Display Name
Vajra
Category
Agent Skills
Sub-Category
Tooling: Orchestrators
Primary Link
https://github.com/thisizmsk-png/vajra
Author Name
Sai Krishna Madavarapu
Author Link
https://github.com/thisizmsk-png
License
MIT
Other License
No response
Description
Agent harness for Claude Code that wraps each run in a verify gate, where a run is not marked done until its tests and build pass. Adds an OS sandbox (Seatbelt on macOS, bubblewrap on Linux), a per-model cost ledger, a tamper-evident replay trace, persistent campaigns, a drift-resistant loop for long autonomous jobs, and a parallel-worktree fleet of 17 specialized agents. Bundles 87 skills and a hardened pre-tool hook that has been through five rounds of red teaming.
Validate Claims
Clone any repo that has a test suite, break one test on purpose, then run /vajra verify-work demo "npm test". The gate runs the tests, exits non-zero, and refuses to mark the run done. Fix the test and rerun to watch it pass. The verify-work script and its exit codes are in scripts/.
Specific Task(s)
Give Claude a small bug fix on a repo with a test suite and require it to prove the fix through the verify gate instead of just claiming done.
Specific Prompt(s)
/vajra verify-work my-fix "npm test" "npm run build" then check that it captured the output and exited non-zero on any failure.
Additional Comments
MIT, solo-built and open source. The verify gate, OS sandbox, cost ledger, and replay trace are the point; the skills and agents are extras. Honest about limits in the README, including that it is Claude Code only and the sandbox is opt-in.
Recommendation Checklist
Display Name
Vajra
Category
Agent Skills
Sub-Category
Tooling: Orchestrators
Primary Link
https://github.com/thisizmsk-png/vajra
Author Name
Sai Krishna Madavarapu
Author Link
https://github.com/thisizmsk-png
License
MIT
Other License
No response
Description
Agent harness for Claude Code that wraps each run in a verify gate, where a run is not marked done until its tests and build pass. Adds an OS sandbox (Seatbelt on macOS, bubblewrap on Linux), a per-model cost ledger, a tamper-evident replay trace, persistent campaigns, a drift-resistant loop for long autonomous jobs, and a parallel-worktree fleet of 17 specialized agents. Bundles 87 skills and a hardened pre-tool hook that has been through five rounds of red teaming.
Validate Claims
Clone any repo that has a test suite, break one test on purpose, then run /vajra verify-work demo "npm test". The gate runs the tests, exits non-zero, and refuses to mark the run done. Fix the test and rerun to watch it pass. The verify-work script and its exit codes are in scripts/.
Specific Task(s)
Give Claude a small bug fix on a repo with a test suite and require it to prove the fix through the verify gate instead of just claiming done.
Specific Prompt(s)
/vajra verify-work my-fix "npm test" "npm run build" then check that it captured the output and exited non-zero on any failure.
Additional Comments
MIT, solo-built and open source. The verify gate, OS sandbox, cost ledger, and replay trace are the point; the skills and agents are extras. Honest about limits in the README, including that it is Claude Code only and the sandbox is opt-in.
Recommendation Checklist