Brain Activity "Lessons" is structurally 0 out of the box — ADR-0018 ships the distillation panel, but ADR-0017 auto-feeder is unbuilt and record-lesson.mjs is manual/unhooked

[sparkling]

Summary

On a default install, the Brain Activity console's "Lessons" metric is structurally always 0, and there is no automatic way to make it non-zero. The panel ships the visualization of a captured → distill → lessons pipeline whose automatic feeder is (per the plugin's own ADRs) not yet built. The only thing that writes to namespace='lessons' is a manual, unhooked script.

This is the plugin behaving contrary to its own stated design intent ("works out of the box"), not a crash. (Distinct from #15, which is the console failing to launch; this is about the console launching and then showing a metric that can never fill on its own.)

The drift (ADR vs code)

• ADR-0018 "Brain Activity panel" — Status: Implemented (ships v3.1.0). It builds the console card with a "compact animated distillation flow strip (captured → distill → lessons)" and a headline Lessons stat, reading namespace='lessons' via GET /api/activity (scripts/onboarding-console.mjs → gatherActivity() → SELECT COUNT(*) … WHERE namespace='lessons').
• ADR-0017 "Recursive per-user learning loop" — Status: Accepted (foundation shipped v3.0.5). Its "Known limits / next layers" section explicitly says the automatic path that would populate lessons — "best-practice → skill promotion (codifying a repeated workflow … as a global skill)" and "retrieval-in-project" — are "designed, not yet built."
• The code confirms the feeder is absent. The only writer to namespace='lessons' is scripts/record-lesson.mjs, and it is wired to zero hooks (grep -c record-lesson plugin/hooks/hooks.json → 0). It is a manual command a user must run by hand.

So a shipped, "Implemented" console (ADR-0018) presents an automatic lessons pipeline whose automatic feeder an "Accepted-but-next-layer" ADR (ADR-0017) says isn't built. The panel outran the wiring.

Reproduce

Fresh install → /ruvnet-brain:configure → Brain Activity card. On any project store that has only auto-captured telemetry:

totals: { memories: 669, lessons: 0 }
namespace breakdown: commands 379 · feedback 215 · session 67 · patterns 5 · cost-tracking 3

Nothing lands in lessons because nothing auto-distills into it; the SONA "Learnings" loop (ADR-0017, working — 4361 patterns) feeds ~/.claude-flow, a different store, not this namespace.

Why it matters

The console's own framing is "the model is the engine — this is the harness driving it" and the loop is described as "shipped as a plugin default … automatically." A headline metric that is structurally 0 for every user until they discover and run a manual script undercuts exactly that "it just works, and you can see it working" story.

The "distillation flow" visualization compounds this: roughly half its width is a real, populated particle-field for ~1,300 captured memories (with a working "see the memories by kind" interaction) and the other half is 5 generic decorative sparkle icons under "0 lessons — the keepers" — not real data standing in for something, but a fallback rendered because there is nothing to show. A deliberately animated, symmetric-looking two-sided visualization where one whole side is structurally guaranteed to render only decoration, for virtually every install, indefinitely.

Additional bug found in the same strip: a self-contradicting "receipt" line

console/activity.js:339:

'<div class="ab-receipt">last compact: <span class="mono">47 msgs → 6 lessons</span> · nothing lost<sup>illustrative — every other number is live</sup></div>'

Two separate problems in this one line:

1. Markup bug: there's no space in the source between lost and <sup>, and no CSS rule for .ab-receipt or sup anywhere in style.css adds one. It renders literally as "nothing lostillustrative — every other number is live" — run together, no separator.
2. Content self-contradiction: this line asserts a specific, concrete-sounding claim — "47 msgs → 6 lessons" — styled with the ab-receipt class, a term this app uses everywhere else specifically to mean real, verified data (mirrors ADR-013's own stated design law: "every number traces to a receipt... no estimated savings"). But this particular "receipt" is fake, disclosed only via a small, easily-missed superscript. It sits directly below the real stat tile reading "0 lessons — the keepers" — so the same card shows "0 lessons" in one place and "6 lessons" (fabricated) in the next, with only that inline caveat resolving the contradiction. A user who doesn't notice the tiny superscript reads two different lesson counts on the same screen.

