chore(deps): update entireio/skills digest to fe5266f

[renovate]

This PR contains the following updates:

Package	Update	Change
entireio/skills	digest	2f9a875 → fe5266f

---

Configuration

📅 Schedule: (UTC)

• Branch creation
  • Between 12:00 AM and 03:59 AM, only on Monday (* 0-3 * * 1)
• Automerge
  • At any time (no schedule defined)

🚦 Automerge: Disabled by config. Please merge this manually once you are satisfied.

♻ Rebasing: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox.

🔕 Ignore: Close this PR and you won't be reminded about this update again.

---

• If you want to rebase/retry this PR, check this box

---

This PR was generated by Mend Renovate. View the repository job log.

[toolhive-release-app]

🛡️ Skill Security Scan Results

✅ entire-explain

• Status: Passed
• Findings: 2
• Allowed (not blocking): 1
  • MANIFEST_MISSING_LICENSE (Allowed: entireio/skills is MIT-licensed at the repo root; upstream does not embed an SPDX license identifier in per-skill SKILL.md frontmatter.)

❌ entire-search

• Status: Failed
• Findings: 5
• Blocking: 3

Blocking issues:

• [LLM_COMMAND_INJECTION] (CRITICAL) The documented execution pattern inserts a query into a double-quoted shell command. If an agent constructs and executes this command through a shell without robust argument handling, user-controlled query text can terminate the quote or trigger shell expansions such as command substitution. This can turn a search request into arbitrary command execution in the agent environment. (SKILL.md)
• [LLM_DATA_EXFILTRATION] (HIGH) The skill searches recorded prompts, transcripts, decisions, and code across repositories, then permits retrieval of an entire session transcript with --full. It also instructs the agent to widen to all repositories whenever an answer might exist elsewhere. These data sources can contain sensitive discussions, code, identifiers, or secrets. The instructions do not require confirmation of repository scope, minimization of transcript retrieval, redaction, or verification that the requesting user is entitled to receive results from every accessible repository. (SKILL.md)
• [LLM_UNAUTHORIZED_TOOL_USE] (HIGH) The skill directs the agent to run every entire command with sandbox_permissions: "require_escalated" on the first attempt in order to access credentials outside the sandbox. This unnecessarily broad elevation applies not only to authentication but also to searches and checkpoint-transcript retrieval, expanding the skill's access to host-managed credentials and data without a per-command necessity or user confirmation. (SKILL.md)

Allowlisted (not blocking):

• MANIFEST_MISSING_LICENSE (Allowed: entireio/skills is MIT-licensed at the repo root; upstream does not embed an SPDX license identifier in per-skill SKILL.md frontmatter.)

❌ entire-session-handoff

• Status: Failed
• Findings: 4
• Blocking: 3

Blocking issues:

• [LLM_COMMAND_INJECTION] (CRITICAL) The checkpoint flow instructs the agent to substitute a user-provided checkpoint ID directly into shell commands and output paths without requiring validation, quoting, or argument-vector execution. If an agent executes these templates through a shell, shell metacharacters in the checkpoint ID can terminate the intended command and execute arbitrary commands in the agent environment. (SKILL.md)
• [LLM_DATA_EXFILTRATION] (HIGH) When worktree matching produces no results, the skill explicitly falls back to the full session list and selects the most recently active session. It then retrieves that session's full transcript and presents a summary to the current user. This can cross worktree, project, or user-context boundaries and expose task details, source-code context, user prompts, tool outputs, or secrets from an unrelated session. The behavior is automatic for bare invocation and does not require the user to identify the intended session. (SKILL.md)
• [LLM_PROMPT_INJECTION] (HIGH) The skill reads raw transcripts containing user, assistant, tool, and function-call content, then directs the agent to derive the next steps from that material and immediately continue work. Session transcripts are untrusted data: they may contain malicious instructions, prior prompt injections, or text impersonating authoritative instructions. The skill does not require treating transcript content as data only, separating factual state from instructions, or obtaining confirmation before executing actions inferred from it. (SKILL.md)

