ARGUS -- Intelligence Platform

CosySim Documentation -- v1.52.0 [2026-03-26]

Browser automation, API surface discovery, RPC registry, and the LiveDebugger.

Overview
CDP / Playwright
LiveDebugger
API Catalog
Token Harvesting
Configuration
Cross-References
Change Log

Overview

ARGUS (Automated Reconnaissance & Google Universal Surveyor) is a living API intelligence platform embedded in CosySim. It systematically maps Google's internal APIs (NotebookLM, Gemini AI, AI Studio, Colab, Apps Script, Workspace) using browser automation and traffic analysis, then stores every discovery in Nexus so all agents can use it immediately.

Core techniques:

Chrome DevTools Protocol (CDP) -- intercept every HTTP request without a proxy
Playwright UI crawlers -- drive every UI flow to trigger every API endpoint
Heap snapshot diffing -- extract rpcids and method names from V8 memory before/after actions
tshark TLS decryption -- decode binary gRPC-web payloads for proto reconstruction
Auto-documentation -- generates API reference Markdown from live captures

Catalog size:

Target	Protocol	Known Items	Coverage
NotebookLM	batchexecute	49 rpcids	67% observed
Gemini	batchexecute	36 rpcids	47% observed
AI Studio (MakerSuite)	gRPC-web	150+ methods	Growing
Google Colab	gRPC-web	10 methods	0% observed
Apps Script	batchexecute	14 rpcids	0% observed
Workspace Gemini	mixed	49 operations	0% observed

See data/argus/registry.json for the versioned live registry.

CDP / Playwright

Architecture

scripts/argus/
+-- config.py                  # Baselines: 49 NLM rpcids, 36 Gemini rpcids, 150+ AIS methods
+-- cdp_bridge.py              # CDP WebSocket client -> Chrome :9222
+-- network_monitor.py         # Capture all requests/responses on all Chrome tabs
+-- tshark_capture.py          # TLS-decrypted packet capture (proto reconstruction)
+-- nexus_sink.py              # Store all discoveries in Nexus (entries + Q&A pairs)
+-- orchestrator.py            # Master controller -- runs all phases, stores results
+-- crawlers/
|   +-- base_crawler.py        # Playwright base: attach/launch, screenshot, CDP handoff
|   +-- nlm_crawler.py         # 14 NLM UI flows (49 rpcid coverage target)
|   +-- gemini_crawler.py      # 10 Gemini UI flows (36 rpcid coverage target)
|   +-- aistudio_crawler.py    # 15 AI Studio UI flows (150+ method coverage target)
+-- decoders/
|   +-- batchexecute.py        # f.req decoder + wrb.fr response frame parser
|   +-- grpc_web.py            # gRPC-web binary frame decoder + proto field extractor
|   +-- heap_diffing.py        # CDP heap snapshot diff -> new API shapes
+-- discovery/
    +-- endpoint_registry.py   # Versioned JSON registry at data/argus/registry.json
    +-- rpcid_detector.py      # Compare live vs known rpcids, fire discovery callbacks
    +-- feature_flag_probe.py  # Enumerate hidden flag IDs via GetFeatureFlags (ozz5Z)
    +-- proto_reconstructor.py # Build .proto stubs from binary frames + bundle field maps

The Crawl Loop

For every UI flow, ARGUS:

Attaches network monitor (CDP Network.enable on all tabs)
Takes a heap snapshot before the action
Performs the action (click, type, navigate) via Playwright
Drains all captured network requests/responses
Takes another heap snapshot
Diffs the heaps -- new strings -- new rpcids/methods
Decodes captured traffic:
- batchexecute -- rpcid + JSON payload
- gRPC-web -- service/method + proto field numbers
Registers everything in the endpoint registry
Stores new discoveries in Nexus

Protocol Details

NotebookLM + Gemini (batchexecute):

POST https://notebooklm.google.com/_/LabsTailwindUi/data/batchexecute
Body: f.req=[[["rpcid","json_payload",null,"generic"]]]
Auth: Session cookies (__Secure-1PSID, __Secure-1PAPISID)
Response: )]}'\n + wrb.fr frames

AI Studio (gRPC-web):

POST https://alkalimakersuite-pa.clients6.google.com/$rpc/google.internal.alkali.applications.makersuite.v1.MakerSuiteService/{Method}
Auth: Authorization: SAPISIDHASH {ts}_{sha1}
Content-Type: application/grpc-web+json OR application/grpc-web+proto

SAPISIDHASH: sha1(f"{timestamp} {SAPISID} {origin}")

gRPC Service Paths

AI Studio:  google.internal.alkali.applications.makersuite.v1.MakerSuiteService/{Method}
Applets:    google.alkali.boq.makersuite.makersuiteappletcontrol.proto.MakersuiteAppletControlService/{Method}
Colab AI:   google.internal.colab.v1.AIService/{Method}
Colab RT:   google.internal.colab.v1.RuntimeService/{Method}
NLM gRPC:   google.internal.labs.tailwind.orchestration.v1.LabsTailwindOrchestrationService/{Method}

TLS Decryption Setup (for proto reconstruction)

Chrome must be launched with SSLKEYLOGFILE for tshark to decrypt HTTPS:

$env:SSLKEYLOGFILE = "C:\Files\Models\CosySim\artifacts\argus\tls\sslkeys.log"
Start-Process "chrome.exe"

Then capture:

python -c "from scripts.argus.tshark_capture import TsharkCapture; tc=TsharkCapture(); tc.start()"
# ... navigate in Chrome ...
python -c "from scripts.argus.tshark_capture import TsharkCapture; tc=TsharkCapture(); tc.stop()"

CDP Requirements

# Check Chrome is exposing CDP
Invoke-RestMethod http://localhost:9222/json

Chrome must be running. ARGUS attaches to the existing session without closing your tabs.

