evolveguard

Catch behavioral drift when a Claude Agent Skill edits itself, before the edit ships.

# PyPI -- Python CLI + library (genuine port)
pip install evolveguard

# npm -- JavaScript/TypeScript CLI + library
npm install -g evolveguard

Neither package is published yet, for two separate, unrelated reasons -- see Status below for the full detail:

• npm is blocked by an account-level constraint: this account's npm 2FA is security-key/passkey-only with no authenticator-app or OTP fallback configured, which blocks the npm publish step independent of the code itself.
• PyPI is blocked by PyPI's own new-project-creation anti-abuse throttle (429 Too many new projects created) on this account, confirmed across repeated upload attempts -- also not a code issue.

Until either clears, clone the repo: npm run build && npm link for the TypeScript CLI, or cd python && pip install -e . for the Python package. See python/README.md for the Python-specific walkthrough.

Claude Code's Agent Skills can be authored by a human, or by an agent itself: /skillify turns a workflow into a SKILL.md, and the auto-memory system in this same environment writes MEMORY.md files that quietly change what an agent does in its next session. None of that gets a regression check today. A skill edit that breaks a working workflow looks exactly like a skill edit that fixes one, until someone notices the agent stopped doing something it used to do.

evolveguard records a baseline of a skill's own capability surface (what tools it's declared or shown to use), then re-derives that surface every time the skill file changes and diffs the result against the baseline: same tool-call sequence, or a flagged drift with a specific reason.

What it does

evolveguard record ./SKILL.md --fixtures ./fixtures.json
# ... skill gets edited, by a human or an agent ...
evolveguard check ./SKILL.md

EvolveGuard v0.1.0 -- Regression Check
skill: monorepo-scanner  baseline: 2026-07-15  fixtures: 1

[DRIFT] fixture: "scan a monorepo"  new tool call: fs.write (baseline had none)
         -> new tool call: fs.write (baseline had none) -- this edit introduces a
            capability the baseline never used

0 PASS, 1 DRIFT, 0 FAIL
exit code 1 (DRIFT blocks merge by default; override with --allow-drift)

That's real output from this repo's own fixtures/labeled-non-breaking-edits/case-03-add-write-capability/ test case, wired to filesystem: read-only -> read-write in the skill's frontmatter.

How it works

evolveguard does not run a live LLM agent, and it does not replay a real conversation transcript. It is a static, deterministic tool by design:

1. record parses a skill file's YAML frontmatter (declared tools, network, filesystem, scope, and any bundled hooks), scans the skill's body text and any hook scripts for static evidence of network calls or filesystem writes, and combines both into a capability surface -- the set of tools the skill is declared or shown to use. Each fixture's expectedToolCalls filters that surface down to the tools the fixture author says it cares about; the result is the recorded baseline.
2. check re-reads the (possibly edited) skill file, re-derives its capability surface with the exact same logic, and re-filters it per fixture.
3. diff compares baseline vs. current per fixture (PASS if the tool set and scopes match, DRIFT with a specific reason otherwise), and separately diffs the whole capability surface so a new capability that no fixture's expectedToolCalls happened to cover still gets caught, instead of silently passing.

evolveguard detects changes in what a skill is declared or shown to be capable of. It can't tell you whether a live agent run would actually behave differently on a given prompt -- that's a real, intentional scope limit. The tradeoff: it just needs a SKILL.md file and a fixtures file, with nothing hosted and no SDK to integrate against, which is also why it runs fully offline in a pre-commit hook or CI job.

How it's different

Braintrust is a general LLM eval and observability platform. It is a strong choice if you're already logging traces from a live agent and want statistical eval scoring across runs, but it needs SDK integration and an eval-definition step per app. evolveguard needs neither: point it at one SKILL.md file and a fixtures JSON, and it works.

agent-eval (this same author's other repo) answers a different, more general question: "did my agent's behavior change between two versions I define," for any agent, framework-agnostic, requiring you to stand up and run both versions yourself. evolveguard answers a narrower, structurally different question, triggered directly by a file diff on SKILL.md/MEMORY.md: did this specific edit change the capability surface the baseline recorded. It parses the skill artifact itself and never asks you to define or run anything live.

| | evolveguard | Braintrust | agent-eval | | -------------- | -------------------------------------------- | ---------------------------------- | --------------------------------- | | Setup | record + check against one file | SDK integration, eval definitions | Define and run two agent versions | | Trigger | SKILL.md/MEMORY.md file diff | Manual eval run | Manual A/B run | | Mechanism | Static capability-surface diff | Live-run trace scoring | Statistical behavior comparison | | Hosted infra | None | Hosted platform | None | | Live LLM calls | None | Yes (scores real runs) | Yes (runs both versions) | | Best for | Self-edited Claude Agent Skills specifically | General LLM app eval/observability | Any agent, generic A/B regression |

