feat(v0.6.5): execution-grounded skill-utility report (adam-skill-utility)

Ranks skills by good:bad outcome co-occurrence (Wilson LB + lift vs
baseline) over the journal's active_skills payloads — the SkillsInjector
(arXiv 2605.29794) execution-grounded utility signal Δ(s), computed from
data already collected, no training.

- reuses adam-score NEGATIVE_SIGNAL_TYPES + entrySeverity (single source of truth)
- registered in install.sh helper-script copy loop
- /reflect pre-step surfaces worst below-baseline skills to the USER as
  advisory (co-occurrence != causation; not fed to the analyst's proposal machinery)
- Test 119 added; full suite 141/141 green
This commit is contained in:
2026-06-02 01:47:40 +01:00
parent c23b09cc09
commit 4d1276a73f
4 changed files with 310 additions and 1 deletions
+272
View File
@@ -0,0 +1,272 @@
#!/usr/bin/env node
// adam-skill-utility.mjs — execution-grounded per-skill utility report.
//
// Inspired by SkillsInjector (arXiv 2605.29794v1), which shows skill injection
// should be driven by execution-grounded *utility* Δ(t,s), not surface keyword
// match — and that some topically-relevant skills actively *lower* success.
// The paper learns Δ(t,s) from rollout outcomes. We don't train anything: the
// adam journal already attaches `active_skills` to both positive outcome events
// (task_completed, clean_recovery, correction_free_streak) and negative ones
// (dead_end, tool_error_loop, …). So we approximate Δ(s) as a co-occurrence
// ratio over the data we already collect.
//
// CAVEAT (honest): this is CO-OCCURRENCE, not causation. A skill active during
// a dead_end did not necessarily cause it. Read the report as "which skills
// correlate with friction", a prompt for review — never as proof.
//
// Metric, per skill active during scored events:
// pos / neg — count of positive / negative outcome events it co-occurred with
// share — pos / (pos+neg)
// lift — share global_baseline (>0 above baseline, <0 below)
// wLB — Wilson 95% lower bound of the positive proportion; ranks
// *reliably* below-baseline skills to the top (sample-aware)
// sevNeg — severity-weighted negative sum (adam SEVERITY_DIVISORS)
// topNeg — dominant negative event type
// Rows sorted worst-first (lowest wLB) so harmful/over-eager skills surface.
//
// CLI:
// adam-skill-utility.mjs [--home <path>] [--input <jsonl-path>]
// [--min <n>] [--days <n>] [--json]
// --min min event count (n) to treat a skill's signal as confident (default 8)
// --days only consider events within the last <n> days (default: all)
// --json emit machine-readable JSON instead of the text table
//
// Reuses adam-utils (jsonl IO) and adam-score (canonical NEGATIVE set +
// severity), so the positive/negative taxonomy stays single-sourced.
import { readFileSync } from "node:fs";
import { join } from "node:path";
import { homedir } from "node:os";
import { readJsonlSafe, listJsonlFiles } from "./adam-utils.mjs";
import { NEGATIVE_SIGNAL_TYPES, entrySeverity } from "./adam-score.mjs";
// Positive outcome signals (mirror adam's vocabulary; task_completed is adam's
// canonical "clean task", the same one adam-score uses for reinforcement).
export const POSITIVE_SIGNAL_TYPES = new Set([
"task_completed",
"clean_recovery",
"correction_free_streak",
]);
export const DEFAULT_MIN_SAMPLE = 8;
function round(x) {
return Math.round(x * 1000) / 1000;
}
// Wilson score interval lower bound for a binomial proportion. Sample-aware:
// a skill with 1 pos / 0 neg does NOT outrank one with 40 pos / 2 neg.
export function wilsonLower(pos, n, z = 1.96) {
if (n <= 0) return 0;
const p = pos / n;
const z2 = z * z;
const denom = 1 + z2 / n;
const center = p + z2 / (2 * n);
const margin = z * Math.sqrt((p * (1 - p) + z2 / (4 * n)) / n);
return (center - margin) / denom;
}
// computeSkillUtility: pure. entries → { baseline, totalPos, totalNeg, min, skills[] }.