Usage

Quick Run (requires Chrome + CDP):

# Full scan
python -m scripts.argus.orchestrator

# Specific target
python -m scripts.argus.orchestrator --target nlm

# Probe NLM feature flags
python -m scripts.argus.orchestrator --probe-flags

# Generate docs from current registry only
python -m scripts.argus.orchestrator --docs-only

Scheduler (automatic weekly):

ARGUS runs automatically via the scheduler:

argus-weekly-scan -- full crawl (all 3 targets), every 7 days
argus-diff-report -- registry diff vs baseline, every 7 days

# Trigger manually
python -m engine.nexus.scheduler_daemon run argus-weekly-scan
python -m engine.nexus.scheduler_daemon run argus-diff-report

Query Results via Nexus:

All discoveries are stored in Nexus under category=argus:

python -m engine.nexus.bridge search "argus rpcid"
python -m engine.nexus.bridge ask "What new NLM rpcids did ARGUS find?"
python -m engine.nexus.bridge search "AI Studio method"

Generated Output Files

File	Description
`data/argus/registry.json`	Versioned endpoint + rpcid registry (all discoveries)
`data/argus/feature_flags.json`	Active NLM feature flag IDs
`data/argus/protos/*.proto`	Reconstructed .proto stubs
`artifacts/argus/har/*/.har`	Raw imported HAR captures and dumps
`artifacts/argus/screenshots/*.png`	Vision/browser screenshots and debug captures
`artifacts/argus/pcap/*.pcapng`	Raw packet captures (tshark)
`docs/NEXUS.md`	Auto-generated NLM API reference
`docs/GEMINI_API_REFERENCE.md`	Auto-generated Gemini API reference
`docs/AISTUDIO_API_REFERENCE.md`	Auto-generated AI Studio API reference

Extending ARGUS

Add a new crawler flow:

# In scripts/argus/crawlers/nlm_crawler.py
async def _new_flow_name(self) -> None:
    """Trigger rpcid XYZ by doing ..."""
    element = await self._page.query_selector("some-selector")
    if element:
        await element.click()
        await asyncio.sleep(1)

# Register in run_flows()
await self.step("new_flow_name", self._new_flow_name)

Add a new discovery pattern:

# In scripts/argus/discovery/rpcid_detector.py
# Add to _is_candidate() or analyse_bundle() to catch new patterns

LiveDebugger -- Real-Time Scene Diagnostics

The LiveDebugger reuses the CDP infrastructure for real-time CosySim scene debugging. While the crawlers map external APIs, the LiveDebugger points the same CDP tooling inward -- attaching to running scene pages to capture console output, inspect the DOM, profile performance, and diagnose UI issues without leaving the terminal or agent context.

Architecture

scripts/argus/
+-- live_debugger.py      # Core async debugger (1149 lines)
+-- cdp_bridge.py         # CDP WebSocket client (used by both crawlers and debugger)
+-- tools/
    +-- debug_scene.py    # CLI tool with 10 subcommands (306 lines)

engine/skills/builtin/
+-- debugger_skills.py    # 14 MCP skills for agent access (600 lines)

Core Capabilities

Console Streaming -- Captures console.log, console.warn, and console.error from the page in real-time via CDP Runtime.consoleAPICalled.
Network Monitoring -- Intercepts all HTTP requests and responses, tracks loading time, and captures request/response bodies.
DOM Inspection -- Executes arbitrary JavaScript in page context via CDP Runtime.evaluate. Query DOM state, check element visibility, read computed styles.
Z-Stack Analysis -- Reports the z-index layering of all positioned elements to diagnose overlay and blocking issues.
Click Testing -- Uses document.elementFromPoint() to verify which element receives clicks at specific coordinates, exposing invisible overlays.
Vision Analysis -- Takes CDP screenshots and can analyse them with LMStudio vision models (Qwen2-VL) for visual regression detection.
Performance Profiling -- Captures performance.timing, resource loading waterfall, and memory usage.
Scene Health Check -- Automated validation: page loads, Socket.IO connects, no console errors, API routes respond.

CLI Usage

# Full diagnostics
python -m scripts.argus.tools.debug_scene --port 5556

# Live console monitoring (Ctrl+C to stop)
python -m scripts.argus.tools.debug_scene --port 5556 --watch

# Execute JavaScript
python -m scripts.argus.tools.debug_scene --port 5556 --eval "document.title"

# DOM queries
python -m scripts.argus.tools.debug_scene --port 5556 --dom "div.ph-director-panel"

# Z-index stack analysis
python -m scripts.argus.tools.debug_scene --port 5556 --z-stack

# Click target at coordinates
python -m scripts.argus.tools.debug_scene --port 5556 --click-test 400,300

# Take screenshot
python -m scripts.argus.tools.debug_scene --port 5556 --screenshot

# Performance metrics
python -m scripts.argus.tools.debug_scene --port 5556 --perf

# List Chrome tabs
python -m scripts.argus.tools.debug_scene --port 5556 --tabs

# Scene health check
python -m scripts.argus.tools.debug_scene --port 5556 --health

MCP Skills (14 skills, pack="debugger")

All skills are exposed to agents via the @skill(pack="debugger") decorator in debugger_skills.py. Each skill wraps the async LiveDebugger core with a sync entry point.

Skill	Description	Key Parameters
`debug_scene`	Full diagnostics snapshot	`port`
`debug_watch`	Live console + network monitor	`port`, `duration`
`debug_console`	Console log capture	`port`, `duration`
`debug_network`	Network traffic capture	`port`, `duration`
`debug_eval`	Execute JS in page	`port`, `expression`
`debug_dom`	Query DOM elements	`port`, `selector`
`debug_z_stack`	Z-index layer analysis	`port`
`debug_click_test`	Check click target	`port`, `x`, `y`
`debug_screenshot`	Capture screenshot	`port`, `output_path`
`debug_click`	Simulate click	`port`, `selector`
`debug_navigate`	Navigate to URL	`port`, `url`
`debug_perf`	Performance metrics	`port`
`debug_list_tabs`	List Chrome tabs	(none)
`debug_health`	Scene health check	`port`

Integration with ARGUS Crawlers

LiveDebugger reuses CDPBridge and CDPSession from cdp_bridge.py -- the same WebSocket client that powers the crawlers.
Same CDP port (localhost:9222) used by crawlers and debugger. Both can coexist on different tabs.
Discoveries from debugging (new endpoints, JS errors) can be stored in Nexus via nexus_sink.py.
Screenshots are stored in data/argus/screenshots/.

