release: v3.3.2 — V-memory was 43% machine output

Asked whether recall still works under 3.x's dogfooding cadence, the measurement
said no.

A run directory carries two machine-generated shapes ending in .md: the per-job
worker prompt the emitter renders, and the spec.md/plan.md a run keeps when its
real spec lives elsewhere. On this repository that was 71 prompts and 44 stubs out
of 267 indexable files — 43% of the corpus — and the prompts are ~40 lines of
near-identical boilerplate.

Measured on a real query, not argued: "scope gate write_allowed violation"
returned one useful result followed by FIVE near-identical worker prompts. After
the exclusion: the PRD's Scope Gate section, a dogfood review, the architecture
doc, the implementation plan, another dogfood review.

Excluded from the INDEX only. They stay in git, stay in the audit trail, stay
readable — they are simply not what anyone means by "what do we know about X".
docs/superpowers/dogfood/** is deliberately kept: hand-written, and the densest
record this project has of what actually went wrong. "Why was a red test floor
merged" now returns that section first.

Index 196 -> 152 files, 1712 chunks, all human prose. Nine selftests pin the
predicate, including that a real spec, a dogfood record and an architecture doc
are never treated as generated.

Known and unchanged: the dense lane holds 0 vectors, so recall is FTS5-only and
the >= 80 vector scale gate has never engaged.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: copeus

.claude-plugin/marketplace.json

@@ -9,7 +9,7 @@
     {
       "name": "superpowers-v",
       "description": "Compound V for Superpowers: triple parallel pre-flight (code archaeology + domain-expert + Context7 library validator), disjoint partitioning, manifest-driven multi-backend dispatch (Claude + Codex + Antigravity + Cursor), git-diff scope enforcement, crash-resumable runs, adaptive tier-based routing, epic mode, research-grounded brainstorming (gated pre-brainstorm recon + batched elicitation), V-memory local-first semantic+lexical recall over docs/superpowers (opt-in pure-python embeddings + a deterministic recall→action bridge), and batched parallel dispatch (Opus default, narrow Sonnet exception)",
-      "version": "3.3.1",
+      "version": "3.3.2",
       "source": "./",
       "author": {
         "name": "Oleg",

.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "superpowers-v",
   "description": "Compound V for Superpowers: triple parallel pre-flight (code archaeology + domain-expert advisor + library/doc validator via Context7), disjoint file partitioning, manifest-driven multi-backend dispatch (Claude + headless Codex + Antigravity + Cursor workers), git-diff scope enforcement, crash-resumable runs, adaptive tier-based routing, epic mode for multi-feature builds, V-memory local-first semantic+lexical recall over docs/superpowers (opt-in pure-python embeddings + a deterministic recall\u2192action bridge), and batched parallel dispatch (Opus default, Sonnet for narrow junior-task carve-out), plus /v:onboard — a project-onboarding command that builds a citation-verified knowledge base + AGENTS.md/CLAUDE.md bridge behind a human gate. Auto-intercepts four transitions: pre-brainstorm recon \u2192 brainstorming \u2192 writing-plans \u2192 execution.",
-  "version": "3.3.1",
+  "version": "3.3.2",
   "author": {
     "name": "Oleg",
     "email": "copeus@gmail.com"

CHANGELOG.md

@@ -4,6 +4,33 @@ All notable changes to **superpowers-v (Compound V)** are documented here.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project uses semantic versioning.
 
+## [3.3.2] - 2026-09-02
+
+### Fixed — V-memory was 43% machine output
+
+Asked whether recall still works under 3.x's dogfooding cadence, the measurement said no.
+
+`docs/superpowers/execution/<run>/` carries two machine-generated shapes that end in `.md`: the per-job worker prompt the emitter renders, and the `spec.md` / `plan.md` a run keeps when its real spec lives elsewhere (the manifest's `spec_path` points into `docs/superpowers/specs/`). On this repository, after a night of dogfooding, that was **71 worker prompts and 44 run-directory stubs out of 267 indexable files — 43% of the corpus**, and the prompts are ~40 lines of near-identical boilerplate.
+
+The effect, measured on a real query rather than argued:
+
+```
+"scope gate write_allowed violation"
+before:  1 useful result, then FIVE near-identical worker prompts
+after:   the PRD's Scope Gate section, a dogfood review, the architecture doc,
+         the implementation plan, another dogfood review
+```
+
+Both are excluded from the **index** now. They stay in git, stay in the audit trail, stay readable — they are simply not what anyone means by *"what do we know about X"*. Recall is evidence for planning and review, and evidence that is 43% copies of one template is worse than a smaller corpus.
+
+Deliberately **not** excluded: `docs/superpowers/dogfood/**`, which is hand-written and the densest record this project has of what actually went wrong. A query for *"why was a red test floor merged"* now returns that section first.
+
+Index: 196 → 152 files, 1712 chunks, all human prose.
+
+### Known
+
+The dense lane holds **0 vectors** — the venv is bootstrapped but nothing has been embedded, so recall is FTS5-only and the `>= 80 vectors` scale gate has never engaged. That is unchanged by this release and unrelated to it; `/v:memory-refresh --bootstrap` is what populates it.
+
 ## [3.3.1] - 2026-09-02
 
 > v3.3.0 was cut automatically mid-session, when its version bump reached `main`, and

scripts/compound-v-memory.py

@@ -243,6 +243,50 @@ def _in_git_worktree(root: str) -> bool:
         return False
 
