Synthetic example, not a measured project result:

AI Context Test — 5 assertions

fact                    expected       mentions  status
────────────────────────────────────────────────────────────────────────
auth-config-path        true                  1  ✗ 1 mismatch
typecheck-script        true                  1  ✗ 1 mismatch
node-version            22.14.0               1  ✓ 1/1 pass
no-vague-language       check                 2  ✓ 2/2 pass
instruction-balance     check                 2  ✓ 2/2 pass
────────────────────────────────────────────────────────────────────────

Mismatches
────────────────────────────────────────────────────────────────────────
auth-config-path        CLAUDE.md:18   true        false
typecheck-script        CLAUDE.md:34   true        false
────────────────────────────────────────────────────────────────────────
✗ 2 mismatch(es) — update the file(s) listed above

npm install -g ctxharness
# or, inside your project:
pnpm add -D ctxharness
pnpm exec ctxharness --help

The CLI uses Node.js even when checking a Python, Go or Rust project. The source workflow gives access to the tagged v0.5.0 implementation independently of npm publication.

Some instructions belong in documentation: architecture choices, team conventions and where to find code. When you choose to state a fact explicitly, CTX Harness can compare supported claims with their local sources. It does not decide which instructions belong in your context.

.ctxharness.yml, minimal starter (one assertion per layer):

version: 1

files:
  include:
    - 'CLAUDE.md'
    - 'AGENTS.md'
    - '.cursorrules'
  exclude:
    - 'node_modules/**'

assertions:
  # L1 — fact drift: auth config path still exists
  - id: auth-config-path
    extractor: fileExists
    extractorArgs:
      path: src/modules/auth/config.ts
    scanner: literalInMd

  # L1 — fact drift: typecheck script matches package.json
  - id: typecheck-script
    extractor: packageScript
    extractorArgs:
      script: type-check
    scanner: literalInMd

  # L2 — instruction quality: no vague language
  - id: no-vague-language
    extractor: constant
    extractorArgs:
      value: check
    scanner: vaguenessPattern

  # L3 — context assembly: hooks are valid
  - id: hook-validity
    extractor: constant
    extractorArgs:
      value: check
    scanner: hookValidity

assertions:
  # allowlist: skip known-intentional mismatches in specific files
  - id: next-version
    extractor: packageJson
    extractorArgs:
      package: next
    scanner: inlineRegex
    scannerArgs:
      pattern: 'Next\.js\s+v?(\d+(?:\.\d+(?:\.\d+)?)?)'
    allowlist:
      - CHANGELOG.md # version history file — intentional old values

  # scopeFiles: restrict an assertion to a subset of files
  - id: instruction-balance
    extractor: constant
    extractorArgs:
      value: check
    scanner: negativeConstraintDensity
    scannerArgs:
      minRatio: 2.0
    scopeFiles:
      include:
        - 'CLAUDE.md'
        - 'AGENTS.md'
      exclude:
        - '.cursorrules' # constraint-only file by design

Version normalization: v22 matches 22.14.0, partial mentions are valid.

Drift scanners (compare against extractor value)

Quality scanners (no extractor value needed, use constant)

Standalone scanners (hookValidity, skillValidity, freshnessScore) bypass files.include and resolve their own targets from the project root. You do not add their paths to files.include, they run once regardless of how many files are in scope.

vaguenessPattern accepts custom patterns via scannerArgs.patterns (array of regex strings).

contextBudget estimates tokens as chars ÷ 4. Designed to run over .claude/rules/**/*.md or CLAUDE.md to catch bloated always-on context files. With followImports: true, it resolves @file.md references recursively up to depth 3 and includes their token footprint in the total.

ruleGlobValidity is designed to run over .claude/rules/**/*.md. By default it fails if a rules file has no YAML frontmatter (meaning it loads at every session with no scoping). Set requirePaths: true to also fail if the frontmatter lacks a paths: field.

freshnessScore works with gitStaleness extractor. gitStaleness returns the commit count since the file was last changed; freshnessScore compares it to your thresholds:

- id: claude-md-freshness
  extractor: gitStaleness
  extractorArgs:
    path: CLAUDE.md
  scanner: freshnessScore
  scannerArgs:
    warnAfter: 20 # ⚠ warn if >20 commits since last edit
    failAfter: 50 # ✗ fail if >50 commits

coverageRatio checks that a fraction of your actual entities (models, routers…) are mentioned in the docs, useful when you can't document everything but want to enforce a minimum:

