10 10 `verboten` is a best-effort Pi extension that keeps model agents from exposing or mutating files whose paths are likely to contain secrets.
11 -It blocks obvious direct filesystem tool calls, filters protected rows from broad content searches, judges ambiguous model-authored shell commands, and gives the model terse preventive instructions.
11 +It blocks obvious direct filesystem tool calls, filters protected rows from broad content searches, and judges ambiguous model-authored shell commands without adding main-agent prompt instructions.
12 12 It does not claim isolation, sandboxing, or complete mediation.
14 14 False positives matter because blocking ordinary engineering work defeats the extension's purpose.
194 194 The repository settings must select `openai-codex/gpt-5.6-luna` as `verboten.judgeModel`.
195 195 The repository settings must append `verboten` last in `ultra.subagentExtensions` after the extension exists.
196 -This ensures Ultra sub-agents receive the same hooks and system instructions.
196 +This ensures Ultra sub-agents receive the same tool-call and tool-result protections.
198 198 ##
Regex matching model 399 -Append a terse Verboten instruction in `before_agent_start` on every agent run.
400 -The instruction must state:
402 -> Never expose protected-file contents, encodings, hashes, or copies to model context or tool output, directly or through a bypass.
403 -> Never create, modify, delete, rename, copy, execute, encode, or transmit protected files.
404 -> Trusted programs, hooks, filters, and credential helpers may consume protected files internally when they neither expose protected data nor mutate protected files.
405 -> Metadata-only listing, finding, status, stat, and existence checks are allowed.
406 -> Broad grep searches rooted outside a protected path are allowed; Verboten filters protected rows.
407 -> Mere textual mentions and exclusion patterns are allowed.
408 -> Never bypass Verboten through shells, symlinks, alternate tools, encoding, or delegation to expose or target protected files.
409 -> Reformulate a false positive only when the new operation avoids protected content and mutation.
411 -Inject every effective regex in XML:
415 - <
pattern><![
CDATA[
regex source]]></
pattern>
419 -Split any literal `]]>` safely across CDATA sections.
420 -Do not put source labels or judge implementation details into the model system prompt.
421 -Keep imperative instructions outside the XML block.
399 +Do not register a `before_agent_start` hook or inject Verboten policy, path summaries, or regexes into the main agent's system prompt.
400 +Keep exact effective regexes in runtime enforcement and the dedicated shell judge's input.
401 +The shell judge retains its own policy prompt, including permission for trusted internal credential consumption.
402 +Tool-call blocks and grep filtering remain unchanged.
403 +Block messages provide guidance when an attempted operation is rejected.
423 405 ##
`/verboten` command 449 431 The root Pi extension package discovers `extensions/verboten/index.ts` normally.
450 432 The user's global `ultra.subagentExtensions` list must include `verboten` last.
451 -Ultra sub-agents therefore receive the same tool hooks and `before_agent_start` instructions.
433 +Ultra sub-agents therefore receive the same tool hooks without main-agent prompt injection.
453 435 Loading order is best effort.
454 436 A later extension can still mutate a tool call or result after Verboten.
489 471 | V6 | Embedded scripts | Bash heredocs, PowerShell here-strings, Python/Node scripts, captured text, invoked text, redirects, containers |
490 472 | V7 | Judge lifecycle | One-tool schema, threshold, same-context contract retry, fallback, timeout, auth failure, cancellation, warning dedupe |
491 473 | V8 | Cache | SHA-256 keys, no raw command storage, model/pattern key changes, 256-entry FIFO eviction, failure non-caching |
492 -| V9 | Prompt and command | XML escaping, every effective regex, terse instructions, source-grouped `/verboten`, untrusted-project notice, no model-context entry |
474 +| V9 | Prompt and command | No main-agent prompt injection, source-grouped `/verboten`, untrusted-project notice, no model-context entry |
493 475 | V10 | Repository integration | Package static check, personal settings, Verboten last in Ultra, README, changelog, full repository checks |
495 477 The permanent shell corpus includes at least:
504 486 - embedded text that looks dangerous but is only assigned, printed, or stored
505 487 - embedded text executed through an interpreter or `Invoke-Expression`
507 -The prompt design has empirical evidence from repeated low-reasoning model experiments.
489 +The shell judge prompt design has empirical evidence from repeated low-reasoning model experiments.
508 490 The explicit content-versus-metadata rubric correctly separated the initial 15-command corpus in both repeats.
509 491 The unambiguous field name `forbiddenOperationProbability` prevented models from treating probability as confidence.
510 492 Explicit heredoc and here-string guidance produced 28 correct classifications across two repeats, while the shorter rubric produced one false positive on captured but unexecuted script text.