Back to search
solo-agent-guards
MarketplaceWorks with Claude CodeGitHub
Guards for running a coding agent alone: prove the test catches the bug, prove the push landed, and prove each status claim.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/robfold--solo-agent-guards)HTML
<a href="/listings/robfold--solo-agent-guards"><img src="/api/listings/robfold--solo-agent-guards/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
Claude Code:
/plugin marketplace add robfold/solo-agent-guards /plugin install solo-agent-guards@solo-agent-guards
Other agents, with the [skills CLI](https://skills.sh):
npx skills add robfold/solo-agent-guards
Or copy the folders in [`skills/`](skills/) into your agent's skills folder (`~/.claude/skills/` for Claude Code).
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 16 lines
- {
- "$schema": "https://anthropic.com/claude-code/marketplace.schema.json",
- "name": "solo-agent-guards",
- "description": "Guards for running a coding agent alone: prove the test catches the bug, prove the push landed, and prove each status claim.",
- "owner": {
- "name": "Rob"
- },
- "plugins": [
- {
- "name": "solo-agent-guards",
- "description": "test-and-prove breaks the guard to show a test catches it. change-integrity checks that only your changes went up and actually landed. claim-audit checks that each status claim rests on the evidence its verb needs.",
- "source": "./",
- "category": "productivity"
- }
- ]
- }
README.md · README · 73 lines
- # solo-agent-guards
- English · [한국어](README.ko.md)
- Agent skills that stop a coding agent from saying "done" when it isn't. For people who run an AI coding agent alone, with no reviewer behind it.
- 
- <sub>Recorded in Claude Code with this plugin loaded, on a small example project. Idle time is cut and long busy stretches are sped up.</sub>
- - **[`test-and-prove`](skills/test-and-prove/SKILL.md): a passing test proves nothing yet.** The agent removes the guard the test protects, runs only that test, sees it fail, and restores the guard. A test that stays green guards nothing and gets rewritten. The agent reads the run count, not the exit code: `go test -run X` and `jest -t X` exit with 0 when nothing ran.
- - **[`change-integrity`](skills/change-integrity/SKILL.md): a successful push proves nothing yet.** The agent checks the remote SHA, the three-dot diff from the base, and what the PR or MR really contains. It stages only its own paths and leaves other sessions' unfinished work alone.
- - **[`claim-audit`](skills/claim-audit/SKILL.md): a success message proves nothing yet.** Before the agent reports "merged", "deployed", "passed", "running in production", or a result number, it re-reads the target and closes four gaps: a merge that is only scheduled, a check that has never failed, a number padded with its own test traffic, and a manual run reported as production. Before it takes over someone else's work, it reads the reviewer's last verdict, not the card's status.
- ## What it looks like
- With `test-and-prove`, the agent reports the break, not just the green run. This is the report from the recording above:
- > | Change to `billing/refunds.py` | Test run | Result |
- > |---|---|---|
- > | `if order.refunded:` → `if False:` | `test_refund_rejects_already_refunded_order` | 1 failed |
- > | `if amount > order.total:` → `if False:` | `test_refund_rejects_amount_over_total_and_leaves_order_untouched` | 1 failed |
- > | `>` → `>=` (an off-by-one that blocks a full refund) | `test_refund_marks_order_refunded` | 1 failed |
- >
- > Each run reported 1 test failing, so the test was selected and failed, not skipped. After putting the code back, `git diff billing/` is empty and the suite passes again (4 passed).
- With `change-integrity`, it reads what landed, not what it meant to push:
- > The push succeeded, but `git diff --stat origin/main...origin/fix-refund` shows 3 files and I changed 2. `config/billing.yml` is another session's unfinished work. I rebuilt the branch from `origin/main` with only my commit and pushed `fix-refund-2`. The new PR shows 2 files.
- With `claim-audit`, it reports what the evidence reaches, not what the tool printed:
- > MR !42 is not merged yet. `glab mr merge` printed "Merged!", but the API says `state: opened`, `merge_when_pipeline_succeeds: true`, `merge_commit_sha: null`: the merge is scheduled for when CI passes. The quote check says 3/3 passed, but it also passes a receipt that quotes nothing, so the check is broken; fixed by hand, the result is 2/3. Yesterday shows 7 paid orders; 4 are mock payments from curl and python-requests and 1 is an internal QA account, so 2 look like customers.
- <sub>The `test-and-prove` report is copied from the recording. The `change-integrity` report is an example with made-up names. The `claim-audit` report is shortened from a test run on a small example project with planted traps.</sub>
- ## Install
- Claude Code:
- ```
- /plugin marketplace add robfold/solo-agent-guards
- /plugin install solo-agent-guards@solo-agent-guards
- ```
- Other agents, with the [skills CLI](https://skills.sh):
- ```
- npx skills add robfold/solo-agent-guards
- ```
- Or copy the folders in [`skills/`](skills/) into your agent's skills folder (`~/.claude/skills/` for Claude Code).
- ## Use
- The skills load on their own when the request matches. `test-and-prove` loads when you ask for tests, coverage, or a flaky-test fix. `change-integrity` loads when you ask the agent to commit, push, or open a PR or MR. `claim-audit` loads right before the agent reports a status or a result number, and when you ask "is that real?". You can also call any of them by name: "write tests for the refund check and prove them".
- `test-and-prove` ends with numbers as the runner printed them, plus the mutation result for each new test. `change-integrity` ends with the remote SHA and the diff the PR or MR really contains. `claim-audit` ends with each claim next to the evidence it was re-read from, or a note that it could not be.
- ## When to use something else
- - **Mutation testing across a whole codebase:** [Stryker](https://stryker-mutator.io/), [mutmut](https://github.com/boxed/mutmut), or [PIT](https://pitest.org/). `test-and-prove` does the same thing by hand, for the one test you just wrote.
- - **Writing the test first:** [`test-driven-development`](https://github.com/obra/superpowers/tree/main/skills/test-driven-development) in obra's superpowers. TDD watches the test fail before the code exists. `test-and-prove` breaks the guard after the code exists, so it also works for tests you add to old code.
- - **Evidence before any claim of success:** [`verification-before-completion`](https://github.com/obra/superpowers/tree/main/skills/verification-before-completion) in obra's superpowers covers this in general. These three skills go deeper on the places where that claim most often goes wrong: tests, pushes, and status reports built on the wrong kind of evidence.
- - **Setting up an isolated workspace:** [`using-git-worktrees`](https://github.com/obra/superpowers/tree/main/skills/using-git-worktrees) in obra's superpowers. `change-integrity` also uses worktrees, but its job is to check what actually went up.
- ## Credits
- The rules came from corrections by the person these skills were first built for. The idea in `test-and-prove` is mutation testing. The command evidence in `change-integrity` links to the git documentation it relies on. Each gap in `claim-audit` comes from a real incident.
- ## License
- [MIT](LICENSE)
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.