chore: add Mythos bug-hunt pipeline + AGENTS.md restructure#86
Conversation
Scaffolds scripts/mythos/ with a four-prompt pipeline (rank, discover, validate, emit) plus portable HOWTO.md, modeled on Anthropic's Claude Mythos (April 2026). Adds a pre-release Mythos delta-pass step scoped by release type (tier-5 only for patches, tier-5+4 for minors, full tier-5 for majors/LTS). Migrates project guidance from CLAUDE.md to AGENTS.md (the tool-neutral canonical location read by Cursor, OpenCode, Claude Code, etc.) and slims CLAUDE.md to a thin pointer. Documents Kani's std::io OOM limitation so future finding reports treat CBMC OOM as "tool cannot reach" rather than "verification succeeded". Bumps examples/wasmtime-loader wasmtime 37 -> 43 to clear the April 2026 CVE patch floor (CVE-2026-34941, 34942, 34943, 35195, et al.). No runtime code change. Trace: skip Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Bumping to 43 breaks the example's `with_context` usage — wasmtime 43's `wasmtime::Error` no longer implements anyhow's `StdError`, so `.with_context(...)` from anyhow::Context doesn't apply. The earlier PR #82 that bumped the main lib to 43 left this example on 37 for this reason. Properly bumping the example requires a port (use wasmtime's own error helpers or `map_err` explicitly). That port is out of scope for this pipeline-scaffolding PR. Trace: skip Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Update — reverted the wasmtime-loader bumpCI ( Root cause: wasmtime 43's Reverted the example's Net effect: PR is now truly zero runtime code change — only |
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
Summary
scripts/mythos/with a four-prompt Mythos-style agentic bug-hunt pipeline (rank → discover → validate → emit) plus portableHOWTO.mdCLAUDE.mdtoAGENTS.md(tool-neutral canonical file); slimsCLAUDE.mdto a thin pointer mandatingAGENTS.mdexamples/wasmtime-loaderwasmtime 37 → 43 to clear the April 2026 CVE patch floorContext
The Mythos pipeline is modeled on Anthropic's Claude Mythos Preview (April 2026). Key discipline: every confirmed finding requires a failing PoC test plus a failing formal-verification harness (Kani/Verus/SMT). When a formal oracle cannot reach a property — e.g., Kani/CBMC OOMs on
std::iosymbolic execution — cite the nearest primitive-layer proof and document the limitation. This is spelled out inAGENTS.mdunder "Kani scope limitation" so future sessions treat CBMC OOM as "tool cannot reach" rather than "verification succeeded".A related PR will follow (PR 2) with the first Mythos finding applied to
signature/sig_sections.rs(silent swallow of cert_count parse errors → certificate-chain downgrade). That PR is decoupled from this one because this PR contains zero runtime code change.Test plan
rust.yml(should — no runtime code change)wasm-signing.ymlend-to-endfuzz.ymlpassesmemory.ymlpassesAGENTS.mdrenders correctly on GitHubscripts/mythos/HOWTO.mdrenders correctlyrivet init --agentsregeneration risk is understood — manually-maintained section inAGENTS.mdis marked; future rivet tooling should preserve it🤖 Generated with Claude Code