Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
37 commits
Select commit Hold shift + click to select a range
614d494
Add tested public-source release audit
VladimirBrejcha Aug 21, 2026
3b4e91e
Fix public-source audit review findings
VladimirBrejcha Aug 21, 2026
698e583
Harden public-source audit edge cases
VladimirBrejcha Aug 21, 2026
a32a2ac
Close public-source audit bypasses
VladimirBrejcha Aug 21, 2026
4a43c63
Harden workflow trust matching
VladimirBrejcha Aug 21, 2026
1f1595d
Parse guarded flow-style workflow keys
VladimirBrejcha Aug 21, 2026
d4eb99a
Resolve remaining audit review edge cases
VladimirBrejcha Aug 21, 2026
85a85d4
Close remaining release audit review gaps
VladimirBrejcha Aug 21, 2026
e630a33
Harden workflow scope and YAML semantics
VladimirBrejcha Aug 21, 2026
1171471
Cover reusable workflows and YAML edge cases
VladimirBrejcha Aug 21, 2026
3d78c37
Harden public audit review edge cases
VladimirBrejcha Aug 21, 2026
b0bbcb8
Address exact-head Codex review findings
VladimirBrejcha Aug 21, 2026
18bcd5e
Close remaining public audit gaps
VladimirBrejcha Aug 21, 2026
5e23fd4
Audit indirect workflow dependencies
VladimirBrejcha Aug 21, 2026
6d664e0
Close indirect workflow audit gaps
VladimirBrejcha Aug 21, 2026
35f5433
Harden derived workflow input checks
VladimirBrejcha Aug 21, 2026
7be2b82
Propagate privileged workflow taint end to end
VladimirBrejcha Aug 21, 2026
d2d4e8a
Audit dynamic workflow runtime inputs
VladimirBrejcha Aug 21, 2026
e2e2488
Audit persisted and shell checkout inputs
VladimirBrejcha Aug 21, 2026
db76cd2
Audit reusable outputs and workflow artifacts
VladimirBrejcha Aug 21, 2026
63e2ae9
Audit public event and runtime execution inputs
VladimirBrejcha Aug 21, 2026
e96c521
Audit remaining privileged workflow inputs
VladimirBrejcha Aug 21, 2026
273bbe4
Close remaining privileged workflow gaps
VladimirBrejcha Aug 21, 2026
4148822
Propagate workflow control taint
VladimirBrejcha Aug 21, 2026
520b6ee
Close reusable and control-flow taint gaps
VladimirBrejcha Aug 22, 2026
bbeb73c
Close evaluator and event-object taint gaps
VladimirBrejcha Aug 22, 2026
28ab49f
Track artifact provenance across execution boundaries
VladimirBrejcha Aug 22, 2026
32aa3a5
Handle heredoc and indirect execution taint
VladimirBrejcha Aug 22, 2026
c066120
Close nested evaluator and preload bypasses
VladimirBrejcha Aug 22, 2026
3462dfd
Close implicit loader and output alias gaps
VladimirBrejcha Aug 22, 2026
770e76d
Preserve execution provenance across shell and jobs
VladimirBrejcha Aug 22, 2026
0134749
Close remaining execution provenance gaps
VladimirBrejcha Aug 22, 2026
848d5e2
Cover remaining dynamic execution sinks
VladimirBrejcha Aug 22, 2026
244dc81
Resolve remaining workflow execution review gaps
VladimirBrejcha Aug 22, 2026
015ac94
Close latest workflow execution review gaps
VladimirBrejcha Aug 22, 2026
bfff6cd
Unify shell function execution analysis
VladimirBrejcha Aug 22, 2026
a22a843
Harden workflow taint and artifact execution analysis
VladimirBrejcha Aug 22, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 26 additions & 5 deletions README.md
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
# 51Code Agent Skills

This repository is the source package for 15 skills maintained by 51Code and
This repository is the source package for 16 skills maintained by 51Code and
the bootstrap for our reviewed machine-global skill baseline.

The repository intentionally has no catalog, registry, generated lock, sync
Expand All @@ -9,14 +9,15 @@ package directly.

## Global baseline

The global baseline supports Codex and Claude Code and contains these 38 skills:
The global baseline supports Codex and Claude Code and contains these 39 skills:

- 51Code-owned: `code-review`, `gemini-files-api`, `harness-engineering`,
`hint-overlay-visual-verification`, `ios-xcodegen`,
`lifecycle-and-side-effects-correctness`, `local-model-serving`,
`mechanism-audit`,
`meeting-transcription`, `silent-pushes-setup`, `spec-creation-updating`,
`swift-testing`, `swiftui-view-refactor`, `xcode-build`, and `xcode-cloud`
`meeting-transcription`, `public-source-release-audit`,
`silent-pushes-setup`, `spec-creation-updating`, `swift-testing`,
`swiftui-view-refactor`, `xcode-build`, and `xcode-cloud`

- Third-party: `swift-concurrency` and the 22 `asc-*` App Store Connect CLI
skills from `rorkai/app-store-connect-cli-skills`
Expand All @@ -37,7 +38,7 @@ idempotent. The script uses `skills@1.5.14`, explicit Git tags or commits, and
explicit skill names. In this CLI, `#ref` selects a Git branch or tag; `@name`
selects a skill and must not be used as a version pin. Sources pinned to a raw
commit are checked out and verified before being passed to the manager as a
local source. Before reporting success, the script verifies all 38 entrypoints
local source. Before reporting success, the script verifies all 39 entrypoints
in the shared and Claude Code manager roots and bootstraps the copied
`gemini-files-api` dependencies in both roots. The Claude Code root honors
`CLAUDE_CONFIG_DIR` when it is set. The script also converts the pinned
Expand All @@ -46,6 +47,26 @@ manager's partial-install result into a non-zero bootstrap failure.
The script installs and reconciles the baseline names only. It does not remove
retired names or unrelated global skills installed outside this baseline.

