Troubleshooting

tok0 is designed to fail loud and fall back safely — if a filter ever errors, the raw command output flows through unchanged. Below are the symptoms you’ll actually see and what to do about each.

”My agent isn’t using tok0.”

Diagnosis ladder:

tok0 doctor          # full diagnostic
tok0 status          # which AI tools are wired
tok0 verify          # are the hooks intact?

The most common cause is that the AI tool was running before tok0 init. Restart the tool (close every window, reopen) so it re-reads its hook config.

If tok0 status shows “not wired” for a tool you use:

tok0 init --tools <tool-name>

If tok0 verify shows a hook signature mismatch, you (or a tool) edited the hook file by hand. Re-run tok0 init to refresh.

”Compression seems off — output is smaller than I expected.”

tok0 profile run "<command>"

tok0 profile shows raw vs compressed bytes plus per-stage timings. It tells you which stage stripped what.

If you suspect the wrong rule is matching:

tok0 rules list | grep <command>
tok0 rules show <name>

A project-local rule overrides user-global which overrides built-in. The first match wins.

”The wrapped command is slow.”

tok0 adds typically <1ms per command. If a command got slow after installing tok0:

1. Verify with the bypass: TOK0_DISABLE=1 <your command>. If still slow, tok0 is not the cause.
2. Check the timeout: tok0 kills any wrapped command at runtime.command_timeout_secs (default 60s). Long-running commands need --timeout or a config bump.
3. Profile: tok0 profile run "<cmd>" will show pipeline time vs command time separately.

”Filter error — falling back to raw output.”

This is a deliberate safety message. A regex in a rule died on unusual input; rather than block your terminal, tok0 emitted the raw output and logged the offending rule.

Recover the failure for debugging:

ls ~/.cache/tok0/snapshots/   # latest failures
tok0 doctor --snapshots       # human summary

Each snapshot is a captured stdout + stderr + the rule name that errored. Open an issue with the snapshot attached if it’s a built-in rule.

”tok0: command not found”

The binary isn’t on your PATH. Confirm it landed:

which tok0
ls -l /usr/local/bin/tok0 ~/.local/bin/tok0 2>/dev/null

If it’s in ~/.local/bin and that directory isn’t on your PATH, add it to your shell rc:

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc
source ~/.zshrc

“Permission denied” on config

tok0 writes to ~/.config/tok0/ and ~/.cache/tok0/. If those exist with the wrong owner (common after running sudo once):

sudo chown -R "$USER" ~/.config/tok0 ~/.cache/tok0

Never run tok0 init with sudo. The bridge writes to user-scoped tool configs (~/.claude/, ~/.gemini/, …) — running as root creates root-owned files in your home directory and breaks every subsequent invocation.

”Database is locked”

Two tok0 processes raced on the meter database. The mpsc writer should serialize writes, so this is rare in practice, but if it happens:

fuser ~/.config/tok0/meter.db   # who has it open?

Kill the offender. The db is durable; you won’t lose history.

”Rule changed but tok0 is using the old version”

The rule cache is mtime-based. Touching the file is enough:

touch ~/.config/tok0/filters/my-rule.toml

If you suspect a stale cache anyway:

rm -rf ~/.cache/tok0/rules-cache

Cache rebuilds on next dispatch; cost is negligible (<1ms).

”I want to see exactly what tok0 sent to my model.”

The --log flag shows it:

TOK0_LOG=trace tok0 git diff 2>tok0.log
cat tok0.log | grep 'filtered_output'

trace logs the byte-exact compressed output before it’s printed to stdout.

”How do I revert everything tok0 did?”

tok0 init --remove                            # remove hooks
tok0 untrust --all                            # revoke all project trusts
rm -rf ~/.config/tok0 ~/.cache/tok0           # nuke state
brew uninstall tok0   # or the manager you used

That’s a complete uninstall. Your AI tool configs are restored to pre-tok0 state.