Suggested fix (either is fine)

1. Wire the automatic path ADR-0017 already scoped: capture → ruflo memory distill → namespace='lessons' on a hook (e.g. SessionEnd, alongside learn-flush.mjs), or surface AgentDB Reflexion critiques (a failed episode + critique is already a retrievable lesson) into the lessons view. Then Lessons fills out of the box.
2. Or make the empty state honest: the Lessons card should say "0 recorded — run record-lesson to capture one; auto-distill is a planned next layer (ADR-0017)" instead of presenting a distillation flow that has no automatic input.

Environment

• ruvnet-brain v3.1.1 (also v3.1.0, where ADR-0018 shipped)
• ADRs: docs/adr/0018-brain-activity-panel.md (Implemented), docs/adr/0017-recursive-per-user-learning-loop.md (Accepted, next-layers unbuilt)
• Node 24, macOS; rUv stack via npx

[sparkling]

Closing this — I filed it too early and mischaracterized it. On re-reading the ADRs:

• ADR-0017 already documents this as a planned next layer, not a defect: its "Known limits / next layers" section explicitly lists retrieval-in-project and best-practice→skill promotion as "designed, not yet built." So the code (panel present, auto-feeder absent) actually agrees with the ADRs — there's no ADR-vs-code drift here.
• Manual record-lesson.mjs may be intentional: its own header distinguishes auto-capture (session transcripts = logging) from lessons (deliberate, structured task/tried/worked/critique). So "Lessons fills only via a deliberate command" could be the design, not a gap.
• The feature shipped yesterday (v3.0.5 / v3.1.0, 2026-07-15). Flagging a documented, one-day-old, planned next-layer as a bug is premature.

Apologies for the noise. Happy to re-open as an enhancement discussion later if the auto-distill layer lands and the empty state still misleads — but as of today this isn't a bug.

[sparkling]

Reopening — this session's investigation goes further than the original closure

I closed this myself in an earlier session as "a documented, planned next-layer, not a bug." A fresh, much deeper investigation this session (prompted by a user pushing back hard on that framing) found enough new, concrete evidence that I'm reopening it. Summary of what's new below — full findings first.

What "distilled lessons" is supposed to be (grounded)

scripts/record-lesson.mjs is a real, working, non-trivial script — not a stub. Full mechanics (read in full this session):

// 1. STORE — native, signal namespace
const out = ruflo(['memory', 'store', '-k', key, '-n', ns, '--value', value]);
// 2. REFINE — the REAL ADR-174 distillation pipeline
const dist = ruflo(['memory', 'distill', 'run']);
// 3. VERIFY — proves the lesson is actually recallable
const search = ruflo(['memory', 'search', '-q', task, '-n', ns]);

Usage: node scripts/record-lesson.mjs --task "..." --tried "..." --worked "..." --critique "..." --outcome success. It genuinely stores a structured (task/tried/worked/critique/outcome) memory, genuinely refines it via ruflo's real memory distill run (ADR-174, Accepted, M0-M5 implemented), and genuinely verifies recall. This is not fake or theater.

Proof the underlying distillation engine (ADR-174) DOES work — just not for this namespace

Checked this project's live .swarm/memory.db this session:

episodes: 329
reasoning_patterns: 329
pattern_embeddings: 329
causal_edges: 303
namespace='lessons': 0

ADR-174's automatic daemon-driven distillation (every 30 min) is genuinely running and has produced 329 real structural patterns from 1,305 raw commands/feedback/session rows. It just has no concept of a curated "Lesson" — that's a separate, ruvnet-brain-specific namespace convention layered on top, and NOTHING decides which pattern is worth promoting into it.

Exhaustive check: is record-lesson.mjs wired to ANYTHING? (No.)

Checked every surface this session:

• plugin/hooks.json (the complete, authoritative hook list — SessionStart, UserPromptSubmit, PreToolUse ×4, PostToolUse, SessionEnd): zero reference.
• Every command (plugin/commands/configure.md, rvcb.md, rvbc.md): zero reference.
• Every skill (ruvnet-brain, brain-score, brain-build, brain-prompt, savings — including the main skill prompt driving this very conversation): zero reference.
• No agents directory exists in this plugin.
• Repo-wide grep for record-lesson across every .md/.mjs/.js/.json/.sh file: exactly one match — the script itself.

