@iceinvein/agent-skills 0.18.3 → 0.20.0

package/skills/sluice/evals/results/2026-09-20T01-52-06-287Z/aggregate-result.json

@@ -0,0 +1,122 @@
+{
+  "schemaVersion": 1,
+  "claudeVersion": "2.1.278",
+  "startedAt": "2026-09-20T01:52:06.287Z",
+  "durationSeconds": 33,
+  "costUsd": 0.11528100000000001,
+  "partial": false,
+  "suite": {
+    "root": "/Users/dikrana/Documents/projects/agent-skills/skills/sluice",
+    "ablation": "none",
+    "caseFilter": "bypass-*",
+    "threshold": 1,
+    "concurrency": 1,
+    "plugins": [
+      {
+        "name": "sluice",
+        "path": "/Users/dikrana/Documents/projects/agent-skills/skills/sluice"
+      }
+    ]
+  },
+  "cases": [
+    {
+      "name": "bypass-question-stays-silent",
+      "dir": "evals/bypass-question-stays-silent",
+      "source": "prose",
+      "promptMarkdown": "What's the difference between an idempotency key and a request ID? I keep seeing both in payment APIs and I'm not sure when each one earns its place.",
+      "runsPerCase": 3,
+      "timeoutSeconds": 300,
+      "maxTurns": 6,
+      "graders": [
+        {
+          "name": "answers-the-question",
+          "type": "llm",
+          "weight": 2,
+          "graderMarkdown": "The reply should answer a conceptual question about two identifiers used in payment APIs.\n\nPASS if the reply explains that an idempotency key is supplied by the caller to make a retried write safe (the server returns the original result instead of performing the operation twice), and that a request ID identifies one call for tracing, logging, or support, without affecting what the server does.\nFAIL if the reply conflates the two, describes only one of them, asks a clarifying question instead of answering, or answers with a plan of work rather than an explanation.",
+          "config": {
+            "criteria": "The reply should answer a conceptual question about two identifiers used in payment APIs.\n\nPASS if the reply explains that an idempotency key is supplied by the caller to make a retried write safe (the server returns the original result instead of performing the operation twice), and that a request ID identifies one call for tracing, logging, or support, without affecting what the server does.\nFAIL if the reply conflates the two, describes only one of them, asks a clarifying question instead of answering, or answers with a plan of work rather than an explanation.",
+            "focus": "last_message"
+          }
+        },
+        {
+          "name": "no-channel-announcement",
+          "type": "regex",
+          "weight": 1,
+          "config": {
+            "target": "last_message",
+            "pattern": "(bypass|fast|main|deep)\\s+channel",
+            "flags": "i",
+            "match": "not_contains"
+          }
+        },
+        {
+          "name": "writes-nothing",
+          "type": "file_exists",
+          "weight": 1,
+          "config": {
+            "path": "**/*",
+            "exists": false
+          }
+        }
+      ],
+      "arms": {
+        "with": [
+          {
+            "score": 1,
+            "passed": true,
+            "turns": 1,
+            "costUsd": 0.11528100000000001,
+            "judgeCostUsd": 0.0030900000000000003,
+            "durationSeconds": 33,
+            "startedAt": "2026-09-20T01:52:06.306Z",
+            "error": null,
+            "tracePath": "/private/tmp/e-cN9KZf/out/trace.jsonl",
+            "skippedPaidGraders": false,
+            "graders": [
+              {
+                "name": "answers-the-question",
+                "passed": true,
+                "weight": 2,
+                "explanation": "judge votes: PASS PASS PASS",
+                "withOnly": false,
+                "scored": true,
+                "judgeVotes": [
+                  true,
+                  true,
+                  true
+                ],
+                "evidence": "They answer different questions, and the cleanest way to keep them straight is cardinality across a retry:\n\n**One idempotency key spans many request IDs.** If your client sends a charge, times out, and retries three times, that's one idempotency key and four request IDs.\n\n## Idempotency key\n\nClient-generated, semantic, part of the API contract. It means: *\"these attempts are the same logical operation — do the work at most once.\"*\n\nThe server has to actually do something with it: store key → (request fingerprint, response), and on a repeat return the stored response instead of re-executing. That implies real machinery:\n\n- **Conflict handling** — same key, different payload is a client bug; reject it (422/409) rather than silently replaying the old response.\n- **In-flight handling** — two concurrent requests with the same key need a lock and a 409, not two charges.\n- **Scoping** — keys must be namespaced per API key/account, or one tenant's UUID collision becomes another tenant's replayed response.\n- **Expiry** — typically 24h (Stripe's window). After that the key is forgotten and a retry executes for real, which is why it's a retry-safety mechanism, not a permanent dedup ledger.\n\nThe subtle client-side requirement: generate the key *before* the first attempt and persist it alongside the pending operation. A key that only lives in the retry loop's memory is lost on process restart, and the retry-after-crash — the exact case you most needed it for — double-charges.\n\n## Request ID\n\nPer-attempt identifier, usually server-generated and echoed in a header (`X-Request-Id`) and in error bodies. It's for correlation: logs, traces, support tickets, \"here's the ID, what happened to my call.\"\n\nIt costs nothing in the request path — no storage, no locking, no conflict semantics — and it is *deliberately different on every attempt*, because distinguishing attempt 1 from attempt 4 is the entire point when you're debugging a retry storm.\n\n## When each earns its place\n\n**Request ID: always.** It's nearly free and the first thing you want when something goes wrong.\n\n**Idempotency key: only for non-idempotent state mutations where a duplicate is expensive.** Creating a charge, a transfer, a refund. You don't need one for:\n\n- `GET`s — already idempotent.\n- `PUT`/`DELETE` against a specific resource path — idempotent by construction.\n- Creates where the client already supplies a natural unique key (your own `order_id` with a uniqueness constraint). That's a dedup mechanism you already have; a second one adds failure modes without adding safety.\n\n## The two ways people get this wrong\n\n**Using the request ID as the idempotency key.** Clients regenerate request IDs per attempt, so every retry looks like a fresh operation. This is the double-charge bug, and it passes testing because it only fires on timeouts.\n\n**Using the idempotency key as the trace ID.** All four attempts collapse into one trace, and you lose the ability to see that there *were* four attempts — which is usually the thing you're trying to diagnose.\n\nOne nicety worth adding if you're building the server side: flag replays explicitly (`Idempotent-Replayed: true`). It lets clients distinguish \"this succeeded just now\" from \"this succeeded earlier and you're seeing a recording,\" which matters for their own logging and for reconciliation."
+              },
+              {
+                "name": "no-channel-announcement",
+                "passed": true,
+                "weight": 1,
+                "explanation": "pattern absent as expected",
+                "withOnly": false,
+                "scored": true
+              },
+              {
+                "name": "writes-nothing",
+                "passed": true,
+                "weight": 1,
+                "explanation": "**/* absent as expected",
+                "withOnly": false,
+                "scored": true
+              }
+            ]
+          }
+        ]
+      },
+      "aggregates": {
+        "score": 1,
+        "passRate": 1
+      }
+    }
+  ],
+  "aggregates": {
+    "casesTotal": 1,
+    "casesPassed": 1,
+    "overallScore": 1,
+    "overallPassRate": 1
+  }
+}

