Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableHabr AI→ original

After a Ban, the Author Launched an Agent Team in Kimi, Codex and OpenCode: Runs from $0.30

After Anthropic's wave of bans, the author of the supertick plugin, in which a "lead" hands out tasks to a team of AI agents, stopped depending on Claude Code. The team now also runs in Kimi Code CLI, Codex CLI and OpenCode. In an acceptance run, where the agents write a slugify function with tests, OpenCode took 15 minutes and cost about $0.30. The main lesson: test every lead model separately, or a single missed rule will break every run.

AI-processed from Habr AI; edited by Hamidun News
After a Ban, the Author Launched an Agent Team in Kimi, Codex and OpenCode: Runs from $0.30
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1Tests grew from 82 to 420, and the changes came to 30 commits and +23 377 lines
Key metric #2Six acceptance runs on live models all ended in PASS
Key metric #3gpt-5.6-sol scored 10 out of 10 on the lead probe, gpt-5.6-luna 9 out of 10

What happened: after Anthropic's wave of bans, a developer realized that keeping all his automation on a single Claude Code subscription was risky. His supertick plugin assembles AI agents into a small team: a "lead" hands out tasks, while workers write code, review it and accept the result. Until now, all of this lived only in Claude Code. The author has now rewritten the plugin so the same team runs in three other environments: Kimi Code CLI, Codex CLI and OpenCode.

Here is an example. You ask the team to add a slugify function to a small Python repository (lowercase, everything except letters and digits replaced by a single hyphen, Cyrillic preserved) along with tests. The coder writes it, the reviewer checks it, the lead accepts the work, and no human is involved. This kind of "acceptance run" in OpenCode took 15 minutes and cost about $0.30.

Why read on: even if you have no use for supertick, the author compiled a list of pitfalls that anyone building agents on top of different CLIs will hit.

How the port works

The lead in supertick is thin and sleeps almost all the time. An external "tick" wakes it, which is an ordinary script with no LLM. The workers are Claude subagents and external CLIs such as cursor and codex. The author rejected the easy route right away: overriding ANTHROPIC_BASE_URL to point Claude Code at DeepSeek or Kimi. He wanted a plugin that works independently of any vendor and even without Claude Code, so that several teams can run on different hosts at the same time.

The architecture is a shared, environment-agnostic core (tick, waiting, engine launch, config, diagnostics, 12 role briefs) with thin adapters on top. The three new environments share a single lead protocol, LEAD.md, with six operations: START, CONTINUE, REPLY, WAIT, ASK, STOP. How exactly to launch a worker is described for each environment in a primitives.md file. In Kimi it is Agent with run_in_background, in Codex it is codex exec, and in OpenCode it is the task tool.

The environments fall into three tiers. Claude Code can do everything, including messages between subagents. Kimi wakes the lead itself after a task finishes. In Codex and OpenCode nobody can wake the lead, so it waits in the foreground and does not end its turn until it delivers the final summary. Porting the lead took roughly three quarters of the total work. The spec estimated about 30 working days to reach the alpha version, but the code actually took three and a half days.

Two techniques are worth borrowing. First, the commands in LEAD.md are written as ready-made bash blocks, and the test executes them as-is on mock engines, so the document and the behavior cannot drift apart. Second, the lead decides from the first line of a report (READY FOR REVIEW, VERDICT, ACCEPT) without reading whole diffs. This is cheap and works even with a model that is not the strongest, but it demands a strict format.

To avoid breaking Claude Code, the author took behavior snapshots of version 1.1.1 and achieved a byte-for-byte match of the build. He replayed the new tick against recordings of 22 past live runs: version 1.1.1 produced 2576 events, and 2.0 produced zero new and zero lost ones.

Pitfalls that are not in the documentation

Each adapter was preceded by a 0.5–0.75-day spike, in which a table of primitives was filled in with real commands and real output. Examples from the list:

  • Kimi: kimi -p does not exit while a background task is alive. The lead started a second wait and forgot about the first, and the run hung until the one-hour timeout. The five-hour window quota is shared across the whole team: two simple runs used it up entirely on the $30 PRO subscription.
  • Codex: codex exec resume without -m picks the default model, not the session's model. You think you are continuing the coder, but it is already a different model.
  • Codex: .git is read-only in the sandbox. By opening write access for commits, the author also opened up hooks, so he added a guard with a .git fingerprint.
  • OpenCode: any permission denial in headless mode aborts the run without a summary, and bash without a timeout kills the command after 120 seconds. When the free Zen limit ran out, the process silently hung for 35 minutes.
  • cursor-agent: it writes an introduction before the verdict, so the first line no longer matches the format. One such case cost an extra review round of about 15 minutes.

Which model to use as the lead

Not every model is fit for the lead role. The free deepseek-v4.1-flash-free, instead of following the protocol, started studying the supertick scripts and a config outside the project, got a permission denial and derailed the run. The paid glm-5.3-flash ($0.15 per 1M input tokens and $0.50 per 1M output tokens) passed the lead probe in 7 minutes. On a 596-line protocol, gpt-5.6-sol scored 10 out of 10 and gpt-5.6-luna 9 out of 10: the lead will break a single missed item in every run.

The author's conclusion: run the probe for each model, not for a model family and not "based on reviews." For this, supertick includes lead-probe: a test bench with mock workers and 9 file-based checks, including "the lead did not touch the code itself." It takes a few minutes.

The overall acceptance picture: six runs on live models ended in PASS, though not always on the first attempt. Kimi with native roles passed on the 4th try in 29 minutes. The other runs took 17 minutes for Claude Code, 15 for OpenCode and 37 for Codex. The practical takeaway: make the lead protocol portable, test each lead model separately, and budget time for the pitfalls of each specific environment.

Story

Key Questions Answered

How the port works?

The lead in supertick is thin and sleeps almost all the time. An external "tick" wakes it, which is an ordinary script with no LLM. The workers are Claude subagents and external CLIs such as cursor and codex. The author rejected the easy route right away: overriding ANTHROPIC_BASE_URL to point Claude Code at DeepSeek or Kimi. He wanted a plugin that works independently of any vendor and even without Claude Code, so that several teams can run on different hosts at the same time.

What are the key details on Pitfalls that are not in the documentation?

Each adapter was preceded by a 0.5–0.75-day spike, in which a table of primitives was filled in with real commands and real output. Examples from the list: Kimi: kimi -p does not exit while a background task is alive. The lead started a second wait and forgot about the first, and the run hung until the one-hour timeout. The five-hour window quota is shared across the whole team: two simple runs used it up entirely on the $30 PRO subscription.

Which model to use as the lead?

Not every model is fit for the lead role. The free deepseek-v4.1-flash-free, instead of following the protocol, started studying the supertick scripts and a config outside the project, got a permission denial and derailed the run. The paid glm-5.3-flash ($0.15 per 1M input tokens and $0.50 per 1M output tokens) passed the lead probe in 7 minutes. On a 596-line protocol, gpt-5.6-sol scored 10 out of 10 and gpt-5.6-luna 9 out of 10: the lead will break a single missed item in every run.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…