Skip to content

Releases: snapsynapse/guidecheck

v0.6.0

10 Jun 19:59

Choose a tag to compare

GuideCheck 0.5.0

10 Jun 05:17

Choose a tag to compare

Audit-driven release. Highlights:

Security

  • package-registry JSON anchors bind the hash to assistant-guide-specific metadata instead of accepting the first sha256 field anywhere in the registry record
  • hosted verification enforces a five-fetch per-request budget with exact fetch deduplication, uses one deterministically selected unbranded content-variation probe, warns on off-domain recommended verifiers, and warns when package-registry assistant-guide URLs disagree with canonical-url

Added

  • version-sync check in make test across every version-bearing surface, including byte-identity of the published .well-known guide copy
  • exact warning pinning on all 68 local-file fixtures (warnings_exact / forbidden_warning_ids), so false-positive warnings fail tests
  • deterministic anchor-channel tests for dns-txt, repository-file, signed-security-txt, and transparency-log
  • first standalone conformance-kit artifact (fixtures, schemas, finding-ids, verifier-conformance) for independent verifier implementations

Changed

  • finding-ids.md is normative for the finding-id registry
  • ADOPTION reframes MCP/A2A as ecosystem integrations of the core profile
  • INTENT records the Level 5 ownership decision (GuideCheck owns the runtime fixture suite and evaluator, gated by pre-level-5 readiness)

Fixed

  • published docs/.well-known/assistant-guide.txt resynced with the repository guide; the new version-sync check makes this drift a test failure

Full changelog: https://github.com/snapsynapse/guidecheck/blob/main/CHANGELOG.md

SHA-256 of the artifacts is in the attached SHA256SUMS files.

GuideCheck 0.4.0

29 May 21:32

Choose a tag to compare

GuideCheck 0.4.0

A security hardening release. It supersedes 0.3.2 (which was version-bumped but
never tagged or published) and covers everything since 0.3.1.

Security

  • detector negation handling rewritten: a negation now suppresses a prohibited
    or encoded-execution pattern only when it directly governs that pattern, so
    inserting do not elsewhere on the line no longer disables the detector
  • marker discipline: action and metadata fences that differ only by surrounding
    whitespace or letter case (for example [ACTION]) are no longer silently
    dropped; they raise a blocking malformed finding so a verifier and a lenient
    agent parser cannot diverge on which actions exist
  • command and class consistency: a command is cross-checked against its declared
    class by command-head analysis; a network fetch piped into an interpreter
    blocks as command.fetch-execute, and under-declared network or code-executing
    commands raise warnings
  • registry anchors: a registry-url is counted toward Level 4 only when its host
    is a recognized independent registry, closing a self-hosted-anchor path

Added

  • finding ids command.fetch-execute, network.command-implies-networked,
    approval.command-implies-required, anchor.registry.unrecognized-host, and
    level4.requires-fetch
  • verifier-conformance.md sections for marker discipline and command/class
    consistency
  • a recorded adversarial review in threat-register.md

Changed

  • the local-file reference verifier now caps the achieved level at Level 3. It
    still checks supplied manifest and anchor evidence for consistency and reports
    level4.requires-fetch, but Level 4 (independent provenance) is assertable
    only by the fetching hosted verifier, matching verifier-conformance.md
    section 6
  • the eval runner imports the primary engine instead of carrying a second copy
    of the checks, so there is one source of truth
  • profile, verifier, hosted verifier, spec, verifier-conformance, examples, and
    public pages now report 0.4.0; guide verifier-conformance ranges move to
    >=0.4.0, <0.5.0 and applies-to to guidecheck 0.4.x
  • the published docs/.well-known/assistant-guide.txt is resynced byte-for-byte
    with the repository assistant-guide.txt (it had drifted at 0.3.1)

Verification

  • make test: 130 eval cases, 66 reference fixtures, 76 contract fixtures,
    6 guide-artifact byte profiles, and the parser-edge, hosted-API, fetch-replay,
    fetch-safety, and CLI-contract suites all pass.
  • Detector changes were re-attacked by an independent adversarial pass before
    release; verified bypasses and false positives were fixed and regression-tested.
  • docs/.well-known/assistant-guide.txt confirmed byte-identical to the
    repository assistant-guide.txt at 0.4.0.

