fix(v0.6.2): A/B volume normalization + memory frontmatter schema

Two issues surfaced by running ADAM's /reflect loop on a large real journal
(4015 entries, 119 sessions) — both caused false/broken auto-apply behavior.

1. A/B over-reported regressions (adam-ab-measure.mjs).
   Regressions were measured on RAW originating-signal counts pre vs post. On a
   busy, growing journal almost every signal count rises post-apply regardless
   of whether the proposal helped — so the loop flagged 9 false "regressions"
   (and would auto-roll-back good proposals). Now the delta is computed on the
   signal's SHARE of total activity (rate = count / window-total). Falls back to
   the raw-count delta when the signal is the only activity in the window
   (preserves prior behavior + all existing A/B tests). Output adds
   raw_delta_pct, pre_total, post_total, normalized for transparency.

2. Memory frontmatter drift (agents/adam.md, SKILL.md).
   The drafting protocol emitted flat `type:`/`originSessionId:` with a prose
   `name`, but the live auto-memory store uses `name` = slug plus a
   `metadata: {node_type, type, originSessionId}` block. Auto-applied memories
   could fail to load/categorize. Protocol + apply-time validation now require
   the live metadata.* schema and cross-checking against an existing file.

Tests: 132 -> 134. New: volume growth (raw +200%) with flat activity-share
classifies neutral, not regressed; a genuine share increase still classifies
regressed.
This commit is contained in:
2026-05-29 12:37:10 +01:00
parent 3a54d7d3e1
commit d929101af4
5 changed files with 109 additions and 20 deletions
+4 -3
View File
@@ -13,7 +13,7 @@ Watches the friction in your coding sessions, clusters the signals via an LLM an
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
[![Version](https://img.shields.io/github/v/release/lukaszraczylo/claude-adam?label=version&color=blue)](https://github.com/lukaszraczylo/claude-adam/releases)
[![Tests](https://img.shields.io/badge/tests-132%20passing-brightgreen.svg)](./adam/tests/run-tests.sh)
[![Tests](https://img.shields.io/badge/tests-134%20passing-brightgreen.svg)](./adam/tests/run-tests.sh)
[![Node](https://img.shields.io/badge/node-22%2B-339933.svg)](https://nodejs.org)
[![Platform](https://img.shields.io/badge/platform-macOS%20%7C%20Linux-lightgrey.svg)]()
@@ -54,7 +54,7 @@ The installer copies files into `~/.claude/`, offers to merge ADAM's hook entrie
Then:
```sh
bash ~/.claude/adam/tests/run-tests.sh # expect: 132 passed, 0 failed
bash ~/.claude/adam/tests/run-tests.sh # expect: 134 passed, 0 failed
# … start a fresh Claude Code session …
/reflect # walks the proposal queue
/reflect --explain # also shows the analyst's clustering trace
@@ -248,11 +248,12 @@ Or pass `--explain` to `/reflect` to render the full trace inline.
│ ├── adam-apply-reinforcement.mjs # reinforcement proposal apply
│ ├── adam-upgrade.mjs # .adam-new file UX (list/diff/accept)
│ └── adam-archive.mjs # post-apply journal cleanup
└── tests/run-tests.sh # 132 isolated tests; never touches live state
└── tests/run-tests.sh # 134 isolated tests; never touches live state
```
## What's new
- **v0.6.2** — two fixes surfaced by running ADAM's loop on a large real journal. **(1) A/B volume normalization** (`adam-ab-measure.mjs`): regressions are now measured on the signal's *share* of total activity (rate = count / window-total), not raw count — so a generally busier journal after an apply no longer masquerades as a regression. Falls back to raw delta when the signal is the only activity in the window (preserves prior behavior + tests); output adds `raw_delta_pct`, `pre_total`, `post_total`, `normalized` for transparency. **(2) Memory frontmatter schema** (`agents/adam.md`, `SKILL.md`): the drafting protocol now emits the live auto-memory shape — `name` = slug + a `metadata: {node_type, type, originSessionId}` block — instead of flat `type:`/`originSessionId:`, so auto-applied memories load and categorize correctly. 134 tests (up from 132).