export function computeSkillUtility(entries, opts = {}) {
const min = Number.isFinite(opts.min) ? opts.min : DEFAULT_MIN_SAMPLE;
const per = new Map();
let totalPos = 0;
let totalNeg = 0;
for (const e of entries || []) {
if (!e || typeof e !== "object") continue;
const isPos = POSITIVE_SIGNAL_TYPES.has(e.type);
const isNeg = NEGATIVE_SIGNAL_TYPES.has(e.type);
if (!isPos && !isNeg) continue;
if (isPos) totalPos++;
else totalNeg++;
const sev = isNeg ? entrySeverity(e) : 0;
const skills = Array.isArray(e.active_skills) ? e.active_skills : [];
for (const slug of skills) {
if (!slug || typeof slug !== "string") continue;
if (!per.has(slug)) {
per.set(slug, { pos: 0, neg: 0, sevNeg: 0, negTypes: {}, recent_ts: null });
}
const s = per.get(slug);
if (isPos) {
s.pos++;
} else {
s.neg++;
s.sevNeg += sev;
s.negTypes[e.type] = (s.negTypes[e.type] || 0) + 1;
}
const ts = typeof e.ts === "string" ? e.ts : null;
if (ts && (!s.recent_ts || ts > s.recent_ts)) s.recent_ts = ts;
}
}
const scored = totalPos + totalNeg;
const baseline = scored ? totalPos / scored : 0;
const skills = [];
for (const [slug, s] of per.entries()) {
const n = s.pos + s.neg;
const share = n ? s.pos / n : 0;
const topNeg = Object.entries(s.negTypes).sort((a, b) => b[1] - a[1])[0];
skills.push({
skill: slug,
n,
pos: s.pos,
neg: s.neg,
share: round(share),
lift: round(share - baseline),
wLB: round(wilsonLower(s.pos, n)),
sevNeg: s.sevNeg,
topNeg: topNeg ? topNeg[0] : null,
lowSample: n < min,
recent_ts: s.recent_ts,
});
}
// Worst-first: lowest Wilson lower bound, then most negatives.
skills.sort(
(a, b) =>
a.wLB - b.wLB ||
b.neg - a.neg ||
(a.skill < b.skill ? -1 : a.skill > b.skill ? 1 : 0),
);
return { baseline: round(baseline), totalPos, totalNeg, min, skills };
}
function parseArgs(argv) {
const args = { home: null, input: null, min: DEFAULT_MIN_SAMPLE, days: null, json: false, help: false };
for (let i = 0; i < argv.length; i++) {
const a = argv[i];
if (a === "--home" && i + 1 < argv.length) args.home = argv[++i];
else if (a === "--input" && i + 1 < argv.length) args.input = argv[++i];
else if (a === "--min" && i + 1 < argv.length) args.min = Number(argv[++i]);
else if (a === "--days" && i + 1 < argv.length) args.days = Number(argv[++i]);
else if (a === "--json") args.json = true;
else if (a === "--help" || a === "-h") args.help = true;
}
return args;
}
function readAllStdin() {
try { return readFileSync(0, "utf8"); } catch { return ""; }
}
function entriesFromText(text) {
const out = [];
for (const line of (text || "").split("\n")) {
if (!line) continue;
try { out.push(JSON.parse(line)); } catch { /* skip */ }
}
return out;
}
// Same gathering strategy as adam-score.mjs: explicit --input, else piped
// stdin (e.g. from adam-window.mjs), else the active journal + rotated files.
function gatherInputEntries(args) {
if (args.input) return readJsonlSafe(args.input);
if (!process.stdin.isTTY) {
const piped = readAllStdin();
if (piped && piped.trim()) return entriesFromText(piped);
}
const home = args.home || join(homedir(), ".claude");
const adamRoot = join(home, "adam");
const sources = [join(adamRoot, "journal.jsonl"), ...listJsonlFiles(join(adamRoot, "journal"))];
const all = [];
for (const p of sources) {
for (const e of readJsonlSafe(p)) all.push(e);
}
return all;
}
function filterByDays(entries, days) {
if (!Number.isFinite(days) || days <= 0) return entries;
// Anchor the window to the newest ts in the data (avoids Date.now()
// nondeterminism and works on historical exports).