Async Pattern

The LiveDebugger is fully async (asyncio). MCP skills use a sync wrapper to bridge the gap:

def _run_async(coro):
    loop = asyncio.new_event_loop()
    try:
        return loop.run_until_complete(coro)
    finally:
        loop.close()

Important: Must use asyncio.new_event_loop() (not asyncio.get_event_loop()) to avoid conflicts with already-running loops and to ensure compatibility with the full test suite.

Testing

python -m pytest tests/test_live_debugger.py -v  # 57 tests

API Catalog

The full RPC registry maintained by ARGUS. All data from live crawls and HAR imports, stored at data/argus/registry.json.

Total baseline operations: 184 | Observed in crawls: 200 | New discoveries: 152

Service	Baseline	Seen	New	Coverage
NotebookLM (batchexecute)	49	33	2	67%
Gemini (BardChatUi)	36	17	0	47%
AI Studio (MakerSuite gRPC)	0	150	150	--
Google Colab (gRPC)	10	0	0	0%
Apps Script (batchexecute)	14	0	0	0%
Workspace Gemini (mixed)	49	0	0	0%
NLM gRPC (proto)	2	0	0	0%
Heap-Discovered (unconfirmed)	24	0	0	0%

NotebookLM rpcids (49 known)

Key rpcids for CosySim integration:

rpcid	Operation	Notes
`bv7rAb`	CreateNotebook
`dlXLMc`	GetNotebook
`UbmKqb`	ListNotebooks
`NbGLKb`	AddSource
`ozz5Z`	GetFeatureFlags	Shared with Gemini
`Bgzyjc`	GenerateFreeFormStreamed	Main chat endpoint
`VHpbob`	GetAudioOverviewStatus
`Ygx6Tb`	StartAudioOverview
`izAoDd`	Add URL/text source	URL, YouTube, Sheets, image, text paste
`o4cbdc`	Register file upload	Returns upload URL + source ID
`rLM1Ne`	Poll source processing	Returns pending source IDs
`hPTbtc`	List sources	All source IDs in a notebook
`LBwxtb`	Delete source / blog post	Multipurpose by payload shape
`CYK0Xb`	Create note (report)	~10k word prompt -> full document
`QA9ei`	Generate audio podcast	30-min deep dive / brief / critique / debate
`gArtLc`	Poll / list artifacts	Get download URL when COMPLETE
`ciyUvf`	Generate flashcards	Instant Q&A pairs from sources
`R7cb6c`	Generate quiz	Multiple choice or true/false
`yyryJe`	Generate mind map	Nested concept JSON tree
`Krh3pd`	Export to Sheets	Returns live Google Sheet URL
`tr032e`	Source summary	Gemini summary of one source
`ub2Bae`	List notebooks	All notebooks in account
`s0tc2d`	Rename notebook	Update display name

See scripts/argus/config.py for the full baseline list.

Gemini rpcids (36 known)

rpcid	Description	Status	Observed	Last Seen
`ku4Jyf`	Code execution request	active	11	2026-03-08
`K4WWud`	Conversation management -- list, create, delete	active	7	2026-03-08
`mMEAEd`	CountTokens	--	0	--
`VUBhEd`	CreateCachedContent	--	0	--
`BgXnQc`	CreateFile	--	0	--
`sPOurf`	DeleteCachedContent	--	0	--
`qVSQ5c`	DeleteFile	--	0	--
`L5adhe`	Draft / edit message -- large state initialization	active	97	2026-03-08
`MaZiqc`	Extension/plugin interaction	active	14	2026-03-08
`ozz5Z`	Feature flags / account state (shared with NLM)	active	7	2026-03-08
`jKHnxe`	GenerateContent	--	0	--
`ESY5D`	Get conversation history / feature flags list	active	72	2026-03-08
`NXpLKc`	Get linked notebooks (cross-product bridge)	active	2	2026-03-05
`XqA3Ic`	Get storybook detail -- fetch specific gem by ID	--	0	never
`sJBwce`	Get subscription tiers -- Pro/Free tier info	--	0	never
`jPv1oc`	GetCachedContent	--	0	--
`ozVbQb`	GetFile	--	0	--
`XqsOBb`	GetModel	--	0	--
`jGArJ`	List my content -- filtered /mystuff	--	0	never
`ZKcapf`	List saved info -- paginated saved content	--	0	never
`HcT8bb`	List storybook gems	--	0	never
`dXH9nb`	ListCachedContents	--	0	--
`mfvMVb`	ListFiles	--	0	--
`k9yDXd`	ListModels	--	0	--
`DYBcR`	Locale / language preferences (shared with NLM)	active	7	2026-03-08
`otAQ7b`	Main chat generation -- send message, get response	active	7	2026-03-08
`CNgdBe`	Model selection / configuration	active	7	2026-03-08
`boaYGb`	ProxyUnaryCall	--	0	--
`GPRiHf`	Response rating / feedback	active	7	2026-03-08
`qpEbW`	Search conversation history	active	11	2026-03-08
`aPya6c`	Session initialization / heartbeat	active	70	2026-03-08
`PCck7e`	Share conversation / gem	active	20	2026-03-08
`r7Bvze`	StreamGenerateContent	--	0	--
`maGuAc`	Upload attachment / file	active	14	2026-03-08
`cYRIkd`	User preferences / settings	active	7	2026-03-08
`o30O0e`	User profile fetch (contacts/identity)	active	7	2026-03-08

AI Studio Methods (150+ observed)

Endpoint: https://alkalimakersuite-pa.clients6.google.com/$rpc/google.internal.alkali.applications.makersuite.v1.MakerSuiteService/{Method}