It is a fully orphaned utility. Not discoverable from the UI, a command, a skill, or a doc — only from reading scripts/ directly.

The severity point (why this isn't just "an unbuilt next layer")

This stat isn't buried — it's the second stat tile on the first card of the console (Brain Activity, MIRROR badge, top of page), labeled "distilled lessons", next to a whole animated "captured → distill → lessons" visualization. Given the exhaustive wiring check above, every single fresh install, on every machine, forever, will show 0 here — there is no ramp-up, no "give it time," no discovery path. This is not a metric that's early; it's a metric that's structurally incapable of ever changing for the overwhelming majority of users, prominently featured as if it were live.

Ruvnet-brain's own UI partially discloses this, inconsistently

console/activity.js:339:

'<div class="ab-receipt">last compact: <span class="mono">47 msgs → 6 lessons</span> · nothing lost<sup>illustrative — every other number is live</sup></div>'

The authors explicitly labeled this one line "illustrative" — a real, honest disclosure that this specific line is a mocked example. But the empty-state text two lines away (activity.js:413) says:

'no distilled lessons yet — they appear once distillation has run'

This actively implies running distillation (which, per ADR-174, already runs automatically every 30 minutes) will eventually populate lessons. It won't — distillation populates reasoning_patterns/episodes, never namespace='lessons'. This specific empty-state copy is the more misleading of the two, and it's the one every user actually sees.

Grounding on "why manual" (not just "unfinished")

record-lesson.mjs's header says it enforces the discipline of rUv's own /remember command (agentdb/plugins/agentdb-memory/commands/remember.md). Pulled that file directly — it's itself flagged in the brain's corpus as "design intent... no implementing source was retrieved — treat as proposed, not confirmed built." So manual, deliberate curation (not silent automatic promotion) does appear to be the ecosystem-wide design philosophy for "is this worth remembering" — which argues the concept of a manual trigger isn't wrong. What's missing isn't automation-by-default; it's any discovery path at all for the manual trigger that does exist.

Suggested directions (not prescribing which — your call)

1. Build the ADR-0017 "next layer" — an auto-feeder that promotes a real reasoning_patterns row to a curated lesson past some confidence/repetition threshold. The real fix, but the larger lift.
2. Short of that: surface record-lesson.mjs somewhere a user can actually find it — a slash command (mirroring rUv's own /remember), a console CTA on the empty state ("record your first lesson →" with the exact command), or at minimum a README/doc mention. Turns an orphaned capability into a usable one without building the automation.
3. Until either lands: soften the empty-state copy so it doesn't imply automatic population, and/or de-emphasize the stat's prominence so it doesn't read as a broken headline metric on every fresh install.

Git history

record-lesson.mjs has exactly one commit — bundled into the v3.2.12 release (2026-07-16), no dedicated commit message, no follow-up wiring it to anything since.

[sparkling]

Follow-up: the visual proportion makes this worse than a stat tile — it's a whole animated element

Beyond the stat tile, the "distillation flow" strip (console/activity.js, the captured → distill → lessons canvas visualization) devotes roughly half its total width to representing the lessons side:

• Left half: a real, populated particle-field of ~1,300 scattered dots representing captured memories — genuinely reflects live data, with a working "see the N memories by kind" interaction.
• Right half: 5 sparkle/star icons under "0 lessons — the keepers" — and since there are 0 real lessons, these aren't even real data points standing in for something; they're a generic decorative fallback rendered for the empty state. The caption underneath ("These survive compaction, restarts, and reboots. Your AI doesn't start over tomorrow.") reads as a confident claim about a capability the layout is, in practice, permanently visualizing as empty for the reasons detailed above.

So this isn't just a stat that reads 0 — it's a deliberately animated, roughly-symmetric two-sided visualization where one whole side is structurally guaranteed to render nothing but decoration, for virtually every install, indefinitely. The visual weight given to "lessons" implies parity with "memories" (equally real, equally alive) that the underlying mechanism cannot deliver without the manual record-lesson.mjs step nothing prompts a user toward.

