IDE extension panel
Pulse Guard is IDE dev-environment protection for keystrokes. Cursor or VS Code: connection health, local prompt compiler, allowlisted verify scripts, Safe Build, and remediation tools.
Capabilities · shipped
AI governance& verify without re-prompting
Control plane for model traffic: six isolated context domains, IDE verify loops, optional Grok in routing presets, signed human-in-the-loop review with Session Replay and SIEM export, Sentry → Vault quarantine, Workspace SSO, and optional post-quantum Vault envelopes — unconfigured integrations degrade; they do not break the rest of the product.
01
Connect
Map the repo, mint an IDE token, and confirm the extension is healthy.
02
Optimize
Compile a local prompt with file attachments and required verify steps — no model call.
03
Verify
Run allowlisted tests or a local build. Pass writes to Vault; repeated fail writes to Hall.
04
Govern
Small Brain routing by default; Big Brain consensus on high drift; shadow then enforce; human-in-the-loop review; SIEM.
Verify loop (Vault / Hall)
Allowlisted local commands write passing results to Vault (verified state memory) and repeated failures to Hall — then unlock the deploy gate.
Model routing + consensus
Default pair is Claude + Gemini. Optional Claude + Grok or Gemini + Grok. High drift can take a three-model majority (Claude + Gemini + Grok). Humans are notified only above the drift threshold.
Defensible token-cost reporting
Metered provider usage, proven avoidance (not estimates), weekly and monthly reports with PDF, and shadow-mode projected spend kept separate from public eco metrics.
Shadow mode → enforcement
Point the SDK at /api/v1 with x-msgf-mode: shadow. Shadow mode is observe-only. Active mode applies the Active Cache, context gating, and live reputation.
Sentry → context quarantine
Sentry crashes match approved records in Vault and mark them quarantined (no silent move to Hall). Operators restore or demote on /admin/ops — runtime monitoring linked to governance memory.
Workspace SSO
Company domains and Google Workspace SSO for corporate seats. Consumer Gmail lands on invite-only — not a silent tenant.
Platform & enterprise
Three-model consensus
Big Brain can take a Claude + Gemini + Grok majority when drift is high. Split votes no longer force human review when a majority agrees.
Signed human-in-the-loop audits
Skip-gateway and arbitration decisions leave HMAC-signed, hash-chained snapshots — a compliance trail of who approved what on /admin/ops.
Deploy gate
CI checks GET /api/msgf/deploy-gate for your project. Ship only when verify is green for that isolation boundary.
Compound tenant isolation
Every vector is scoped by project and subpath — agencies and monorepos keep client and app memory from crossing silos.
Post-quantum envelopes
Optional hybrid post-quantum envelopes (X25519 + ML-KEM-768) for Vault secrets and ML-DSA-65 authorship certificates. HTTPS post-quantum only when the load-balancer SSL policy is enabled — not Pulse, Redis, or Supabase.
Admin ops console
One /admin/ops surface: audit timeline, Session Replay, security event log, most-used resources, model fitness, diff impact, tenant budgets, SIEM export, trusted-OSS bulk review, and Sentry quarantine.
Platform governance audit
Hashed resource-usage rankings, unified audit events, prompt-session forensics, and human-in-the-loop gates — policy failures and harm always open review; budgets never bypass safety.
Budgets + SIEM export
Hard monthly spend caps and rapid-retry circuit breakers before gateway dispatch, plus async OpenTelemetry JSON webhooks to your SIEM — primary paths never wait on export.
Context partitions
Six isolated policy domains
MSGF maps the active project into six isolated local slices inside a hidden .msgf/ directory so prompts do not mix contexts:
P1 · Rules
Immutable rules and security checks before a request reaches the model.
P2 · Flow
Pipeline and execution order, including verify-before-ship.
P3 · Identity
Identity, roles, and tenant isolation.
P4 · State
Runtime telemetry and active session memory.
P5 · Context
Workspace context stays sharded so prompts do not carry unused files.
P6 · Isolation
Vault is verified state memory. Hall stores rejected outcomes.
IDE extension
MSGF Pulse Guard for Cursor & VS Code
Pulse Guard is IDE dev-environment protection for keystrokes. It is not the proxy. The native extension ships an IDE panel: connect once, compile targeted prompts, run allowlisted verify scripts, and sync outcomes to your token-cost dashboard — including routing presets (Claude / Gemini / Grok pairs).
Local prompt compiler
Describe the task once. MSGF builds a sharded, attachment-ready prompt with required verify steps — locally, with no server-side model call.
Verify scripts
Allowlisted verify commands in .msgf/run-scripts.json. Re-run tests without regenerating the prompt. Counters appear on the token-cost dashboard.
Safe Build
One-click local build/test with async preflight. Pass syncs to verify-result; fail triggers low-cost remediation — not a blind incident dump.
IDE extension panel
Connection status, prompt compiler, verify scripts, Safe Build, and advanced Pulse / remediation tools in one Cursor or VS Code panel.
Verify loop
Approved on pass · rejected on repeat failure
When verify passes with a linked context pack, MSGF writes a positive record to the Vault (verified state memory). Repeated failures on the same command pattern dedupe into the Hall (rejected-outcome store) after three strikes. Safe execution uses execFile with an allowlisted command set (no shell injection from tampered script files).
Routing
Small Brain by default · Big Brain on drift
Small Brain is cost-optimized local routing. Big Brain is cross-model consensus and arbitration. Tenant verification defaults to Claude + Gemini (unanimous). Compliance paths can switch to Claude + Grok or Gemini + Grok. When logic drift is high, Big Brain can run a three-model majority (Claude + Gemini + Grok) and only opens human notification on high original drift, no majority, or a security non-human signal. Prompt compile stays local.
Security
Optional post-quantum Vault envelopes
Long-lived secrets can use hybrid KEM envelopes (X25519 + ML-KEM-768) and authorship proofs can use ML-DSA-65 certificates when enabled. Transport still relies on platform TLS; we do not claim HTTPS is post-quantum until your load balancer negotiates a PQ key exchange.
Integrations
Sentry, Workspace SSO, and SIEM
Sentry issues match approved Vault records and quarantine poisoned context on /admin/ops. Google Workspace SSO and company domains gate corporate seats. SIEM is an OpenTelemetry JSON webhook to your sink — not a Splunk marketplace app. Unconfigured panels show as unconfigured.
Token cost
Defensible spend on the dashboard
After you map a project and run the IDE loop, your governance dashboard shows grouped counters: IDE verify (pass/fail/Vault/Hall), script reruns, local prompt packs, ingest hash skips, and routing mix. A 24-hour rollup separates MSGF cloud tokens from context savings you can defend to finance.
Remediation
IDE build failures without a full Pulse
dev-event handles IDE build failures with a Vault lexical match or a single low-cost model call — not the full telemetry and consensus chain. Traffic stays on the cost-efficient path unless drift truly escalates.