Residual risks

Recorded in threat-register.md (Adversarial review 2026-05-29): heuristic
command analysis is best-effort, verifier fingerprinting is not fully defeated,
the hosted checker is a modest request-proxy amplifier, anchor extraction can
pick a decoy hash, and the version string is still duplicated across files.

GuideCheck 0.3.1

27 May 01:38

Choose a tag to compare

Patch release for GuideCheck positioning and ecosystem integration documentation. Adds MCP/A2A integration notes, a database MCP server Level 3 example guide, homepage trust-boundary positioning, and updates first-party profile/verifier version strings to 0.3.1. Conformance semantics remain unchanged.

GuideCheck 0.3.0

24 May 18:42

Choose a tag to compare

GuideCheck 0.3.0 adds Level 4 verifier coverage, Level 5 readiness reporting, hosted public-web hardening warnings, clearer guide-score messaging, and Level 4 package-registry adoption examples. See CHANGELOG.md for details.

GuideCheck v0.2.0

22 May 05:01

Choose a tag to compare

Human-Verifiable Assistant Guide profile 0.2.0 and verifier-conformance profile 0.2.0.

This release resolves all open questions from the spec, the verifier-conformance profile, and the roadmap, restructures the spec into an adoption-first document set, and bumps the version to reflect constraint tightenings.

Added

  • ADOPTION.md — practical on-ramp: conformance ladder, level-by-level path, guide-author checklist.
  • operator-guide.md — non-normative defense-in-depth practices.
  • Public append-only transparency log as an independent cross-channel provenance anchor.
  • Reading guide and linked contents at the top of spec.md.

Changed

  • code-executing actions now require explicit approval at Level 3.
  • Guide copies served at both the well-known path and the repository root must be byte-identical.
  • repository-url is defined as the source repository root.
  • Staleness keys off the publisher's valid-until; no arbitrary last-reviewed threshold.
  • Reference verifier and eval harness updated to match.

Removed

  • The Open Questions sections from spec.md and verifier-conformance.md; resolved and future items now live in roadmap.md.

Full detail in CHANGELOG.md. Conformance is not safety: a verifier confirms form, the human confirms meaning.

v0.1.0

22 May 03:59

Choose a tag to compare

Initial draft for review.

Specification

  • Human-Verifiable Assistant Guide profile for assistant-guide.txt
  • core artifact, one-artifact bounded-task scope, canonical well-known path
  • strict ASCII byte profile, 8 KiB size cap, 120-byte line and 400-line limits
  • disallowed constructs and Markdown-as-text clarification
  • required sections at Level 3, compact verification instruction at Level 1, assistant invocation prompt content
  • guide metadata block with normative fences, version-range syntax, and field set
  • sidecar manifest provenance model and cross-channel hash publication (DNS TXT, package registry, public repository file, signed security.txt)
  • action classification with seven classes including code-executing, structured [action] blocks, command field restrictions, runner semantics
  • stop-and-ask conditions and canonical approval phrasing
  • threat model, untrusted content handling, integrity-versus-instruction fetch distinction, hard ban on chained guides
  • public information safety and risky pattern guidance
  • five-level conformance ladder including Level 5 runtime-enforced execution
  • discovery surfaces, HTTPS serving requirements, verifier requirements, verifier output schema
  • residual threats and operator defense-in-depth checklist
  • locale handling and final ASCII-only position

Companion documents

  • Verifier Conformance Profile defining public-web and local-file evaluation modes, fetch safety, SSRF defenses, level calculation, output schema, and fixture suite conformance
  • design rationale capturing the reasoning behind the 8 KiB cap, ASCII-only profile, sidecar manifest, cross-channel publication, hard chained-guide ban, and other decisions
  • threat register enumerating network, hosting, provenance, verifier, runtime, user, and availability risk classes
  • JSON Schema for the manifest and verifier output

Project

  • designated standard primary verifier at https://guidecheck.org/verify
  • canonical site at https://guidecheck.org/