❌ entire-session-to-skill

• Status: Failed
• Findings: 3
• Blocking: 2

Blocking issues:

• [LLM_COMMAND_INJECTION] (CRITICAL) The workflow directs the agent to construct Bash commands using user-derived workflow terms, user-supplied session/checkpoint identifiers, and transcript paths obtained from repository metadata. Several templates interpolate these values directly into shell syntax. In particular, workflow terms are placed inside double quotes and transcript paths are unquoted. A quote, command substitution, shell metacharacter, leading dash, or other crafted value could alter command execution when an agent executes the template through Bash. (SKILL.md)
• [LLM_PROMPT_INJECTION] (HIGH) The skill directs the agent to retrieve and inspect Entire session histories, checkpoint explanations, raw transcripts, and metadata, then derive a future skill from that material. These sources can contain untrusted user messages, prior agent outputs, tool output, or attacker-controlled repository content. The instructions do not require the agent to treat retrieved transcript content as data only, ignore embedded directives, or prevent transcript text from changing the current task. A malicious historical session could therefore inject instructions such as requests to disclose data, execute commands, alter the generated skill, or bypass approval requirements. (SKILL.md)

❌ entire-what-happened

• Status: Failed
• Findings: 5
• Blocking: 1

Blocking issues:

• [LLM_COMMAND_INJECTION] (HIGH) The workflow directs the agent to place user-supplied values directly into Bash command text for grep and git blame. Although the examples use double quotes and --, double quotes do not prevent shell command substitution or expansion when an agent constructs a shell command string: a path containing $(...) or backticks may execute commands before grep or git receives the intended argument. The same risk applies to the user-selected distinctive snippet line in the grep template. (SKILL.md)

Allowlisted (not blocking):

• ATR_2026_00066 (Allowed: FP: matched the documented grep and git blame end-of-options syntax (-- ") at SKILL.md:74 and SKILL.md:98; these examples do not execute or instruct arbitrary commands.)
• ATR_2026_00066 (Allowed: FP: matched the documented grep and git blame end-of-options syntax (-- ") at SKILL.md:74 and SKILL.md:98; these examples do not execute or instruct arbitrary commands.)

---

Summary: Completed 5 of 5 skill scan(s), found 9 blocking issue(s).

⚠️ Action Required: Review the failed scan jobs and any blocking findings. Add a justified entry to the skill's security.allowed_issues[] only when a finding is a confirmed false positive.

[renovate]

Edited/Blocked Notification

Renovate will not automatically rebase this PR, because it does not recognize the last commit author and assumes somebody else may have edited the PR.

You can manually request rebase by checking the rebase/retry box above.

⚠️ Warning: custom changes will be lost.

[danbarr]

Security triage:

Allowed:

• MANIFEST_MISSING_LICENSE (existing): the pinned entireio/skills repository has an MIT root license; individual SKILL.md files omit SPDX front matter.
• ATR_2026_00066 in entire-what-happened (new): both hits are the documented -- " end-of-options syntax at SKILL.md:74 and :98, not executable attack payloads.

Still blocking (no exclusion):

• entire-search: LLM_COMMAND_INJECTION, LLM_DATA_EXFILTRATION, LLM_UNAUTHORIZED_TOOL_USE — unsafe shell interpolation, unrestricted history/transcript retrieval, and blanket credential elevation are material risks.
• entire-session-handoff: LLM_COMMAND_INJECTION, LLM_DATA_EXFILTRATION, LLM_PROMPT_INJECTION — unquoted checkpoint use, automatic cross-context transcript fallback, and untrusted transcript continuation lack mitigations.
• entire-session-to-skill: LLM_COMMAND_INJECTION, LLM_PROMPT_INJECTION — user-derived shell interpolation and untrusted transcript-derived instructions remain unsafe.
• entire-what-happened: LLM_COMMAND_INJECTION — quoted shell templates still permit expansion when constructed through a shell.

Validated entire-what-happened against pinned commit fe5266f; reprocessing the trusted raw scan now allows only the two ATR_2026_00066 occurrences and leaves the material command-injection finding blocking.