Vajra — Development skill for Claude Code
Agent harness for Claude Code: verify-gate, OS sandbox, fleet mode, cost ledger, replay trace.
How to install Vajra
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open thisizmsk-png/vajra and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Vajra does
Agent harness for Claude Code: verify-gate, OS sandbox, fleet mode, cost ledger, replay trace.
Alternatives in Development
- Verify Theme — Run the full Graphite quality gate (drift, tokens, contrast, YAML) and report pass/fail 452 ★
- Exception — Signed gate-exception registry — replace ad-hoc --admin / --no-verify bypasses with an auditable, signed, expi 83 ★
- Skill Eval Harness — Agent Skill evaluation harness for paired variants, trace artifacts, and runner adapters 77 ★
README
Turn Claude Code into a harness you can run unattended.
Every run is verify-gated (done means the tests passed), OS-sandboxed, and replayable, with a per-run cost ledger. Not another skill pack. A harness that verifies its own work.
[](LICENSE) [](https://docs.claude.com/en/docs/claude-code) [](#install-in-30-seconds) [](SECURITY.md) [](https://github.com/thisizmsk-png/vajra)
One real Claude Code session, sped up: a spec, a plan you trim, code, the tests proven green, the cost of the run, the replay trace, and a parallel fleet.
What you get on every run
**VerifyGate: done = tests passed** · OS-sandboxed (Seatbelt / bubblewrap) · replayable runs · per-run cost ledger · red-teamed x5 · 17-agent fleet · 87 skills
VerifyGate is the point. It runs your tests and build, captures the real output, and refuses to call a run done until they pass. Checks fail, the run is not done, and in CI it exits non-zero so a broken agent change cannot merge. Everything else (the routing, the sandbox, the ledger, the replay trace, the fleet) exists to make that one line trustworthy.
/vajra verify-work my-change "npm test" "npm run build"
# exit 0 all checks passed, walkthrough capturedRelated Skills
Maestro Preset
Multi-agent orchestration preset for the DeepSeek Harness (DSH): cross-plane supervisor over Orca / dais (ex-z
Evidence Verify Gate
You are auditing a proposed skill edit for cluster 0. The edit was inferred from a cluster of FAILING agent tr
Hindcast
Browse, search & replay every Claude Code session on your Mac. Local-first session viewer, timeline scrubber &
Fux Verify
Replay a fux receipt and report whether the answer still reproduces.
Validation Gating
Turn proposed harness changes into candidate-first changes with explicit replay, regression, and rollback expe
Secure Delivery
The ship gate for anything a customer or another machine will run — secrets, egress, lethal trifecta, taint, s
Related Agents
Autopilot Implementer
TDD implementer for autopilot's fast/cost-efficient mode. Implements ONE PLAN.md work package test-first, runs
Protocol Replay Adversary
Use this agent for adversarial analysis of protocol-state attacks — lease replay across windows, nonce-ledger
Iteration Guard
The verify gate of the Logic-Lens iteration loop. Given a baseline iteration and a candidate iteration, compar