Auth: Authorization: SAPISIDHASH <ts>_<sha1> + session cookies

Format: gRPC-web with binary proto encoding OR JSON mode

Cancel

Method	Status	Observed	Last Seen
`CancelCodeAssistantOfflineGeneration`	--	0	never
`CancelTuningJob`	--	0	never

Check

Method	Status	Last Seen
`CheckCloudProjectForTermsOfService`	--	never
`CheckCloudRunService`	--	never
`CheckImage`	--	never

Count

Method	Status	Observed	Last Seen
`CountSessionTurns`	--	0	never
`CountTokens`	active	21	2026-03-05

Create

Method	Status	Observed	Last Seen
`CreateApplet`	--	0	never
`CreateCloudApiKey`	--	0	never
`CreateCloudProject`	--	0	never
`CreateCloudRunService`	--	0	never
`CreateContextCache`	--	0	never
`CreateDataset`	--	0	never
`CreateGitHubRepository`	--	0	never
`CreateInteraction`	--	0	never
`CreatePrompt`	active	5	2026-03-05
`CreateSession`	--	0	never
`CreateTunedModel`	--	0	never

Delete

Method	Status	Last Seen
`DeleteApplet`	--	never
`DeleteCloudApiKey`	--	never
`DeleteCloudRunService`	--	never
`DeleteContextCache`	--	never
`DeleteDataset`	--	never
`DeletePrompt`	--	never
`DeleteSession`	--	never
`DeleteTunedModel`	--	never
`DeleteUploadedFile`	--	never

Export / Fetch

Method	Status	Observed	Last Seen
`ExportDataset`	--	0	never
`FetchMetricTimeSeries`	active	37	2026-03-08

Generate

Method	Status	Observed	Last Seen
`GenerateAccessToken`	active	142	2026-03-17
`GenerateCodeAssistantSuggestionChips`	active	34	2026-03-05
`GenerateContent`	active	7	2026-03-05
`GenerateFunctionCallAnswer`	--	0	never
`GenerateGitHubCommitMessage`	--	0	never
`GenerateImage`	--	0	never
`GenerateTitle`	active	5	2026-03-05
`GenerateVideo`	--	0	never

Get

Method	Status	Observed	Last Seen
`GetAnnouncementBanner`	--	0	never
`GetApplet`	active	18	2026-03-08
`GetAppletCloudRunServiceLogs`	--	0	never
`GetAppletOutputMetadata`	--	0	never
`GetCodeAssistantSnapshot`	active	10	2026-03-05
`GetDataset`	--	0	never
`GetExtension`	--	0	never
`GetFeatureFlags`	--	0	never
`GetGenerateVideoOperation`	--	0	never
`GetGitHubAuthStatus`	--	0	never
`GetGitHubSettings`	--	0	never
`GetGitHubStatus`	--	0	never
`GetGroundingPassage`	--	0	never
`GetImFeelingLuckyOptions`	--	0	never
`GetLoggingContext`	active	62	2026-03-08
`GetModel`	--	0	never
`GetPrepayEligibility`	--	0	never
`GetProjectUsageLimit`	--	0	never
`GetPrompt`	--	0	never
`GetSample`	--	0	never
`GetSession`	--	0	never
`GetSessionTurn`	--	0	never
`GetSharedPrompt`	--	0	never
`GetStarterPrompts`	--	0	never
`GetSurvey`	--	0	never
`GetTunedModel`	--	0	never
`GetTuningJob`	--	0	never
`GetUploadedFile`	--	0	never
`GetUserPreferences`	active	62	2026-03-08
`GetUserRestrictions`	active	15	2026-03-08
`GetVersionInfo`	--	0	never

Import

Method	Status	Observed	Last Seen
`ImportGitHubRepository`	--	0	never
`ImportProject`	--	0	never

List

Method	Status	Observed	Last Seen
`ListAppletRunConfigurations`	--	0	never
`ListAppletTemplates`	--	0	never
`ListApplets`	active	4	2026-03-05
`ListBillingAccounts`	--	0	never
`ListCloudApiKeys`	active	8	2026-03-05
`ListCloudProjects`	active	12	2026-03-08
`ListCodeAssistantConfigurations`	active	62	2026-03-08
`ListCodeAssistantFeatures`	active	36	2026-03-08
`ListCodeAssistantOfflineGenerations`	active	18	2026-03-08
`ListCodeGenSuggestionCards`	active	10	2026-03-05
`ListContextCaches`	--	0	never
`ListDatasets`	--	0	never
`ListDriveApplets`	--	0	never
`ListExtensions`	--	0	never
`ListGitHubRepositories`	--	0	never
`ListImportedProjects`	active	9	2026-03-08
`ListModels`	active	78	2026-03-08
`ListPromos`	active	4	2026-03-05
`ListPrompts`	active	125	2026-03-08
`ListRecentApplets`	active	74	2026-03-08
`ListSessionTurns`	--	0	never
`ListSessions`	--	0	never
`ListTunedModels`	--	0	never
`ListTuningJobs`	--	0	never
`ListUnsetAppletSecrets`	active	44	2026-03-08
`ListUploadedFiles`	--	0	never

Stream

Method	Status	Observed	Last Seen
`StreamBidiGenerateContent`	--	0	never
`StreamCodeAssistantOfflineGeneration`	active	26	2026-03-05
`StreamExtractVideoFrames`	--	0	never
`StreamGenerateContent`	--	0	never

Update

Method	Status	Last Seen
`UpdateCloudProject`	--	never
`UpdateCloudRunService`	--	never
`UpdateDataset`	--	never
`UpdateProjectUsageLimit`	--	never
`UpdatePrompt`	--	never

