site: npm 0.7.0 marketing bump (2026-09-25)

site/harnesses/index.html

@@ -85,7 +85,7 @@ <h2 class="harness-operator__title">Agent Looper Grok Bot</h2>
           </p>
         </div>
         <div class="harness-operator__body">
-          <p>Grok Bot is the Grok operator that runs Agent Looper on the user's computer. It is not a <code>--runtime</code> enum value. Distinct from the Grok 4.7 model that Cursor uses as judge.</p>
+          <p>Grok Bot is the Grok operator that runs Agent Looper on the user's computer. It is not a <code>--runtime</code> enum value. Distinct from the Grok 4.7 model that Cursor uses as judge by default.</p>
           <h3>How to use</h3>
           <ol>
             <li>Freeze <code>GOAL.md</code> and <code>verify.sh</code> (or <code>pnpm exec agent-loop-prompt --out .cursor/loops/my-task</code>).</li>
@@ -113,9 +113,10 @@ <h3>Cursor</h3>
             <p class="harness-card__chips">
               <span class="harness-chip">Worker <code>composer-2.5</code></span>
               <span class="harness-chip">Judge Grok 4.7</span>
+              <span class="harness-chip"><code>grok-4.6</code> / <code>grok-4.5</code> ok</span>
             </p>
             <p class="harness-card__run"><code>pnpm exec agent-loop run .cursor/loops/my-task --runtime cursor --review-gate</code></p>
-            <p class="harness-card__note">Export <code>CURSOR_API_KEY</code> or run under Doppler. Run <code>pnpm exec agent-loop-init</code> once. Set <code>costPreset: "cursor"</code> in <code>loop.json</code> to stay on Cursor for both worker and judge. Repo defaults that pin OpenCode <code>escalateModel</code> are not applied on <code>--runtime cursor</code> — judge uses <code>reviewRuntime</code> / <code>reviewModel</code>.</p>
+            <p class="harness-card__note">Export <code>CURSOR_API_KEY</code> or run under Doppler. Run <code>pnpm exec agent-loop-init</code> once. Default judge <code>reviewModel</code> is <code>grok-4.7</code> when the worker is Cursor; Composer 2.5 is the only Cursor worker. Set <code>costPreset: "cursor"</code> in <code>loop.json</code> to stay on Cursor for both worker and judge. Repo defaults that pin OpenCode <code>escalateModel</code> are not applied on <code>--runtime cursor</code> — judge uses <code>reviewRuntime</code> / <code>reviewModel</code>.</p>
           </article>
 
           <article class="harness-card" id="dsh">
@@ -200,10 +201,11 @@ <h3>Codex</h3>
             <p class="harness-card__job"><code>@openai/codex-sdk</code> and <code>codex</code> CLI — ChatGPT / OpenAI BYO.</p>
             <p class="harness-card__chips">
               <span class="harness-chip">Worker <code>gpt-5.6-luna</code> → <code>gpt-5.6-terra</code></span>
+              <span class="harness-chip">Optional <code>gpt-6-astra</code></span>
               <span class="harness-chip">Judge any runtime, optional</span>
             </p>
             <p class="harness-card__run"><code>pnpm exec agent-loop run .cursor/loops/my-task --runtime codex --review-gate</code></p>
-            <p class="harness-card__note">Install <code>@openai/codex-sdk</code>; ensure <code>codex</code> CLI is on PATH.</p>
+            <p class="harness-card__note">Install <code>@openai/codex-sdk</code>; ensure <code>codex</code> CLI is on PATH. Codex judge default stays <code>gpt-5.6-sol</code>; <code>gpt-6-astra</code> is listed but not the default judge.</p>
           </article>
 