- **v0.6.1** — new `file_reread` signal (MOSS §1 harness self-modification, proposed and approved through ADAM's own `/reflect` loop). Consecutive Reads of the same file at different `offset`/`limit` escaped `retry_loop`'s arg-hash dedup and leaked into `tool_error_loop`; `file_reread` now catches them (same file ≥3× in the 10-event window, offset-agnostic, guarded against double-counting byte-identical reads). Fully wired: detection (`adam-observe.mjs`), 14-day window (`adam-window.mjs`), severity divisor 3 (`adam-score.mjs`), file-basename clustering (`adam-batch.mjs`), and the analyst rubric/spec. 132 tests (up from 126).
- **v0.6.0** — review hardening. Struggle signals now emit `active_skills`, so `silent_drift`'s primary cluster key and the §5b skill-attribution sub-clustering (+1 rubric bonus) actually fire (both were silently dead). `proposal_fingerprint` is now deterministically computable via `adam-cooldown.mjs --compute` instead of asking the LLM analyst to hand-compute a djb2 hash; spec now mandates a *stable* cluster id so fingerprints reproduce across runs. `reinforcement` proposals are correctly excluded from A/B tracking (the spec previously contradicted itself). `adam-nudge.mjs` pending-upgrade check now mirrors the full install set (`adam-utils`/`adam-batch`/`adam-rollback` were missing). Doc/test-count drift corrected. 126 tests (up from 114).
- **v0.5.0** — MOSS-grounded self-evolution (arXiv 2605.22794). Transcript capture: `context_window` field on struggle signals captures 8 surrounding events for evidence-based diagnosis. Two-stage analysis pipeline: diagnose+plan → inter-stage validation → implement (§3.3). Evidence batching via `adam-batch.mjs`: pre-clusters journal into coherent failure batches (§3.1). Pre-apply verification: 4-check deterministic gate before auto-apply (§3.4). Auto-rollback via `adam-rollback.mjs`: reverts regressed proposals detected by A/B measurement, creates regression nudges (§3.5). Harness self-modification: new `harness_edit` proposal type lets ADAM propose edits to its own scripts with test-suite-gated apply (§1 Table 1). Keypoint matrix: 5 capability dimensions scored per batch for structured evaluation (§4.2). 114 tests (up from 94).
+44 -7
View File
@@ -3,11 +3,19 @@
//
// Reads ~/.claude/adam/ab-tracking.jsonl (one line per auto-apply event,
// written by adam-self-improvement/SKILL.md), then for each entry old enough
// (>= --min-age-days; default 7) compares signal counts in the 7-day window
// BEFORE applied_at against the 7-day window AFTER applied_at across the
// (>= --min-age-days; default 7) compares the originating signal in the 7-day
// window BEFORE applied_at against the 7-day window AFTER applied_at across the
// full journal corpus (active + rotated). Surfaces regressions so /reflect
// can flag proposals that made things worse.
//
// Volume normalization: when the windows contain other (non-originating)
// activity, the delta is computed on the signal's SHARE of total activity
// (rate = count / total), not its raw count — so a generally busier journal
// after apply does not masquerade as a regression. When the signal is the only
// activity in the windows, it falls back to the raw-count delta. Output carries
// both `delta_pct` (drives status) and `raw_delta_pct` + `normalized` for
// transparency.
//
// CLI:
// adam-ab-measure.mjs [--home <path>] [--format json|table] [--min-age-days N]
//
@@ -92,31 +100,60 @@ export function computeDeltas(entries, journal, opts = {}) {
const preStart = appliedAt - windowDays * DAY_MS;
const postEnd = appliedAt + windowDays * DAY_MS;
// preCount/postCount = originating-signal occurrences; preTotal/postTotal =
// ALL journal entries in the window (the activity denominator).