let maxMs = 0;
for (const e of entries) {
const ms = e && typeof e.ts === "string" ? Date.parse(e.ts) : NaN;
if (Number.isFinite(ms) && ms > maxMs) maxMs = ms;
}
if (!maxMs) return entries;
const cutoff = maxMs - days * 86400000;
return entries.filter((e) => {
const ms = e && typeof e.ts === "string" ? Date.parse(e.ts) : NaN;
return Number.isFinite(ms) ? ms >= cutoff : false;
});
}
function pad(s, w) {
s = String(s);
return s.length >= w ? s : s + " ".repeat(w - s.length);
}
function padL(s, w) {
s = String(s);
return s.length >= w ? s : " ".repeat(w - s.length) + s;
}
function renderText(report) {
const { baseline, totalPos, totalNeg, min, skills } = report;
const lines = [];
lines.push("adam skill-utility report — execution-grounded Δ(skill) proxy");
lines.push(
`baseline positive-rate ${(baseline * 100).toFixed(1)}% ` +
`(${totalPos} positive / ${totalNeg} negative outcome events) min-sample n≥${min}`,
);
lines.push("CAVEAT: co-occurrence, not causation. Worst-first. ⚠ = below baseline with n≥min.");
lines.push("");
const head =
pad("skill", 44) + padL("n", 5) + padL("pos", 6) + padL("neg", 6) +
padL("share", 8) + padL("lift", 8) + padL("wLB", 7) + padL("sevNeg", 8) +
" " + pad("topNeg", 18) + "flag";
lines.push(head);
lines.push("-".repeat(head.length));
for (const s of skills) {
const below = s.lift < 0 && !s.lowSample;
const flag = below ? "⚠" : s.lowSample ? "·(low n)" : "";
lines.push(
pad(s.skill, 44) +
padL(s.n, 5) +
padL(s.pos, 6) +
padL(s.neg, 6) +
padL((s.share * 100).toFixed(0) + "%", 8) +
padL((s.lift >= 0 ? "+" : "") + (s.lift * 100).toFixed(0) + "%", 8) +
padL(s.wLB.toFixed(2), 7) +
padL(s.sevNeg, 8) +
" " +
pad(s.topNeg || "-", 18) +
flag,
);
}
return lines.join("\n");
}
function main() {
const args = parseArgs(process.argv.slice(2));
if (args.help) {
process.stdout.write(
"usage: adam-skill-utility.mjs [--home <path>] [--input <jsonl-path>] " +
"[--min <n>] [--days <n>] [--json]\n",
);
process.exit(0);
}
try {
let entries = gatherInputEntries(args);
entries = filterByDays(entries, args.days);
const report = computeSkillUtility(entries, { min: args.min });
if (args.json) {
process.stdout.write(JSON.stringify(report) + "\n");
} else {
process.stdout.write(renderText(report) + "\n");
}
process.exit(0);
} catch (e) {
process.stderr.write(`adam-skill-utility error: ${e.message}\n`);
process.exit(1);
}
}
if (import.meta.url === `file://${process.argv[1]}`) {
main();
}
+28
View File
@@ -18,6 +18,7 @@ APPLYREIN="$REAL_HOME/.claude/adam/scripts/adam-apply-reinforcement.mjs"
UPGRADE="$REAL_HOME/.claude/adam/scripts/adam-upgrade.mjs"
BATCH="$REAL_HOME/.claude/adam/scripts/adam-batch.mjs"
ROLLBACK="$REAL_HOME/.claude/adam/scripts/adam-rollback.mjs"
SKILLUTIL="$REAL_HOME/.claude/adam/scripts/adam-skill-utility.mjs"
TMP_HOME="$(mktemp -d -t adam-test.XXXXXX)"
trap 'rm -rf "$TMP_HOME"' EXIT INT TERM
@@ -37,6 +38,7 @@ APPLYREIN_RUN(){ HOME="$TMP_HOME" node "$APPLYREIN" "$@" --home "$TMP_HOME/.clau
UPGRADE_RUN() { HOME="$TMP_HOME" node "$UPGRADE" "$@"; }
BATCH_RUN() { HOME="$TMP_HOME" node "$BATCH" "$@"; }
ROLLBACK_RUN(){ HOME="$TMP_HOME" node "$ROLLBACK" "$@"; }
SKILLUTIL_RUN(){ HOME="$TMP_HOME" node "$SKILLUTIL" "$@"; }
PASS=0
FAIL=0
@@ -2148,6 +2150,32 @@ else
fi
rm -f "$ROOT/proposals/"*rb-ab-001* "$ROOT/applied/"*rb-ab-001* "$ROOT/ab-tracking.jsonl" "$ROOT/active-nudges.json"
# --- Test 119: adam-skill-utility ranks friction-correlated skills below baseline ---
echo "Test 119: adam-skill-utility computes per-skill good:bad utility (execution-grounded Δ)"
reset_state
SU_INPUT="$TMP_HOME/su-input.jsonl"
{