           <article class="harness-card" id="muse">
@@ -237,7 +239,7 @@ <h3>Claude</h3>
               <span class="harness-chip">Judge any runtime, optional</span>
             </p>
             <p class="harness-card__run"><code>pnpm exec agent-loop run .cursor/loops/my-task --runtime claude --review-gate</code></p>
-            <p class="harness-card__note">Ensure <code>claude</code> 2.1.169+ is on PATH and <code>claude login</code> has been run. Typical mix: cheap worker + Claude as judge (default <code>opus</code>). See <a href="https://github.com/dancingteeth/agent-looper/blob/main/docs/claude-runtime.md">docs/claude-runtime.md</a>.</p>
+            <p class="harness-card__note">Ensure <code>claude</code> 2.1.169+ is on PATH and <code>claude login</code> has been run. <code>opus</code> / <code>sonnet</code> / <code>fable</code> aliases track Claude Code latest-per-family. Typical mix: cheap worker + Claude as judge (default <code>opus</code>). See <a href="https://github.com/dancingteeth/agent-looper/blob/main/docs/claude-runtime.md">docs/claude-runtime.md</a>.</p>
           </article>
         </div>
       </section>

site/harnesses/index.md

@@ -10,7 +10,7 @@ You say what to build and how to know it's done. It keeps a coding agent working
 
 [Add to Grok Bot](https://x.ai/bot/AETdGbRRNWfckrRGv22LD)
 
-Grok Bot is the Grok operator that runs Agent Looper on the user's computer. It is not a `--runtime` enum value. Distinct from the Grok 4.7 model that Cursor uses as judge.
+Grok Bot is the Grok operator that runs Agent Looper on the user's computer. It is not a `--runtime` enum value. Distinct from the Grok 4.7 model that Cursor uses as judge by default.
 
 ### How to use
 
@@ -29,7 +29,7 @@ Install once: `pnpm add -D @dancingteeth/agent-looper`, then add the SDK or CLI
 IDE subscription via `@cursor/sdk`.
 
 - Worker: `composer-2.5`
-- Judge (when worker is Cursor): Grok 4.7
+- Judge (when worker is Cursor): `grok-4.7` (`grok-4.6` / `grok-4.5` still allowed)
 - Run: `pnpm exec agent-loop run .cursor/loops/my-task --runtime cursor --review-gate`
 
 Export `CURSOR_API_KEY` or run under Doppler. Run `pnpm exec agent-loop-init` once. Set `costPreset: "cursor"` in `loop.json` to stay on Cursor for both worker and judge. Repo defaults that pin OpenCode `escalateModel` are not applied on `--runtime cursor` — judge uses `reviewRuntime` / `reviewModel`.
@@ -48,7 +48,7 @@ DeepSeek Harness CLI — worker is `dsh --profile headless`; `dsh-agent-looper`
 
 `@cline/sdk` — `cline-pass` for subscription quota, `cline` for credits.
 
-- `cline-pass` worker: `cline-pass/deepseek-v4.1-flash` → `qwen3.7-plus`
+- `cline-pass` worker: `cline-pass/deepseek-v4.1-flash` → `qwen3.7-plus` (`deepseek-v4-flash` still allowed)
 - `cline` worker: `deepseek/deepseek-chat` → `qwen/qwen3-coder-plus`
 - Judge: any runtime, optional
 - Run: `pnpm exec agent-loop run .cursor/loops/my-task --runtime cline-pass --review-gate` (or `--runtime cline` for credits)
@@ -57,7 +57,7 @@ DeepSeek Harness CLI — worker is `dsh --profile headless`; `dsh-agent-looper`
 
 `@opencode-ai/sdk` and `opencode` CLI — Go quota by default, or BYOK through OpenRouter, Vercel AI Gateway, Ollama, or another OpenAI-compatible router.
 
-- Go worker: `opencode-go/deepseek-v4.1-flash` → `qwen3.7-plus`
+- Go worker: `opencode-go/deepseek-v4.1-flash` → `qwen3.7-plus` (`deepseek-v4-flash` still allowed)
 - BYOK: `openrouter/…`, `openrouter/…:free`, `vercel/…`, `ollama/…`
 - Judge: any runtime, optional
 - Run: `pnpm exec agent-loop run .cursor/loops/my-task --runtime opencode --review-gate`
@@ -75,7 +75,8 @@ DeepSeek Harness CLI — worker is `dsh --profile headless`; `dsh-agent-looper`
 `@openai/codex-sdk` and `codex` CLI — ChatGPT / OpenAI BYO.
 
