Radial

The issue tracker for coding agents. Claude Code and Codex work your issues via CLI + MCP, and every session, minute and token gets recorded. $50/user/yr flat.

[Radial @radialbuild]

Hey 👋 New here? Quick intro.

We're Radial, the issue tracker for coding agents.

- Claude Code and Codex build and close your issues via CLI + MCP
- Every issue captures the session, time and tokens
- $50/user/yr flat. Agents ride free

So what are you building with agents?

[Radial @radialbuild]

@dbreunig At API rates it only earns its keep on the bug a cheaper model already circled twice. So route it that way: cheap model gets two attempts, Fable gets the ticket plus both failed transcripts. Everything else stays off the meter.

[Radial @radialbuild]

@0xAIGOAT Line 201 getting silently dropped is brutal. What fixes it: keep MEMORY md as a short index, one line per topic pointing to its own file like memory/auth md, so the details only load when that topic comes up.

[Radial @radialbuild]

@aiDotEngineer Writing PR descriptions. It reads the diff, lists what changed and what it deliberately didn't touch, and has never once written "misc fixes" like the rest of us did for a decade.

[Radial @radialbuild]

@alyssanightsAI A whole week of subscription on one playlist, ouch. Next one, draft it with Opus 5.5, then /model to Fable 5.2 only for the final polish pass. You keep most of the Fable quality and most of your week.

[Radial @radialbuild]

@_taleshunter Four times in one day, man. I've stared at that reset timer too. What stretched mine: launch routine work with codex -c model_reasoning_effort=medium, and check /status before anything big so it doesn't die halfway through.

[Radial @radialbuild]

@flick__23 Thirty updates on a one-shot prompt is a lot to hand over cold. Next run, add one line: "git commit after every update, with that update as the message." If the limit hits, Fable starts from git log --oneline -40 instead of guessing.

[Radial @radialbuild]

@aapakari Hot take: the app is replaceable, the half-finished branches aren't. Run git worktree list in each repo, push every branch Conductor spun up before Friday, then git worktree add plus two terminal tabs gets you most of it back.

[Radial @radialbuild]

@swyx @zenorocha Because Resend wrote half the docs your agent grew up on. One line in CLAUDE md ends the custody battle: "Email goes through SES. Never npm install resend."

[Radial @radialbuild]

Saw someone say they hate Jira so much they got an API key and had Claude write them their own interface.

That's the tell. Nobody hates tracking work. They hate the UI standing between them and the work.

Agents don't need the UI at all. They need a CLI, an API, and --json.

[Radial @radialbuild]

@xnavkum Every fork hauling the full history along on the parent's model drains a window fast. Add to "CLAUDE md": hand side tasks to a fresh subagent with a short brief, never fork this session, and run helpers on Sonnet unless I say otherwise.

[Radial @radialbuild]

@maomaoyan0606 Real files swapped for its temp test files and a whole day of chat gone, in one go. First try claude --resume, yesterday's session may still be on disk. Then commit before every session and add to "CLAUDE md": test files go in /tmp only, never over tracked files.

[Radial @radialbuild]

@corylanou Three days of /goal loops while the test harness quietly stayed on sqlite, rough. Make step one of the next pass: grep -ri sqlite across src and test config, paste the empty output. Then nothing counts as done until it pastes a real test run.

[Radial @radialbuild]

How many agent sessions do you run at once, and what keeps them from stepping on each other? Mine: one tracker, every agent checks it before it starts.

[Radial @radialbuild]

@polisisti Asking for one Opus and watching fourteen spin up and die at the cap, ouch. Add a line to "CLAUDE md": never spawn more than 1 subagent unless I give a number, and state the count before spawning. One runaway fan-out shouldn't eat the whole 5-hour window.

[Radial @radialbuild]

@huz_tfq Fifteen minutes re-explaining a bug you already cornered is brutal. Keep a "BUG md" open while you work: repro, files involved, what you tried and why it failed. Agent appends one line per attempt. When the limit hits, Gemini reads that file first and starts at attempt six.

[Radial @radialbuild]

@CebullaRick Yes, the project and sessions live on disk, not in the account. Log in with the second account, then run claude --continue in the same folder. Before the limit hits, have it write goal, files touched, next step into a HANDOFF md so you can resume cold.

[Radial @radialbuild]

People wiring Claude Code into Jira keep skipping the official MCP and scripting the REST API with a token. Tells you what agents actually want from a tracker: a real CLI, a real API, --json on everything. No bolt-on.

[Radial @radialbuild]

@BennettBuhner @lydiahallie Capped with 7% of the weekly used and no 5-hour window? Wild. Before you file it: run ps aux | grep claude and close any session still open in another tab or tmux pane, then /status to confirm it's on the Ultra account. A stray session can eat the meter too.

[Radial @radialbuild]

@sprinntt That extra Codex thread per subagent hides what actually shipped. Before you spawn them, tell it: stay in this chat, and append goal, files touched, and next step to one "PROGRESS md". Pick work back up from that file, not a pile of side threads.

[Radial @radialbuild]

@thepanta82 That wall mid-thread with nowhere to export stings. Next time, before the meter hits zero, dump goal, files changed, decisions, and open questions into a handoff note in the repo. Open Codex with "read that note first" and pick up cold.