package/skills/sluice/evals/results/2026-09-20T01-52-06-287Z/report.html

@@ -0,0 +1,324 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Eval report — sluice</title>
+</head>
+<body>
+<style>
+:root{--plane:#f9f9f7;--surface:#fcfcfb;--ink:#0b0b0b;--ink-2:#52514e;--ink-3:#6b6a64;--hairline:rgba(11,11,11,.10);--grid:#e1e0d9;--inset:rgba(11,11,11,.04);--accent:#2a78d6;--base-fill:#6b6a64;--delta-good:#006300;--good:#067d06;--warning:#8a6100;--critical:#d03b3b;
+  color-scheme:light dark;
+}
+@media (prefers-color-scheme:dark){:root{--plane:#0d0d0d;--surface:#1a1a19;--ink:#ffffff;--ink-2:#c3c2b7;--ink-3:#898781;--hairline:rgba(255,255,255,.10);--grid:#2c2c2a;--inset:rgba(255,255,255,.05);--accent:#3987e5;--base-fill:#898781;--delta-good:#0ca30c;--good:#0ca30c;--warning:#fab219;--critical:#e06c6c;color-scheme:dark}}
+:root[data-theme="dark"]{--plane:#0d0d0d;--surface:#1a1a19;--ink:#ffffff;--ink-2:#c3c2b7;--ink-3:#898781;--hairline:rgba(255,255,255,.10);--grid:#2c2c2a;--inset:rgba(255,255,255,.05);--accent:#3987e5;--base-fill:#898781;--delta-good:#0ca30c;--good:#0ca30c;--warning:#fab219;--critical:#e06c6c;color-scheme:dark}
+:root[data-theme="light"]{--plane:#f9f9f7;--surface:#fcfcfb;--ink:#0b0b0b;--ink-2:#52514e;--ink-3:#6b6a64;--hairline:rgba(11,11,11,.10);--grid:#e1e0d9;--inset:rgba(11,11,11,.04);--accent:#2a78d6;--base-fill:#6b6a64;--delta-good:#006300;--good:#067d06;--warning:#8a6100;--critical:#d03b3b;color-scheme:light}
+*{box-sizing:border-box}
+body{margin:0;background:var(--plane);color:var(--ink);
+  font:14px/1.55 system-ui,-apple-system,"Segoe UI",sans-serif;
+  -webkit-text-size-adjust:100%}
+.wrap{max-width:880px;margin:0 auto;padding:40px 24px 64px;display:flex;flex-direction:column;gap:20px}
+.eyebrow{font-size:11px;font-weight:600;letter-spacing:.08em;text-transform:uppercase;color:var(--ink-3)}
+header h1{margin:2px 0 0;font-size:24px;font-weight:600;line-height:1.25;text-wrap:balance}
+.meta{display:flex;flex-wrap:wrap;gap:6px 14px;margin-top:8px;color:var(--ink-2);font-size:13px}
+.meta .num{font-variant-numeric:tabular-nums lining-nums}
+.banner{display:flex;align-items:center;gap:8px;padding:10px 14px;border-radius:8px;
+  border:1px solid var(--hairline);background:var(--surface);font-size:13px}
+.banner .chip-warn{color:var(--warning);font-weight:600}
+.tiles{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:12px}
+.tile{background:var(--surface);border:1px solid var(--hairline);border-radius:10px;padding:14px 16px;
+  display:flex;flex-direction:column;gap:4px}
+.tile .label{font-size:11px;font-weight:600;letter-spacing:.06em;text-transform:uppercase;color:var(--ink-3)}
+.tile .value{font-size:26px;font-weight:600;font-variant-numeric:tabular-nums lining-nums;line-height:1.1}
+.tile.hero .value{font-family:Georgia,"Times New Roman",serif;font-weight:400;font-size:48px}
+.tile .sub{font-size:12px;color:var(--ink-2);font-variant-numeric:tabular-nums}
+.tile .value.delta-pos{color:var(--delta-good)}
+.tile .value.delta-neg{color:var(--critical)}
+.toolbar{display:flex;gap:8px;justify-content:flex-end}
+.toolbar button{font:12px system-ui,-apple-system,"Segoe UI",sans-serif;color:var(--ink-2);
+  background:var(--surface);border:1px solid var(--hairline);border-radius:6px;padding:4px 10px;cursor:pointer}
+.toolbar button:hover{color:var(--ink)}
+.toolbar button:focus-visible{outline:2px solid var(--accent);outline-offset:1px}
+.case{background:var(--surface);border:1px solid var(--hairline);border-radius:10px;padding:18px 20px;
+  display:flex;flex-direction:column;gap:10px}
+.case-regressed{border-left:3px solid var(--critical)}
+.verdict{margin:6px 0 0;font-size:15px}
+.verdict .delta{font-size:15px}
+.flag{font-size:11px;font-weight:600;color:var(--warning);white-space:nowrap}
+.star{color:var(--warning);font-size:.45em;vertical-align:super;line-height:0;cursor:help}
+.legend summary{cursor:pointer;font-size:12px;font-weight:600;letter-spacing:.04em;text-transform:uppercase;color:var(--ink-2)}
+.legend-list{margin:8px 0 0;padding-left:20px;display:flex;flex-direction:column;gap:5px;font-size:13px;color:var(--ink-2)}
