chain

Attack-path chaining

Findings are triaged independently; the highest-impact issues are often a path assembled from individually-unremarkable bugs (a low info-leak + a medium auth gap + a candidate SSRF ⇒ critical RCE). This searches for those paths — it doesn't just restate confirmed findings. Requires ≥2 live findings in .kuzushi/findings.json (run the hunts first); richer with a threat model + code-graph built (they supply the assets + reachability the search keys off).

1. Run node "${CLAUDE_PLUGIN_ROOT}/scripts/cmd/chain-prepare.mjs" --target "<repo root>". If there are fewer than 2 live findings, tell the user to run more hunts and stop. Read prepPath — it carries the findings (incl. sub-threshold leads), the crown-jewel context.assets, the attacker-reachable context.entryPoints, and a context.reachability summary.
2. Search entry→asset paths (forward from entry points, backward from assets, meet in the middle) where each step is enabled by a finding; compose sub-threshold primitives where the composition is what makes it critical. Order members, name the data/control link between steps, and use each fingerprint verbatim. Write { chains: [...] } (with kind/entryPoint/asset) to draftPath. If nothing genuinely composes, write { "chains": [] } — don't force a chain.
3. Run the assembleCommand (finalize). It validates each chain (≥2 real members, ordered narrative ≥120 chars), escalates severity to at least the max member's (composed impact is never under-reported), writes .kuzushi/chains.json, and attaches a chains ref onto each member (status unchanged). Report the paths highest-impact first; /report renders them.

When NOT to use

• Before ≥2 live findings exist — run /sweep / /threat-hunt / /taint-analysis first.
• To discover new bugs — chaining only connects existing findings; name missing links as gaps.

Rationalizations to Reject

• "All low/candidate — not worth chaining." → Composing sub-threshold primitives into a critical is the whole point; severity-gating the links misses the chains that matter most.
• "Same bug class, so they chain." → A link needs one finding's effect to satisfy the next's precondition — name that edge or drop it. Co-occurrence ≠ composition.