Propose a patch

Evidence

The sandbox observed nothing beyond the static findings; the proposal cites those.

diff --git a/claude-plugin/scripts/session-init.js b/claude-plugin/scripts/session-init.js
--- a/claude-plugin/scripts/session-init.js
+++ b/claude-plugin/scripts/session-init.js
@@ -57,1 +57,0 @@
-// hook used to write there — "CLAUDE.md was modified", "hooks look dark",
diff --git a/claude-plugin/scripts/adopt.js b/claude-plugin/scripts/adopt.js
--- a/claude-plugin/scripts/adopt.js
+++ b/claude-plugin/scripts/adopt.js
@@ -39,1 +39,0 @@
-// `followLink` is opt-in, and ONLY the CLAUDE.md read-modify-write passes it.
@@ -941,1 +940,0 @@
-'Edit CLAUDE.md manually to opt in.';
@@ -970,1 +968,0 @@
-else lines.push('[code-graph] CLAUDE.md block already up-to-date — no write');
@@ -1021,1 +1018,0 @@
-'  Refusing rather than guessing: both verbs write to your CLAUDE.md.\n');
diff --git a/claude-plugin/scripts/cli-entry.js b/claude-plugin/scripts/cli-entry.js
--- a/claude-plugin/scripts/cli-entry.js
+++ b/claude-plugin/scripts/cli-entry.js
@@ -18,1 +18,0 @@
-* they write `<cwd>/CLAUDE.md`, `<cwd>/.claude/` and `~/.cache/code-graph`, and

Zero-context hunks: apply with git apply --unidiff-zero fix.patch on the tagged source of 0.193.0. If the behavior is intended, declare it (host under permissions.network, a line in the README) instead of removing it; the next crawl scores declarations.

Issue / PR body

Title: code-graph: Instructs the agent to modify its persistent memory or instruction files (KillerSkills verdict caution 66)

## What the honeypot observed
KillerSkills ran **[email protected]** in an isolated sandbox (decoy credentials, recording proxy, syscall trace; engine 1.8.1) and scored it **CAUTION 66/100**. Summary: no network, touches nothing outside the project, writes no memory.

- See the findings below.

Full trace and findings: https://killerskills.ai/skills/npm/code-graph/0.193.0

## Findings with a location
- `claude-plugin/scripts/session-init.js:57` · Instructs the agent to modify its persistent memory or instruction files — claude-plugin/scripts/session-init.js:57 · Instructs the agent to modify its persistent memory or instruction files · informational: not an instruction file
- `claude-plugin/scripts/adopt.js:1021` · Instructs the agent to modify its persistent memory or instruction files — claude-plugin/scripts/adopt.js:1021 · Instructs the agent to modify its persistent memory or instruction files · informational: not an instruction file
- `claude-plugin/scripts/adopt.js:39` · Instructs the agent to modify its persistent memory or instruction files — claude-plugin/scripts/adopt.js:39 · Instructs the agent to modify its persistent memory or instruction files · informational: not an instruction file
- `claude-plugin/scripts/cli-entry.js:18` · Instructs the agent to modify its persistent memory or instruction files — claude-plugin/scripts/cli-entry.js:18 · Instructs the agent to modify its persistent memory or instruction files · informational: not an instruction file
- `claude-plugin/scripts/adopt.js:970` · Instructs the agent to modify its persistent memory or instruction files — claude-plugin/scripts/adopt.js:970 · Instructs the agent to modify its persistent memory or instruction files · informational: not an instruction file
- `claude-plugin/scripts/adopt.js:941` · Instructs the agent to modify its persistent memory or instruction files — claude-plugin/scripts/adopt.js:941 · Instructs the agent to modify its persistent memory or instruction files · informational: not an instruction file

## Proposed change
The attached patch removes the 6 lines above. If the behavior is intended, the fix is to declare it instead: list the host under `permissions.network` (or the MCP manifest) and say in the README why the package needs it; the next crawl will re-score it with the declaration taken into account.

## How this will be verified
Nothing is "verified safe" by assertion. When a new version is published, KillerSkills re-runs it in the honeypot and the verdict page updates; you can trigger it early from the verdict page ("Re-scan after I publish a fix"). Reply here or at [email protected] if you believe the verdict is wrong; disputes are reviewed by a person and the outcome is published on the verdict page.

_Proposed through killerskills.ai/threats by a user of the package, under their own GitHub account. The evidence is public; the package was never modified by KillerSkills._