What portable-artifacts 1.0.0 adds

Four changes: what each does, how it was built, and what is still unproven

The verdict

Numbers hold, builds cost less, plans get answers

Good

This release adds one-line numbers, cheap-model routing, a plan mode and the docs for all three.

  • Every test suite passed before release.
  • Copilot CLI ran Haiku 5.5 in 19 of 19 measured runs.
  • Codex and VS Code Copilot Chat are untested.

Numbers

What the branch measured#

Plan tasks done
16
Smoke runs passed
9 of 9
Motion gate
10 of 10
Runs on Haiku 5.5
19 of 19
Themes now
9
Plugin version
1.0.0

Features

Four features, one branch#

  • A. Numbers on one line

    A number and its unit such as 42% or 1,204 ms no longer split across lines in explainers, decks, dashboards and motion.

  • B. Cheap-model routing

    Skills hand the build to an artifact-builder agent that runs a cheaper model on Copilot and Codex.

  • C. Plan mode

    A plan preset in make-explainer where reviewers answer decisions in the page and return a JSON file to the agent.

  • D. Docs and release

    READMEs, feature pages and counts updated; plugin version raised from 0.8.1 to 1.0.0 for the CI guard.

A: how

A: one token finder, many renderers#

shared/nowrap.mjs finds number-plus-unit tokens and skips code, pre, kbd and samp. The explainer wraps each hit in a nowrap span; motion binds the gap with a non-breaking space through an exact-match server route.

KPI tiles size their font by value length with container-query units. Long values wrap inside the tile instead of spilling. check.mjs measures a real browser and fails on a split or an overflow.

B: how

B: how the cheap model is picked#

  1. Generate agent files

    bin/agents.mjs writes an artifact-builder agent: Haiku 5.5 with GPT-5.4 nano as fallback on Copilot, gpt-6-luna on Codex, low effort.

  2. Install them

    bin/install.mjs writes the files to the user-level agent folders; --project and --agents-only narrow it. After npx skills add, doctor.mjs --fix builder_agent writes the same files.

  3. Tell the skill to hand off

    All five SKILL.md files carry a Cheaper model block with a loop guard so the sub-agent does not delegate again.

  4. Measure it

    eval/run-eval.mjs --agent ran 9 smoke runs and the 10-prompt motion gate. Copilot logs confirmed Haiku in every run.

C: how

C: plan mode has four moving parts#

  • Plan preset and sections

    Five new section types: decision, approval, question, verdict and review, laid out decision-first with the verdict last.

  • Two themes

    mono is techie with a cyan accent. brief is executive. The explainer now has 9 themes.

  • Page runtime

    plan.js keeps answers in the browser and offers Copy, Download (plan-answers.json) and Save as HTML, a sticky button that saves the page with the answers inside.

  • Answer checker

    answers.mjs validates the answers file, or a saved page, against a closed schema and quotes reviewer text only as data, never as instructions.

Gaps

Known gaps left open#

SeverityRiskMitigation
Low Wide glyphs (W, M, CJK) under 16 characters can spill a KPI tile check.mjs reports it.
Low Stale-answers notice stays visible after the first new answer Parked for a follow-up.
Low --force in inline code breaks after the hyphen on every explainer page Parked for a follow-up.
Low docs/catalog.html search terms miss the new features Parked for a follow-up.

Process

How the release was built#

  1. Spec and plan

    A design spec and a 16-task plan were written and committed under docs/superpowers/.

  2. One agent per task

    Each task went to a fresh implementer, then a spec and quality review, then fix rounds. The controller made every commit.

  3. Whole-change review

    The strongest model reviewed the whole change, one fix wave followed, and a scoped re-review came back clean.