chore(preflight): bump skill integrity pin for v4 + document new signing defenses - #316
Merged
Conversation
…iant) The vaultpilot-preflight skill (separate repo, lives under ~/.claude/skills/) gains Invariant #8 in v4 — agent-side rules for free-form message signing (sign_message_btc / sign_message_ltc and the contacts-CRUD message signers). v4 also corrects Invariant #7's overreach: v3 listed contacts CRUD as the *only* legitimate trigger of an on-device sign-MESSAGE prompt, which would have caused agents to false-positive refuse on legitimate sign_message_btc/ltc calls. v4 carves both pre-existing tools into the legitimate-trigger list. A one-liner under Invariant #1 also clarifies that Safe propose / approve / execute are standard EVM transactions (eth_sendTransaction of approveHash / execTransaction), not typed-data signing — so they need no special agent handling. The skill content change shifts both the SHA-256 of SKILL.md and the in-file integrity sentinel. Update the pin in this MCP's source so the on-each-flow integrity check the MCP delegates to the agent succeeds against the new skill version. - SHA-256: 21c8c60ac26528732dbe0b40b4be7e4790607db1881489fe4cd1751f75536cd6 (old, v3) → cd689838314a700dfff80d4c881bf51190cd6c71c747f3152e7cab8d943df2cc (new, v4) - Sentinel fragment B: _v3_ → _v4_ - Sentinel fragment C: 2d3b876b38550fe5 → 7655818578c7a044 Coordinated with the matching PR in szhygulin/vaultpilot-security-skill#5. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The address-book PR (#284) and the pre-existing sign_message_btc / sign_message_ltc tools together expand the signing surface beyond sign-transaction flows. SECURITY.md previously covered only the contacts blob signing trade-off (path C) and did not describe the free-form public sign-message tools at all. This commit: - Updates the contacts "EVM signing trade-off" paragraph to point at skill v4 Invariant #7 as the agent-side defense (decoration, tamper warning, sign-message-during-prepare anomaly check, post- add_contact verify cross-check), correctly attributing the contacts JSON preimage as NOT subject to character-by-character matching (its user-side defense is the recognizable domain prefix + JSON shape, per the skill's own carve-out). - Adds a new "Free-form message signing" section describing the public sign_message_btc / sign_message_ltc tools, the substitution threat, and skill v4 Invariant #8's MESSAGE-PREVIEW block as the agent-side defense (the user verifies the previewed string against the device screen character-by-character). - Notes that EVM has no equivalent public free-form signer (personal_sign is restricted to the contacts signer) and eth_signTypedData_v4 remains excluded from the WC namespace. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
szhygulin
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
The vaultpilot-preflight skill (separate repo, lives under
~/.claude/skills/) gains Invariant #8 in v4 covering free-form message signing, and corrects an Invariant #7 overreach from v3. This PR carries the matching SHA-256 + sentinel pin bump insrc/index.tsand the correspondingSECURITY.mdupdate describing the new agent-side defenses.Coordinated with szhygulin/vaultpilot-skill#5.
What changed in the skill (v3 → v4)
Invariant Release v0.2.0 #8 (NEW) — free-form message-signing discipline. Address-book v1.0 (feat(contacts): address book v1.0 — BTC + EVM #284) introduced contacts-CRUD message signing; pre-existing
sign_message_btc/sign_message_ltctools also produce BIP-137 sigs over arbitrary UTF-8. v4 specifies the agent-side preview discipline for the free-form tools: render an explicitMESSAGE-PREVIEWblock with the exact UTF-8 string before the device prompt, user matches character-by-character, treat any deviation as✗ MESSAGE-PREVIEW MISMATCH — DO NOT SIGN.(Contacts CRUD signs a structurally-fixed JSON preimage that is NOT subject to char-by-char matching — its user-side defense is the recognizable domain prefix.)Invariant Fix Compound dust rendering and add tx simulation #7 carve-out fix. v3 said the contacts-CRUD tools are the only VaultPilot tools that legitimately trigger an on-device sign-MESSAGE prompt. That was wrong —
sign_message_btcandsign_message_ltc(pre-existing public MCP tools) also legitimately trigger sign-message prompts. v3 would have caused agents to false-positive refuse on legitimate calls. v4 adds both to the legitimate-trigger list.Invariant Add CI workflow: typecheck + tests on Node 20/22 #1 (Safe note). One-liner clarifying Safe propose / approve / execute are standard EVM transactions (
eth_sendTransactionofapproveHash/execTransaction), not typed-data signing — so they need no special agent handling. Reinforces the existing decision to keepeth_signTypedData_v4excluded from the WC namespace.Pin changes in this PR
21c8c60ac26528732dbe0b40b4be7e4790607db1881489fe4cd1751f75536cd6cd689838314a700dfff80d4c881bf51190cd6c71c747f3152e7cab8d943df2cc_v3_→_v4_2d3b876b38550fe5→7655818578c7a044The fragmented-pin format (A + B + C concatenation) keeps the literal sentinel string out of agent context — see the explanatory comment in
src/index.tsat the pin site.SECURITY.md changes
add_contactverify cross-check), explicitly noting the contacts JSON preimage is NOT subject to character-by-character matching.sign_message_btc/sign_message_ltc)" section: describes the public sign-message surface, the substitution threat (concrete: user asks "I own bc1q...mine", MCP forwards "I authorize transfer of all funds to bc1q...attacker"), and skill v4 Invariant Release v0.2.0 #8'sMESSAGE-PREVIEWblock as the agent-side defense.personal_signrestricted to the contacts signer) andeth_signTypedData_v4remains excluded from the WC namespace.Test plan
npm install && npm run buildclean in worktreesrc/index.ts(grep)~/.claude/skills/vaultpilot-preflight/SKILL.mdmatches new pin🤖 Generated with Claude Code