This release fixes issues for SREs watching stability and regressions.
✓ No known CVEs patched in this version
Topics
+13 more
Affected surfaces
Summary
AI summaryFixed false budget_exit from misclassified no-spend proof runs.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Security | Low |
Root tarball no longer exposes vendored internal CLI bin surface inside dist/vendor/cli/package.json. Root tarball no longer exposes vendored internal CLI bin surface inside dist/vendor/cli/package.json. Source: llm_adapter@2026-06-06 Confidence: high |
— |
| Bugfix | Medium |
No-spend proof runs now complete honestly without misclassifying verifier-only pass as no_code_change. No-spend proof runs now complete honestly without misclassifying verifier-only pass as no_code_change. Source: llm_adapter@2026-06-06 Confidence: high |
— |
| Bugfix | Medium |
Portable Windows Codex execution correctly handles npm-style .cmd shims. Portable Windows Codex execution correctly handles npm-style .cmd shims. Source: llm_adapter@2026-06-06 Confidence: high |
— |
| Bugfix | Medium |
Hosted OpenAI defaults target the hosted endpoint without needing MARTIN_OPENAI_BASE_URL. Hosted OpenAI defaults target the hosted endpoint without needing MARTIN_OPENAI_BASE_URL. Source: llm_adapter@2026-06-06 Confidence: high |
— |
| Bugfix | Medium |
Claude live runs honor provider permissions and no longer force‑add --dangerously-skip-permissions. Claude live runs honor provider permissions and no longer force‑add --dangerously-skip-permissions. Source: llm_adapter@2026-06-06 Confidence: low |
— |
| Bugfix | Low |
Claude live runs respect provider permissions and stop forcing `--dangerously-skip-permissions`. Claude live runs respect provider permissions and stop forcing `--dangerously-skip-permissions`. Source: granite4.1:30b@2026-06-06-audit Confidence: low |
— |
| Refactor | Low |
Public CLI guidance consistently uses martin-loop across help, tour, onboarding, receipts, and MCP config hints. Public CLI guidance consistently uses martin-loop across help, tour, onboarding, receipts, and MCP config hints. Source: llm_adapter@2026-06-06 Confidence: high |
— |
Full changelog
Fixed
- No-spend proof runs now complete honestly -
martin-loop run ... --proofno longer misclassifies a clean verifier-only pass asno_code_change, which previously cascaded into a falsebudget_exit. - Portable Windows Codex execution - Windows command resolution now handles npm-style
.cmdshims correctly, including realcodex.cmdand verifier commands launched through npm wrappers. - Hosted OpenAI defaults -
--engine openainow targets the hosted OpenAI endpoint by default, so local proofs do not requireMARTIN_OPENAI_BASE_URLjust to avoid a localhost misroute. - Claude live runs honor provider permissions again - the public Claude adapter no longer force-adds
--dangerously-skip-permissionson live governed runs.
Changed
- Public CLI guidance now consistently uses
martin-loopin help, tour, onboarding, workflow receipts, and MCP config hints. - The public root tarball no longer exposes the vendored internal CLI bin surface inside
dist/vendor/cli/package.json.
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About Keesan12/Martin-Loop
All releases →Related context
Related tools
Earlier breaking changes
- vmcp-v0.1.3 martin_status uses oneOf for selector exclusivity, latest as const.
- vmcp-v0.1.3 maxIterations and maxTokens modeled as integers in tool schemas.
- vmcp-v0.1.3 Tool schemas enforce additionalProperties: false on public contracts.
- vmcp-v0.1.3 Packaged artifacts now require and ship server.json alongside package.json.
Beta — feedback welcome: [email protected]