1nA5WYd3

save it before a README you never opened starts giving orders

[Tidyowl @tidyowl]

holy sh*t. Claude Code reads every subagent's report before Claude does. and it never deletes a word

Anthropic's docs: a subagent can read files, web pages and command output you never saw. so its final report gets scanned for anything shaped like an instruction aimed at your main conversation

The idea: the scan doesn't decide what's malicious. it makes sure a fake instruction can't pass for a real one

nothing to turn on. this is the line you'll see on top of a flagged report:

[harness: subagent output matched instruction-shaped pattern(s): …

here's what it does:

→ text that imitates Claude Code's own output, like a <system-reminder> tag or a line starting with Human: or Assistant:, gets a backslash so it reads as plain text
→ a report that imitates those tags or mentions bypassPermissions or --dangerously-skip-permissions gets that marker line on top
→ nothing is removed or reworded. the permission mentions stay exactly as written
→ every report reaches Claude under a header saying its instructions carry no authority from you

three things that catch people:

> the scan doesn't judge whether anything is malicious. it only marks the shape
> a tool call the report talks Claude into still goes through your permission checks and sandbox. the scan doesn't replace them
> it's not a substitute for limiting what the subagent can reach in the first place

Scanned, not judged. Escaped, not deleted. Permissions still decide.

the setup
> a background subagent's report arrives inside a completion notification marked as an automated event, not as a message from you
> give research subagents only the tools they need. the scan marks text, the tools list limits what they can do
> see the marker line? read that part of the report yourself before you approve anything it suggests
> a backslash in front of <system-reminder> in a report is the scan working, not a typo

paste this into Claude Code ↓

"Audit my subagents for what they can read and reach:
1. List every subagent in .claude/agents/ and the tools each one has
2. Mark the ones that read web pages, issues or files from outside this repo
3. For those, suggest the smallest tool list that still does the job, and whether they need Bash at all
4. Tell me what you'll do when a report comes back with the harness marker line
Show me every change as a diff first. No edits until I say go"

↳ https://t.co/9b1nA5WYd3

save it before a README you never opened starts giving orders

[Tidyowl @tidyowl]

@0xlatiuk Imagine dominating three games by a combined 45 runs and still losing the Series. Baseball has a funny way of reminding you that winning big doesn’t matter if you lose the close ones. Mazeroski made sure nobody forgot Game 7.

[Tidyowl @tidyowl]

holy sh*t.
every Claude Code subagent forgets everything the moment it finishes. one line in its file fixes that

Anthropic's docs: give a subagent the memory field and it gets its own folder that survives across conversations

The idea: your reviewer shouldn't rediscover the same codebase quirks on every PR. let it keep notes

you add one line to the subagent's frontmatter:

memory: project

here's the split:

→ user: ~/.claude/agent-memory/<agent>/ for lessons it should carry into every project
→ project: .claude/agent-memory/<agent>/ for knowledge you commit and share with the team. the docs' recommended default
→ local: .claude/agent-memory-local/<agent>/ for project knowledge that stays out of git

what changes when it's on:

→ its system prompt gets instructions for reading and writing that folder
→ it starts with the first 200 lines or 25KB of its MEMORY․md, whichever comes first
→ Read, Write and Edit switch on automatically so it can keep its own notes

three things that catch people:

> it's part of auto memory. turn auto memory off and the memory field does nothing. no instructions, no memory tools
> past 200 lines or 25KB, the rest of MEMORY․md doesn't load. the subagent is told to curate it, so let it
> having memory isn't the same as using it. ask it to check its memory before the task and save what it learned after

One line to remember. Project to share. Ask it to save.

the setup
> project is the default to start with: the notes live in version control next to the code
> put the memory instructions in the subagent's own file so it keeps notes without being asked
> local keeps project notes out of git. user carries them across every repo you work in
> before a review: "check your memory for patterns you've seen before"

paste this into Claude Code ↓

"Give my subagents a memory:
1. List every file in .claude/agents/ and tell me which ones would gain from memory and which should stay stateless
2. For the ones that should remember, add memory: project and a short instruction to read memory before work and save learnings after
3. Check my settings for autoMemoryEnabled or CLAUDE_CODE_DISABLE_AUTO_MEMORY, because either one switches all of this off
4. Draft what each agent's first MEMORY․md should say
Show me every change as a diff first. No edits until I say go"