 - Worker: `gpt-5.6-luna` → `gpt-5.6-terra`
-- Judge: any runtime, optional
+- Optional: `gpt-6-astra` (listed; not the default judge)
+- Judge: any runtime, optional (Codex-native judge defaults to `gpt-5.6-sol`, not `gpt-6-astra`)
 - Run: `pnpm exec agent-loop run .cursor/loops/my-task --runtime codex --review-gate`
 
 ### Muse (`--runtime muse`)
@@ -88,7 +89,7 @@ DeepSeek Harness CLI — worker is `dsh --profile headless`; `dsh-agent-looper`
 
 ### Claude (`--runtime claude`)
 
-PATH `claude` CLI — Claude Code subscription. `--safe-mode` so the harness prompt is the only instruction source (strips project hooks and auto-memory). Not on `costPreset` minmax. See [docs/claude-runtime.md](https://github.com/dancingteeth/agent-looper/blob/main/docs/claude-runtime.md).
+PATH `claude` CLI — Claude Code subscription. `--safe-mode` so the harness prompt is the only instruction source (strips project hooks and auto-memory). Not on `costPreset` minmax. `opus` / `sonnet` / `fable` aliases track Claude Code latest-per-family. See [docs/claude-runtime.md](https://github.com/dancingteeth/agent-looper/blob/main/docs/claude-runtime.md).
 
