Add the ToolApprovalGuard middleware package - #17
Merged
Conversation
`PIIRedactor::defaultDetectors()` and `PromptInjectionGuard::$patterns` were both protected, so nothing else could reuse them without duplicating detection logic. Moves the detector list to `PIIRedactor\Detectors\DefaultDetectors::all()` and the injection patterns to `InjectionGuardDefaults::patterns()`, with both call sites delegating to the new seams. The pattern strings are byte-identical, which matters because they appear in log context and are passed to custom callbacks as `$detection['pattern']`. Extracting the detectors exposed that the credit card detector was not usable on its own. It emitted every 13 to 19 digit run and relied on `PIIRedactor` applying the Luhn check afterwards in `shouldKeepDetection()`. The check now lives in the detector's own validator closure, matching how the URL detectors already validate, so a caller taking the detector gets the whole behaviour. Detection results are unchanged. `PIIRedactor` loses its redundant filtering branch and its private checksum copy, 109 lines lighter. BREAKING: removes the protected `PIIRedactor::passesLuhn()` method. Only affects code that subclassed `PIIRedactor` and called it directly.
`ScansApprovalDecisions` already walked the arguments edited when resolving a paused run. The tool calls the model proposed carry arguments of exactly the same shape, at the other end of the same approval cycle. Adds `pendingApprovalSegments()`, reusing the existing dot-path walk and keying each segment by the pending approval ID.
Intercept inspects what goes out, but never saw what comes back in. Tool results, retrieved documents and conversation history do not pass through the middleware pipeline, which is where indirect prompt injection lives. The effect of a successful injection almost always surfaces as a tool call. `ToolApprovalGuard` inspects the tool calls an agent proposes while pausing for approval, before they are surfaced for review. It checks tool allow and deny lists, scans arguments for PII and secrets as an exfiltration signal, and scans them for injection patterns. It reuses the PII Redactor detectors and the Injection Guard patterns rather than keeping a second copy. This is the first middleware to act on the response rather than the prompt, since the tool calls it guards are proposed by the model. On a streamed run it cannot block, because the caller has already received the streamed text by the time approvals are known, so block degrades to logging. The tool has still not executed, so a logged proposal continues to require human approval. Actions are block and log only. Rewriting a proposed tool call would desynchronise the paused turn the provider recorded.
The middleware works without it via internal defaults. This adds the section to the publishable config so the options are discoverable and can be set globally.
veeqtoh
changed the title
ToolApprovalGuard middleware package
veeqtoh
force-pushed
the
feat/tool-approval-guard
branch
from
August 1, 2026 15:36
8e261c5 to
846bb8a
Compare
veeqtoh
force-pushed
the
feat/tool-approval-guard
branch
from
August 6, 2026 16:17
846bb8a to
cac7ff7
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation and Context
Intercept inspects what goes out, but never saw what comes back in. Tool results, retrieved documents and conversation history do not pass through the middleware pipeline, which is where indirect prompt injection lives. The effect of a successful injection almost always surfaces as a tool call.
ToolApprovalGuardinspects the tool calls an agent proposes while pausing for approval, before they are surfaced for review. It checks tool allow and deny lists, scans arguments for PII and secrets as an exfiltration signal, and scans them for injection patterns. It reuses the PII Redactor detectors and the Injection Guard patterns rather than keeping a second copy.This is the first middleware to act on the response rather than the prompt, since the tool calls it guards are proposed by the model. On a streamed run it cannot block, because the caller has already received the streamed text by the time approvals are known, so block degrades to logging. The tool has still not executed, so a logged proposal continues to require human approval.
Actions are
blockandlogonly. Rewriting a proposed tool call would desynchronise the paused turn the provider recorded.Dependencies
Depends on
promptphp/intercept-injection-guardpromptphp/intercept-pii-redactorpromptphp/intercept-supportTest Instructions
Run
composer test