+.case-head{display:flex;align-items:baseline;gap:10px;flex-wrap:wrap}
+.case-head h2{margin:0;font-size:16px;font-weight:600}
+.case-head .spacer{flex:1}
+.case-score{font-size:15px;font-weight:600}
+.mono{font:12px "SF Mono",ui-monospace,Menlo,Consolas,monospace}
+.num{font-variant-numeric:tabular-nums lining-nums}
+.muted{color:var(--ink-3);font-size:12px}
+.meter{display:inline-block;position:relative;width:120px;height:6px;border-radius:4px;background:var(--grid);
+  vertical-align:middle}
+.meter>span{display:block;height:100%;border-radius:4px;max-width:100%}
+.meter .tick{position:absolute;top:-2px;bottom:-2px;width:2px;background:var(--ink-3);border-radius:1px}
+.m-accent>span{background:var(--accent)}
+.m-base>span{background:var(--base-fill)}
+.delta{font-size:12px;font-weight:600;font-variant-numeric:tabular-nums}
+.delta-pos{color:var(--delta-good)}
+.delta-neg{color:var(--critical)}
+.delta-zero{color:var(--ink-3)}
+details.section{border-top:1px solid var(--grid);padding-top:10px}
+details.section>summary{cursor:pointer;font-size:12px;font-weight:600;letter-spacing:.04em;
+  text-transform:uppercase;color:var(--ink-2);list-style-position:outside;margin-left:2px}
+details.section>summary:hover{color:var(--ink)}
+details.section[open]>summary{margin-bottom:8px}
+.md{display:flex;flex-direction:column;gap:8px;background:var(--inset);border-radius:8px;
+  padding:12px 14px;overflow-wrap:break-word}
+.md>:first-child{margin-top:0}
+.md h1,.md h2,.md h3,.md h4,.md h5,.md h6{margin:4px 0 0;font-size:1em;font-weight:600;line-height:1.3}
+.md p,.md ul,.md ol,.md blockquote,.md table,.md pre,.md hr{margin:0}
+.md ul,.md ol{display:flex;flex-direction:column;gap:4px;padding-left:20px}
+.md blockquote{border-left:2px solid var(--hairline);padding-left:10px;color:var(--ink-2)}
+.md :not(pre)>code{background:var(--inset);padding:1px 4px;border-radius:4px;
+  font:.92em "SF Mono",ui-monospace,Menlo,Consolas,monospace}
+.md pre{background:var(--inset);border:1px solid var(--hairline);padding:10px 12px;border-radius:6px;
+  overflow-x:auto;font:12px/1.5 "SF Mono",ui-monospace,Menlo,Consolas,monospace}
+.md pre code{background:none;padding:0;font:inherit}
+.md table{width:100%;border-collapse:collapse}
+.md th,.md td{padding:5px 8px;text-align:left;vertical-align:top;border-bottom:1px solid var(--grid)}
+.md th{font-weight:600;color:var(--ink-2)}
+.md a{color:var(--accent);text-decoration:none}
+.md a:hover{text-decoration:underline}
+.grader-def{display:flex;flex-direction:column;gap:6px;padding:8px 0}
+.grader-def+.grader-def{border-top:1px solid var(--grid)}
+.grader-def-head{display:flex;align-items:baseline;gap:8px}
+.grader-name{font:13px "SF Mono",ui-monospace,Menlo,Consolas,monospace;font-weight:600}
+.badge{font-size:10px;font-weight:600;letter-spacing:.04em;text-transform:uppercase;color:var(--ink-2);
+  border:1px solid var(--hairline);border-radius:999px;padding:1px 8px}
+.config{display:flex;flex-direction:column;gap:2px;background:var(--inset);border-radius:8px;padding:10px 14px}
+.config code{font:12px "SF Mono",ui-monospace,Menlo,Consolas,monospace;overflow-wrap:anywhere}
+.arm{display:flex;flex-direction:column;gap:8px;padding:6px 0}
+.arm+.arm{border-top:1px dashed var(--grid);margin-top:4px;padding-top:12px}
+.arm-head{display:flex;align-items:baseline;gap:10px;flex-wrap:wrap}
+.arm-label{font-size:13px;font-weight:600}
+.run{border:1px solid var(--grid);border-radius:8px;padding:10px 12px;display:flex;flex-direction:column;gap:8px}
+.run-head{display:flex;align-items:baseline;gap:10px;flex-wrap:wrap}
+.run-title{font-size:12px;font-weight:600;color:var(--ink-2)}
+.run-error .explanation{color:var(--ink-2)}
+.note{font-size:12px;color:var(--ink-2)}
+.graders{display:flex;flex-direction:column;gap:4px}
+details.grader{border-radius:6px}
+details.grader>summary{cursor:pointer;display:flex;align-items:baseline;gap:8px;padding:3px 4px;
+  border-radius:6px;list-style:none}
+details.grader>summary::-webkit-details-marker{display:none}
+details.grader>summary::before{content:'▸';font-size:10px;color:var(--ink-3);flex:none;
