Platform Test Matrix

Ground-truth verification of every AI platform Quint claims to support. Aspiration vs reality.

Platform Test Matrix

As of 2026-05-03, Quint has only ever been end-to-end tested against Claude Code. Every other “supported” platform is a fingerprint entry that has never been exercised. This document is the plan to change that.

The Six Detection Layers

Every platform Quint claims to detect is detectable through some combination of these signals. The strength of detection = which layers fire.

Layer	Source	Strength	Spoofable?
1. Code-signing	macOS ES (teamID, signingID)	Cryptographic	No (verified by macOS kernel)
2. Dedicated domain	MITM destination host	Cryptographic	No (agent API endpoint)
3. Custom HTTP headers	MITM request headers	Strong	Yes (trivially)
4. System prompt fingerprint	Parsed LLM request body	Strong	Yes (just a string)
5. Process name / path	ES exec event	Weak	Yes (rename binary)
6. User-Agent string	MITM request header	Weak	Yes (any HTTP client)

A strong detection is at least one cryptographic signal + one supporting signal. Everything else is best-effort.

The Registry, Audited

Inventory of every platform in internal/agentdetect/fingerprints.go:44-258:

#	Platform	TeamID	SigningID	Domain	Headers	Prompt	UA	Strength
1	claude-code	✅ Q6L2SF6YDW	✅ com.anthropic.claude-code	—	—	✅	✅	🟢 strong
2	cursor	✅ VDXQ22DGB9	✅	✅ *.cursor.sh	✅ 5 headers	✅	✅	🟢 strong
3	copilot	✅ VS Code UBF8T346G9	✅	✅ api.githubcopilot.com	✅ 5 headers	✅	✅	🟢 strong
4	windsurf	—	—	✅ *.codeium.com	✅ 2 headers	✅	✅	🟡 medium
5	aider	—	—	—	—	✅	✅	🟡 medium
6	cline	—	—	—	—	✅	✅	🟡 medium
7	continue	—	—	—	✅ 1 header	✅	✅	🟡 medium
8	codex	—	—	—	—	✅	✅	🟡 medium
9	goose	—	—	—	—	✅	✅	🟡 medium
10	gemini-cli	—	—	—	—	✅	✅	🟡 medium
11	amp	—	—	—	—	✅	✅	🟡 medium
12	kiro	—	—	—	—	✅	✅	🟡 medium
13	augment	—	—	—	—	✅	✅	🟡 medium
14	zed	—	—	✅ zed.dev	—	—	✅	🔴 weak
15	opencode	—	—	—	—	—	✅	🔴 weak
16	pearai	—	—	—	—	—	✅	🔴 weak
17	trae	—	—	—	—	—	✅	🔴 weak
18	void	—	—	—	—	—	✅	🔴 weak
19	devin	—	—	—	—	—	✅	🔴 weak

Desktop AI applications — all missing

Not in the registry, but expected by every CISO after CrowdStrike/Microsoft Agent 365 marketing:

Platform	Signal available	Priority
Claude Desktop (Anthropic)	code-signing + anthropic.com domain	🔴 must-have
ChatGPT Desktop (OpenAI)	code-signing + chatgpt.com domain	🔴 must-have
Gemini Desktop (if exists yet)	code-signing	🟡 nice-to-have
M365 Copilot Desktop (Microsoft)	code-signing + office.com	🟡 nice-to-have

The Verification Checklist

For each platform, verify all seven checks. A platform is validated only when all seven are green.

#	Check	How to verify	Pass criteria
1	ES sees process exec	curl -s localhost:8080/debug/es-events | jq '.[] | select(.pid==<PID>)'	PID appears in event stream within 5s of launch
2	Agent identified	curl -s localhost:8080/debug/sessions | jq '.[] | select(.pid==<PID>) | .platform'	Returns expected platform name, not unknown
3	Code-signing verified (if applicable)	.signingID field on session	Matches registry entry
4	NE intercepts LLM traffic	curl -s localhost:8080/debug/flows | jq '.[] | select(.pid==<PID>)'	Flow exists with relay: true
5	LLM parser succeeds	curl -s localhost:8080/api/sessions/timeline?pid=<PID>	Returns tool_call entries, not PARSE_FAILED warnings
6	Session reaches cloud	curl https://api.quintai.dev/v1/sessions -H 'Authorization: Bearer $TOK' | jq '.[] | select(.session_id==\"<PID>-<TS>\")'	Session row exists in Postgres
7	Events attributed	Same endpoint + /events	All tool calls have the correct agent_id stamped

The First-Pass Manual Test (3 platforms)

Per the audit, the three most strategic platforms to validate right now are:

1. Cursor 🟢 strong detection claim

Why first: Second-most common developer choice, strongest competitive story (we have TeamID + headers + domain + prompt = 4 signals), and Cursor’s Connect RPC protocol is a distinct test of the llmparse router’s protocol detection.

