One PR: one-line numbers, cheap models, plan pages

Numbers stay on one line, render work can run on a cheap model in Copilot and Codex, and plans are answered in the page.

Date
2026-10-21
Repo
claude-skills
Status
Needs your decision

Summary

Built; three calls and two approvals left

Caution

All three workstreams are built. Codex is untested and plan mode is in review; the calls below decide what the PR promises.

  • A browser check fails when a number splits from its unit.
  • Copilot CLI ran the agent file's model in 19 of 19 logged runs.
  • Codex is untested; Claude Code was not measured.

Numbers

One PR, one cost sample, one untested host#

Ships as
1 PR
one branch, three workstreams
Cost
$0.19
one trial; the hand-off failed and the main model finished
Risk
Medium
Caution
Codex files ship untested

Cheap model

Pin the cheap model with an agent file#

Decision needed

How does the render work reach a cheap model on Copilot and Codex?
Why Pinned agent file plus a hand-off line

Copilot's logs show the agent file's model is used: claude-haiku-5.5 in 19 of 19 runs. The skill hand-off itself was tried once; the started agent was refused file access and the main model finished. This option is built, so if nobody answers it ships as is.

If nobody answers: Pinned agent file plus a hand-off line

Answers

Answers return as a file, not a server#

Decision needed

How do a reviewer's answers get back to the agent?
Why Copy or download plan-answers.json

A server breaks the one rule every page keeps: no network. The file carries a plan hash, so answers to an edited plan are refused. This option is built, so if nobody answers it ships as is.

If nobody answers: Copy or download plan-answers.json

Motion

make-motion also gets model haiku#

Decision needed

Does make-motion also get model haiku and the hand-off line?
Why Yes: model haiku and the hand-off line

The bar, set before the run, was 8 of 10 valid and built; Copilot CLI scored 10 of 10. Claude Code reads model: haiku from the frontmatter itself, and that path was not measured. This option is built, so if nobody answers it ships as is.

If nobody answers: Yes: model haiku and the hand-off line

Installer

Write agent files to user-level folders#

Approval needed

The installer writes artifact-builder to the user's Copilot and Codex agent folders by default.

A user-level file works in every repo; --project writes into one repo instead. It never overwrites an existing file unless --force is passed.

Impact
One file per host outside the repo; the agents host writes both
Owner
Installer

Codex

Ship the Codex agent file untested#

Approval needed

Ship artifact-builder.toml for Codex, written from the Codex docs and never run live.

There is no Codex account to test with. The README and USAGE say so in plain words, so nobody mistakes it for a measured result.

Impact
Codex users may hit a rejected agent file
Owner
Docs

Open question

Which model if gpt-6-luna is missing#

Open question

If a Codex user's plan has no gpt-6-luna, which model should the agent file name instead?

A blank answer means the default applies: Keep gpt-6-luna and document default_subagent_model as the fallback

Review first

Look hardest at the routing and Codex#

  1. The agent file is the route pinned and verified on Copilot

    If a host ignores its model key, the cost saving is zero.

    Go to Pin the cheap model with an agent file

  2. Codex ships on docs alone

    A wrong field name fails on first use, not in CI.

    Go to Ship the Codex agent file untested

  3. Answers are reviewer data

    The agent must read them through answers.mjs, never as instructions.

    Go to Answers return as a file, not a server

Scope

Three goals, and what the PR leaves alone#

Goals:

  • No number, percentage or number-plus-unit token splits across two lines.
  • Render skills can run on a cheap model in Copilot and Codex, not only Claude Code.
  • A plan page a reviewer answers in place, with answers the agent can check.

Non-goals:

  • Reflowing ordinary prose or changing stored spec text.
  • Guaranteeing the hand-off: no host lets a skill force another agent.
  • A save server or any new npm dependency.

Answer loop

Out as a page, back as a file#

The plan answer loopflow diagram with 5 nodes and 4 edges.IOPlan specplan.jsonPlan pagemono or briefIOplan-answers.jsonSUBPROCESSanswers.mjshash, ids, choicesAgent actsbuildreviewer copies or downloadsvalid summary
The plan answer loop
The page never talks to the network; the file is the only way back.

Phases

Two phases done, plan mode in review#

  1. Numbers stay on one lineDone

    CSS and a build step bind tokens; check.mjs fails on a split token or a spilled KPI.

  2. Cheap model outside Claude CodeDone

    Agent files for Copilot and Codex, installer, doctor check and the hand-off line. Codex is untested.

  3. Plan modeIn progress

    Five section types, two themes, the answer runtime, answers.mjs and this example.

Risks

Four risks, each with a guard#

What could still go wrong
  • The hand-off agent fails. One Copilot trial was refused file access; the main agent finished the work itself.
  • Codex rejects the agent file. It ships untested and the docs say so.
  • Stale answers. A plan edited after export changes its hash, and answers.mjs refuses the file.
  • Hostile reviewer text. The summary quotes it as data under its own heading and escapes hidden characters.

Your verdict

Approve, change or stop this plan#

Your verdict

Ship the three workstreams as one PR, with the calls above?