+  transition:transform .12s ease}
+details.grader[open]>summary::before{transform:rotate(90deg)}
+details.grader>summary:hover{background:var(--inset)}
+details.grader>summary:focus-visible{outline:2px solid var(--accent);outline-offset:1px}
+.grader-body{padding:6px 8px 8px 24px;display:flex;flex-direction:column;gap:6px}
+.chip{font-size:11px;font-weight:600;border-radius:999px;padding:1px 8px;white-space:nowrap}
+.chip-pass{color:var(--good);border:1px solid currentColor}
+.chip-fail{color:var(--critical);border:1px solid currentColor}
+.chip.chip-warn{color:var(--warning);border:1px solid currentColor}
+.explanation{margin:0;font-size:13px;color:var(--ink-2);white-space:pre-wrap;overflow-wrap:break-word}
+.kv{display:flex;gap:8px;font-size:12px;color:var(--ink-3)}
+.votes{font-variant-numeric:tabular-nums;letter-spacing:.1em}
+pre.evidence{margin:0;background:var(--inset);border:1px solid var(--hairline);border-radius:6px;
+  padding:8px 10px;overflow-x:auto;max-height:320px;
+  font:12px/1.5 "SF Mono",ui-monospace,Menlo,Consolas,monospace;white-space:pre-wrap;overflow-wrap:break-word}
+footer{color:var(--ink-3);font-size:12px;text-align:center;padding-top:8px}
+@media (prefers-reduced-motion:no-preference){
+  details.grader>summary,.toolbar button{transition:background .12s ease,color .12s ease}
+}
+@media (prefers-reduced-motion:reduce){
+  details.grader>summary::before{transition:none}
+}
+@media print{
+  :root,:root[data-theme="dark"],:root[data-theme="light"]{--plane:#f9f9f7;--surface:#fcfcfb;--ink:#0b0b0b;--ink-2:#52514e;--ink-3:#6b6a64;--hairline:rgba(11,11,11,.10);--grid:#e1e0d9;--inset:rgba(11,11,11,.04);--accent:#2a78d6;--base-fill:#6b6a64;--delta-good:#006300;--good:#067d06;--warning:#8a6100;--critical:#d03b3b;color-scheme:light}
+  body{background:#fff}
+  .toolbar{display:none}
+  .case,.run,.grader-def{break-inside:avoid}
+  pre.evidence{max-height:none}
+}
+</style>
+<div class="wrap">
+<header>
+<div class="eyebrow">Plugin eval report</div>
+<h1>sluice</h1>
+
+<div class="meta">
+<span class="mono">.</span>
+<span>Claude Code v2.1.278</span>
+<span class="num">2026-09-20 01:52 UTC</span>
+<span class="num">33s</span>
+<span class="num">$0.12</span>
+<span class="num">1 runs</span>
+
+
+
+<span>judge default</span>
+<span class="num">threshold 100%</span>
+<span>filtered: --case bypass-*</span>
+</div>
+</header>
+
+<section class="tiles">
+<div class="tile hero"><span class="label">Suite score</span><span class="value">100%</span><span class="sub">mean of per-case scores</span></div>
+
+<div class="tile"><span class="label">Cases</span><span class="value num">1</span><span class="sub">1 of 1 ≥ 100% threshold</span></div>
+<div class="tile"><span class="label">Perfect runs</span><span class="value num">100%</span><span class="sub">runs where every grader passed</span></div>
+</section>
+<details class="section legend">
+<summary>How to read this report</summary>
+<ul class="legend-list">
+<li>A run's score is the weighted fraction of its graders that passed; a "perfect" run passed every grader.</li>
+<li>A case's score is the mean of its runs; a case passes when its score is at or above the 100% threshold (the tick on each case bar).</li>
+<li>The suite score is the mean of the per-case scores.</li>
+<li>Judge votes are independent samples of the LLM judge; the majority decides pass or fail.</li>
+<li><strong>No baseline arm was run</strong> (ablation off) — this report shows absolute scores only and cannot say whether the plugin caused them. Re-run with <code>--ablation with-without</code> to measure the plugin’s effect.</li>
+
+</ul>
+</details>
+<div class="toolbar"><button type="button" data-act="expand">Expand all</button><button type="button" data-act="collapse">Collapse all</button></div>
+<article class="case" id="case-1">
+<div class="case-head">
+<h2>bypass-question-stays-silent</h2>
+<span class="muted mono">evals/bypass-question-stays-silent</span>
+<span class="spacer"></span>
+
+
+
+<span class="meter m-accent" aria-hidden="true"><span style="width:100.0%"></span><i class="tick" style="left:100.0%"></i></span>