Install path: Download from https://cursor.com/download, sign in, open a folder, trigger an AI chat.

Known unknowns:

• Does Cursor use Connect RPC or plain HTTPS in May 2026? (May have changed from the identifyFromProtocol code path in fingerprints.go:444-467)
• Do the 5 custom x-cursor-* headers still exist?
• Is cursor.com or *.cursor.sh the current API domain?

2. GitHub Copilot in VS Code 🟢 strong detection claim

Why second: Literally every Fortune 500 enterprise dev has this. If Quint can’t see Copilot, the pitch to 80% of buyers falls apart. Strong signal claims (VS Code TeamID + api.githubcopilot.com + 5 headers + UA) — maximum validation value.

Install path: VS Code + GitHub Copilot extension + valid subscription, open a file, trigger a Copilot completion or Copilot Chat.

Known unknowns:

• Does Copilot Chat use the same endpoints as inline completions?
• Are the copilot-integration-id / editor-version headers still present in 2026 versions?
• Does the VS Code teamID UBF8T346G9 attribute correctly, or does it over-match to any VS Code extension?

3. Claude Desktop 🔴 NOT IN REGISTRY — must add before testing

Why third: Quint claims “20+ platforms” but has zero desktop AI coverage. CrowdStrike and Microsoft both explicitly support Claude Desktop. This is the biggest demo-day gap. Adding Claude Desktop is ~1 hour of fingerprint work + validation.

Install path: Download from https://claude.ai/download, sign in, open a chat (this is NOT the CLI).

What to add to fingerprints.go:

{
    name:         "claude-desktop",
    teamID:       "Q6L2SF6YDW",   // Same Anthropic team as claude-code
    signingIDs:   []string{"com.anthropic.claudefordesktop"},
    processNames: []string{"claude"},
    processPathPatterns: []string{
        "Claude.app", "/Claude",
    },
    // System prompt likely differs from claude-code — needs capture
    uaPatterns: []string{"claude-desktop"},
},

Known unknowns: Entire signal set. This is discovery work, not verification.

The Output — What to Log

Every test run should produce a test-results.yml with:

platform: cursor
version: "0.43.2"
macos_version: "15.5"
tested_at: "2026-05-03T14:22:11Z"
checks:
  es_process_exec: pass
  agent_identified: pass
  code_signing: pass
  ne_intercepts: pass
  llm_parser: pass     # or "fail: PARSE_FAILED on openai_responses format"
  cloud_session: pass
  events_attributed: pass
signals_fired:
  - teamID
  - signingID
  - domain
  - headers (3/5)
  - prompt
  - UA
notes: "x-cursor-client-version no longer present in v0.43"

Check these results into docs/operations/test-results/ — they become the source of truth for the platform coverage table, replacing the aspirational list.

Escalation Path When Something Fails

Failure	Likely cause	File to investigate
ES doesn’t see exec	ES buffer overflow, ColdStart race	internal/eslistener/ + memory project_ingestion_gaps.md
Agent identified as unknown	Fingerprint mismatch (platform changed)	internal/agentdetect/fingerprints.go
NE doesn’t intercept	Not in NE domain list, or HTTP/2 reject	internal/netdetect/ + memory project_ingestion_gaps.md
Parser returns PARSE_FAILED	Wrong format router, new API shape	internal/llmparse/router.go + specific parser file
Cloud session missing	Session ID mismatch, ingest 4xx	Check /debug/flows and CloudWatch ingest logs
Wrong agent_id attributed	Detection priority ordering, parent-cascade	internal/agentdetect/detector.go

Expand the Matrix, Honestly

Once the first 3 platforms are validated, the dashboard “Platform Coverage” page should stop listing 19 names and start listing:

• ✅ Validated — test passed within the last 30 days (link to test-results.yml)
• 🟡 Aspirational — fingerprint exists, never tested
• 🔴 Known broken — tested and failed, has an open bug

This is what CISOs want to see. “We actually tested these 5 platforms yesterday” is infinitely more credible than “we support 20+.”

Related

• Test harness CLI: quint test-platforms — automated runner for this checklist
• Ingestion gaps — known issues that might cause platform tests to fail
• Platform coverage strategy — the broader platform expansion roadmap