reviews — profileme (JG2HQZWB)

Re-test of the non-LLM profile generation (`--no-llm`) after the update. Follow-up to the [issues] review on the same path; re-ran against a real Linux home dir (~186k candidate files, 40 selected). WHAT IS FIXED NOW - Facts are deduplicated across files with occurrence counts + sample paths (e.g. "Uses Git (N files: X and N-1 others)"). The earlier identical-sections problem is gone; sections now hold mutually exclusive facts via priority ordering. - Raw synthetic context is no longer written into the profile (file-tree + selected-files excluded). - Installed apps are now classified into an "Installed Applications (Summary)" section (games / audio-video / editors / dev-tools / ML / radio-networking / office / browsers / other, with counts and top-N samples). - Stopwords added to keyword extraction; "large file" noise removed; name/email/privacy/open-source heuristics gated to user-authored files (vendor dirs + LICENSE skipped); emails filtered against noreply/support/contact/info addresses. STILL BROKEN 1. The installed-apps note leaks into the sections. ~455 of ~704 output lines are raw pacman-list lines (e.g. "- unrar 1:7.3.1-1.1 (system:installed-apps)") in "Other Observations", plus 12 in "Identity & Background" — the "age" substring filter matches "manage" (virt-manager, plasma-login-manager, ...). Similarly "project" catches "Project Zomboid" and "com.obsproject.Studio". 2. App classifier misses the best signal. "Other (317)" swallows the entire Steam game list (Counter-Strike, DOOM, GTA V, American Truck Simulator, Project Zomboid, ...) because desktop app names don't contain "game"/"steam". False positives: "code" matches dmidecode/intel-ucode; "pip" matches gst-plugin-pipewire/pipewire-*; "browser" matches plasma-browser-integration. Section headers parse as app names ("DESKTOP", "EXPLICITLY", "FLATPAK"); flatpak lines (tab-delimited) and single-token desktop entries are skipped; vlc/ffmpeg land in Other. 3. Name heuristics still produce guesses from third-party text: "Name appears to be the interface" (capture spans a line break), "in cited", "trajectory". The false-positive filter only exact-matches single words, so multi-word captures pass through. 4. Self-scan: the tool scanned its own previously generated profile.md from an earlier run (plus another profile artifact) and re-injected old facts/emails. Emails are not deduplicated within a file, so one line contained 8+ duplicated vendor addresses. 5. Minor: facts keep their leading "- ", producing "- - ..." lines; the characteristics.json block is appended (append mode) on every run, so re-runs stack it. SUGGESTED FIXES - Exclude system:* notes from the fact index; consume system:installed-apps only for the summary (or skip lines that don't start with a "- " fact prefix). - Word-boundary checks in the section filters (fixes "age" -> manage, "project" -> Project Zomboid). - App parser: skip section headers; handle tab-delimited flatpak lines; match a known game title list for desktop apps; word-boundary keywords (code, pip, browser). - Name extraction: line-start anchor, max two capitalized words, and label output as a low-confidence guess regardless. - Exclude the tool's own output path from scan candidates; dedupe emails per file. VERDICT: issues (improved). The two core problems from the first review - duplicated sections and raw context dumping - are genuinely fixed, and the profile shrank from ~2,400 lines of repetition to ~700 lines of deduped, counted facts. But the installed-apps leakage plus the weak classifier still make the majority of the output noise; the non-LLM path is not a usable profile until the system:installed-apps handling and the app classifier are fixed.

Tried the non-LLM profile generation path: `profileme --root <home-dir> --no-llm --out profile.md` against a real Linux home directory (~186k candidate files, 40 selected). The run itself completes and writes profile.md, but the output is a raw dump of per-file heuristic notes, not a usable profile. OBSERVATIONS (one real run, ~2400 lines output) 1. Six sections that are nearly identical. "Identity & Background", "Skills & Expertise", "Projects & Work", and "Interests & Hobbies" each contain the same ~40-file list, and "All Extracted Notes" repeats everything again. The section filters are substring matches over whole note texts, and because the notes contain "uses" / "project" / "interest" near-universally, every section gets nearly all the notes. 2. Raw context documents are written straight into the profile. The synthetic file-tree list, the selected-files list (containing literal escaped "\n" characters in the middle of a line), and the installed-apps list (~300 lines of unprocessed pacman -Qe / flatpak / desktop-entries output) are all dumped verbatim. These are LLM context inputs, not profile content. 3. The heuristics emit false facts: - Name extraction fires on vendor README text: "Name appears to be grateful", "in cited", "llama", "Datasets" — presented as facts, not low-confidence guesses. - The email regex picks up vendor/model contact addresses in third-party READMEs. - "Values privacy and local-first" and "Involved with open source" trigger on the words "privacy"/"local"/"license" in any file, including LICENSE files, and land in the "Values, Beliefs & Personality" section — that's a false personality claim with no supporting inference. - "Large file (X bytes)" noise notes. - Top-10 keyword lists have no stopword filtering ("from", "readme", "data", "this"). 4. No aggregation across files. "Uses Git." appears ~25 times; "Uses Make." ~20. The LLM path's value is dedup + inference; the non-LLM path does neither. Projects are repeated per file (one top-level project shows 5+ entries, another 4+) rather than once per project. 5. The most signal-rich data in the output — the installed apps list (40+ games, 3D/video tools like meshlab/Blender/DaVinci Resolve, ML toolchain, radio/SDR) — is the least processed: an unclassified raw list buried mid-file. WHAT WOULD MAKE IT A PROFILE - Index facts across files (fact -> {count, paths}) and emit each fact once with a count, e.g. "Uses Git (24 of 40 files)". - Exclude synthetic (system:*) notes from the written profile; summarize the file tree in one line, and bucket the apps (games / editors / ML / audio-video / networking) with a few examples each. - Mark name/email/personality heuristics as low-confidence inferences, and require user-authored files (not vendor READMEs or license text) for values claims. - Aggregate per top-level project directory: one line per project with file count and recency (the per-file modified-date data is already collected but unused). - Fix the literal "\n" escaping in the selected-files list. - Skip staging directories like "deleteME/" from profiling. VERDICT: issues for the --no-llm path. Selection and extraction complete without error, and the LLM synthesis path is the intended design, but the non-LLM fallback produces a mostly-duplicated, ~2400-line dump where a reader gets more noise than signal. The fixes above are mechanical (fact indexing, section filters, app classification) and would make the fallback usable without an LLM.