ShipGuard
Ship with confidence. ShipGuard finds bugs before your users do — and gets smarter every time.
Say what you want. ShipGuard keeps the agent inside it.
Powerful reasoning models sometimes turn an ambiguity — or their own suggestion — into permission
to change more than you asked. ShipGuard prevents that automatically. Your request is the allowed
change; everything else stays as it is. If two interpretations would produce materially different
results, the agent asks before acting.
There is no command to learn. Talk to Claude or Codex in ordinary language, in any language:
“Change only the zoom-photo prompt. Keep everything else exactly as it is.”
“Give the agent more freedom for this task.”
“Disable scope protection until I turn it back on.”
“Protect the scope strictly again.”
“Is strict scope protection active?”
Strict scope protection starts enabled for every model and reasoning effort on supported hosts.
Choose one of three modes in the chat: protected, relaxed for this task only, or relaxed across
future tasks and sessions until you turn protection back on. If you only say “turn it off,” the
agent asks which duration you mean.
You can also ask Claude or Codex to test one or several model/effort combinations:
“Test whether Astra at high, max and ultra respects my scope, with and without strict protection.”
ShipGuard replies in the chat with PASS, FAIL, or UNAVAILABLE for each configuration.
Version 2.18.4
Prepared from 5 non-merge commits since 2.18.3,
through 61f4ecbcd2f3. See the changelog
for the complete commit range.
Consolidated release 2.16.0
Goal-linked pre-review, verifier protocol hardening, an optional E2E evidence adapter, prompt audits
and grounded dialogue reviews are now integrated with private communication memory.
See capabilities, validation limits and native updates.
Private communication memory
Ask ShipGuard to study your authorized conversations and update a private contextual lexicon.
When an expression remains ambiguous, the mission hook points the agent to that local memory.
Personal profiles stay outside the plugin; updating ShipGuard distributes the method, while
profile distribution remains separate. No automatic transcript collection or permission inference.
See communication memory.
Start with your task
Describe what you need in ordinary language. These examples lead to different kinds of evidence:
| You need to… | Say… | ShipGuard path |
|---|
| Understand existing interfaces before editing | “Review the existing interfaces before I change this.” | sg-pre-review |
| Check a completed change | “Check whether my change is ready to ship.” | sg-ship |
| Look for implementation bugs | “Audit this change for technical bugs.” | sg-code-audit |
| Check a workflow rule or invariant | “Check that this workflow still respects its rules.” | sg-logic-audit |
| Compare behavior before and after | “Compare this process before and after my change.” | sg-process-check |
| Verify what changed on screen | “Test the affected screens.” | sg-visual-run |
| Capture a bug you can reproduce in the browser | “Record this browser flow as a test.” | sg-record |
Each path reports the evidence it obtained and what remains unverified. You can ask for one path directly; checking a change does not require running every path.
How it works
code audit → check contracts and invariants → simulate behavior → confirm on screen → human review
ShipGuard closes the loop between static analysis, intended behavior, and visual reality. Code audit finds implementation bugs. Logic Audit checks workflows, state machines, retries, transactions, and algorithms against declared contracts and invariants. Process Check compares before/after behavior. Visual tests verify what users see. You decide what ships.
Status
| Module | Status | Notes |
|---|
| Visual E2E Debugger | 🟢 Stable | discover → run → review → fix loop |
| Code Audit | 🟢 Updated in 2.9.0 | report-only default, fix-safety tiers, run isolation, evidence gates |
| Logic Audit | 🟢 Stable | detected automatically by SG-SHIP; contract/invariant checks and counterexamples |
| Macro Recorder | 🟢 Stable | record → replay via YAML manifests |
| Self-Improving Engine | 🟡 Experimental | sg-improve + sg-scout, evolving |
| Review Dashboard | 🟢 Stable | HTML generation, Findings tab, annotations, monitor |
| Recette CLI | 🟢 New in 2.5.0 | shipguard init/serve/crawl/run/review — deterministic, exit codes 0/1/2/3 |
| Scope protection | 🟢 Updated in 2.14.0 | All models by default; three conversational protection modes |
| CI/CD Integration | 🔴 Planned | the CLI's stable exit codes are the intended CI entry point |
⚠️ Requires Claude Code or Codex. Visual flows also require agent-browser. Some flows are experimental and evolving fast.
Six AI-powered modules. Use one, some, or all six. No test files to write.
📸 Visual E2E Debugger
Auto-discover routes, generate tests, mark bugs on screenshots — AI traces to source code and fixes automatically.
“Check the screens affected by my changes.”
This is the beginning of the README. Read the rest at the original. Read the full README ↗
