Skip to content
CodeVetter docs
Esc
↑↓navigate↵open⌘Jpreview
On this page

Failed approaches and reusable mistakes

Things that broke, why, and what not to try again.

Failed approaches and reusable mistakes

Each entry is a thing that actually went wrong in this repo, the root cause, and the constraint it leaves behind. Link here when rejecting a similar idea.

Dual package manager (npm + pnpm) lockfile drift

  • What broke: Cloudflare Pages build failed on feat/landing-page-overhaul (2026-05-02) even though GitHub CI passed.
  • Root cause: a root package-lock.json and a nested desktop pnpm-lock coexisted; @saas-maker/eslint-config was absent from the pnpm lockfile. Pages ran pnpm install --frozen-lockfile against a stale lockfile.
  • Fix: regenerated pnpm-lock.yaml, deleted the npm lockfile, committed the lockfile in isolation.
  • Constraint: one package manager (pnpm) across all CI workflows. Do not reintroduce package-lock.json. The repo pins packageManager: pnpm@10.33.2.
  • See: PROJECT_STATUS.md 2026-07-11 desloppification sweep.

Wrong Cloudflare Pages output directory

  • What broke: after the lockfile fix, Pages still produced an empty/wrong build.
  • Root cause: Pages destination_dir was set to dist while the desktop Vite config writes to out (apps/desktop/vite.config.ts outDir: "out"). Pages was also pointed at apps/desktop instead of apps/landing-page-astro.
  • Fix: reconfigure Pages root_dir: apps/landing-page-astro, destination_dir: dist. The Astro site writes to dist; the desktop app writes to out.
  • Constraint: Pages only ever deploys apps/landing-page-astro. Never point it at apps/desktop. See operations/landing-deploy.md.

@tauri-apps/plugin-sql was never the DB layer

  • What broke: docs claimed plugin-sql was the SQLite layer; it was a dead dependency.
  • Root cause: the Rust backend has used rusqlite all along; plugin-sql was a leftover from an earlier prototype.
  • Fix: removed plugin-sql in the 2026-07-11 sweep.
  • Constraint: DB is Rust-internal via rusqlite. Do not re-add plugin-sql. See architecture/data-model.md.

Claude usage double-counted ~2.2×

  • What broke: all Claude token/cost numbers were inflated ~2.2× (measured 103–134% per month).
  • Root cause: the indexer summed the usage object of every Claude JSONL line, but Claude Code writes one line per content block and each repeats the same final usage — 50%+ of usage lines are byte-identical repeats.
  • Fix: adapter dedups usage by (message.id, requestId) with the last key persisted per session (cc_sessions.last_usage_key); one-time backfill re-scans on-disk transcripts.
  • Constraint: never sum Claude JSONL usage lines without the (message.id, requestId) dedup. See knowledge/learnings/telemetry-and-indexing.md.

Codex cumulative-token inflation

  • What broke: one Codex session booked 61.5B tokens / $35k (true: 391M / ~$220); “today” read ~$12.9k.
  • Root cause: Codex reports session-cumulative token totals; the incremental indexer was adding that running total every pass.
  • Fix: tokens_absolute flag so cumulative tokens are SET not added; one-time fix_codex_token_totals repair re-reading each Codex file.
  • Constraint: Codex tokens are cumulative — use SET semantics, not ADD.

Multi-model Claude sessions billed to the last model

  • What broke: a 211MB session with 17k opus-4-7 messages + 1.6k fable-5 messages billed $3.6k entirely to fable.
  • Root cause: session-level model_used is last-model-wins.
  • Fix: per-message session_model_usage table populated by the indexer + one-time streaming backfill over existing Claude JSONL.
  • Constraint: by-model attribution must use per-message model splits, not session-level model_used.

Indexer burning ~95% of a core

  • What broke: sustained background indexer CPU burn.
  • Root cause: subagent sidechain transcripts shared the parent’s sessionId, collapsing onto one DB row so each was re-parsed + archive -replaced every pass; the skip compared drift-prone nanosecond mtime strings.
  • Fix: skip on exact byte-offset == file-size; key sidechains by unique per-file id; migrate the offset backlog; repair FTS sync UUID handling. Steady-state index pass 87s → 1.9s.
  • Constraint: sidechain transcripts need their own key; mtime comparison must be byte-offset-based, not nanosecond-string-based.

tauri-driver native e2e never supported macOS

  • What broke: the native e2e path was dead weight.
  • Root cause: tauri-driver never actually supported macOS in our setup.
  • Fix: removed in the 2026-07-11 sweep; Playwright chromium against the Vite dev server is the e2e path.
  • Constraint: do not reintroduce tauri-driver. See development/testing.md.

More lessons

See also LESSONS.md for older entries and DECISIONS.md for the decision log. When a new failure happens, add an entry here in the same shape (what broke / root cause / fix / constraint).

Was this page helpful?