+<span class="num case-score">100%</span>
+</div>
+
+<details class="section" open>
+<summary>Results</summary>
+<section class="arm">
+<div class="arm-head">
+<span class="arm-label">Runs</span>
+<span class="meter m-accent" aria-hidden="true"><span style="width:100.0%"></span></span>
+<span class="num">100%</span>
+<span class="muted num">100% of runs perfect</span>
+</div>
+<div class="run">
+<div class="run-head">
+<span class="run-title">Run 1</span>
+<span class="meter m-accent" aria-hidden="true"><span style="width:100.0%"></span></span>
+<span class="num">100%</span>
+<span class="muted num">1 turns · $0.12 · 01:52:06 UTC</span>
+</div>
+
+<div class="graders"><details class="grader">
+<summary><span class="chip chip-pass">✓ pass</span> <span class="grader-name">answers-the-question</span> <span class="muted">×2</span> </summary>
+<div class="grader-body">
+<p class="explanation">judge votes: PASS PASS PASS</p>
+<div class="kv"><span>Judge votes</span><span class="votes">✓ ✓ ✓</span></div><div class="kv"><span>Evidence (what the judge was shown; for an image, a description of what was sent)</span></div><pre class="evidence">They answer different questions, and the cleanest way to keep them straight is cardinality across a retry:
+
+**One idempotency key spans many request IDs.** If your client sends a charge, times out, and retries three times, that&apos;s one idempotency key and four request IDs.
+
+## Idempotency key
+
+Client-generated, semantic, part of the API contract. It means: *&quot;these attempts are the same logical operation — do the work at most once.&quot;*
+
+The server has to actually do something with it: store key → (request fingerprint, response), and on a repeat return the stored response instead of re-executing. That implies real machinery:
+
+- **Conflict handling** — same key, different payload is a client bug; reject it (422/409) rather than silently replaying the old response.
+- **In-flight handling** — two concurrent requests with the same key need a lock and a 409, not two charges.
+- **Scoping** — keys must be namespaced per API key/account, or one tenant&apos;s UUID collision becomes another tenant&apos;s replayed response.
+- **Expiry** — typically 24h (Stripe&apos;s window). After that the key is forgotten and a retry executes for real, which is why it&apos;s a retry-safety mechanism, not a permanent dedup ledger.
+
+The subtle client-side requirement: generate the key *before* the first attempt and persist it alongside the pending operation. A key that only lives in the retry loop&apos;s memory is lost on process restart, and the retry-after-crash — the exact case you most needed it for — double-charges.
+
+## Request ID
+
+Per-attempt identifier, usually server-generated and echoed in a header (`X-Request-Id`) and in error bodies. It&apos;s for correlation: logs, traces, support tickets, &quot;here&apos;s the ID, what happened to my call.&quot;
+
+It costs nothing in the request path — no storage, no locking, no conflict semantics — and it is *deliberately different on every attempt*, because distinguishing attempt 1 from attempt 4 is the entire point when you&apos;re debugging a retry storm.
+
+## When each earns its place
+
+**Request ID: always.** It&apos;s nearly free and the first thing you want when something goes wrong.
+
+**Idempotency key: only for non-idempotent state mutations where a duplicate is expensive.** Creating a charge, a transfer, a refund. You don&apos;t need one for:
+
+- `GET`s — already idempotent.
+- `PUT`/`DELETE` against a specific resource path — idempotent by construction.
+- Creates where the client already supplies a natural unique key (your own `order_id` with a uniqueness constraint). That&apos;s a dedup mechanism you already have; a second one adds failure modes without adding safety.
+
+## The two ways people get this wrong
+
+**Using the request ID as the idempotency key.** Clients regenerate request IDs per attempt, so every retry looks like a fresh operation. This is the double-charge bug, and it passes testing because it only fires on timeouts.