for i in 1 2 3 4 5; do echo "{\"ts\":\"2026-05-20T0$i:00:00Z\",\"session\":\"sSU\",\"type\":\"task_completed\",\"active_skills\":[\"goodskill\"]}"; done
for i in 1 2 3 4 5; do echo "{\"ts\":\"2026-05-20T1$i:00:00Z\",\"session\":\"sSU\",\"type\":\"dead_end\",\"count\":8,\"active_skills\":[\"badskill\"]}"; done
} > "$SU_INPUT"
su_out=$(SKILLUTIL_RUN --input "$SU_INPUT" --json --min 3 2>/dev/null)
su_check=$(echo "$su_out" | node -e '
let buf=""; process.stdin.on("data",d=>buf+=d).on("end",()=>{
try {
const p=JSON.parse(buf);
const bad=p.skills.find(s=>s.skill==="badskill");
const good=p.skills.find(s=>s.skill==="goodskill");
const ok = bad && good && bad.lift<0 && good.lift>0 && p.skills[0].skill==="badskill" && bad.neg===5 && good.pos===5;
console.log(ok?"ok":"bad:"+JSON.stringify({bad,good,first:p.skills[0]&&p.skills[0].skill}));
} catch(e){ console.log("parse-error:"+e.message); }
});')
if [ "$su_check" = "ok" ]; then
echo " PASS: badskill below baseline + ranked worst-first, goodskill above"; PASS=$((PASS+1))
else
echo " FAIL: skill-utility ranking wrong ($su_check)"; FAIL=$((FAIL+1))
fi
rm -f "$SU_INPUT"
echo
echo "Results: $PASS passed, $FAIL failed"
[ "$FAIL" = "0" ]
+2 -1
View File
@@ -126,7 +126,8 @@ copy_file "$SRC/adam/scripts/adam-archive.mjs" "$DEST/adam
copy_file "$SRC/adam/scripts/adam-upgrade.mjs" "$DEST/adam/scripts/adam-upgrade.mjs"
# v0.3.3 helper scripts — invoked from SKILL.md / hooks / analyst flow
for _adam_script in adam-utils adam-window adam-explain adam-nudge-eligibility adam-cooldown \
adam-score adam-ab-measure adam-apply-reinforcement adam-batch adam-rollback; do
adam-score adam-ab-measure adam-apply-reinforcement adam-batch adam-rollback \
adam-skill-utility; do
copy_file "$SRC/adam/scripts/${_adam_script}.mjs" \
"$DEST/adam/scripts/${_adam_script}.mjs"
run "chmod +x \"$DEST/adam/scripts/${_adam_script}.mjs\""
+8
View File
@@ -81,6 +81,14 @@ node ~/.claude/adam/scripts/adam-batch.mjs --input /tmp/adam-windowed-journal.js
This groups entries by (signal_type, cluster_key) and reports per-batch metadata including `has_context_window` (whether transcript evidence is attached). If the script fails: log stderr, pass `null` to the analyst (graceful degradation — analyst falls back to raw journal clustering).
**Skill utility** (execution-grounded selection signal, in the spirit of SkillsInjector arXiv 2605.29794 — utility Δ(s), not surface match): compute per-skill good:bad outcome ratios over the windowed journal:
```bash
node ~/.claude/adam/scripts/adam-skill-utility.mjs --input /tmp/adam-windowed-journal.jsonl --json > /tmp/adam-skill-utility.json 2> /tmp/adam-skill-utility.log
```
This ranks skills by how often they co-occur with positive (`task_completed`, `clean_recovery`, `correction_free_streak`) vs negative outcome events, surfacing skills below the baseline positive rate (with sufficient sample) — advisory candidates for description disambiguation or archival. **CO-OCCURRENCE, NOT CAUSATION**: display the worst 3 below-baseline skills (`lift < 0`, not low-sample) to the *user* as a one-line advisory before listing proposals (e.g. `skill-utility: chezmoi 9% pos n=85, ghostty-config 14% pos n=50, …`). Do NOT feed this into the analyst's proposal machinery or auto-draft skill-archival from it — the human decides. If the script fails: log stderr, skip (best-effort).
### 2. Dispatch the analyst (two-stage pipeline)
MOSS §3.3: "A single prompt asked to diagnose, plan, implement, verify, and decide overloads context and produces lower-quality output than a sequenced flow." The analyst is dispatched in two stages with a validation gate between them.