A row has its own agent and model

[suleimansh]

Each row of the Automations page gets its own coding agent and its own model.

👤

• Every row of a project ran on one model (mine is Opus) and always on Claude Code. In my last 150 runs /update-tickets, a simple copying job that may run every 15 minutes, cost $0.27 a run on Opus, and /work-queue, real coding, $1.33.
• A row has its own model, picked like everything else on a row: the skill may name the one it is made for, and I may pick my own on my machine. A row with no pick uses the scheduler's model, as today.
• A row has its own agent too: Claude Code or Codex.
• Both are one more line each in the row's Edit panel, and the row's line shows them.

🤖 automated · Opus 5.5, effort high

What a person sees

• The Edit panel has a new part, "What it runs on": an Agent menu and a Model menu. The row's line says them: "Claude Code, Haiku 5.5", or "Codex".
• The Model menu lists the picked agent's own models. Its first choice is what follows when nobody picks: "As the skill says (Haiku 5.5)", "The scheduler's model (Opus 5.5)", or "Codex's own default".
• Picking another agent starts its model from that first choice: a model belongs to one agent.
• An agent that cannot run the row is greyed, and the panel says why.

What a skill may say

schedule:
  every: 15m
  agent: claude-code
  model: haiku

• Both keys are optional. agent is claude-code or codex; model is the id that agent knows the model by.
• A person's pick on their machine wins: agent-scheduler agent <command> <claude-code|codex|skill> and agent-scheduler model <command> <id|skill>. agent-scheduler model <id>, with one word, is still the scheduler's own model.
• A model belongs to one agent. A skill's model is a model of the agent the skill names, of Claude Code when it names none. A person's model is kept with the agent it was picked for, and holds only while the row runs on that agent. So a row whose agent changes later, by a pick or because its skill reached the other folder, never sends one agent's model to the other.
• agent-scheduler model <command> with the id forgotten is refused: a command's name, or skill, is not taken for the scheduler's own model.

What Codex needed

• The part that starts a run already took an agent and a model; the scheduler now passes each row's own.
• Codex reads a project's skills from .agents/skills, Claude Code from .claude/skills. A row runs on an agent only when its skill is in that agent's folder. A skill that is only in .agents/skills used to be listed and never started; it now runs on Codex. A row made with "only on this machine" is no skill, so either agent runs it.
• "Its skill is where a run's checkout starts" is asked for the folder of the agent the run is on.
• Whether the agent's command-line tool is installed and logged in is asked once per agent, per look.

My picks, to correct if wanted

• Nothing can read how much of Codex's quota is left. A Codex row starts without that check, and the panel says so. A Claude row keeps the check, now with its own model, so a model's own weekly window counts for the rows that use it.
• Codex prices no turn, so a Codex run records no cost. The panel says so too.
• The key is agent, one letter from agents (how many at once). A slip is refused in words: "agents is a whole number from 1 to 99", "agent is one of claude-code, codex".

Known

• With no quota check, a Codex row whose shell line keeps printing work, and whose runs fail at once (Codex out of quota), starts again on every look. A Claude row has the same gap for failures the quota check does not see. Nothing here holds a row back after a failed run.
• A skill kept as two different files in the two folders is scheduled by the .claude/skills copy alone. Where one folder links to the other, as in this repository, there is one file.
• A skill that names an agent whose folder does not hold it is listed and says "Cannot start on …" until the skill is put there or a person picks the other agent.
• A row made with "New automation" is saved in .claude/skills alone, so it runs on Codex only once it is linked into .agents/skills by hand; Remove then leaves that link.

Checked

• On a rig of my own (own port, own settings folder): a row whose skill names agent: codex ran on Codex and answered; a row whose skill names model: haiku ran on Claude Code on Haiku for $0.002; another model picked in the panel was saved; a row whose skill is not in .agents/skills refused Codex in words.
• Codex does run a skill sent as /name from .agents/skills, also with the front matter our scheduled skills carry, and it hands the word after the command to the skill (/name word). Tried with codex-cli 0.144.4.
• Full root CI alone, the scheduler's tests once more without my git settings, 17 deliberate breaks caught by the tests, and one fresh reviewer: nine findings; the real ones are fixed with a test each, the rest are under Known.

DECISIONS.md

In packages/agent-scheduler/DECISIONS.md. Two bullets said that a scheduled run is on Claude Code, and that a row whose skill is only in Codex's folder is never started. This pick replaces them:

The schedule is read from both folders a coding agent reads skills from, .claude/skills first. A coding agent runs a command only when the command's skill is in that agent's own folder, .claude/skills for Claude Code and .agents/skills for Codex: sent to the other agent, the command finds no skill. A command whose skill is only in one folder runs on that folder's agent unless its skill or a person names the other, and a row set to an agent that cannot read its skill is listed, not started, and says so. Picked over reading .claude/skills alone, where such a skill would have no row and nothing would say why.

A row has its own coding agent and its own model, picked like its pace: its skill may name the ones it is made for (agent, model), and each person may pick their own for their machine, in the state file. With no pick a run is on Claude Code, on the scheduler's model. A model is one agent's: a person's model is kept with the agent it was picked for and holds only while the row runs on that agent, and a skill's model is a model of the agent the skill names, of Claude Code when it names none. Picked over one model and one agent for every row of a project, where a simple copying job that may run every 15 minutes cost as much a turn as real coding.

The scheduler's own model is per user, in the state file, set with model <id>, opus when unset: what a run on Claude Code starts on where neither its row's skill nor its person names another. Picked over a line in the tracked schedule alone, so each person controls what their own machine spends their quota on.

One bullet is new, for my pick about Codex's quota:

Only Claude's quota can be read. A run on Codex starts without the quota question, and the page says so; its cost is not recorded either, since Codex prices no turn. Picked over holding Codex rows back until a reading exists: nothing would start them.

And two bullets name the change in passing: the tick's order ("then the quota, for a run on Claude Code") and the dashboard part's commands (agent, model).

Logic files not updated

By the person's choice the LOGIC.md files are not updated in this pull request (#2038). They are brought in line later, in one pass.