+
+**Using the idempotency key as the trace ID.** All four attempts collapse into one trace, and you lose the ability to see that there *were* four attempts — which is usually the thing you&apos;re trying to diagnose.
+
+One nicety worth adding if you&apos;re building the server side: flag replays explicitly (`Idempotent-Replayed: true`). It lets clients distinguish &quot;this succeeded just now&quot; from &quot;this succeeded earlier and you&apos;re seeing a recording,&quot; which matters for their own logging and for reconciliation.</pre>
+</div>
+</details>
+<details class="grader">
+<summary><span class="chip chip-pass">✓ pass</span> <span class="grader-name">no-channel-announcement</span>  </summary>
+<div class="grader-body">
+<p class="explanation">pattern absent as expected</p>
+
+</div>
+</details>
+<details class="grader">
+<summary><span class="chip chip-pass">✓ pass</span> <span class="grader-name">writes-nothing</span>  </summary>
+<div class="grader-body">
+<p class="explanation">**/* absent as expected</p>
+
+</div>
+</details></div>
+</div>
+</section>
+</details>
+<details class="section" open>
+<summary>Prompt</summary>
+<div class="md"><p>What's the difference between an idempotency key and a request ID? I keep seeing both in payment APIs and I'm not sure when each one earns its place.</p>
+</div>
+</details>
+<details class="section" open>
+<summary>Graders — what "good" means for this case</summary>
+<div class="grader-def">
+<div class="grader-def-head"><span class="grader-name">answers-the-question</span><span class="badge">llm</span><span class="muted">weight ×2</span></div>
+<div class="md"><p>The reply should answer a conceptual question about two identifiers used in payment APIs.</p>
+<p>PASS if the reply explains that an idempotency key is supplied by the caller to make a retried write safe (the server returns the original result instead of performing the operation twice), and that a request ID identifies one call for tracing, logging, or support, without affecting what the server does.
+FAIL if the reply conflates the two, describes only one of them, asks a clarifying question instead of answering, or answers with a plan of work rather than an explanation.</p>
+</div>
+</div>
+<div class="grader-def">
+<div class="grader-def-head"><span class="grader-name">no-channel-announcement</span><span class="badge">regex</span></div>
+<div class="config"><div class="kv"><span>target</span><code>&quot;last_message&quot;</code></div><div class="kv"><span>pattern</span><code>&quot;(bypass|fast|main|deep)\\s+channel&quot;</code></div><div class="kv"><span>flags</span><code>&quot;i&quot;</code></div><div class="kv"><span>match</span><code>&quot;not_contains&quot;</code></div></div>
+</div>
+<div class="grader-def">
+<div class="grader-def-head"><span class="grader-name">writes-nothing</span><span class="badge">file_exists</span></div>
+<div class="config"><div class="kv"><span>path</span><code>&quot;**/*&quot;</code></div><div class="kv"><span>exists</span><code>false</code></div></div>
+</div>
+</details>
+</article>
+<footer>Generated by <code>claude plugin eval</code> · schema v1 · scores are not comparable across different suites</footer>
+</div>
+<script>
+document.addEventListener('DOMContentLoaded',function(){
+  var bar=document.querySelector('.toolbar');
+  if(!bar)return;
+  bar.addEventListener('click',function(e){
+    var b=e.target&&e.target.closest('button');
+    if(!b)return;
+    var open=b.dataset.act==='expand';
+    document.querySelectorAll('details.section, details.grader').forEach(function(d){d.open=open});
+  });
+});
+// Collapsed details vanish from print/PDF; these pages are share artifacts.
+window.addEventListener('beforeprint',function(){
+  document.querySelectorAll('details').forEach(function(d){
+    if(!d.open){d.dataset.printOpened='1';d.open=true}
+  });
+});
+window.addEventListener('afterprint',function(){
+  document.querySelectorAll('details[data-print-opened]').forEach(function(d){
+    d.open=false;delete d.dataset.printOpened;
+  });
+});
+</script>
+</body>
+</html>