let preCount = 0;
let postCount = 0;
let preTotal = 0;
let postTotal = 0;
for (const je of journal || []) {
if (!je || typeof je !== "object") continue;
if (!sigSet.has(je.type)) continue;
const t = tsMs(je);
if (Number.isNaN(t)) continue;
if (t >= preStart && t < appliedAt) preCount++;
else if (t >= appliedAt && t < postEnd) postCount++;
const inPre = t >= preStart && t < appliedAt;
const inPost = t >= appliedAt && t < postEnd;
if (!inPre && !inPost) continue;
if (inPre) preTotal++; else postTotal++;
if (!sigSet.has(je.type)) continue;
if (inPre) preCount++; else postCount++;
}
let status;
let deltaPct;
let rawDeltaPct = null;
let normalized = false;
if (preCount === 0) {
status = "no_baseline";
deltaPct = null;
} else {
deltaPct = ((postCount - preCount) / preCount) * 100;
rawDeltaPct = Math.round(((postCount - preCount) / preCount) * 10000) / 100;
// Volume normalization: when the windows contain non-originating activity,
// compare the signal's SHARE of activity (rate), not its absolute count —
// otherwise a generally busier post-window masquerades as a regression.
// No background (signal IS the only activity) → fall back to raw delta,
// preserving prior behavior.
const hasBackground = (preTotal - preCount) + (postTotal - postCount) > 0;
if (hasBackground && postTotal > 0) {
const preRate = preCount / preTotal; // preTotal >= preCount > 0
const postRate = postCount / postTotal;
deltaPct = ((postRate - preRate) / preRate) * 100;
normalized = true;
} else {
deltaPct = ((postCount - preCount) / preCount) * 100;
}
// Round to 2 dp for stable comparison + presentation.
deltaPct = Math.round(deltaPct * 100) / 100;
if (deltaPct <= IMPROVED_PCT) status = "improved";
else if (deltaPct >= REGRESSED_PCT) status = "regressed";
else status = "neutral";
}
out.push({ ...base, pre_count: preCount, post_count: postCount, delta_pct: deltaPct, status });
out.push({
...base,
pre_count: preCount, post_count: postCount,
pre_total: preTotal, post_total: postTotal,
raw_delta_pct: rawDeltaPct, normalized,
delta_pct: deltaPct, status,
});
}
return out;
}
+44
View File
@@ -2014,6 +2014,50 @@ done
assert_grep "$ROOT/journal.jsonl" '"type":"retry_loop"' "3x byte-identical reads emit retry_loop"
assert_no_grep "$ROOT/journal.jsonl" '"type":"file_reread"' "byte-identical reads NOT double-counted as file_reread (sameToolArgs>=RETRY guard)"
# --- Test 112: A/B volume normalization — busier journal does NOT fake a regression ---
echo "Test 112: A/B volume-normalized (raw +200% but flat share → neutral)"
reset_state
applied_at_ms=$(node -e 'console.log(Date.now() - 14*86400000)')
cat > "$ROOT/ab-tracking.jsonl" <<EOF
{"applied_at":$applied_at_ms,"proposal_id":"ab-vol-001","proposal_type":"memory","target_skill":"vol","proposal_fingerprint":"fpV","originating_signals":[{"type":"correction","count":2,"session_ids":["sV"]}],"pre_window_days":7}
EOF
> "$ROOT/journal.jsonl"
# pre window: 2 correction + 8 dead_end (rate 0.2)
for i in 1 2; do ts=$(node -e "console.log(new Date(Date.now()-(15+$i*0.2)*86400000).toISOString())"); echo "{\"ts\":\"$ts\",\"session\":\"sV\",\"type\":\"correction\"}" >> "$ROOT/journal.jsonl"; done
for i in 1 2 3 4 5 6 7 8; do ts=$(node -e "console.log(new Date(Date.now()-(15+$i*0.1)*86400000).toISOString())"); echo "{\"ts\":\"$ts\",\"session\":\"sV\",\"type\":\"dead_end\"}" >> "$ROOT/journal.jsonl"; done
# post window: 6 correction + 24 dead_end (rate 0.2 — share unchanged, raw count +200%)
for i in $(seq 1 6); do ts=$(node -e "console.log(new Date(Date.now()-(8+$i*0.1)*86400000).toISOString())"); echo "{\"ts\":\"$ts\",\"session\":\"sV\",\"type\":\"correction\"}" >> "$ROOT/journal.jsonl"; done