 - Worker: `sonnet` → `opus`
 - Judge: any runtime, optional

site/index.html

@@ -107,7 +107,7 @@
             "name": "How do Agent Looper worker and judge presets work?",
             "acceptedAnswer": {
               "@type": "Answer",
-              "text": "Named presets: minmax (efficiency — cheap capable worker, strongest included judge), balanced (spend more on the worker, same judge), cursor (stay on Cursor: Composer + Grok). Or you, your agent, or Looper wires a pair from what's installed. The pair stays for the whole loop. Not Auto. On runtime: cursor, repo defaults that pin OpenCode Go escalateModel are not applied — Composer plus another judge uses reviewRuntime / reviewModel, not escalateModel."
+              "text": "Named presets: minmax (efficiency — cheap capable worker, strongest included judge), balanced (spend more on the worker, same judge), cursor (stay on Cursor: Composer + Grok — default judge grok-4.7). Or you, your agent, or Looper wires a pair from what's installed. The pair stays for the whole loop. Not Auto. On runtime: cursor, repo defaults that pin OpenCode Go escalateModel are not applied — Composer plus another judge uses reviewRuntime / reviewModel, not escalateModel."
             }
           },
           {
@@ -171,7 +171,15 @@
             "name": "What do the spend numbers mean?",
             "acceptedAnswer": {
               "@type": "Answer",
-              "text": "Watch and the report card show two numbers when they differ: list (public API rates, including prompt-cache) and billed (what the runtime invoice says). $0 on a subscription quota is billed $0, not free. Budget caps use billed when you are on PAYG and list when the invoice is $0. OpenCode Go and Cline Pass model lists and list prices come from models.dev; unknown but well-formed slugs soft-gate — warn at loop start, show $0 until priced. Budget caps still reject unpriced worker or escalate models."
+              "text": "Watch and the report card show two numbers when they differ: list (public API rates, including prompt-cache) and billed (what the runtime invoice says). $0 on a subscription quota is billed $0, not free. Budget caps use billed when you are on PAYG and list when the invoice is $0. OpenCode Go and Cline Pass model lists and list prices come from models.dev — sync picks up Grok 4.7, GPT-6 Luna, MiMo V2.6 Flash/Pro, and Space Bunny Free; unknown but well-formed slugs soft-gate — warn at loop start, show $0 until priced. Budget caps still reject unpriced worker or escalate models."
+            }
+          },
+          {
+            "@type": "Question",
+            "name": "How do I harden verify between runs?",
+            "acceptedAnswer": {
+              "@type": "Answer",
+              "text": "Before you add another line to verify.sh or REVIEWS.md, walk a diverse sample of past runs under .cursor/loop-exports/ so the scoreboard matches real failures — not one lucky green. Procedure: docs/unknowns-preflight.md on the repo. Freeze during a run is unchanged."
             }
           }
         ]
@@ -342,7 +350,7 @@ <h2>Which coding agents does Agent Looper work with?</h2>
           <article class="tui-answer tui-answer--presets">
             <h2>How do Agent Looper worker and judge presets work?</h2>
             <p>
-              Named presets: minmax (efficiency — cheap capable worker, strongest included judge), balanced (spend more on the worker, same judge), cursor (stay on Cursor: Composer + Grok). Or you, your agent, or Looper wires a pair from what's installed. The pair stays for the whole loop. Not Auto. On <code>runtime: cursor</code>, repo defaults that pin OpenCode Go <code>escalateModel</code> are not applied — Composer plus another judge uses <code>reviewRuntime</code> / <code>reviewModel</code>, not <code>escalateModel</code>.
+              Named presets: minmax (efficiency — cheap capable worker, strongest included judge), balanced (spend more on the worker, same judge), cursor (stay on Cursor: Composer + Grok — default judge <code>grok-4.7</code>). Or you, your agent, or Looper wires a pair from what's installed. The pair stays for the whole loop. Not Auto. On <code>runtime: cursor</code>, repo defaults that pin OpenCode Go <code>escalateModel</code> are not applied — Composer plus another judge uses <code>reviewRuntime</code> / <code>reviewModel</code>, not <code>escalateModel</code>.
             </p>
           </article>
           <article class="tui-answer tui-answer--budget">
@@ -395,7 +403,13 @@ <h2>What if verify fails because my environment is broken?</h2>
           <article class="tui-answer tui-answer--spend">
             <h2>What do the spend numbers mean?</h2>
             <p>
-              Watch and the report card show two numbers when they differ: <strong>list</strong> (public API rates, including prompt-cache) and <strong>billed</strong> (what the runtime invoice says). <code>$0</code> on a subscription quota is billed <code>$0</code>, not “free.” Budget caps use billed when you are on PAYG and list when the invoice is <code>$0</code>. OpenCode Go and Cline Pass model lists and list prices come from <a href="https://models.dev/api.json">models.dev</a>; unknown but well-formed slugs soft-gate — warn at loop start, show <code>$0</code> until priced. Budget caps still reject unpriced worker or escalate models.
+              Watch and the report card show two numbers when they differ: <strong>list</strong> (public API rates, including prompt-cache) and <strong>billed</strong> (what the runtime invoice says). <code>$0</code> on a subscription quota is billed <code>$0</code>, not “free.” Budget caps use billed when you are on PAYG and list when the invoice is <code>$0</code>. OpenCode Go and Cline Pass model lists and list prices come from <a href="https://models.dev/api.json">models.dev</a> — sync picks up Grok 4.7, GPT-6 Luna, MiMo V2.6 Flash/Pro, and Space Bunny Free; unknown but well-formed slugs soft-gate — warn at loop start, show <code>$0</code> until priced. Budget caps still reject unpriced worker or escalate models.
+            </p>
+          </article>
+          <article class="tui-answer tui-answer--exports">
+            <h2>How do I harden verify between runs?</h2>
+            <p>
+              Before you add another line to <code>verify.sh</code> or <code>REVIEWS.md</code>, walk a diverse sample of past runs under <code>.cursor/loop-exports/</code> so the scoreboard matches real failures — not one lucky green. Procedure: <a href="https://github.com/dancingteeth/agent-looper/blob/main/docs/unknowns-preflight.md">docs/unknowns-preflight.md</a>. Freeze during a run is unchanged.
             </p>
           </article>
         </div>
@@ -410,6 +424,7 @@ <h2>How it works</h2>
           Optional <code>setup.sh</code> (or <code>setup</code> in <code>loop.json</code>) runs once before the first worker — setup failure does not spawn a worker.
           A fresh worker loops until that check passes — optional judge / <code>reviewGate</code> only on serious findings.
           After each visit the harness restores frozen specs if a worker edited them — and removes any frozen basename they planted (like <code>setup.sh</code>) so it cannot carry into the next run.
+          Between runs, label export packs before you tighten the check (<a href="https://github.com/dancingteeth/agent-looper/blob/main/docs/unknowns-preflight.md">unknowns preflight</a>).
           Progress lives in git and files, not chat memory.
         </p>
         <div class="grid two-col">

site/index.md

@@ -12,7 +12,7 @@ Agent Looper uses the coding agents you already pay for: Cursor, Cline, OpenCode
 
