[HTMLGuyLLC]

Fixes every finding from the audit of the areas the 2026-11-01 audit never reached, except the two Ideas.

Scoring changes (RULES_VERSION 13, so everyone's history is rescored once)

• Messages from other agents, scheduled tasks and coordinators in Claude Code no longer count as the person's.
• A rebuild now drops what older rules counted in a chat that holds none of the person's messages under the new rules. Before, it treated that chat as missing and put the old counts back.
• Polite closers and routine requests ("No, that's all, thanks", "Nah, it's cool", "Same thing for the footer please") aren't corrections. Real ones still are ("Same thing, it still crashes", "Not working on Safari", "No, that's not needed, remove it").
• "You're right" after a plain question isn't an admission. Statement-then-question corrections still are.
• Added context ("by the way…") never counts against the AI, as README says.
• A ? in code isn't the AI asking. Mistake categories and several prompt signals misfire less.

Sources: queued messages with images, ChatGPT attachments, and Codex handoff and voice threads are read. Exports stream instead of loading whole. Lines are capped at 64 MB. Pasted blocks are stripped before scoring.

Redaction (REDACTION_VERSION 3): any *_KEY with a key-like value, pass:, Google and Slack tokens. ISO dates are kept, and numbers that only start like a date are still hidden. Code such as primary_key: id and test output like Pass: 5 are left alone.

Lessons: hints and examples match what the checker credits.

Tests: no test touches the real CODEX_HOME or Claude history. New tests pin retention, WAL truncation, re-redaction, two-source Forget and consent cutoffs. CI also runs Rust and Vitest under Pacific/Kiritimati and Pacific/Pago_Pago.

Not fixed: Haiku 5.5 is still priced at its low rate. Usage is stored summed, so a request's size is unknown. dirs, sha2 and zbus stay on Tauri's versions, since upgrading would only add second copies.

Checked locally

• 755 Rust tests in UTC, UTC+14 and UTC-11. 698 Vitest tests in all three.
• clippy, fmt, Biome, tsc, actionlint, and gitleaks on the diff.
• An independent review found 8 regressions in the first pass. All are fixed, with tests.
• Rescoring a copy of real data: old rules gave 9,500 messages, 243 corrections and 1,224 tasks. New rules give 9,447, 240 and 1,241.

🤖 Generated with Claude Code