Always-on anti-hedging policy for the Pi coding agent — auto-fix instead of caveat. [](https://www.npmjs.com/package/token-phish-blocker) [](https://pi.dev/packages/token-phish-blocker) [](./LICENSE) [](https://github.com/getdatasurge/token-phish-blocker/actions/workflows/ci.yml) ## Install ``` pi install npm:token-phish-blocker ``` Or browse it on the [pi.dev package catalog](https://pi.dev/packages/token-phish-blocker). From git, if you're not using npm: ``` pi install git:github.com/getdatasurge/token-phish-blocker ``` Try it once without installing: ``` pi -e npm:token-phish-blocker ``` ## What it does Agents love to hedge: "one caveat — I didn't handle the empty-array case," "just a heads up, I didn't wire up error logging." That's a fixable gap phrased as commentary instead of fixed. Token Phish Blocker injects a 4-rule policy into the system prompt on every turn: stop hedging, fix trivial and blocking issues before calling the task done, stay inside the current task's blast radius, and log every such auto-fix through a mandatory tool call. The log call drives a live counter in the footer, so the policy being "on" is directly observable — not just asserted. That's the whole thing. 30 seconds, no config, no runtime deps. ## The four rules | Rule | What it means | | --- | --- | | 1. No hedging | Banned caveat phrasing ("one thing I didn't do", "just a heads up", ...) — fix it instead of mentioning it. | | 2. Auto-implement | Trivial fixes and fixes that block the task from actually working land before you say "done" — no permission-asking. | | 3. Stay in scope | Auto-fixes are limited to the blast radius of the current task. Anything bigger still gets surfaced, plainly. | | 4. Log every auto-fix | Every rule-2 fix calls `token_phish_blocker_log` with a one-line description — mandatory, and it's the audit trail. | Full text: [docs/POLICY.md](./docs/POLICY.md). ## The audit trail Every auto-fix appends a tab-separated line to `~/.pi/agent/token-phish-blocker.log`: ``` ISO-timestampcwddescription ``` Example: ``` 2026-08-12T03:14:07.221Z /home/user/projects/api added missing null check in parseConfig() that would have crashed on empty input ``` ## The status line The moment a session starts, the footer shows the counter at zero — that's your proof the policy is injected and active: ``` 🎣 TPB · 0 fixes ``` Each `token_phish_blocker_log` call ticks it up live, so by the end of a session you can see exactly how many caveats became fixes instead: ``` 🎣 TPB · 3 fixes ``` | Situation | What you see | | --- | --- | | Fresh session | `🎣 TPB · 0 fixes` — policy injected, nothing logged yet | | After auto-fixes | `🎣 TPB · N fixes`, updating the moment each fix lands | | Headless (`pi -p "..."`) | No status line — the audit log still gets every entry | | `TPB_NO_STATUS=1` | Counter hidden — policy and audit log keep working | Nothing to install or configure — the status-line helper ships inside the extension and registers itself on session start. ## Customizing the policy Copy [docs/POLICY.md](./docs/POLICY.md) to `~/.pi/agent/token-phish-blocker-policy.md` and edit it. The override file replaces the built-in policy verbatim. Delete it to fall back to the built-in policy. ## FAQ **Does it work in omp?** Yes — `omp install npm:token-phish-blocker`. **Does it phone home?** No. Zero network calls, zero runtime dependencies. The only writes are to the local log file. **How do I turn it off?** `pi remove npm:token-phish-blocker`, or disable it per-project via `pi config`. **How do I hide the status counter?** Set `TPB_NO_STATUS=1` in your environment. The policy injection and audit log keep working. --- MIT © [getdatasurge](https://github.com/getdatasurge) · [issues](https://github.com/getdatasurge/token-phish-blocker/issues)