KnowledgeHarvester.ts hardcodes a macOS auto-memory path — zero notes graduate on Linux

[stratofax]

Component: PAI/TOOLS/KnowledgeHarvester.ts (memory → KNOWLEDGE graduation tool, v5)
Affected source: Releases/v5.0.0/.claude/PAI/TOOLS/KnowledgeHarvester.ts @ 2fde1bb
Severity: High on Linux — the tool runs, exits 0, and graduates nothing. Silent total failure.

Summary

On Linux, KnowledgeHarvester reads auto-memory notes from a hardcoded macOS path
(-Users-${USER}--claude). That directory never exists on Linux (the real slug is
-home-${USER}--claude), so the source directory is missing, the scan returns empty, and
zero memory notes ever graduate to MEMORY/KNOWLEDGE/ — with no error and a success exit code.

Bug — Hardcoded macOS auto-memory path

Lines 39–45:

const CURRENT_USER = process.env.USER;
if (!CURRENT_USER) {
  // ...guard...
}
const AUTO_MEMORY_DIR = path.join(HOME, ".claude", "projects",
  `-Users-${CURRENT_USER}--claude`, "memory");

Root cause: the slug -Users-${CURRENT_USER}--claude is the macOS form of the ~/.claude path
(/Users/<name>/.claude). On Linux the home is /home/<name>, so the real auto-memory directory is
…/projects/-home-<name>--claude/memory. The hardcoded slug points at a path that does not exist on Linux.

Expected: the harvester resolves the auto-memory directory from the actual ~/.claude path on any platform.

Actual (Linux): fs.existsSync(AUTO_MEMORY_DIR) is false → collectAutoMemory() returns [] →
no candidates → graduation count is 0, exit 0, no warning.

Impact: the memory→KNOWLEDGE graduation pipeline is a silent no-op on every Linux PAI v5 install.
(Pre-v5 installs do not ship KnowledgeHarvester at all — it is a v5-only tool — so this is a v5-Linux
regression, not a historical bug.) Because it fails silently with a success exit code, it reads as
"nothing to graduate" rather than "broken."

Proposed fix

Derive the slug from the live ~/.claude path — the same cross-platform pattern the sibling
SessionHarvester.ts already uses in this release (see Related below) — and drop the now-unused
CURRENT_USER guard:

const AUTO_MEMORY_SLUG = path.join(HOME, ".claude").replace(/[\/\.]/g, "-");
const AUTO_MEMORY_DIR = path.join(HOME, ".claude", "projects", AUTO_MEMORY_SLUG, "memory");

This yields -Users-<name>--claude on macOS and -home-<name>--claude on Linux from the same code.

Verification (two independent Linux installs)

• Install A — Linux, PAI 5.x (empirical): with the fix applied, all 23 local auto-memory notes
  graduated successfully where the unpatched tool graduated 0.
• Install B — Linux (Ubuntu), PAI 4.0.3 (path-derivation corroboration): KnowledgeHarvester.ts
  is not present on this 4.0.3 install (v5-only tool), so this is not a harvest reproduction.
  What was checked: the real auto-memory dir on this host is .../projects/-home-<user>--claude/memory
  (confirmed to exist); the buggy upstream slug resolves to -Users-<user>--claude, which does not
  exist on this Linux host; the proposed fix's slug-derivation resolves to -home-<user>--claude,
  matching the real directory. Independently corroborates the Linux slug mismatch and confirms the fix
  yields the correct path on a second Linux box.

Related issues & precedent

• KnowledgeHarvester is single-PAI-instance — per-project memory dirs (~/.claude/projects/<slug>/memory/) invisible without PAI_DIR override #1170 (open) touches the same AUTO_MEMORY_DIR line from a different angle — per-project memory
  dirs invisible / single-PAI-instance scope — and quotes the -Users-${USER}--claude slug as-is. This
  issue is the platform dimension: on Linux that slug resolves to a path that does not exist, so the
  harvester silently graduates nothing. The glob-based fix proposed in KnowledgeHarvester is single-PAI-instance — per-project memory dirs (~/.claude/projects/<slug>/memory/) invisible without PAI_DIR override #1170 would incidentally resolve
  this too; filing separately so the Linux-platform symptom is tracked and linkable.
• SessionHarvester.ts: PROJECTS_DIR hardcoded — cannot harvest from archives or sibling project dirs #1256 (closed/completed) fixed the same class of hardcoded-projects-dir bug in the sibling
  SessionHarvester.ts.
• In-repo inconsistency: in this same v5.0.0 release, SessionHarvester.ts (lines 33–36) already
  derives the slug cross-platform (CLAUDE_DIR.replace(/[\/\.]/g, "-"), with an explicit macOS/Linux
  comment), while KnowledgeHarvester.ts (lines 39–45) hardcodes -Users-. The fix simply brings
  KnowledgeHarvester in line with its sibling.
• Companion: KnowledgeHarvester.ts classifier over-matches the People domain via substring matching #1367 — a separate, platform-independent classifyDomain precision bug found and fixed
  in the same pass.

Suggested labels

bug, platform:linux, tool:KnowledgeHarvester

[danielmiessler]

Hey @stratofax, thanks for raising this, and sorry it sat for a while.

We're changing how LifeOS ships. Instead of cloning a full ~/.claude directory and running it as a complete system, LifeOS is becoming a skill you install through an agentic installer. The installer hands integration to your own AI, which reads your actual machine (your OS, your paths, your harness) and wires the hooks and system prompt in where they belong.

That's aimed right at what you hit here. The old "one directory, one layout, hope it matches your setup" approach is exactly what broke for so many people, and the new model should handle it far better because your AI does the integration per machine instead of us guessing.

So we're closing this in prep for that release. If it still bites you once the skill-based version is out, reopen or file a fresh one and we'll jump on it. Appreciate you taking the time.