@iceinvein/agent-skills 0.18.2 → 0.19.0

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/scripts/session-start.sh

@@ -31,12 +31,63 @@ if [ ! -t 0 ]; then
 fi
 cwd="$PWD"
 source=""
+sid=""
 if [ -n "$input" ] && command -v jq >/dev/null 2>&1; then
 	got="$(printf '%s' "$input" | jq -r '.cwd // empty' 2>/dev/null || true)"
 	[ -n "$got" ] && cwd="$got"
 	source="$(printf '%s' "$input" | jq -r '.source // empty' 2>/dev/null || true)"
+	sid="$(printf '%s' "$input" | jq -r '.session_id // empty' 2>/dev/null || true)"
 fi
 
+# The baseline for stop-guard.sh's entry check. That guard asks whether the tree
+# moved while this session held it, and this hook runs at the only moment the
+# answer is still "not yet". Taken for every session, run or no run: a session
+# that routes nothing is exactly the one the guard is there to catch.
+#
+# It lives beside the transcripts rather than in the tree, because a session
+# that opens in someone's repo should not leave a directory behind in it, and
+# because a baseline is spent the moment its session ends. A stamp that is
+# missing, unreadable or stale costs the guard its check and nothing else, so
+# every failure here is silent.
+stamp_baseline() {
+	local sid="$1" src="$2" tree digest
+	[ -n "$sid" ] || return 0
+	case "$sid" in */* | .*) return 0 ;; esac
+	tree="$(git -C "$cwd" rev-parse --show-toplevel 2>/dev/null)" || return 0
+	[ -n "$tree" ] || return 0
+
+	local dir="${CLAUDE_CONFIG_DIR:-${HOME:-}/.claude}/sluice/stamps"
+	mkdir -p "$dir" 2>/dev/null || return 0
+
+	# This hook fires again on compact, resume and clear, all carrying the
+	# session id the first one carried, and they are not the same event.
+	#
+	# A compact happens inside a session that never let go of the tree, so its
+	# baseline still stands; re-stamping there would move it onto the work
+	# already done and read a half-finished session as an untouched tree, and a
+	# session long enough to compact is the one this exists for.
+	#
+	# A resume reopens a session that had stopped, and a clear throws away what
+	# it was doing. Both leave the old baseline describing a tree from some
+	# unbounded time ago, and everything that happened to it since, by a
+	# colleague or an editor or another session, would be charged to whoever
+	# reopens it. Those two start again, and give back the nudge the previous
+	# sitting may have spent.
+	case "$src" in
+		resume | clear) rm -f "$dir/$sid" "${dir%/stamps}/nudged/$sid" 2>/dev/null || true ;;
+		*) [ -e "$dir/$sid" ] && return 0 ;;
+	esac
+
+	# A stamp outlives nothing but its session, and the harness never deletes
+	# one. Without this the directory grows for the life of the machine.
+	find "$dir" "${dir%/stamps}/nudged" -type f -mtime +7 -delete 2>/dev/null || true
+
+	digest="$(bash "$here/tree-snapshot.sh" "$tree" 2>/dev/null)" || return 0
+	[ -n "$digest" ] || return 0
+	printf '%s\n%s\n' "$tree" "$digest" >"$dir/$sid" 2>/dev/null || true
+}
+stamp_baseline "$sid" "$source"
+
 # No gate of its own beyond the directory existing: status.sh anchors on the
 # main worktree of whatever tree it is given, which is what lets a session
 # opened in a subdirectory or a linked worktree find the run, and on a tree