Other

Method	Status	Observed	Last Seen
`AuthenticateGitHub`	--	0	never
`BidiGenerateContent`	--	0	never
`BulkDeleteSessionTurns`	--	0	never
`CodeAssistant`	--	0	never
`CodeAssistantOffline`	active	28	2026-03-05
`ComputeStagedGitHubDiff`	--	0	never
`ConnectApplet`	--	0	never
`DisconnectApplet`	--	0	never
`EmbedContent`	--	0	never
`EnhancePrompt`	--	0	never
`GeminiSpeechToText`	--	0	never
`GoogleSearch`	--	0	never
`LoadBundledApplet`	--	0	never
`LoadCodeAssistantInteractionHistory`	active	18	2026-03-08
`LoadCodeAssistantSnapshots`	--	0	never
`LoadDriveApplet`	--	0	never
`Log`	active	26	2026-03-05
`ProvisionAndInitializeApplet`	active	28	2026-03-05
`ProxyStreamedCall`	--	0	never
`ProxyUnaryCall`	active	4	2026-03-05
`ProxyUnaryFileApiCall`	--	0	never
`PushNewCommit`	--	0	never
`QueryCodeSearch`	--	0	never
`RecordSessionTurnFeedback`	--	0	never
`RecordSurveyResponse`	--	0	never
`RerunTuningJob`	--	0	never
`SaveApplet`	active	26	2026-03-05
`SaveDriveApplet`	--	0	never
`SharePrompt`	--	0	never
`StoreRecentApplet`	active	41	2026-03-08
`UpgradeAndDisablePrepay`	--	0	never
`UploadScs`	--	0	never
`batchGenerateContent`	--	0	never

Google Colab Methods (10 known)

Endpoint: https://colab.research.google.com/$rpc/google.internal.colab.v1.{Service}/{Method}

Auth: Session cookies + X-Goog-AuthUser header

Format: gRPC-web binary proto

Method	Status	Last Seen
`AgentCreateTask`	--	never
`AgentQuerySuggestions`	--	never
`AgentQueryTask`	--	never
`AgentUpdateTask`	--	never
`CompleteCode`	--	never
`ExecuteCell`	--	never
`GetRuntimeProxyToken`	--	never
`GetUserInfo`	--	never
`ListAssignments`	--	never
`SmartPaste`	--	never

Apps Script rpcids (14 known)

Endpoint: https://script.google.com/_/AppsMakerFrontendUi/data/batchexecute

Format: Same batchexecute f.req encoding as NLM/Gemini

rpcid	Description	Status
`pEig0e`	Execute a named function in the script project	--
`OQOG2e`	Get all files in the script project	--
`LuHlxe`	Get current editor state/mode	--
`AvwHP`	Get extended project metadata with container info	--
`NFMk7c`	Get project metadata (name, dates, owner)	--
`yFXSbd`	Get project revision history with tour hints	--
`UvGaob`	Get project settings and configuration	--
`AJ6bre`	Initialize page/view state	--
`zzomTc`	List project version history with pagination	--
`OOPYjd`	List script execution history with status filters	--
`KKLVD`	List script triggers (time-driven, event-driven)	--
`toGAmc`	Save code content to a script file	--
`GXx9jd`	Save/update project with full metadata	--
`ivJzse`	Update cursor position in code editor	--

Google Workspace Operations (49 known)

Hosts: appsgenaiserver-pa.clients6.google.com, docs.google.com, sheets.google.com, drive.google.com

Auth: API key + session cookies OR SAPISIDHASH

Format: REST JSON, gRPC-JSON transcoding, or batchexecute

Cloud Search

Method	Description
`query_search`	Cross-workspace semantic search query

Docs Gemini

Method	Description
`help_me_create`	Generate document content from a prompt
`match_style`	Match generated content style to existing document

Drive Gemini

Method	Description
`ai_overview_search`	Semantic search across Drive files using AI Overviews
`ask_gemini`	Ask Gemini a question about Drive files

Drive V2Internal

Method	Description
`copy_file`	Copy a file in Drive (template duplication)
`export_file`	Export a Workspace file in a specified format
`get_file`	Get file metadata from Drive
`get_permissions`	List permissions on a Drive file
`insert_permission`	Add/modify sharing permissions
`list_files`	List/search files in Drive
`trash_file`	Move file to trash
`update_file`	Update file metadata (title, description, parents)
`upload_file`	Upload file to Drive with metadata (multipart)

People Stack

Method	Description
`autocomplete`	Autocomplete people/contacts for sharing and @mentions
`autocomplete_alt`	Alternative API key for people autocomplete (load balancing)
`warmup`	Pre-warm people autocomplete service

Sheets BigQuery

Method	Description
`createDataSourcePivotTableOnNewSheet`	Create a pivot table backed by a data source on a new sheet
`enableAllDataSourcesExecution`	Enable execution for all data source connections
`getBigQueryProjects`	List BigQuery projects accessible from Sheets
`insertDataSourceSheet`	Insert a new sheet backed by a data source
`newDataSourceSpec`	Create a new data source specification (BigQuery, Looker)
`refreshAllDataSources`	Refresh all connected data sources

Sheets Extended

Method	Description
`external_data_batch`	Batch fetch external data for multiple cell ranges
`get_prefs`	Get/set session preferences for spreadsheet editing
`get_revision_history`	Get version history/revisions for a spreadsheet
`save`	Save spreadsheet changes with commands bundle

Sheets Gemini

Method	Description
`columnsmith_execute`	AI-driven column transformation via Gemini on cell ranges
`external_data_fetch`	Fetch and inject external data into sheet cells

Sheets REST

Method	Description
`save`	Save document changes (internal Sheets RPC)
`scripts_getitems`	Get Apps Script items bound to the spreadsheet
`scripts_uiready`	Signal that the Apps Script UI is loaded and ready

Workspace Analytics

Method	Description
`create`	Create a new analytics session/event
`ping`	Lightweight activity heartbeat ping

Workspace Gemini

Method	Description
`get_settings`	Get current Gemini settings for the user
`list_gems`	List available Gemini models and capabilities
`quota_summary`	Get Gemini API usage quota summary
`stream_generate`	Stream-based Gemini text generation for Workspace apps
`update_settings`	Update user Gemini preferences

Workspace Support

Method	Description
`addons_list`	List installed Workspace add-ons and extensions
`async_data`	Fetch async data for Workspace integrations
`doc_sync`	Real-time document collaboration sync
`fetch_recommendation`	Fetch AI-powered feature recommendations
`fetch_recommendations_batch`	Batch fetch multiple AI recommendations
`peoplestack_autocomplete`	People autocomplete for sharing and collaboration
`prewarm`	Pre-warm Gemini AI models before generation
`scripts_ui`	Apps Script UI integration for document automation
`waa_ping`	Workspace analytics and activity tracking ping
`workspace_batch`	Batch multiple Workspace UI operations

Token Harvesting

Authentication Model

CosySim's external API layer is built on a single insight: every Google service accepts the same browser session cookies. One HAR capture from a logged-in account yields all the credentials needed to talk to Drive, Sheets, Colab, NotebookLM, AI Studio, and Apps Script programmatically -- no OAuth dance, no service account, no API key quota.