 ## How do Agent Looper worker and judge presets work?
 
-Named presets: minmax (efficiency — cheap capable worker, strongest included judge), balanced (spend more on the worker, same judge), cursor (stay on Cursor: Composer + Grok). Or you, your agent, or Looper wires a pair from what's installed. The pair stays for the whole loop. Not Auto. On `runtime: cursor`, repo defaults that pin OpenCode Go `escalateModel` are not applied — Composer plus another judge uses `reviewRuntime` / `reviewModel`, not `escalateModel`.
+Named presets: minmax (efficiency — cheap capable worker, strongest included judge), balanced (spend more on the worker, same judge), cursor (stay on Cursor: Composer + Grok — default judge `grok-4.7`). Or you, your agent, or Looper wires a pair from what's installed. The pair stays for the whole loop. Not Auto. On `runtime: cursor`, repo defaults that pin OpenCode Go `escalateModel` are not applied — Composer plus another judge uses `reviewRuntime` / `reviewModel`, not `escalateModel`.
 
 ## How does Agent Looper keep cost down for indie builders?
 
@@ -50,11 +50,15 @@ When the check fails because something on your machine is missing or broken —
 
 ## What do the spend numbers mean?
 
-Watch and the report card show two numbers when they differ: **list** (public API rates, including prompt-cache) and **billed** (what the runtime invoice says). `$0` on a subscription quota is billed `$0`, not “free.” Budget caps use billed when you are on PAYG and list when the invoice is `$0`. OpenCode Go and Cline Pass model lists and list prices come from [models.dev](https://models.dev/api.json); unknown but well-formed slugs soft-gate — warn at loop start, show `$0` until priced. Budget caps still reject unpriced worker or escalate models.
+Watch and the report card show two numbers when they differ: **list** (public API rates, including prompt-cache) and **billed** (what the runtime invoice says). `$0` on a subscription quota is billed `$0`, not “free.” Budget caps use billed when you are on PAYG and list when the invoice is `$0`. OpenCode Go and Cline Pass model lists and list prices come from [models.dev](https://models.dev/api.json) — sync picks up Grok 4.7, GPT-6 Luna, MiMo V2.6 Flash/Pro, and Space Bunny Free; unknown but well-formed slugs soft-gate — warn at loop start, show `$0` until priced. Budget caps still reject unpriced worker or escalate models.
+
+## How do I harden verify between runs?
+
+Before you add another line to `verify.sh` or `REVIEWS.md`, walk a diverse sample of past runs under `.cursor/loop-exports/` so the scoreboard matches real failures — not one lucky green. Procedure: [docs/unknowns-preflight.md](https://github.com/dancingteeth/agent-looper/blob/main/docs/unknowns-preflight.md). Freeze during a run is unchanged.
 
 ## How it works
 
-Your agent writes `GOAL.md` and a deterministic check. Optional `setup.sh` (or `setup` in `loop.json`) runs once before the first worker — setup failure does not spawn a worker. A fresh worker loops until the check passes. After each visit the harness restores frozen specs if a worker edited them — and removes any frozen basename they planted (like `setup.sh`) so it cannot carry into the next run.
+Your agent writes `GOAL.md` and a deterministic check. Optional `setup.sh` (or `setup` in `loop.json`) runs once before the first worker — setup failure does not spawn a worker. A fresh worker loops until the check passes. After each visit the harness restores frozen specs if a worker edited them — and removes any frozen basename they planted (like `setup.sh`) so it cannot carry into the next run. Between runs, label export packs before you tighten the check ([unknowns preflight](https://github.com/dancingteeth/agent-looper/blob/main/docs/unknowns-preflight.md)).
 
 ## How is Agent Looper different from looping in chat?