[stuinfla]

🤖 Automated issue-fix run (issue-fix.mjs) — a human reviews before anything merges.

What I verified

I read the issue body and both comments (including the self-closure/reopen history) and checked every concrete claim against the code in this worktree at 7df905a (branch issue-fix/16, off main):

• record-lesson.mjs is genuinely unhooked. grep -c record-lesson plugin/hooks/hooks.json → 0. Confirmed it's a real, non-trivial script (store → ruflo memory distill run → verify-recall), not a stub — and confirmed via repo-wide grep it's referenced nowhere else (no command, no skill, no doc).
• The markup bug is real. console/activity.js:339 (pre-fix) was '... · nothing lost<sup>illustrative...' — no separating space, and no CSS rule adds one, so it renders as nothing lostillustrative — every other number is live run together.
• The content self-contradiction is real. That same line asserts a concrete-looking 47 msgs → 6 lessons, styled with the ab-receipt class the app otherwise uses for real receipts, directly under a stat tile reading the real 0 lessons — the keepers. The only disclosure is a small superscript.
• The misleading empty-state copy is real. console/activity.js:413 (pre-fix) read no distilled lessons yet — they appear once distillation has run — this implies distillation (which, per ADR-174, already runs automatically every 30 min) will eventually populate this namespace. It won't; nothing promotes a reasoning_patterns row into namespace='lessons'.

I did not re-verify the .swarm/memory.db episode/pattern counts cited in the second comment (329 episodes, etc.) — that's local runtime state on the reporter's machine, not something checkable from a fresh worktree, and it isn't load-bearing for the fix below.

What I drafted, and why I'm not committing it

I drafted the smaller of the two fixes the issue itself proposes ("or make the empty state honest") — a copy-only change, no behavior change, no auto-feeder design decision:

--- a/console/activity.js
+++ b/console/activity.js
@@
-    '#card-activity .ab-receipt .mono{font-size:13px}',
-    '#card-activity .ab-receipt sup{font-size:12px;color:var(--muted);margin-left:6px;font-style:italic;letter-spacing:0}',
@@
-      '<div class="ab-receipt">last compact: <span class="mono">47 msgs → 6 lessons</span> · nothing lost<sup>illustrative — every other number is live</sup></div>' +
+      '<div class="ab-receipt">distillation keeps what proves durable and discards the rest — nothing is lost, and every kept lesson traces back to a real session</div>' +
@@
-      : '<p class="ab-fallback">no distilled lessons yet — they appear once distillation has run</p>';
+      : '<p class="ab-fallback">0 recorded — lessons don\'t fill in on their own yet. Run <span class="mono">node scripts/record-lesson.mjs --task "..."</span> to capture one; automatic distillation into lessons is a planned next layer (ADR-0017).</p>';

This: (1) fixes the markup bug by removing the fabricated 47 msgs → 6 lessons example instead of just patching the missing space, which also resolves the self-contradiction with the real "0 lessons" stat shown elsewhere on the same card; (2) makes the empty state honest about the manual step and points at the real command, instead of implying auto-population; (3) removes the two CSS rules that only styled the content just deleted.

I'm not committing it, because both required gates are currently red on a clean checkout of main (7df905a), before any change of mine — I confirmed this by stashing the fix and re-running both commands against unmodified HEAD:

• node scripts/sync-version.mjs --check fails: 5 surfaces (package.json, data/manifest.json, kb/package.json, primer/ruvnet-primer.md, explainer/index.html) are drifted at 3.3.0 against the source-of-truth 3.3.1 — it looks like the most recent commit ("Nightly brain refresh v3.3.1", 7df905a) bumped the source of truth but didn't propagate it to those files.
• npx vitest run tests/unit fails: 5 files / 7-8 tests red, independent of my change — tests/unit/version-tag-vs-field.test.mjs fails on the same version drift; tests/unit/verify-interface.test.mjs and tests/unit/version-bump-gate.test.mjs fail on 10s timeouts in shell-fixture tests that look environment-specific to this sandboxed worktree (they weren't touched by my fix and aren't related to console/activity.js).