SAPISIDHASH (all Google services)

Every Google property uses the same SAPISID cookie pattern:

import hashlib, time

ts = str(int(time.time()))
origin = "https://aistudio.google.com"  # or drive.google.com, docs.google.com, etc.
digest = hashlib.sha1(f"{ts} {sapisid} {origin}".encode()).hexdigest()
auth = f"SAPISIDHASH {ts}_{digest}"
# Plus SAPISID1PHASH and SAPISID3PHASH if those cookies are present

Account Pool

Account credentials live in data/accounts/pool.json. Import an account with:

from engine.integrations.google_account_pool import get_account_pool
pool = get_account_pool()
pool.import_from_har("path/to/capture.har", name="nihilistcod", services=["colab", "drive", "notebooklm"])

Account Tiers

Account	Tier	Notes
nihilistcod	Free	Can set `[2]` tier marker (client-side gating)
knack112358	Pro	Full Pro tier access

SDK Architecture

Three layers connect agents to external services:

SDK clients -- standalone, auth-encapsulated HTTP clients for each service
Artifact Bus -- a unified routing layer that moves artifacts between services
Skills -- @skill-decorated wrappers that expose every bus operation to LLM agents

LLM Agents / CosySim Skills
             |
    Artifact Bus (engine/integrations/artifact_bus.py)
      |     |     |     |
  Drive  Sheets  Colab  NLM SDK
             |
    GoogleAccountPool (data/accounts/pool.json)

Google Drive Client

File: engine/integrations/google_drive_client.py

Auth: SAPISIDHASH with origin = "https://drive.google.com". All requests target clients6.google.com.

Method	Purpose	Endpoint
`list_files(folder_id, query)`	List files, optionally filtered	`GET /drive/v3/files`
`get_file_metadata(file_id)`	Get metadata for one file	`GET /drive/v2beta/files/{id}`
`download_file(file_id)`	Download raw bytes	`GET /drive/v3/files/{id}?alt=media`
`download_text(file_id)`	Download and decode UTF-8	same
`upload_file(name, content, mime_type, folder_id)`	Create or update a file	`POST /upload/drive/v3/files?uploadType=multipart`
`create_folder(name, parent_id)`	Create a folder	`POST /drive/v3/files`
`find_or_create_folder(name, parent_id)`	Idempotent folder upsert	list + create
`delete_file(file_id)`	Delete file or folder	`DELETE /drive/v3/files/{id}`
`upload_text_to_cosysim_folder(name, content, subfolder)`	Upload to `CosySim/{subfolder}/`	multipart create
`make_file_accessible_to_notebooklm(file_id)`	Grant anyone-reader	`POST /drive/v3/files/{id}/permissions`
`get_shareable_link(file_id)`	Build shareable URL	(string build)

from engine.integrations.google_drive_client import get_drive_client

drive = get_drive_client()                    # round-robin from pool
drive = get_drive_client("nihilistcod")       # specific account

Google Sheets Client

File: engine/integrations/gsheets_client.py

Auth: SAPISIDHASH with origin = "https://docs.google.com". Sheets v4 at sheets.googleapis.com + Drive v3 at clients6.google.com.

Method	Purpose
`create_sheet(title, folder_id)`	Create new spreadsheet
`get_metadata(sheet_id)`	Get tabs list and properties
`read_rows(sheet_id, range_, include_headers)`	Read as list of dicts
`read_raw(sheet_id, range_)`	Read as list of lists
`append_rows(sheet_id, rows, sheet_name)`	Append rows, auto-header
`write_rows(sheet_id, rows, sheet_name, start_row)`	Overwrite from row N
`clear_sheet(sheet_id, sheet_name)`	Clear all values
`create_from_data(title, rows, folder_id)`	Create + populate in one call
`export_as_csv(sheet_id, sheet_name)`	Export to CSV string
`list_sheets(sheet_id)`	List tab names
`add_sheet_tab(sheet_id, tab_name)`	Add new tab
`make_public(sheet_id)`	Grant anyone-reader
`get_shareable_url(sheet_id)`	Build `?usp=sharing` URL

NLM Direct Client

File: engine/integrations/nlm_direct_client.py

See Nexus for full rpcid catalog.

Two-endpoint architecture:

Endpoint	Purpose	Auth
`GenerateFreeFormStreamed`	Multi-turn notebook Q&A (ask / streaming ask)	cookies + f.req
`batchexecute`	All studio ops: create, generate, export, manage	cookies + rpcids

Both require a bl (build label) and f.sid (session fingerprint) extracted from the NLM homepage HTML.

Audio types: AUDIO_DEEP_DIVE=1 (30 min), AUDIO_BRIEF=2 (5 min), AUDIO_CRITIQUE=3, AUDIO_DEBATE=4.

Colab Runtime Client

File: engine/integrations/colab_client.py

Auth: SAPISIDHASH with origin = "https://colab.research.google.com". All RPC calls go to colab.clients6.google.com/$rpc/google.internal.colab.v1.*.

AI Agent API (Gemini 3.1 Pro) -- three-call cycle:

Method	RPC Endpoint	Purpose
`create_task()`	`AIService/AgentCreateTask`	Allocate a task UUID
`update_task(task_id, context)`	`AIService/AgentUpdateTask`	Load context (notebook content, code)
`query_task(task_id)`	`AIService/AgentQueryTask`	Poll for the response

Kernel Execution: WebSocket-based cell execution using the Jupyter messaging protocol over wss://{runtime_url}/api/kernels/{kernel_id}/channels.