Quickstart

# 1. Record a baseline against a skill and its labeled fixtures
evolveguard record ./skills/my-skill/SKILL.md --fixtures ./fixtures/my-skill.json
# writes ./skills/my-skill/.evolveguard-baseline.json

# 2. Edit the skill (by hand, or let an agent edit it)

# 3. Check for drift
evolveguard check ./skills/my-skill/SKILL.md
# writes ./evolveguard-report.json, exits 1 if drift was found

A fixtures file is a JSON array of labeled prompts and the tool-call shapes each one is expected to touch:

[
  {
    "id": "scan-a-monorepo",
    "prompt": "scan a monorepo",
    "expectedToolCalls": [{ "tool": "fs.read" }, { "tool": "fs.write" }]
  }
]

expectedToolCalls is optional -- omit it and the fixture is treated as exercising the skill's entire capability surface. scopeMatches (a glob) narrows a tool to a specific filesystem scope, e.g. { "tool": "fs.write", "scopeMatches": "./workspace/**" }.

CLI command reference

Generated from the actual --help output of the built CLI.

Usage: evolveguard [options] [command]

Regression-testing CLI for self-edited Claude Agent Skills (SKILL.md,
MEMORY.md) -- golden-transcript record/replay against a skill's own declared
and inferred capability surface, zero hosted infrastructure.

Options:
  -V, --version                  output the version number
  -h, --help                     display help for command

Commands:
  record [options] <skillPath>   Record a golden-transcript baseline for a
                                 skill against a set of labeled fixtures
  check [options] <skillPath>    Replay the fixtures from a baseline against
                                 the current (possibly edited) skill and report
                                 drift
  report [options] [reportPath]  Print a previously generated
                                 evolveguard-report.json
  mcp                            [coming soon] Expose record/check/report as
                                 MCP tools for a coding agent to call
                                 mid-session
  help [command]                 display help for command

Usage: evolveguard record [options] <skillPath>

Record a golden-transcript baseline for a skill against a set of labeled
fixtures

Arguments:
  skillPath          path to the SKILL.md or MEMORY.md file to baseline

Options:
  --fixtures <path>  path to a fixtures JSON file (array of {id, prompt,
                     expectedToolCalls?})
  --baseline <path>  path to write the baseline file (default:
                     <skill-dir>/.evolveguard-baseline.json)
  --json             output structured JSON instead of human-readable text
                     (default: false)
  -h, --help         display help for command

Usage: evolveguard check [options] <skillPath>

Replay the fixtures from a baseline against the current (possibly edited) skill
and report drift

Arguments:
  skillPath          path to the SKILL.md or MEMORY.md file to check

Options:
  --baseline <path>  path to the baseline file (default:
                     <skill-dir>/.evolveguard-baseline.json)
  --report <path>    path to write the report file (default:
                     "./evolveguard-report.json")
  --allow-drift      exit 0 even if drift is detected (drift is still reported)
                     (default: false)
  --json             output structured JSON instead of human-readable text
                     (default: false)
  -h, --help         display help for command

Usage: evolveguard report [options] [reportPath]

Print a previously generated evolveguard-report.json

Arguments:
  reportPath  path to the report file (default: "./evolveguard-report.json")

Options:
  --json      output structured JSON instead of human-readable text (default:
              false)
  -h, --help  display help for command

Exit codes: 0 all fixtures PASS and no surface-level drift, 1 at least one DRIFT was found (pass --allow-drift to still exit 0 while still reporting it), 2 a usage error or a file that failed to parse.

Agent-native usage

Every subcommand supports --json for structured output an agent can parse directly:

evolveguard check ./SKILL.md --json

{
  "schemaVersion": 1,
  "skillName": "monorepo-scanner",
  "results": [
    {
      "id": "scan-a-monorepo",
      "verdict": "DRIFT",
      "changes": [
        /* ... */
      ]
    }
  ],
  "surfaceChanges": [],
  "summary": { "pass": 0, "drift": 1, "total": 1 },
  "exitCode": 1
}

evolveguard mcp is documented but not implemented yet -- see the CLI reference above. Until it ships, call record/check/report --json directly as a subprocess from your coding agent.

Library API

evolveguard also exports a programmatic API for the same pipeline, if you'd rather integrate it into your own tooling than shell out to the CLI. Both distributions expose the same functions and the same JSON-compatible file format -- a baseline recorded with one CLI can be checked with the other (see docs/concepts.md).