 
+# --------------------------------------------------------------------------- #
+# WHAT IS PROSE, AND WHAT IS EMITTER OUTPUT THAT HAPPENS TO END IN .md
+#
+# `docs/superpowers/execution/<run>/` is a run's audit trail, and two things in it
+# are machine-generated: the per-job worker prompt that `compound-v-emit-workflow.py`
+# renders, and the `spec.md` / `plan.md` a run carries when its real spec lives
+# elsewhere (`spec_path` in the manifest usually points into
+# `docs/superpowers/specs/`).
+#
+# Measured on this repository 2026-09-02, after a night of dogfooding: 71 worker
+# prompts and 44 run-directory stubs out of 267 indexable files — **43% of the
+# recall corpus**, and the prompts are ~40 lines of near-identical boilerplate.
+# A query for "scope gate write_allowed violation" returned five worker prompts in
+# its top six, burying the one document that actually explains the scope gate.
+#
+# Recall is evidence for planning and review. Evidence that is 43% copies of one
+# template is worse than a smaller corpus, so these are excluded from the INDEX.
+# They stay in git, stay in the audit trail, and stay readable — they are simply not
+# what anyone means by "what do we know about X".
+#
+# NOT excluded: `docs/superpowers/dogfood/**` (written by hand, and the densest
+# record this project has of what actually went wrong), and every run-directory file
+# that is not one of these two shapes.
+_EXEC_PREFIX = DOCS_REL.rstrip("/") + "/execution/"
+
+
+def is_generated_run_artifact(rel: str) -> bool:
+    """True for emitter-rendered prose inside a run directory."""
+    rel = str(rel or "").replace("\\", "/")
+    if not rel.startswith(_EXEC_PREFIX):
+        return False
+    tail = rel[len(_EXEC_PREFIX):]
+    parts = tail.split("/")
+    if len(parts) < 2:
+        return False
+    rest = parts[1:]
+    # <run>/jobs/<job>.prompt.md — rendered by render_worker_prompt()
+    if len(rest) == 2 and rest[0] == "jobs" and rest[1].endswith(".prompt.md"):
+        return True
+    # <run>/spec.md and <run>/plan.md — the manifest's spec_path/plan_path point at
+    # the real ones; these are the run's local copies or stubs.
+    return len(rest) == 1 and rest[0] in ("spec.md", "plan.md")
+
+
 def tracked_files(root: str):
     """Repo-relative *.md/*.jsonl under docs/superpowers, GIT-TRACKED only.
 
@@ -267,7 +311,9 @@ def tracked_files(root: str):
                 )
                 if roots.returncode == 0:
                     rels += [p for p in roots.stdout.decode("utf-8", "replace").split("\0") if p]
-                return sorted(r for r in rels if r.endswith(".md") or r.endswith(".jsonl"))
+                return sorted(r for r in rels
+                              if (r.endswith(".md") or r.endswith(".jsonl"))
+                              and not is_generated_run_artifact(r))
         except (OSError, subprocess.SubprocessError):
             pass
         return []  # fail closed: inside git but ls-files failed — index nothing, never untracked
@@ -276,7 +322,9 @@ def tracked_files(root: str):
     for dirpath, _dirs, files in os.walk(docs_abs):
         for f in files:
             if f.endswith(".md") or f.endswith(".jsonl"):
-                rels.append(os.path.relpath(os.path.join(dirpath, f), root))
+                rel = os.path.relpath(os.path.join(dirpath, f), root)
+                if not is_generated_run_artifact(rel):
+                    rels.append(rel)
     return sorted(rels)
 
 
@@ -1239,6 +1287,27 @@ def _fake_q(texts):
         shutil.rmtree(tmp, ignore_errors=True)
 
     # onboarding: doc_type_for clean labels for root files (no filename leak)
+    # --- generated run artefacts are not prose (3.3.2) --------------------
+    D = DOCS_REL.rstrip("/")
+    check("worker prompt is generated",
+          is_generated_run_artifact(D + "/execution/r1/jobs/impl.prompt.md"))
+    check("run-dir spec stub is generated",
+          is_generated_run_artifact(D + "/execution/r1/spec.md"))
+    check("run-dir plan stub is generated",
+          is_generated_run_artifact(D + "/execution/r1/plan.md"))
+    check("a real spec is NOT generated",
+          not is_generated_run_artifact(D + "/specs/2026-01-01-thing-design.md"))
+    check("a dogfood record is NOT generated",
+          not is_generated_run_artifact(D + "/dogfood/2026-01-01-run.md"))
+    check("an architecture doc is NOT generated",
+          not is_generated_run_artifact(D + "/architecture/native-mechanisms.md"))
+    check("a run-dir file that is neither shape is NOT generated",
+          not is_generated_run_artifact(D + "/execution/r1/notes.md"))
+    check("a nested jobs path outside a run dir is NOT generated",
+          not is_generated_run_artifact(D + "/jobs/impl.prompt.md"))
+    check("a non-execution path is never generated",
+          not is_generated_run_artifact("README.md"))
+
     check("doc_type root agents", doc_type_for("AGENTS.md") == "agents")
     check("doc_type root claude", doc_type_for("CLAUDE.md") == "claude")
     check("doc_type root conventions", doc_type_for("CONVENTIONS.md") == "conventions")