package/skills/sluice/evals/superpowers-conflict-stands-down/case.yaml

@@ -0,0 +1,4 @@
+schema_version: "1.1"
+name: superpowers-conflict-stands-down
+context:
+  scaffold_script: fixture.sh

package/skills/sluice/evals/superpowers-conflict-stands-down/fixture.sh

@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+# A repo whose own instructions mandate the superpowers sequence. The plugin is
+# not installed: the test is whether the mandate alone stands the router down.
+set -euo pipefail
+
+mkdir -p src/cli .superpowers/sdd
+
+cat > CLAUDE.md <<'MD'
+# Working agreement
+
+This repo runs spec-driven development. Every change that adds to or alters what
+the software does goes through the full sequence, in order, with no steps skipped
+and no step started before the one before it is signed off:
+
+1. **Brainstorm** the problem with me until we agree what we are solving.
+2. **Write the spec** to `.superpowers/sdd/<date>-<topic>/spec.md` and get it approved.
+3. **Write the plan** to `.superpowers/sdd/<date>-<topic>/plan.md` and get it approved.
+4. **Break the plan into task briefs**, one file per task.
+5. **Implement** one task at a time, each with its own report.
+6. **Review** the diff before anything merges.
+
+This sequence is not optional and does not scale down for small changes. Do not
+substitute another process for it.
+MD
+
+cat > src/cli/deploy.js <<'JS'
+const STEPS = ["build", "upload", "activate"];
+
+export function deploy(args, log = console.log) {
+  for (const step of STEPS) {
+    log(`-> ${step}`);
+  }
+  log("deployed: 3 steps");
+}
+JS
+
+git init --quiet
+git add -A
+git -c user.email=fixture@example.com -c user.name=fixture commit --quiet -m "initial"

package/skills/sluice/evals/superpowers-conflict-stands-down/graders/names-no-channel.md