- id: prisma-model-coverage
  extractor: prismaModelList
  extractorArgs:
    path: src/server/db/prisma/schema.prisma
  scanner: coverageRatio
  scannerArgs:
    minRatio: 0.5 # at least 50% of models mentioned
  valueAllowlist:
    - MigrationVersion # internal model, not required in CLAUDE.md

Local JSON includes source paths and diagnostic details. Review it before sharing. Use the versioned portable profile to omit paths, names, raw messages and credentials. It includes declared coverage and static sizes by default:

ctxharness doctor --host both --scope project --export portable > doctor.portable.json
ctxharness doctor --host claude --scope project --coverage --budget --format json

Byte counts measure discovered source content; the named character estimate does not establish model input or savings. Validate artifacts with the core contracts and fixtures. Opaque references do not guarantee anonymization.

ctxharness config-snapshot --host claude --scope project --output first.json
ctxharness config-snapshot --host claude --scope project --output second.json
ctxharness config-diff first.json second.json

Snapshot output never overwrites an existing file; structural comparison does not update the baseline or measure performance. These are separate from the documentation snapshot and diff commands. See the full guide.

ctxharness doctor --host claude --scope project --tools rg,gh --tool-path /opt/example/bin --tool-environment declared-client --format json

The example PATH is caller-supplied. Absent environment evidence stays unknown, and no tool is executed or installed. Authentication and client permissions remain unknown even when an executable is found. See tool options and limits.

ctxharness fix finds every assertion where the actual version differs from expected on a specific line and shows what it would change. Pass --apply to write the files:

$ ctxharness fix
CLAUDE.md:13  prisma-version  7.5 → 7.7.0
CLAUDE.md:42  next-version    15.2.0 → 15.3.1

Run ctxharness fix --apply to write changes.

This detects file paths, npm scripts, and version numbers mentioned in the file and checks each against your codebase. Paths and scripts first, because those are the claims most likely to silently break agent behavior when they drift.

Scanning CLAUDE.md...

  src/config/auth     fileExists   ✗ path not found (moved to src/modules/auth/config.ts?)
  npm run typecheck   packageScript  ✗ script not found in package.json
  Node.js 22.14.0    version      ✓ matches .nvmrc

2 issues found. Run with --suggest-config to generate .ctxharness.yml.

scan follows supported local @file.md imports up to depth 3 within caller-selected roots. Physical files are deduplicated, with provenance edges retained. Unsupported, missing, out-of-boundary or capped reads remain incomplete. See read boundaries.

npx ctxharness scan CLAUDE.md --suggest-config   # generate a starter .ctxharness.yml
npx ctxharness scan CLAUDE.md --exit-zero        # warn without blocking (hooks / CI)

The detector filters out common false positives: Claude Code slash commands (/plan, /ship), URL route patterns (/api/chunk, /about/), and template placeholders ({slug}, [owner]).

scan vs run: scan discovers claims without configuration and exits 1 when drift is found; --exit-zero makes drift warning-only. run enforces the assertions in .ctxharness.yml and exits 1 on failure. Input errors remain errors in either mode.

Husky hook, the post-merge template automatically picks the right mode:

# .husky/post-merge (generated by ctxharness init --hooks)
if [ -f ".ctxharness.yml" ]; then
  ctxharness check          # blocking — full enforcement
else
  ctxharness scan --exit-zero  # informational — zero-config discovery
fi

# Save baseline after initial setup
ctxharness snapshot

# In CI: compare against the committed baseline
ctxharness diff     # exit 1 if score dropped

Snapshots are saved to .ctxharness/snapshots/ with timestamp. Commit the latest snapshot file to use diff in CI.

run, check, score, and the YAML assertion mode of doctor auto-record a drift score to ~/.ctxharness/history.jsonl. The configuration doctor does not record a trend entry. The trend command shows your documentation score history:

ctxharness trend

Trend — myproject (8 runs)

  Sparkline   ▃▅▆▇▇▇██
  Direction   ↑ improving  (+14 pts over 8 runs)
  Avg Score   91/100

  Date                    Score      G      Pass   Fail   Time
  ──────────────────────────────────────────────────────────────
  May 06, 14:23:01       100/100    S      5      0      42ms
  May 06, 11:45:22        98/100    S      5      0      38ms
  May 05, 09:12:08        87/100    B      4      1      55ms
  ...

Direction is computed by comparing the average of the first third of runs against the last third: improving (delta > 3 pts), worsening (delta < -3 pts), or flat.

ctxharness trend --all          # all projects
ctxharness trend --limit 50     # last 50 runs (default: 20)
ctxharness trend --project api  # specific project name

In CI, use --no-trend to skip recording, useful when you only want trend data from your main branch, not every PR run:

ctxharness run --no-trend

# Dry-run (default): preview what would be added
ctxharness populate

# Write changes to .ctxharness.yml
ctxharness populate --apply

Typical workflow: run ctxharness init once to bootstrap the config, then run ctxharness populate --apply any time you update your AI docs and want to capture new claims.