My instructions require both gates to pass before I commit, and fixing that drift/those timeouts is unrelated to issue #16's subject and out of scope for a surgical fix here — so per my instructions I'm falling through to the no-code-change path rather than either committing on red gates or scope-creeping into unrelated repo maintenance.

What a human needs to do

1. Unblock the gates first (unrelated to this issue): run node scripts/sync-version.mjs to resync the 5 drifted surfaces to 3.3.1 (or cut a new patch version), and look into why verify-interface.test.mjs/version-bump-gate.test.mjs are timing out — worth checking whether they need network/ruflo access this worktree doesn't have.
2. On this issue specifically, decide between the two directions the issue itself lays out — the drafted copy fix above (small, ships now) vs. building the real ADR-0017 auto-feeder (larger, needs a product decision on what promotes a reasoning_patterns row to a curated lesson and at what threshold — genuinely not something to guess at automatically).
3. If the copy fix above is wanted, it can be applied directly from this comment once the gates are green.

[stuinfla]

You're right, and this is an accurate read of the drift: ADR-0018 ships the Lessons panel, but ADR-0017's automatic feeder (captured → distill → lessons) isn't wired yet, and record-lesson.mjs is the only writer — manual and unhooked. So on a fresh install the metric is structurally 0 with no automatic path to fill it. That's the panel showing a pipeline whose feeder doesn't run — exactly as you describe.

I'm keeping this open honestly rather than closing it, because the real fix is a feature, not a one-liner, and it's actively in progress: the auto-distillation path (reflexion.mjs — AgentDB's ReflexionMemory: store what happened with what failed and why, distill into the lessons namespace) is being built now. Two things will land from it: (1) an automatic feeder hooked to session boundaries so lessons accumulates without record-lesson.mjs being run by hand, and (2) an honest 0-state on the panel in the meantime, so a fresh install reads "lessons distill as the brain learns — none yet" instead of a bare 0 that looks like a defect.

Tracking it here until both land. Thanks for pinning the ADR-vs-code gap precisely — that's exactly the kind of report that keeps this honest.

[stuinfla]

Confirmed — you're right, and this is a real design gap, not a crash. Verified live: the Brain Activity card reads namespace='lessons', and the ONLY writer to that namespace is scripts/record-lesson.mjs, which is manual and unhooked. learn-flush.mjs (the SessionEnd hook) accumulates learnings into ~/.claude-flow but does not feed the lessons namespace the card counts. So out of the box, Lessons is structurally 0 — exactly as you describe, and exactly the ADR-0018-panel / ADR-0017-feeder mismatch you cite.

This one is genuine feature work (the auto-distill-to-lessons feeder from ADR-0017's 'next layers'), not a one-line fix — and I'm deliberately not going to hack the metric non-zero with noise just to make the card look alive. The correct fix is to hook session-end distillation to write a real, deduplicated lesson to namespace='lessons' only when a session produced one, so the count reflects genuine learning. Keeping this open and tracked as the real feature it is; will link the PR here.

(Sibling issues #17-#21 are fixed, verified, and closed. Thanks for the exceptionally precise reports — the file:line + ADR citations made these fast to confirm.)

[stuinfla]

@sparkling — this is the kind of issue we want more of. You didn't just spot the empty metric; you caught the part that actually corrodes trust: a fabricated "receipt" (47 msgs → 6 lessons) styled with the exact ab-receipt class this app uses everywhere to mean real, verified data — sitting one line under the genuine "0 lessons." Two different lesson counts on the same card, reconciled only by a superscript most people would never notice. And you nailed the markup bug too (nothing lostillustrative run together).

Fixed in c46d288 (v3.4.8-dev):

• The fabricated receipt is deleted outright — no /api/activity field backs a "last compact" count, so there was nothing honest to put there.
• The 5 decorative "keeper" sparkles that rendered even at zero lessons are gone (nStars = 0 when there are no real lessons).
• The empty state is now honest: "0 recorded — auto-distill is a planned next layer (ADR-0017); capture one with record-lesson." No invented number survives anywhere in that strip.

Bigger than the fix: the maintainer turned your catch into a hard product rule — it can never present a fabricated number as real — and we swept the entire surface for others (found and removed several). Thank you for holding the line on honesty; you were exactly right. 🙏