Colab Notebook Builder

File: engine/integrations/colab_notebook_builder.py

build_and_run() pipeline:

Get/assign Colab runtime
Create Jupyter kernel session
Optional: prepend a cell that loads a Drive file as input data
Ask AI agent to create cells for task_description
Execute cells with self-repair (up to 3 retries per failing cell)
For each chain_prompt: ask AI for follow-up cells with prior outputs injected
Save notebook JSON to Drive
Store output summary in Nexus

Specialised pipelines: training_notebook(), research_to_notebook(), data_analysis_notebook().

Colab GPU Manager

File: engine/integrations/colab_gpu_manager.py

Tier	CU/hour	VRAM (GB)	RAM (GB)	Best For
FREE	0.0	0	12.7	CPU-only tasks
T4	0.5	16	12.7	Inference, embeddings, <3B LoRA
L4	1.2	22.5	53.0	3-13B LoRA, vLLM server, video gen
A100	6.0	40	83.5	7-34B LoRA, image fine-tune
H100	7.0	80	83.5	34B+ LoRA, full fine-tune

Budget state persists to data/accounts/cu_budget.json.

Colab Venv Manager

File: engine/integrations/colab_venv_manager.py

Drive-backed venv pattern: packages installed once (~3 GB), stored on Drive, activated by prepending site-packages to sys.path at notebook start.

DRIVE_MOUNT_PATH   = "/content/drive"
COSYSIM_DRIVE_ROOT = "/content/drive/MyDrive/CosySim"
VENV_PATH          = "/content/drive/MyDrive/CosySim/.venv"
OUTPUTS_PATH       = "/content/drive/MyDrive/CosySim/outputs"
MODELS_PATH        = "/content/drive/MyDrive/CosySim/models"
DATASETS_PATH      = "/content/drive/MyDrive/CosySim/datasets"

Artifact Bus

File: engine/integrations/artifact_bus.py

The bus abstracts all transport logic between services.

Service Enum: LOCAL, DRIVE, COLAB, NLM, SHEETS, NEXUS

Route Matrix:

From -> To	Transport
LOCAL -> DRIVE	multipart upload
LOCAL -> NLM	via Drive (make_public + add_source_url)
LOCAL -> NEXUS	read text + add_entry
DRIVE -> NLM	make_public + add_source_url
DRIVE -> COLAB	kernel cell that mounts Drive + copies file
DRIVE -> NEXUS	download_text + add_entry
DRIVE -> SHEETS	download JSON/CSV + create_sheet + append_rows
COLAB -> DRIVE	kernel reads file via base64 + upload
COLAB -> NLM	via Drive intermediary
COLAB -> SHEETS	Colab -> Drive then Drive -> Sheets
COLAB -> NEXUS	kernel reads text + add_entry
NLM -> COLAB	inject content as Python variable in kernel
NLM -> DRIVE	local audio download + upload
NLM -> NEXUS	generate_flashcards + add_qa (or add_entry)
SHEETS -> NLM	add spreadsheet URL as source
SHEETS -> NEXUS	read_rows + JSON + add_entry

Methods: handoff() (single hop), pipeline() (multi-hop), full_knowledge_loop() (compound workflow).

Apps Script as Serverless Compute

Google Apps Script (GAS) is a serverless JavaScript runtime built into every Google account. Key properties:

Free -- unlimited execution time for personal use, 6-minute max per execution (30 min for Workspace accounts)
Native Workspace access -- SpreadsheetApp, DriveApp, GmailApp, CalendarApp work without OAuth, without credentials
Web App deployment -- any script deployed as public HTTPS endpoint
Time triggers -- ScriptApp.newTrigger().timeBased().everyHours(4) is cron, for free
UrlFetchApp -- outbound HTTP with custom headers including Cookie and Authorization

Every Google account is another GAS environment. An account pool of 10 accounts is 10 independent scheduled runtimes.

GAS Template Library (planned templates/gas/):

Template	Purpose
`webhook_receiver.js`	Receives POST from CosySim scheduler, dispatches actions, POSTs results back
`nlm_caller.js`	Calls NLM batchexecute with session cookies from Script Properties
`drive_processor.js`	Processes Drive folder files on schedule, calls NLM, tracks in Sheet
`nexus_ingestor.js`	Reads Q&A pairs from Sheet, POSTs to CosySim Nexus API

Client-Side Research Targets

Target	Value
NLM Quota Counter (`remainingQueries` in heap)	Rotate accounts before 429s
Model Override in batchexecute payload	Force Gemini 2.5 Pro for specific operations
AI Studio Model ID Override	Switch frontier models without UI
Feature Flag IDs 400-1200	Gate premium generation capabilities on free accounts

Configuration

ARGUS is configured via scripts/argus/config.py which contains:

Baseline rpcid lists for all targets (NLM, Gemini, AI Studio)
CDP connection settings (localhost:9222)
Output paths for registry, screenshots, and proto stubs
Crawl flow definitions per target
Feature flag probe ranges

Scheduler tasks are defined in the CosySim scheduler config:

argus-weekly-scan -- full crawl
argus-diff-report -- registry diff vs baseline

Generic API Discovery Engine (v1.50)

ARGUS now includes a general-purpose API discovery engine that analyzes any HAR file or V8 heap snapshot — not limited to Google services.

