Skip to content

OpsContext for AI Agents — 2.0.0

Latest
Latest

Choose a tag to compare

View all tags
@FASTPROD FASTPROD released this 11 Jun 13:30
· 209 commits to main since this release
v2.0.0
1bf2bf4

Strategic pivot release. The package previously known as @compr/contextengine-mcp ships as @compr/opscontext-mcp starting with this version. Same code, sharper positioning — the ops + compliance layer Claude Code can't grow natively.

📦 npm: https://www.npmjs.com/package/@compr/opscontext-mcp
🚨 Migration: @compr/contextengine-mcp is deprecated on the registry with a pointer at this package. Existing installs keep working; new installs should use the new name.

Three new pillars

🔭 Operational visibility

Read-only collectors snapshot your live infra so AI agents see what's actually running, not stale code:

  • PM2 process state, nginx config, Docker containers
  • Git working-tree drift across multiple repos, port-conflict detection
  • Crontab, redacted .env, composer, systemd, shell history
  • Cross-project search through everything in one query

🧾 Tamper-evident audit log

Hash-chained JSONL at ~/.contextengine/audit.log — compliance evidence out of the box:

  • SOC2 CC7.2 (audit logging) + ISO 27001 A.12.4.1 (event logs)
  • audit_verify walks the chain; mutating any record breaks verification at the mutated index
  • 19 adversarial tests covering tamper / splice / forgery / corrupt-JSON-graceful-failure

🛡️ Policy-as-code git hooks

Declarative .contextengine/policy.json with four sections:

  • secret_patterns (with paths glob scoping — e.g. JWT pattern only applied to docs/sessions/**)
  • doc_coverage (diff-aware, replaces wall-clock staleness gates that taught their own bypass)
  • deploy_verify_hosts (encodes "DEPLOY = VERIFY LIVE")
  • bypass_tokens (documented escape hatches with reason + TTL — alternative to undocumented --no-verify)
  • Pre-commit hook now layers gitleaks (~150 patterns) + scoped policy patterns + 17 inline CE patterns
  • contextengine hook secret-scan / hook doc-coverage CLI for CI use

Plus

  • 🪝 Native Claude Code integration: install-skill adds an opscontext skill to ~/.claude/skills/; sync-claude-md maintains a managed block in CLAUDE.md so the latest snapshot reaches the agent at every session start with zero MCP calls; auto-discovery of ~/.claude/projects/*/memory/ brings Claude's own auto-memory into search results.
  • 🔐 Ed25519-signed PRO licenses: server now signs license payloads with Ed25519. Client verifies with the embedded public key. Closes a real revenue leak (the old "signature" was a recomputable hash never verified). Backward compatible — legacy SHA-256 licenses grandfathered until 2026-08-15.
  • 📦 @huggingface/transformers moved to optionalDependencies: cold install drops from 547 MB to 120 MB. BM25 keyword search ships always; semantic re-ranking is opt-in.
  • 📤 export-learnings --project X: cross-client confidentiality for consultants.
  • 🧹 101 → 122 kB tarball despite three new substantive modules (audit, policy, hooks). Sourcemaps + obfuscation theater + dead artifacts removed.

What's NOT changed

  • Storage paths (~/.contextengine/), repo policy path (.contextengine/policy.json), env vars (CONTEXTENGINE_HOME), config name (contextengine.json), TypeScript code identifiers — all preserved so existing users lose nothing.
  • Bin aliases: opscontext (primary) + opscontext-mcp + contextengine + contextengine-mcp (back-compat).

Stats

  • 20 MCP tools, 21 CLI subcommands
  • 182 / 182 tests passing
  • Build clean, server side also clean
  • 1,000+ accumulated learnings preserved

Full changelog

See CHANGELOG.md.

Session arc

The complete story of the pivot + ship is in docs/sessions/SESSION_05_2026-06-11.md, including the production path bug caught and fixed during the Ed25519 deploy.

Built by FASTPROD (PROD LLC). BSL-1.1 license — see LICENSE.

Assets 2
Loading