Add redact_secrets builtin hook + before_llm_call transform#2577
Merged
dgageot merged 3 commits intodocker:mainfrom Apr 28, 2026
Merged
Conversation
Adds a single agent-level switch (redact_secrets: true) that brackets every leak vector for known credentials shapes (GitHub PATs, AWS keys, Stripe / Slack / GitLab tokens, JWTs, private keys, ...): - pkg/secretsscan: detection + redaction, ruleset derived from the MIT-licensed docker/mcp-gateway/pkg/secretsscan. - pkg/hooks/builtins/redact_secrets: pre_tool_use builtin that scrubs secret material from ToolInput before tools see it. - pkg/runtime/redact_secrets: before_llm_call message transform that scrubs the same patterns from outgoing chat content (Content, MultiContent, and prior-turn ToolCall.Arguments). - AgentConfig.RedactSecrets / agent.WithRedactSecrets / teamloader wiring so a single YAML knob enables both halves. Assisted-By: docker-agent
Behavior is unchanged. The redact_secrets feature still scrubs the same patterns from tool arguments (pre_tool_use builtin) and outgoing chat messages (before_llm_call transform), gated on the same agent flag. Concretely: - pkg/secretsscan: drop the hasKeyword method (just a free function inlined at the two call sites), drop the dead defensive branch in redactWithRule, use re.SubexpIndex instead of looping SubexpNames. - pkg/hooks/builtins/redact_secrets: drop the unused RedactSecretsName type alias, fold redactMap and redactValue into one redactAny function, drop the slice-identity preservation that wasn't observable at the executor boundary. - pkg/runtime/redact_secrets: fold redactMessageSecrets and redactSingleMessage into a single redactMessage, replace the hand-rolled copy-on-write loop with slices.Clone (Redact returns the same string on no-match, so allocations only happen when there's actually something to scrub anyway). - Trim essay-length docstrings throughout (fields, options, tests). Net: -203 lines. mise lint + mise test stay green. Assisted-By: docker-agent
…nctionCall scrub
Five fixes from a self-review pass:
1. Bug: pre_tool_use builtin returned the entire ToolInput map in
UpdatedInput. Hooks for one event run concurrently and aggregate
via shallow maps.Copy in config order, so emitting unchanged keys
would clobber a sibling hook's modifications. Fix: only emit keys
whose value actually changed (new redactToolInput helper). This
also removes the unsafe top-level type assertion on redactAny's
any-typed return value.
2. Security gap: the runtime transform missed two text-bearing
surfaces that round-trip to providers:
- chat.Message.ReasoningContent (sent back to Anthropic, Bedrock,
and DeepSeek as a thinking block)
- the legacy singular chat.Message.FunctionCall.Arguments
(still emitted by the OpenAI provider when set)
A secret in either of those fields would round-trip to the next
LLM call. Now both are scrubbed; the FunctionCall pointer is
deep-copied so the caller's history stays untouched.
3. Performance: secretsscan.Redact called strings.ToLower(out) and
compiledRules() once per rule (~85 iterations), which on long
inputs dominated the work. Hoisted both out of the loop. Sound
because RedactionMarker contains no rule keyword (locked in by
the new TestRedactionMarkerIsNotASecret) so a stale lower-case
only ever produces extra-cautious regex runs, not false negatives.
Allocations on a 9KB clean message: ~85 -> 2.
4. Test gap: added TestRedactionMarkerIsNotASecret to lock in the
safety property the idempotence claim relies on (the marker must
not match any rule, even when surrounded by arbitrary text).
5. Added BenchmarkRedactCleanInput / BenchmarkRedactWithSecret as a
regression guard for docker#3.
Tests + lint: green. Behavior on clean inputs and previously-tested
secrets is unchanged.
Assisted-By: docker-agent
aheritier
approved these changes
Apr 28, 2026
Contributor
aheritier
left a comment
aheritier
left a comment
There was a problem hiding this comment.
LGTM but missing docs updates
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
2 participants
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.
Adds a single agent-level switch —
redact_secrets: true— that brackets every place where credential material can leak from a docker-agent run to a third party. It scrubs known secret patterns (GitHub PATs, AWS keys, Stripe / Slack / GitLab tokens, JWTs, private keys, Docker Hub PATs, …) at two enforcement points:MessageTransformwalks every text-bearing field of the outgoing chat history (Content,ReasoningContent,MultiContenttext parts, the legacy singularFunctionCall.Arguments, and eachToolCalls[i].Function.Arguments) and replaces detected secrets with[REDACTED].pre_tool_usebuiltin hook (redact_secrets) walksToolInputrecursively (into nested maps and slices) and emits anUpdatedInputwith only the keys whose value actually changed.Detection mirrors the MIT-licensed
docker/mcp-gateway/pkg/secretsscanruleset (re-licensed under Apache-2.0 in attribution) and is kept fast by a keyword pre-filter in front of every regex.Files
pkg/secretsscan/— new package: ruleset,ContainsSecrets,Redact, tests, benchmarks.pkg/hooks/builtins/redact_secrets.go— thepre_tool_usebuiltin and its registration.pkg/runtime/redact_secrets.go— thebefore_llm_callMessageTransform(gated onAgent.RedactSecrets()).pkg/agent/{agent.go,opts.go},pkg/teamloader/teamloader.go,pkg/config/latest/types.go— theredact_secrets: trueagent flag plumbing.agent-schema.json+examples/redact_secrets.yaml— schema entry and a runnable example.Try it
```yaml
agents:
root:
model: openai/gpt-5-mini
instruction: You are a helpful assistant.
redact_secrets: true
toolsets:
- type: shell
```
Now any secret pasted by the user — or generated by a previous turn's tool — is replaced with `[REDACTED]` before it reaches the model OR a tool process. The surrounding text is preserved (`--token=[REDACTED]` rather than the whole arg disappearing).
Notes
Tests
`mise lint` and `mise test` are green. The new tests cover detection parity with upstream, redaction span correctness, idempotence, multi-secret handling, marker safety, the pre_tool_use builtin's recursion + "only changed keys" output, agent-flag wiring, and the runtime transform's per-surface scrubbing + non-mutation contract.
Assisted-By: docker-agent