TypeScript:

import {
  recordBaseline,
  replaySkill,
  diffAll,
  writeBaseline,
  readBaseline,
} from 'evolveguard';

const baseline = recordBaseline('./SKILL.md', './fixtures.json');
writeBaseline('./.evolveguard-baseline.json', baseline);

// ... skill gets edited ...

const saved = readBaseline('./.evolveguard-baseline.json');
const replay = replaySkill('./SKILL.md', saved);
const report = diffAll(saved, replay);

See src/evolveguard/index.ts for the full exported surface: parseSkillFile, deriveCapabilitySurface, loadSkill, buildFixtureSnapshots, loadFixtures, recordBaseline, replaySkill, diffFixture, diffAll, diffSurface, writeBaseline, readBaseline, writeReport, readReport, plus the shared types.ts interfaces.

Python (pip install evolveguard):

from evolveguard import record_baseline, replay_skill, diff_all, write_baseline, read_baseline

baseline = record_baseline("./SKILL.md", "./fixtures.json")
write_baseline("./.evolveguard-baseline.json", baseline)

# ... skill gets edited ...

saved = read_baseline("./.evolveguard-baseline.json")
replay = replay_skill("./SKILL.md", saved)
report = diff_all(saved, replay)

See python/README.md for the Python-specific walkthrough and the same exported surface under evolveguard/__init__.py.

False-positive rate

No accuracy claim ships without the command that produced it. Run:

npx vitest run src/evolveguard/benchmark.test.ts

against fixtures/labeled-non-breaking-edits/ -- a small, hand-labeled corpus of real before/after SKILL.md pairs (wording tweaks and typo fixes labeled non-breaking; a filesystem-scope widen, a new write capability, and a hook script gaining a network call labeled breaking). As of this commit: 0% false positives (0 of 2 non-breaking cases flagged as drift). The corpus is small and will grow as more real skill edits are reported; see fixtures/labeled-non-breaking-edits/README.md.

What is evolveguard, and why does it exist

evolveguard is a command-line tool and TypeScript library that detects capability drift in Claude Agent Skill files (SKILL.md) and Claude Code auto-memory files (MEMORY.md) after they are edited, by a human or by an agent. It works by parsing a skill's declared frontmatter scope and any static evidence of network or filesystem-write behavior in its body text and bundled hook scripts, snapshotting that as a baseline, and re-deriving the same snapshot after an edit to diff against it. It exists because Claude Code's Agent Skills ecosystem lets skills and memory files change an agent's behavior without a human necessarily reviewing every edit for regression, and no existing tool checks that specific artifact shape without requiring SDK integration or a live agent run.

FAQ

Does evolveguard call an LLM? No. Record and check are both fully static and deterministic -- see "How it works" above.

Can it catch a change in what an agent actually decides to do on a given prompt? No, not directly. It catches changes in a skill's declared or inferred capability surface (the tools it's allowed or shown to use), which is a proxy for behavior, not a live trace of it. If you need to score actual agent runs, use Braintrust or agent-eval.

Does it work with MEMORY.md files, which have no frontmatter? Yes. A file with no frontmatter is parsed with an empty declared scope, so its capability surface comes entirely from static evidence found in the body text.

Is this a general agent-evolution framework? No. See "How it's different" above -- evolveguard deliberately does not build or host a self-evolving agent framework; it only tests skill/memory edits that already happened.

Is this published on npm or PyPI yet? Not yet on either -- see Status for the two separate, unrelated blockers.

Status

This is a v0.1 release: a small, focused addition to the existing Claude Agent Skills ecosystem. It ships fully MIT-licensed with no proprietary tier, as two independent, equally first-class packages -- both fully built, tested, and publish-ready, blocked on two separate account-level issues, neither a code problem:

• PyPI (evolveguard, Python) -- a genuine independent port, not a wrapper around the Node binary (see python/README.md). Wheel and sdist are built, inspected, and verified end to end from a fresh venv install. The first twine upload on this account is blocked by PyPI's own new-project-creation anti-abuse throttle (429 Too many new projects created), confirmed across repeated upload attempts. It will publish as soon as that throttle clears; no code change is needed. Clone the repo and run cd python && pip install -e . in the meantime.
• npm (evolveguard, TypeScript) -- npm pack --dry-run passes clean and package.json metadata is publish-ready, but the actual npm publish is blocked on a separate account-level constraint: this npm account's 2FA is security-key/ passkey-only, with no authenticator-app or OTP fallback configured, and npm currently requires completing that 2FA challenge interactively to publish. Clone the repo and run npm run build && npm link for the TypeScript CLI locally in the meantime.

Contributing

See CONTRIBUTING.md.

Security

See SECURITY.md.

License

MIT. See LICENSE.