for i in $(seq 1 24); do ts=$(node -e "console.log(new Date(Date.now()-(8+$i*0.05)*86400000).toISOString())"); echo "{\"ts\":\"$ts\",\"session\":\"sV\",\"type\":\"dead_end\"}" >> "$ROOT/journal.jsonl"; done
out=$(ABMEASURE_RUN --format json 2>/dev/null)
if echo "$out" | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{const a=JSON.parse(b);const e=a.find(x=>x.proposal_id==="ab-vol-001");process.exit(e&&e.normalized===true&&e.raw_delta_pct===200&&e.status==="neutral"?0:1)})'; then
echo " PASS: volume growth normalized → neutral (raw +200%)"; PASS=$((PASS+1))
else
echo " FAIL: volume normalization wrong (got: $out)"; FAIL=$((FAIL+1))
fi
rm -f "$ROOT/ab-tracking.jsonl"
# --- Test 113: A/B genuine rate regression still flagged ---
echo "Test 113: A/B genuine share increase → regressed"
reset_state
applied_at_ms=$(node -e 'console.log(Date.now() - 14*86400000)')
cat > "$ROOT/ab-tracking.jsonl" <<EOF
{"applied_at":$applied_at_ms,"proposal_id":"ab-vol-002","proposal_type":"memory","target_skill":"vol2","proposal_fingerprint":"fpV2","originating_signals":[{"type":"correction","count":2,"session_ids":["sV2"]}],"pre_window_days":7}
EOF
> "$ROOT/journal.jsonl"
# pre: 2 correction + 8 dead_end (rate 0.2)
for i in 1 2; do ts=$(node -e "console.log(new Date(Date.now()-(15+$i*0.2)*86400000).toISOString())"); echo "{\"ts\":\"$ts\",\"session\":\"sV2\",\"type\":\"correction\"}" >> "$ROOT/journal.jsonl"; done
for i in 1 2 3 4 5 6 7 8; do ts=$(node -e "console.log(new Date(Date.now()-(15+$i*0.1)*86400000).toISOString())"); echo "{\"ts\":\"$ts\",\"session\":\"sV2\",\"type\":\"dead_end\"}" >> "$ROOT/journal.jsonl"; done
# post: 6 correction + 6 dead_end (rate 0.5 — share up → genuine regression)
for i in $(seq 1 6); do ts=$(node -e "console.log(new Date(Date.now()-(8+$i*0.1)*86400000).toISOString())"); echo "{\"ts\":\"$ts\",\"session\":\"sV2\",\"type\":\"correction\"}" >> "$ROOT/journal.jsonl"; done
for i in 1 2 3 4 5 6; do ts=$(node -e "console.log(new Date(Date.now()-(8+$i*0.07)*86400000).toISOString())"); echo "{\"ts\":\"$ts\",\"session\":\"sV2\",\"type\":\"dead_end\"}" >> "$ROOT/journal.jsonl"; done
out=$(ABMEASURE_RUN --format json 2>/dev/null)
if echo "$out" | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{const a=JSON.parse(b);const e=a.find(x=>x.proposal_id==="ab-vol-002");process.exit(e&&e.normalized===true&&e.status==="regressed"?0:1)})'; then
echo " PASS: genuine share increase → regressed"; PASS=$((PASS+1))
else
echo " FAIL: genuine regression missed (got: $out)"; FAIL=$((FAIL+1))
fi
rm -f "$ROOT/ab-tracking.jsonl"
echo
echo "Results: $PASS passed, $FAIL failed"
[ "$FAIL" = "0" ]
+16 -9
View File
@@ -250,10 +250,12 @@ Required structure:
```markdown
---
name: <human-readable name, ≤80 chars>
description: <one-line description used to decide future relevance — be specific, ≤200 chars>
type: user | feedback | project | reference
originSessionId: <session_id from journal entries that fed this cluster>
name: <slug — snake_case, MUST equal the target filename without `.md`, e.g. feedback_go_test_cache>
description: "<one-line used to decide future relevance — be specific, ≤200 chars>"
metadata:
node_type: memory
type: user | feedback | project | reference
originSessionId: <session_id from journal entries that fed this cluster>
---
<Body content per type, see CLAUDE.md memory schema:
@@ -263,12 +265,17 @@ originSessionId: <session_id from journal entries that fed this cluster>
- reference: pointer to external system + what's there.>
```
The frontmatter MUST match the live auto-memory schema exactly: `name` is the
slug (NOT a prose title), and `node_type`, `type`, `originSessionId` live under
a `metadata:` block (verify against an existing file in the target memory dir
before drafting — match its shape).
