fix(security): redact tool outputs and patch vulnerable dependencies
Merged
OK
5d ago+473+473+012345678901234567890123456789-85-85-01234567890123456789
What and why
- Resolve the seven dependency alerts with scoped overrides for
fast-uri,@hono/node-server,glob, andminimatch. - Redact project-tool results, serialized JSON, custom
toModelOutputtext, and tool/engine errors before model input. Redact strings before truncation, and apply the live secret ledger to AI trace names, scope metadata, input, and output. - Redact and bound collection-time and runtime skip reasons before reporters receive them, including skips beside an existing failure.
- Remove polynomial regex scanning from stack-frame parsing and provider endpoint trimming; retain single-slash file URL support. Preserve ACP forwarding failures in the runner's redacted diagnostics while keeping exception text out of HTTP responses. Normal client disconnects remain harmless.
- Add regression coverage, documentation, consumer-skill updates, and patch changesets.
flowchart LR
T[Tool output] --> S[Serialize custom JSON output]
S --> R[Redact provider input]
T --> A[Serialize and redact AI trace]
Issue
- Maintainer-requested security cleanup. Private reports remain tracked in the repository's Security tab.
- Web redirect-header reports and three test-fixture scanning findings remain outside this PR. No GitHub alerts or advisories were dismissed.
Verified
Ran it locally: yes
- Built the packages and ran the same new tool-output, skip-reason, and AI-trace regression tests against a source archive of current
mainand this branch.
| main | this branch |
|---|---|
| Tool-output regression fails | Passes, including null-prototype dictionaries, URL serialization, and custom output callbacks |
| Skip-reason regression fails report-schema validation | Passes; JSON, Markdown, and JUnit contain redacted, bounded reasons |
| AI-trace redaction regression fails | Passes with the live ledger applied during serialization |
| ACP forwarding exceptions are sent in the HTTP response | The HTTP response stays generic; the runner stops the session and reports the redacted cause |
pnpm audit reports seven advisories | Zero vulnerabilities |
- Ran the built CLI in a scratch project with a custom tool returning a URL, another throwing an error, and
--ai-trace. The fake provider rejects raw sentinel values. unbox-ai found the redacted marker in the trace and no raw sentinel matches.
✓ redacts custom tools and trace > agent.act "read both tools" · 2 model calls
Test Files 1 passed (1)
Tests 1 passed (1)
pnpm check; 55 focused core tests cover the final review changes. The preceding full core run passed 4,584 tests with 11 skipped. GitHub, ACP, BrowserStack, and TestMu suites also passed. Public declaration comparison: all 25 files unchanged.- All 47 ACP tests passed, including real HTTP streaming/disconnect tests and startup/active-turn error propagation. A built-CLI failure probe confirmed the cause appears in terminal and JSON output with its synthetic secret redacted.
MODEL_PROVIDER_FAILED: The ACP agent session failed: step-tool request forwarding failed: <secret:unsupportedMethod> HTTP method is unsupported.
- Full testbed through the built CLI: 588 passed, 1 skipped. Deterministic web benchmark: 91 passed.
- Expo fingerprint generation succeeded with the updated glob dependencies; the patched Hono server passed a request/response smoke test.
Checklist
- Changeset added for published-package changes.
- Breaking change: not applicable.
- Matching docs and consumer skill updated.
- Public API or wire contract change: not applicable.
- Engine contract change: not applicable.