v0.1.0-spike.2
Pre-releaseSpike pre-release v0.1.0-spike.2
Experimental runmark — local, deterministic workspace/script facts for AI-agent shell calls. Does not execute commands or decide allow/ask/deny.
No clone required. Schema and flags may change.
What's new since v0.1.0-spike.1
- Claude Code hook adapter:
runmark hook claudejoinsrunmark hook codex; both share one PreToolUse+Bash pipeline and install viainstall.sh --with-claude/--with-codex. - Context auto-injection: adapters read a bounded
package.json/Makefilefrom the event cwd, sonpm run build/makeexpand into real facts instead ofcontext_missing(disable withRUNMARK_HOOK_CONTEXT=0). - cd tracking:
cd sub && rm -rf .now resolves against the post-cd directory; subshells never leak changes; failure branches keep the pre-cd directory. - fd redirects:
2>,2>&1,&>,&>>parse correctly — descriptor duplication never fabricates a write fact. - Env substitution: caller-supplied env values resolve
$VARpaths, withcaller_contextprovenance and pre-substitution evidence kept traceable. - More command families: git (push/reset/clean/rm/checkout/restore), write family (tee/touch/truncate/ln/rmdir), sed, find, xargs, tar/zip/unzip, pip/cargo/yarn/bun/npx installs — all evidence-backed and gold-fixture locked.
- Honest unknowns: non-blocking unknowns now surface in RunmarkFacts; unsupported commands downgrade coverage instead of claiming completeness.
Install
macOS / Linux
curl -fsSL https://github.com/phaethix/runmark/releases/download/v0.1.0-spike.2/install.sh | bash
# optional: user-level Codex hook
curl -fsSL https://github.com/phaethix/runmark/releases/download/v0.1.0-spike.2/install.sh | bash -s -- --with-codex
# optional: user-level Claude Code hook
curl -fsSL https://github.com/phaethix/runmark/releases/download/v0.1.0-spike.2/install.sh | bash -s -- --with-claude
export PATH="$HOME/.local/bin:$PATH"
runmark versionWindows (PowerShell, CLI only)
irm https://github.com/phaethix/runmark/releases/download/v0.1.0-spike.2/install.ps1 | iex
runmark versionPreToolUse hooks on Windows are unreliable today (shell often does not fire hooks). Use analyze on Windows; use macOS/Linux for hook integration.
No-Codex user scenarios
curl -fsSL https://github.com/phaethix/runmark/releases/download/v0.1.0-spike.2/demo-user-scenarios.sh | bashWalks NL intents → agent Bash → naive guard vs runmark (with/without workspace snapshot). Does not need Codex.
Try CLI
runmark analyze 'echo hi > out.txt' --cwd logical://workspace --format text
cat > /tmp/rm-ctx.json <<'EOF'
{"cwd":"logical://workspace","files":{"package.json":"{\"scripts\":{\"build\":\"rm -rf dist\"}}"}}
EOF
runmark analyze 'npm run build' --context-file /tmp/rm-ctx.json --format textHook trial
After install.sh --with-codex or --with-claude: enable hooks in the client config if needed, then ask it to run a simple shell write. Expect PreToolUse + facts injected as additionalContext. Default behavior is context only (no deny/ask).
Assets
| File | Platform |
|---|---|
| runmark-darwin-arm64 / runmark-darwin-amd64 | macOS |
| runmark-linux-arm64 / runmark-linux-amd64 | Linux |
| runmark-windows-amd64.exe / runmark-windows-arm64.exe | Windows |
| install.sh / install.ps1 | installers |
| demo-user-scenarios.sh | demo |
| SHA256SUMS.txt | checksums |
Built from e6923f7. See README for status and non-goals.