Assertions can return three states: pass, warn, or fail. warn is counted as 0.5 in the score, useful for staleness checks where you want early signal without blocking CI:

⚠ 1 warn   — claude-md-freshness: 35 commits since last edit

ctxharness doctor categorizes all issues by layer, shows a per-layer score, and suggests next actions:

ctxharness doctor

L1  Doc Drift           ██████████  100/100
L2  Instruction Quality ████████░░   80/100
L3  Context Assembly    ██████░░░░   60/100

Score: 80/100  Grade: B

Issues:
  L2  no-vague-language   AGENTS.md:14   vague pattern found: "be careful"
  L3  hook-validity        .claude/settings.json   hook entry has empty matcher

import { definePlugin, loadPlugin } from '@florianbruniaux/ctxharness-core'

const myPlugin = definePlugin({
  extractors: [
    {
      name: 'myExtractor',
      fn: (root, args) => {
        // read ground truth from codebase, return a string
        return '1.2.3'
      },
    },
  ],
  scanners: [
    {
      name: 'myScanner',
      fn: (filePath, expectedValue, args) => {
        // check the file, return ScanResult[]
        return [{ status: 'pass', line: 0, actual: expectedValue }]
      },
    },
  ],
})

loadPlugin(myPlugin)

Use extractor: myExtractor and scanner: myScanner in your .ctxharness.yml like any built-in.

From a source checkout, copy a preset into the project you want to check:

cp templates/presets/t3.yml /path/to/project/.ctxharness.yml

.claude/settings.json vs settings.local.json

Claude Code follows a two-file convention for project settings:

• settings.json, committed to the repo. Contains config that should work for anyone who clones the project: hook definitions, permission rules, shared assertions. Use relative paths for hook commands (.claude/hooks/my-hook.sh, not /Users/yourname/...).
• settings.local.json, gitignored. Contains machine-specific or personal overrides: your own keybindings, local path overrides, personal MCP servers.

hookValidity validates settings.json. If a hook command contains an absolute path (e.g. /Users/yourname/...), it returns status: warn, those paths break for every other contributor.

The same layering applies to CLAUDE.md files: project-level goes in the repo root (committed), personal goes in ~/.claude/CLAUDE.md (your machine only). Claude Code memory docs.

Context engineering, why accuracy matters

• Context Engineering, Simon Willison, June 2025. Why "context engineering" is a better term than "prompt engineering" and what it means in practice.
• The Rise of Context Engineering, LangChain, June 2025. "Most of the time when an agent is not performing reliably, the underlying cause is that the appropriate context has not been communicated to the model."
• Context Engineering for Large Codebases, Packmind, April 2026. Documents "context drift", stale instruction files referencing deprecated frameworks cause agents to silently generate code using wrong patterns.

What stale context does to LLMs

• Contextual Drag: How Errors in the Context Affect LLM Reasoning, arXiv, Feb 2026. Wrong context causes 10-20% performance drops across 11 models and 8 reasoning tasks. Self-refinement makes it worse, not better.
• Knowledge Conflicts for LLMs: A Survey, EMNLP 2024. Temporal knowledge conflicts (outdated context vs. model knowledge) are a primary source of factually wrong outputs. LLMs may generate code using deprecated function signatures from older library versions.
• Lost in the Middle, Stanford / ACL 2024. Relevant information placed in the middle of long contexts is systematically under-weighted by LLMs. Instruction files that accumulate stale content push critical facts into this dead zone.
• Your Agent's Context Is a Junk Drawer, Augment Code, Feb 2026. Documents "context collapse", agents forget earlier constraints when context grows stale and unmanaged.

The CLAUDE.md / AGENTS.md problem specifically

• Writing a Good CLAUDE.md, HumanLayer, Nov 2025. "Don't include code snippets — they will become out-of-date quickly." Direct practitioner warning on content drift.
• New Research Reassesses the Value of AGENTS.md Files, InfoQ, March 2026. ETH Zurich study: LLM-generated context files reduce task success by 3% on average and increase inference costs by 20%+. Authors recommend limiting instructions to non-inferable details, exactly the facts ctxharness verifies.
• When AGENTS.md Backfires, Feb 2026. Only 14.5% of agent context files include security instructions. LLM-generated files reduced task success in 5 of 8 evaluation settings.