Release 0.6.0: publish the polish loop

README.md

@@ -56,6 +56,18 @@ Prefer Claude? Use `--agent byokit:claude --model claude-sonnet-5-5` in step 2 w
 - **See the market first.** `--competitors` finds 3–5 real products of the brief's kind, captures their public pages through the same robots/terms review as a reference, and shows them beside the options with their sources and every refusal.
 - **Keep the original.** Refine into a new option, or send the choices to [sign-off](https://github.com/umeranjum17/sign-off) for someone's pick.
 
+**Polish one direction.** Pick one option you already generated and refine only that direction. An independent judge scores each version, and the loop stops at 9/10, at `--max-rounds` (default 5), or after two rounds without improvement. It does not guarantee 9/10: a version below the bar is reported as below the bar, not as finished.
+
+```sh
+first-draft-axi polish notes-1.html "A notes app" \
+  --out-dir notes-polished \
+  --max-rounds 4 \
+  --feedback "Warmer paper background"
+first-draft-axi polish-feedback notes-polished "Make the sidebar narrower"
+```
+
+`polish-feedback <run-folder> "<request>"` adds cumulative feedback while a run is in progress; the next round judges it. The original file is untouched.
+
 `first-draft-axi init` also installs the [agent skill](skills/first-draft-axi/SKILL.md) for Claude Code, Codex and pi in your repo. Commands return JSON, so your coding agent can drive the same workflow.
 
 ## Proof, not a promise of perfect design
@@ -85,7 +97,7 @@ Same brief, same model (Claude Sonnet 5.5), 24 October 2026 — First Draft on t
 
 One run per cell, scored by the review worker, not customers and not an independent judge; before our fixes First Draft scored **4** on both of these. **[How good is it?](docs/quality.md)** has all four briefs side by side, the method, the limits, what we fixed and the rerun command.
 
-- [Published npm releases](https://www.npmjs.com/package/first-draft-axi?activeTab=versions): the clean install path was checked with the packed **0.5.0** build; the release workflow publishes after merge.
+- [Published npm releases](https://www.npmjs.com/package/first-draft-axi?activeTab=versions): the clean install path was checked with the packed **0.6.0** build; the release workflow publishes after merge.
 - [CI](https://github.com/umeranjum17/first-draft/actions/workflows/ci.yml): builds and drives CLI/browser checks with local provider stand-ins, without paid generation. The **25 October 2026** local run passed **74 tests, zero failures, zero skips**.
 - [Comparison contract](CONTRIBUTING.md): wrapper additions need paired rendered evidence; our own tool must not quietly lower the plain model's output bar.
 

docs/index.html

@@ -208,7 +208,7 @@ <h3>Hand it to your coding agent.</h3>
       </section>
     </main>
 
-    <footer><span>first-draft <span class="version">0.5.0</span> · Open source · MIT · Design engine: <a href="https://github.com/umeranjum17/open-design">OpenDesign</a> · Apache-2.0</span><a href="https://github.com/umeranjum17/first-draft#reference">CLI reference ↗</a></footer>
+    <footer><span>first-draft <span class="version">0.6.0</span> · Open source · MIT · Design engine: <a href="https://github.com/umeranjum17/open-design">OpenDesign</a> · Apache-2.0</span><a href="https://github.com/umeranjum17/first-draft#reference">CLI reference ↗</a></footer>
   </div>
   <script>document.addEventListener('click', (e) => { const b = e.target.closest('[data-copy]'); if (b) navigator.clipboard?.writeText(b.dataset.copy).then(() => b.setAttribute('aria-label', 'Copied'), () => {}); });</script>
 </body>

package.json

@@ -1,6 +1,6 @@
 {
   "name": "first-draft-axi",
-  "version": "0.5.0",
+  "version": "0.6.0",
   "description": "Pick a design direction before writing app code. A CLI for developers and coding agents.",
   "keywords": [
     "ai",

skills/first-draft-axi/SKILL.md

@@ -8,7 +8,7 @@ description: "Drive first-draft-axi to prototype and explore designs: several se
 Use it by judgement: bigger redesigns, new screens, landing pages, or "show me 3 options to pick from". A small UI fix (a colour, a margin, a label) does not need it; edit the code.
 Every command prints JSON; its `help` list says what to do next, and a failure is `{"error": {"code", "message"}, "help": [...]}`: read the message, do not retry blindly.
 
-## The 5 commands
+## The 7 commands
 
 1. **init**, once per repo (safe to re-run; writes only inside the repo):
    `first-draft-axi init`
@@ -28,6 +28,12 @@ Every command prints JSON; its `help` list says what to do next, and a failure i
    `first-draft-axi post .first-draft/designs/settings.html --title "Which settings layout?" --recommend a --option a="Grouped list" --option b="Tabs" --option c="Single scroll"`
    Copies the options to the board's `mocks/`, starts the view, and each option card gets an Open link to its live Play page (desktop and phone urls are in the output). The board is the nearest `.sign-off/` above the folder, or `--board <board.json>`, or `$FIRST_DRAFT_BOARD`. You choose `--recommend` (the option you would pick); a second post on the same title needs `--key <new-key>`. Tests, journeys and demos ask nobody: `--evidence --media shot.png` instead of `--recommend`.
    On a machine without a tailnet address pass `--host <address>` (or `FIRST_DRAFT_VIEW_HOST`).
+6. **polish**, to refine only the one direction the user picked (an independent judge scores every version; run it in the background, it takes several rounds):
+   `first-draft-axi polish .first-draft/designs/settings-2.html "A settings page for a podcast app" --out-dir .first-draft/designs/settings-polished --max-rounds 4 --feedback "Calmer billing section"`
+   Stops at a score of 9/10, at `--max-rounds` (default 5), or after two rounds without improvement; it does not guarantee 9/10. Each round prints a `polish_round` line (`score`, `accepted`, `file`, `gaps`), and `polished` holds the result. A version below 9 is below the bar: report its score and gaps, never call it finished. The original file is untouched.
+7. **polish-feedback**, to add the user's request to a polish run that is still going:
+   `first-draft-axi polish-feedback .first-draft/designs/settings-polished "Make the tabs larger"`
+   Feedback is cumulative and each later round is judged against it. A run that has stopped refuses it: start a new `polish`.
 
 For reference-grounded screen comparison: `first-draft-axi benchmark --product "AI chat app" --screen "phone chat" --input screens.json --width 390 --theme light --out benchmark.html`. The manifest must contain at least one current/candidate and is an array of `{role,name,file,source,permission,captured_at,theme}` captures; roles are `current`, `candidate`, `reference`, `counterexample`, paths resolve beside the manifest, capture time may be null if unknown. Search reuses public robots/terms capture; every refusal remains in the sheet/report. Positive references and targets must match the requested theme; negative examples may differ. The self-contained sheet scales original images to one width without recolouring them; adjacent JSON records per-axis reference observations, cited capture IDs and target scores (1–10 or null when not comparable). Every current/candidate is judged separately by the configured independent subscription judge, never a prose-only bar. Public acquisition currently supports light HTML captures only, not app-store/press image extraction. The central resolver requires an authenticated engine and a judge on a different engine, named in the output. `--no-discovery` and `--no-judge` are private test plumbing only (`FIRST_DRAFT_TEST_BENCHMARK=1`), not user modes. Missing references, unavailable judges and failed reviews publish no new sheet/report, including no partial scores; credential failures give one sign-in step. An earlier complete result remains unchanged.