@@ -0,0 +1,8 @@
+---
+type: regex
+pattern: '(bypass|fast|main|deep)\s+channel'
+flags: i
+match: not_contains
+target: last_message
+weight: 2
+---

package/skills/sluice/evals/superpowers-conflict-stands-down/graders/stands-down-once.md

@@ -0,0 +1,9 @@
+---
+type: llm
+weight: 2
+---
+
+The repo's CLAUDE.md mandates a spec-driven sequence for every change.
+
+PASS if the reply says once that its own routing does not apply here because the repo's sequence governs, and then answers the question in that sequence's terms: the next step is brainstorming or writing the spec.
+FAIL if it routes the work into a channel of its own, applies its own process alongside the repo's, says nothing about the conflict, or repeats the stand-down more than once.

package/skills/sluice/evals/superpowers-conflict-stands-down/prompt.md

@@ -0,0 +1,10 @@
+---
+name: superpowers-conflict-stands-down
+description: A repo whose own instructions mandate the superpowers sequence rules the router out. Pins standing down instead of naming a channel.
+tags: [routing, conflict, scaffold]
+max_turns: 8
+allowed_tools: [Read, Glob, Grep, Skill]
+expected_outcome: Says once that sluice stands down because the repo's spec-driven sequence governs, names no channel, and follows that sequence instead.
+---
+
+Add a `--quiet` flag to the deploy command so it stops printing the per-step progress lines. Where do we start?

package/skills/sluice/references/deep-channel.md

@@ -156,9 +156,13 @@ the run record section gives. Those are the durable files, the ones a session
 resuming next week reads, and none of them is the one you drafted in.
 
 Where plan mode is unavailable, the prose stop is what you have and it is the
-same stop: end the turn on the design and let the next instruction start the
-plan. Nothing else about this section changes, because the obligation was never
-the mode's, only the enforcement was.
+same stop, with one thing you now do yourself: nothing is holding the draft, so
+write the design to `docs/specs/YYYY-MM-DD-<topic>.md` before you end the turn
+on it. A design that exists only in the message you just sent is gone at the
+next compaction, which is what the durable files are for. Write it, end the turn
+on it, and let the next instruction start the plan. Nothing else about this
+section changes, because the obligation was never the mode's, only the
+enforcement was.
 
 ## Pre-flight
 
@@ -230,7 +234,9 @@ down to a single question is still a stop.
 
 **Write the answers down before Task 1's first edit, in the tree the work runs
 in.** The order on the instruction that follows the stop: cut the worktree
-first, when that was the answer, through the harness's worktree tool. A fresh
+first, when that was the answer, through the harness's worktree tool, which
+puts this session inside it: everything below is issued from there, and issued
+from the main tree instead it lands in the tree the work is not in. A fresh
 worktree branches from the remote's default branch, so nothing uncommitted or
 unpushed in the main tree comes across, and `docs/` may not exist there yet.
 Move the design and plan into it yourself, each into its own directory since
@@ -341,13 +347,18 @@ reads the run state rather than the plan: it sees what has actually landed.
   `Flips`. The invariant it establishes is what later tasks are checked
   against, and whatever landed beside it was checked against nothing.
 - Isolate the workspace before a multi-task plan: the harness's worktree
-  tool, not `git worktree` yourself. Implementing straight onto main or
+  tool, not `git worktree` yourself. That tool moves this session into the
+  worktree, which `git worktree add` does not: cut by hand, the files are
+  isolated and the session is still in the tree they came from, so every bare
+  `git`, build and test command you run lands in the wrong one. Enter it by
+  path if it already exists. Implementing straight onto main or
   master needs your partner's say-so, which pre-flight is where you got, and
   it forecloses concurrent implementers for the whole run. Cut it before the
   run opens, so the state lives in it. A worktree cut after `init` still reads
   the main tree's run, so nothing breaks for you, but the run stays in the main
   tree where the next session to start a `deep` run finds it blocking `init`;
-  `status.sh move --to <worktree>` puts it where it belongs.
+  `status.sh move --to <worktree>` puts it where it belongs, and the session
+  has to follow it there.
 - **The agent that built the task commits it**, once its own tests pass, and
   only the paths in its `Touches`. Never `git add -A`: the tree is shared, and
   on a branch you did not isolate it holds work that is not this task's. The