Pay for the plan, not the typing

stepAgents has let you name a different agent for propose, review, apply and verify since 0.42. What is new is a reason to actually do it: the four stages do not cost the same thing to run well, and they do not need to.

Deciding what a change should do and reading whether the result is right both call for judgement. Writing the diff that gets there is comparatively mechanical - which is not an insult to the work, it is why a cheap, fast model does it about as well as an expensive one. So the obvious split is: a capable agent on propose and review, and a cheap one on apply. This is that split, what it costs at today's prices, and - the part worth reading past the headline number - exactly what this project's own telemetry can and cannot prove about the savings.

The four stages take different things seriously

A harness.json entry accepts agent, model, effort, budget and customAgent - and only propose, review, apply and verify accept a stepAgents entry at all; archive and git are mechanical, with nothing to configure. A model is only accepted for the four adapters whose registry entry declares a --model flag: claude-cli, copilot-cli, claude-cli-acp and copilot-cli-acp. budget is maxCostUsd or maxAiCredits, whichever unit the chosen agent's own capabilities accept - the other is rejected.

{
  "stepAgents": {
    "propose": { "agent": "claude-cli-acp", "model": "claude-opus-5-5", "effort": "high", "budget": { "maxCostUsd": 5 } },
    "review":  { "agent": "claude-cli-acp", "model": "claude-opus-5-5", "effort": "high", "budget": { "maxCostUsd": 3 } },
    "apply":   { "agent": "deepseek-cli-acp" },
    "verify":  { "agent": "copilot-cli-acp", "effort": "medium", "budget": { "maxAiCredits": 200 } }
  },
  "timeout": { "maxStageSeconds": 2700 }
}

propose and review above pin claude-opus-5-5 at high effort, with a dollar ceiling each. claude-fable-5-1 is the other reasonable choice for either - cheaper per token, still a model built for judgement rather than throughput; which of the two is worth its price on your own changes is something to measure, not assume. apply carries no model, no effort and no budget at all, and that is not an oversight: deepseek-cli-acp accepts none of the three. Its model is a session option this project's driver never sends, so DeepSeek's own default runs; nothing here can raise or lower how hard it thinks.

What the harness can prove, and what it flatly cannot see

This is the part a "save money" article usually skips, and it is the most useful thing this one has to say.

Two of the three agents in this split convert cleanly to a ceiling you can actually set. claude-cli-acp reports a cost figure directly - measured in this repository on 2026-09-05, one stage: 60 input, 8,262 output and 1,693,507 cache tokens for $1.57, and budget.maxCostUsd compares against exactly that number. copilot-cli-acp reports no dollar figure, but it reports AI Credits, which are not a proxy this project invented - they are the unit Copilot itself bills in, 1 credit = $0.01 - so budget.maxAiCredits is a real ceiling in real money, just denominated differently. Two agents, two currencies, both enforceable.

deepseek-cli-acp is the one genuine blind spot. Measured 2026-09-23, its one usage figure is a context gauge - tokens used against the model's window - never a cost, a credit count or even a token split. budget has no mechanism to set on it, in either unit, because there is nothing to compare a ceiling against.

So the honest description of the harness.json above is: three of the four stages carry a leash denominated in real money - two in dollars, one in AI Credits - and the cheap one does not. timeout.maxStageSeconds still bounds how long apply can run, because elapsed time needs no report from the agent - but nothing in this product will tell you, or stop you, if a deepseek-cli-acp stage is burning through a prepaid balance faster than expected. That balance lives entirely on DeepSeek's side. Watch it there.

Two more limits worth knowing before you lean on the cheap stage. Neither copilot --acp nor dsh (DeepSeek's CLI) asks for permission before writing a file or running a command, despite the Agent Client Protocol supporting the mechanism - both were observed completing writes and shell commands here without ever raising session/request_permission. Cheaper is not the same as better-supervised; the review gate between apply and verify is what is actually watching this stage, not the permission channel. And dsh needs Node 22.18+ on the 22 line, or 24.2+ - below that floor it exits with code 0 before answering, silently. If your repository pins an older Node for its own tooling the way this one pins 22.11 with Volta, that is the first place to look when an apply stage using DeepSeek produces nothing.

What this costs, at today's prices

These are vendor list prices, read 2026-09-24, not a figure this project measured. Nothing here converts one into the other, and the harness has no opinion about which subscription you carry - that lives entirely outside it, in each CLI's own login.

Claude Pro is $20/month billed monthly, $17 annually; Claude Max starts at $100/month for the 5x tier. Copilot Pro is $10/month, including $10 of AI Credits at 1 credit = $0.01 - which is also where budget.maxAiCredits's minimum of 30 comes from: thirty cents is the smallest ceiling the product will accept, because Copilot itself will not meter anything finer. DeepSeek has no subscription tier at all: deepseek-flash (what this project's own measurement on 2026-09-22 recorded as DeepSeek-V4-Flash - names move faster than documentation; check what dsh reports before you trust either) prices output at $0.60 per million tokens off-peak, $1.20 during DeepSeek's own weekday peak windows. Five dollars of prepaid credit is, at off-peak rates, upward of eight million output tokens - many multiples of what one apply stage moves by the only measurement available here (1.7 million tokens, all kinds, on a claude-cli-acp stage). Read that as "a lot", not as a number of stages: different models tokenize and work differently, and this project has no apply-stage measurement on DeepSeek to compare it against directly.

The arithmetic says $65 a month, and it survives the correction: GitHub's own current price for Copilot Pro is $10, not the $20 an older plan once charged, and the saving lands in the same place regardless, because the difference comes out of the Claude Max to Claude Pro step either way.

What this article is not

Not a benchmark. Nothing here measured whether deepseek-cli-acp produces work as good as claude-cli-acp on the same apply stage - that is a real, answerable question, and it is not this one. Not a recommendation for your specific changes: an architecture decision or a security-sensitive apply stage may be exactly the case where the expensive agent belongs on every stage, ceilings included. And not a claim that Codex or Gemini would be a worse or better third leg of this - both remain agents this project has never run against a real binary, on any stage.

Try it

The full field reference for stepAgents, model, effort and budget is in HARNESS.md. Which agents report what, in which unit, is in LIMITS.md. The code is at github.com/VeryComplexAndLongName/OpenSpec-UI, where the repository and packages keep the name OpenSpec-UI.

Where each claim comes from

• stepAgents, its four fields, which stages accept it, and which agents accept a model: HARNESS.md, "stepAgents".
• The agent capability table (model, effort, spending-cap unit, run against a real binary): HARNESS.md, "Agents, models, effort, and spending caps".
• Which agents report usage, in which unit, and the $1.57 and 1,701,829-token measurements: LIMITS.md, "Which agents report usage".
• deepseek-cli-acp's context gauge, its measurement date, and that it reports no cost, no credits and no token split: LIMITS.md, the same section, and the archived change a-run-can-outgrow-its-context.
• Neither copilot --acp nor dsh asking for permission, and the Node 22.18+/24.2+ floor: HARNESS.md, "ACP effort and budget capabilities", and the article A new agent, and nothing else moved.
• Claude Pro and Max prices: claude.com/pricing, read 2026-09-24.
• Copilot Pro's price and its AI Credits: GitHub's own announcement, read 2026-09-24.
• DeepSeek's per-token prices and peak windows: DeepSeek's own pricing page, read 2026-09-24.