## Public-source safety

`public-source-release-audit` includes the repository's deterministic safety
gate. It audits committed files and the Git index, including filenames,
symlink targets, common encoding variants, high-confidence credentials,
private-key blocks, Git LFS pointers, and public-workflow trust boundaries.
Stage the intended publication candidate before relying on a local result:

```bash
node public-source-release-audit/scripts/public-source-release-audit.mjs \
--repo .
```

Before making a repository public or closing a suspected leak, run the same
gate with `--history` from a complete clone. Add `--github OWNER/REPO` and one
`--required-check NAME` per expected check to verify live GitHub visibility,
secret scanning, push protection, GitHub Actions check binding, branch rules,
and runner isolation. The gate's fixture suite runs as part of
`scripts/verify.sh`.

## Updating the baseline

1. Update the 51Code-owned skill folders in this repository.
Expand Down
11 changes: 11 additions & 0 deletions THIRD_PARTY_NOTICES.md
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,17 @@ registry-local forks below also retain their reviewed upstream MIT notices.
- Upstream notice: `Copyright (c) 2026 Thomas Ricouard`
- Local state: modified maintained fork

## public-source-release-audit scanner

- Source: [Codex Autopilot public safety audit](https://github.com/fiveonecode/autopilot/blob/9dd7ecb8a1aecb3d757b935a970991fd3461f5f4/agent-harness/src/lib/public-safety-audit.ts)
- Reviewed upstream commit: `9dd7ecb8a1aecb3d757b935a970991fd3461f5f4`
- Upstream notice: `Copyright (c) 2026 Codex Autopilot contributors`
- Local state: dependency-free JavaScript port with a repository-release wrapper

The scanner is distributed under the MIT License reproduced in this
repository's [LICENSE](LICENSE). Keep local behavior changes fixture-backed and
reconcile future upstream changes deliberately.

The global bootstrap also installs `swift-concurrency` directly from its pinned
upstream repository. It is not vendored here; its upstream license and notices
remain authoritative. Impeccable remains an optional project-local install
Expand Down
1 change: 1 addition & 0 deletions bootstrap.sh
Original file line number Diff line number Diff line change
Expand Up @@ -39,6 +39,7 @@ owned_skills=(
local-model-serving
mechanism-audit
meeting-transcription
public-source-release-audit
silent-pushes-setup
spec-creation-updating
swift-testing
Expand Down
67 changes: 67 additions & 0 deletions public-source-release-audit/SKILL.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,67 @@
---
name: public-source-release-audit
description: Audit a repository before public publication when committed files, Git history, workflow runner trust, credentials, private machine paths, or GitHub protection settings could leak or weaken the public source boundary. Use for public-repository launches, release readiness, repository visibility changes, or investigation of a suspected public-source leak; do not use it as a substitute for product security review.
---

# Public Source Release Audit

Use the bundled gate as the deterministic oracle, then apply judgment only to
the residual public meaning that a scanner cannot decide.

## Run The Gate

From the repository being audited, resolve this skill's installed directory
and run:

```bash
node <skill-directory>/scripts/public-source-release-audit.mjs --repo .
```

This scans `HEAD` and the Git index, not dirty working-tree replacements. Stage
the intended publication candidate before relying on the result.

Before a public launch or after a suspected leak, use a complete, non-shallow
clone with all public refs and Git LFS objects present, then add `--history`.
History mode fails closed for shallow or grafted repositories.

For live GitHub controls, add the canonical repository and every required check:

```bash
node <skill-directory>/scripts/public-source-release-audit.mjs \
--repo . --github OWNER/REPO --required-check CHECK_NAME
```

The GitHub mode requires evidence for public visibility, secret scanning, push
protection, strict default-branch status checks bound to GitHub Actions,
deletion and force-push protection, zero ruleset bypass actors, and no
self-hosted runner available to the public repository. Inability to read
required evidence is a failure, not a clean result.

## Interpret Results

- Any error finding blocks publication. Do not bypass it by scanning a dirty
replacement, dropping history, or weakening a rule.
- The gate never prints matched credential or path values. Preserve that
redaction in reports, issues, logs, and review comments.
- Workflow advisories identify mutable action refs, `pull_request_target`, or
dynamic runner selection that needs threat-model review. Use
`--fail-on-warning` when the repository has adopted those stricter policies.
- A historical credential requires credential revocation and a decision about
history remediation. Removing it only from the current tree is not closure.
- Fetch and verify missing Git LFS objects; do not allowlist an unaudited
pointer.

## Finish The Human Audit

The gate detects high-confidence credentials, private-key blocks, non-generic
machine homes, sensitive filenames and symlink targets, encoding variants,
and mechanical repository controls. It cannot decide whether ordinary prose,
screenshots, customer facts, private repository links, internal task names, or
business context are approved for publication. Review the staged diff, commit
messages, workflows, generated artifacts, and public-facing metadata for those
semantic leaks before declaring the source safe.

Report the exact command, audited ref, whether history and live GitHub evidence
were included, error/advisory rule IDs, remediation, and remaining uncertainty.
Do not change GitHub settings, rewrite history, revoke credentials, push, or
publish without the authorization required for that separate action.
4 changes: 4 additions & 0 deletions public-source-release-audit/agents/openai.yaml
Original file line number Diff line number Diff line change
@@ -0,0 +1,4 @@
interface:
display_name: "Public Source Release Audit"
short_description: "Audit repositories before public release"
default_prompt: "Use $public-source-release-audit to verify this repository's committed source, history, workflow trust, and live GitHub protections before publication."
Loading