Constraints:
- Frontmatter fields `name`, `description`, `type` are **required**. Skill enforces this at apply time.
- `originSessionId` is required — must be a `session` value from one of the cluster's journal entries.
- Top-level `name` + `description` and nested `metadata.node_type` (always `memory`) + `metadata.type` are **required**. Skill enforces this at apply time.
- `metadata.originSessionId` is required — must be a `session` value from one of the cluster's journal entries.
- ≤50 LOC of body content. Surgical.
- Slug (used in `target` path filename) must not collide with any existing memory file.
- For `type=feedback` and `type=project`, body MUST contain `**Why:**` and `**How to apply:**` lines (CLAUDE.md memory schema).
- `name`/slug (also the `target` path filename) must not collide with any existing memory file.
- For `type: feedback` and `type: project`, body MUST contain `**Why:**` and `**How to apply:**` lines (CLAUDE.md memory schema).
## Diagnosis drafting protocol (required for every proposal)
@@ -509,7 +516,7 @@ MOSS's core thesis: "routing, hook ordering, state invariants, and dispatch live
2. `cross_session_evidence == true` (≥5 occurrences across ≥3 sessions)
3. `auto_apply_eligible: false`**always**. Harness edits are never auto-applied.
4. `blast_radius: high`
5. Proposal includes a `# Test verification` section with the command `bash ~/.claude/adam/tests/run-tests.sh` and the expected result "132 passed, 0 failed" (or current pass count). The skill runs this test before applying.
5. Proposal includes a `# Test verification` section with the command `bash ~/.claude/adam/tests/run-tests.sh` and the expected result "134 passed, 0 failed" (or current pass count). The skill runs this test before applying.
6. Change is surgical: ≤30 LOC diff, single file.
7. `# Diagnosis` reconstructs the causal chain from harness-level behavior (not from text-artifact behavior). The mismatch must name a specific code path (function, regex, threshold) in the target file.
+1 -1
View File
@@ -300,7 +300,7 @@ Before writing any proposal:
- For `skill_new`: confirm the slug doesn't collide with any existing skill in `~/.claude/skills/`. If it does, refuse and ask user to rename.
- For `skill_edit`: confirm the diff is append-only (no `-` lines that remove existing content) and that target SKILL.md exists. When auto-applying, ALSO re-verify the eligibility gate steps in §3 (cooldown, blacklist, byte cap) before any `Edit` call — never trust frontmatter alone.
- For `skill_edit` with `auto_apply_eligible: true`: confirm `contradiction_flag` is absent or null in frontmatter. Refuse auto-apply if `contradiction_flag` is set with any non-empty value (treat the agent's flag as a hard veto on auto-apply; user can still manually approve in walk-the-queue if they disagree with the heuristic).
- For `memory`: confirm `# Proposed change` body starts with `---` frontmatter containing required fields `name`, `description`, `type`, `originSessionId`. Refuse if frontmatter missing — agent must redraft per the Memory drafting protocol.
- For `memory`: confirm `# Proposed change` body starts with `---` frontmatter matching the live auto-memory schema — top-level `name` (the slug) + `description`, plus a `metadata:` block with `node_type: memory`, `type`, and `originSessionId`. Cross-check the shape against an existing file in the target memory dir. Refuse if frontmatter is flat (`type:`/`originSessionId:` at top level) or missing the `metadata:` block — agent must redraft per the Memory drafting protocol.
- For `harness_edit`: confirm `auto_apply_eligible: false` (never auto-apply). Confirm `confidence ≥ 5`. Confirm `# Test verification` section names the test command. Confirm diff is ≤30 LOC and targets a single allowed harness file (see `agents/adam.md` §"Harness self-modification"). Run test suite before AND after applying — revert on any regression.
- Confirm `source_entries` is present in proposal frontmatter as a non-empty list (used for archive). Warn (do not refuse) if missing — legacy proposals from before v0.2.0 won't have it.