README
monero-review
Automated security review of monero-project/monero pull requests. Created by the MAGIC Monero Fund, please donate to help support this effort.
A GitHub runner fetches a new PR's head commit once every 30 minutes, has Claude read the diff against a Monero-specific skill, and files what it finds as an issue here. Every finding is attacked by other agents for accuracy and reachability.
Run a review
On GitHub Actions, which files the result as an issue:
gh workflow run review.yml --repo xmrack/monero-review -f pr=11155
Run locally, leaves the result in reviews/:
./review-local.sh 11155
| flag | does |
|---|---|
-f pr=<n> / -f pr=sweep | one PR, or the next unreviewed ones. sweep is the default, and it takes only pull requests targeting master; by number reviews any branch |
-f mode=deep | the escalation. Requires a PR number; pr=sweep is refused |
-f model=<id> | pins the model. Default auto, the same model on both tiers |
DEEP=1 ./review-local.sh | deep, locally |
TIER=single ./review-local.sh | the single-reviewer fallback |
Deep issues carry a deep-review label. Standard reviews are unlabelled,
because unlabelled is what ordinary looks like.
To stop the scheduled sweeps:
gh variable set REVIEW_PAUSED --body 1 --repo xmrack/monero-review
--body 0 resumes. Reviewing a PR by number still works while paused.
Two tiers
Every PR the sweep picks up gets a standard scan. I can manually launch a deep review of a PR.
| standard | deep | |
|---|---|---|
| units the diff is split into | at most 4 | at most 8 |
| researchers | one per unit, all its weakness classes | one per unit per weakness class |
| a pass across the unit seams | no | yes |
| a second look at every unit | no | yes |
| verifier angles per candidate | 2 (reachability, introduced-by-this-diff) | 3 |
| an advocate for a candidate one vote short | no | yes |
| findings that are one defect merged into one entry | yes | yes |
| chosen by | default, always | a human, by PR number |
Both run .claude/workflows/monero-deep-scan.js under different profiles.
Every candidate goes to a panel whose votes are counted in JavaScript rather than argued in prose, and the count lowers severity. What survives is checked for duplicates.
Where things are
References, shared by every skill:
.claude/references/writing.md— how a report is written. Orwell's six rules and the ASD-STE100 writing rules, applied to a security finding, with the words to cut and a pass to run over the draft..claude/references/monero/— how the Monero codebase works: architecture, six end-to-end flows, per-subsystem notes, the macro families that make grep lie, errors and concurrency, build and tests.README.mdthere is the index..claude/skills/monero-security-review/references/— the other half: what to suspect, where the tree says what the code is.
Skills:
monero-standard-review/— the automatic review. Every PR on the queue gets it.monero-deep-review/— A more indepth review which must be manually requested.monero-security-review/— (LEGACY) the single-reviewer fallback, for a session that cannot be granted the agent tools. Not a tier, and nothing in CI routes to it.monero-review-refute/is its adversarial second pass and runs only behind it.
Agents and orchestration — .claude/agents/ and .claude/workflows/. One
workflow script serves both tiers, keyed on profile: what they share is the
coverage arithmetic, the candidate schema and the vote counting, which are
exactly the parts that must not drift. .claude/agents/monero-context.md is
the context contract every agent opens with.
See what people are saying
Similar repos
Languages
Python
44.0%
JavaScript
40.0%
Shell
16.0%
README
monero-review
Automated security review of monero-project/monero pull requests. Created by the MAGIC Monero Fund, please donate to help support this effort.
A GitHub runner fetches a new PR's head commit once every 30 minutes, has Claude read the diff against a Monero-specific skill, and files what it finds as an issue here. Every finding is attacked by other agents for accuracy and reachability.
Run a review
On GitHub Actions, which files the result as an issue:
gh workflow run review.yml --repo xmrack/monero-review -f pr=11155
Run locally, leaves the result in reviews/:
./review-local.sh 11155
| flag | does |
|---|---|
-f pr=<n> / -f pr=sweep | one PR, or the next unreviewed ones. sweep is the default, and it takes only pull requests targeting master; by number reviews any branch |
-f mode=deep | the escalation. Requires a PR number; pr=sweep is refused |
-f model=<id> | pins the model. Default auto, the same model on both tiers |
DEEP=1 ./review-local.sh | deep, locally |
TIER=single ./review-local.sh | the single-reviewer fallback |
Deep issues carry a deep-review label. Standard reviews are unlabelled,
because unlabelled is what ordinary looks like.
To stop the scheduled sweeps:
gh variable set REVIEW_PAUSED --body 1 --repo xmrack/monero-review
--body 0 resumes. Reviewing a PR by number still works while paused.
Two tiers
Every PR the sweep picks up gets a standard scan. I can manually launch a deep review of a PR.
| standard | deep | |
|---|---|---|
| units the diff is split into | at most 4 | at most 8 |
| researchers | one per unit, all its weakness classes | one per unit per weakness class |
| a pass across the unit seams | no | yes |
| a second look at every unit | no | yes |
| verifier angles per candidate | 2 (reachability, introduced-by-this-diff) | 3 |
| an advocate for a candidate one vote short | no | yes |
| findings that are one defect merged into one entry | yes | yes |
| chosen by | default, always | a human, by PR number |
Both run .claude/workflows/monero-deep-scan.js under different profiles.
Every candidate goes to a panel whose votes are counted in JavaScript rather than argued in prose, and the count lowers severity. What survives is checked for duplicates.
Where things are
References, shared by every skill:
.claude/references/writing.md— how a report is written. Orwell's six rules and the ASD-STE100 writing rules, applied to a security finding, with the words to cut and a pass to run over the draft..claude/references/monero/— how the Monero codebase works: architecture, six end-to-end flows, per-subsystem notes, the macro families that make grep lie, errors and concurrency, build and tests.README.mdthere is the index..claude/skills/monero-security-review/references/— the other half: what to suspect, where the tree says what the code is.
Skills:
monero-standard-review/— the automatic review. Every PR on the queue gets it.monero-deep-review/— A more indepth review which must be manually requested.monero-security-review/— (LEGACY) the single-reviewer fallback, for a session that cannot be granted the agent tools. Not a tier, and nothing in CI routes to it.monero-review-refute/is its adversarial second pass and runs only behind it.
Agents and orchestration — .claude/agents/ and .claude/workflows/. One
workflow script serves both tiers, keyed on profile: what they share is the
coverage arithmetic, the candidate schema and the vote counting, which are
exactly the parts that must not drift. .claude/agents/monero-context.md is
the context contract every agent opens with.
See what people are saying
Similar repos
Languages
Python
44.0%
JavaScript
40.0%
Shell
16.0%