↳ https://t.co/9b1nA5WYd3

save it before your reviewer finds the same bug for the third time

[Tidyowl @tidyowl]

holy sh*t. Claude Code has 5 different ways to run agents in parallel, and Anthropic just documented which one to use when

the docs boil it down to three questions: who coordinates the work, do the workers need to talk, and do they touch the same files

The idea: don't pick the one you saw in a thread. pick the one where the right party holds the plan

the fastest way to watch parallel agents work is one command (it's subagents plus worktrees, packaged):

/batch migrate src/ from JavaScript to TypeScript

here's the split:

→ subagents: Claude hands a side task off inside your session and gets a summary back. for searches and logs you won't read twice
→ agent view: you dispatch independent sessions with claude agents and check back later. research preview
→ agent teams: a lead plans, assigns and keeps workers in sync. experimental, off until CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
→ dynamic workflows: a script holds the plan and cross-checks the results. for a 500-file migration or a codebase-wide audit
→ projects: one ongoing conversation that starts threads for you, in the cloud or on your machine. public beta on Pro and Max

three things that catch people:

> agent teams don't put teammates in separate worktrees. give each one its own set of files or they collide
> subagents report back to the session that spawned them. agent view sessions report only to you
> every worker in every approach is a full Claude session, so token usage multiplies

Claude delegates. You dispatch. A script decides.

the setup
> /batch researches the repo, splits the change into 5 to 30 units and shows you the plan before anything runs
> once you approve, each unit gets its own background subagent in its own worktree, runs the tests and publishes its change
> /tasks lists background work in this session, finished subagents included
> /workflows shows each run, its phase and how many agents are done

paste this into Claude Code ↓

"I need to [describe the task]. Before you start:
1. Tell me who should coordinate it: you with subagents, me with agent view, a team lead, or a workflow script
2. Tell me whether the workers need to talk to each other, and whether any of them touch the same files
3. If files overlap, tell me how you'll keep them apart: worktrees, or one owner per set of files
4. Estimate how many agents this needs and why
Show me the plan first. Nothing runs until I say go"

↳ https://t.co/onu1RrRGI3

save it before you spin up a team for a job one subagent could do

[Tidyowl @tidyowl]

holy sh*t. one failed agent in a Claude Code workflow reruns every agent that started after it. even the ones that already finished

Anthropic's docs spell it out: A, B, C and D start in that order. B fails. you relaunch. A comes back from cache. B, C and D all run again

The idea: a workflow saves every agent's result, but replays the run in start order. the first agent that failed or changed breaks the cache for everyone after it

you resume a paused run like this:

/workflows → select the run → p

here's how a relaunch behaves:

→ finished agents return their saved result, up to the first one whose prompt changed
→ a failed agent runs again, and so does everything the script started after it
→ stopping one agent with x counts as a failure. stopping the whole run doesn't
→ agents that were still running when you stopped start over from scratch

three things that catch people:

> swap the order of two agents in the script and everything from the swap onward runs again
> a stalled agent restarts up to 5 times with a clean transcript. files it already changed stay changed, and its tokens stay in the run's total
> hit your usage limit and the run waits for the reset instead of failing. only in an interactive session on a Claude subscription, with autoContinueAtUsageLimit on, a reset within 24 hours, and twice at most

Saved in order. One failure reruns the rest. Two waits, then it fails.

the setup
> inside parallel() or pipeline(), a dead agent leaves null and the run keeps going. awaited directly, it ends the whole run
> a slow but healthy agent? give it stallMs on its agent() call, or set CLAUDE_ASYNC_AGENT_STALL_TIMEOUT_MS for all of them
> claude -p, the Agent SDK and background sessions don't wait out a usage limit. the agent just fails
> before v2.1.305, finished agents could rerun even when you relaunched an unedited script. update first

paste this into Claude Code ↓

"Before I run this workflow:
1. List the agents in the order the script starts them, and mark the ones most likely to fail
2. For each risky one, tell me how many finished agents would rerun if it failed, and suggest an order that reruns less
3. Check that every risky agent sits inside parallel() or pipeline(), so one failure can't end the run
4. Tell me whether this session would pause or fail at my usage limit, and why
Show me the script diff first. No edits until I say go"

↳ https://t.co/fPeOfeh1Q3

save it before your next 40-agent run dies at agent 12

[Tidyowl @tidyowl]

https://t.co/F0NCjyItQQ