@akasecurity/ai-tc-codex 0.9.14 → 0.9.16

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -806,18 +806,33 @@ watching out for Codex going forward."
806
806
  Before you finish, confirm every AKA_SHOW region on the path you took was
807
807
  relayed to the user. If you summarized one instead of pasting it, paste it now.
808
808
 
809
- ## Known limitation
810
-
811
- Live PreToolUse/PostToolUse redaction currently covers `Bash` shell calls only —
812
- Codex does not yet fire these hooks for `apply_patch` (file-write) calls (an
813
- upstream Codex hook-coverage gap). File-write **content is not scanned at all**
814
- on Codex — not live, and not after the fact either. The consent-gated historical
815
- scan this wizard offers in steps 1/3 reads conversation messages only, and the
816
- tool-call reconciler records a write's changed paths and byte sizes, never the
817
- bytes themselves. So a secret written into a file through `apply_patch` is
818
- neither redacted before it lands nor reported afterwards; only the path is
819
- recorded. If the user asks why a secret in a file edit wasn't caught, say so
820
- plainly rather than implying it is picked up somewhere else.
809
+ ## Known limitations
810
+
811
+ Live scanning covers `Bash` shell calls and `apply_patch` file writes before
812
+ they run, and the output of shell calls, Codex's web tool and MCP tools after
813
+ they return. That includes the calls a code-mode `exec` makes: Codex fires the
814
+ hooks for each tool the script calls (confirmed on codex-cli 0.160.0). The
815
+ script itself fires no hook, so it is seen only through those calls, and web
816
+ and MCP tool **inputs** are not scanned on the way in.
817
+
818
+ The full text of every `apply_patch` call is **recorded in the local store**,
819
+ with the values AKA detects masked, whether or not anything was found, the
820
+ same as file writes on the other assistants AKA covers. That record is what a
821
+ later re-scan checks against. Shell commands and tool output are recorded only
822
+ when something is detected. Old bodies are cleared only if local body expiry is
823
+ turned on in the workspace settings. A very large tool result is scanned up to
824
+ a bound (5 million characters or 2,000 pieces, and about 7 seconds), past which
825
+ the rest reaches the model unscanned.
826
+
827
+ On a Codex release that fires these hooks for `Bash` alone, file-write
828
+ **content is not scanned at all** — not live, and not after the fact either.
829
+ The consent-gated historical scan this wizard offers in steps 1/3 reads
830
+ conversation messages only, and the tool-call reconciler records a write's
831
+ changed paths and byte sizes, never the bytes themselves. So there a secret
832
+ written into a file through `apply_patch` is neither redacted before it lands
833
+ nor reported afterwards; only the path is recorded. If the user asks why a
834
+ secret in a file edit wasn't caught, say so plainly rather than implying it is
835
+ picked up somewhere else.
821
836
 
822
837
  A redact policy cannot mask a **shell command**, because rewriting one silently
823
838
  changes what runs. What happens in its place is the workspace's **redact fallback** (`redactFallback` in