CLI

# HAR analysis
python -m scripts.argus.analyze har path/to/file.har            # Console report
python -m scripts.argus.analyze har path/to/file.har --json      # Machine-readable JSON
python -m scripts.argus.analyze har path/to/file.har --report    # Markdown intelligence report

# Heap snapshot analysis
python -m scripts.argus.analyze heap path/to/snapshot.heapsnapshot

# Batch + comparison
python -m scripts.argus.analyze dir path/to/har_folder/          # Analyze all HARs
python -m scripts.argus.analyze compare file1.har file2.har      # Diff two captures
python -m scripts.argus.analyze heap-diff before.heap after.heap # Diff two heaps

Protocol Auto-Detection

The analyzer classifies every request by content, not domain:

Protocol	Detection Method
batchexecute	`/_/*/data/batchexecute` URL or `f.req=` in body
gRPC-web	`$rpc/` in URL or `application/grpc-web` Content-Type
GraphQL	`query`/`mutation` in JSON body or `/graphql` URL
REST JSON	`application/json` Content-Type
WebSocket	`wss://` URL scheme
Protobuf	`application/x-protobuf` Content-Type

What It Discovers

All unique endpoints grouped by domain/service
Auth schemes: Bearer, API key, Cookie, SAPISIDHASH, custom headers
Tokens (redacted): API keys, JWTs, session tokens
GraphQL operations: query/mutation names, variable shapes
Rate limits: 429 responses, Retry-After headers
Service groups: auto-detected from domain names
Feature flags: Statsig, LaunchDarkly, custom flag endpoints
WebSocket endpoints and protocols

Architecture

scripts/argus/analyzers/
├── __init__.py              # Package exports
├── data_types.py            # Report dataclasses (HARAnalysisReport, etc.)
├── protocol_detector.py     # Auto-detect protocol from URL/headers/body
├── har_analyzer.py          # Core HAR analysis engine
└── heap_analyzer.py         # V8 heap snapshot analyzer

scripts/argus/
├── analyze.py               # CLI entry point
└── clients/                 # Target-specific API exploration clients
    └── sesame_client.py     # Sesame AI interactive explorer

Target Clients

ARGUS can auto-generate exploration clients from HAR intelligence. Example:

# Sesame AI Explorer (built from HAR analysis)
python -m scripts.argus.clients.sesame_client flags      # 27 feature gates
python -m scripts.argus.clients.sesame_client domains     # Email domain gate testing
python -m scripts.argus.clients.sesame_client configs     # 14 dynamic configs
python -m scripts.argus.clients.sesame_client user        # Profile + roles
python -m scripts.argus.clients.sesame_client agents      # 5 agent-service instances
python -m scripts.argus.clients.sesame_client protocol    # WebSocket agent protocol spec
python -m scripts.argus.clients.sesame_client interactive  # Interactive REPL
python -m scripts.argus.clients.sesame_client full        # Everything

Sesame AI Findings (2026-03-25)

Finding	Detail
Employee domains	`@sesame.com`, `@sesameai.com` unlock 12 extra feature gates (19/27)
Meta partnership	`@meta.com` gets 1 extra gate (8/27 vs normal 7/27)
Staff configs	`webrtc_log_level: "info"`, `show_toggle: true`
Agent services	5 instances (agent-service-0 through 4), behind Google IAP
WebSocket protocol	13 message types, WebRTC SDP/ICE, character selection (Maya, Miles)
Public bucket	`sesame-dev-public` on GCS — no auth required for reads
Firebase project	`sesame-ai-demo`, API key `AIza_REDACTED`
User roles	`['USER', 'EMAIL_VERIFIED']`, `moderation_status: 'ALLOWED'`
Audio	44100 Hz, raw WebRTC Opus, TURN servers at `34.134.236.52:3478`

Cross-References

Nexus -- all discoveries stored in Nexus for agent access
Architecture -- system layers, engine/integrations placement
Nexus -- NotebookLM full rpcid documentation
Gemini API Reference -- auto-generated Gemini catalog
AI Studio API Reference -- auto-generated AI Studio catalog

Change Log

Version	Date	Description
v1.50.1	2026-03-25	Generic API Discovery Engine: HAR/heap analyzers, protocol auto-detection, Sesame AI explorer CLI, interactive REPL, domain gate testing
v1.50	2026-03-22	Merged EXTERNAL_APIS.md into unified ARGUS doc; added full API catalog, SDK clients, Artifact Bus, Colab/GPU/Venv managers, Apps Script, Workspace operations
v1.49	2026-03-19	Added ARGUS catalog summary with crawl statistics
v1.45	2026-03-15	LiveDebugger section added (14 MCP skills, CLI tool, async pattern)
v1.0	2026-03-01	Initial ARGUS doc -- CDP crawlers, decoders, discovery pipeline

FilesExpand file tree

ARGUS.md

Latest commit

History

ARGUS.md

File metadata and controls

ARGUS -- Intelligence Platform

Table of Contents

Overview

CDP / Playwright

Architecture

The Crawl Loop

Protocol Details

gRPC Service Paths

TLS Decryption Setup (for proto reconstruction)

CDP Requirements

Usage

Generated Output Files

Extending ARGUS

LiveDebugger -- Real-Time Scene Diagnostics

Architecture

Core Capabilities

CLI Usage

MCP Skills (14 skills, pack="debugger")

Integration with ARGUS Crawlers

Async Pattern

Testing

API Catalog

NotebookLM rpcids (49 known)

Gemini rpcids (36 known)

AI Studio Methods (150+ observed)

Cancel

Check

Count

Create

Delete

Export / Fetch

Generate

Get

Import

List

Stream

Update

Other

Google Colab Methods (10 known)

Apps Script rpcids (14 known)

Google Workspace Operations (49 known)

Cloud Search

Docs Gemini

Drive Gemini

Drive V2Internal

People Stack

Sheets BigQuery

Sheets Extended

Sheets Gemini

Sheets REST

Workspace Analytics

Workspace Gemini

Workspace Support

Token Harvesting

Authentication Model

SAPISIDHASH (all Google services)

Account Pool

Account Tiers

SDK Architecture

Google Drive Client

Google Sheets Client

NLM Direct Client

Colab Runtime Client

Colab Notebook Builder

Colab GPU Manager

Colab Venv Manager

Artifact Bus

Apps Script as Serverless Compute

Client-Side Research Targets

Configuration

Generic API Discovery Engine (v1.50)

CLI

Protocol Auto-Detection

What It Discovers