Versions and Limitations¶
The atlas is a detailed snapshot, not a timeless specification. Its strongest claims are limited to one file digest and one platform.
Captured version¶
| Property | Value |
|---|---|
| Version | 2.1.177 |
| Platform | macOS arm64 |
| Artifact SHA-256 | eb0730351be2f02b482b1855870f5877489085aac86b0c4c1db4e458d9e40ed9 |
| Entry-module SHA-256 | 45cb1eaa2b7e274ce87b1df0a1729f017ac06fffe782fac8acb42ab186243573 |
| Capture date | 2026-07-08 |
| Latest reported by installer at capture | 2.1.204 |
The capture host retained three versioned executables:
| Version | Size | SHA-256 |
|---|---|---|
2.1.174 |
224,051,232 | 20c5380b4423be9963c510f5464cc1f443235a9b4423179f9c01f28021b81bad |
2.1.175 |
224,216,352 | 6b75bf132c866ed409bf913c318ca32011e73ffb12d3cd67ecc37bc4ee9ec65d |
2.1.177 |
225,124,512 | eb0730351be2f02b482b1855870f5877489085aac86b0c4c1db4e458d9e40ed9 |
These identities are recorded in installed-versions.json, but only 2.1.177 has a committed full inventory and anchor set. The older files have not yet been interpreted as a behavioral version diff.
Major limitations¶
Platform coverage¶
Only the macOS arm64 native build was parsed. Linux, Windows, x86_64, WSL, and package-manager distributions may have different native modules, sandbox implementations, paths, signatures, and integration code.
Server-side opacity¶
The artifact does not reveal model implementation, server authorization, account flags, remote policy, retention, or provider infrastructure. Endpoint and client-schema anchors are not server source.
Controlled runtime coverage only¶
Published dynamic evidence comes from narrow synthetic scenarios against the
exact artifact: a loopback model provider, temporary homes and projects, dummy
credentials, disposable settings, a fixture MCP server, and bounded filesystem
effects. The probes exercised provider streaming, a Read → Bash tool loop,
session persistence, one scalar settings key, sibling command hooks, MCP stdio,
explicit extension discovery, permission rules, and sandbox containment. They
did not log into production accounts, contact Anthropic services, exercise
remote control, or establish universal defaults.
Bundling and minification¶
Most application logic is one bundled CJS module with no source map in the Bun graph. Descriptive reconstruction boundaries are ours. An anchor can identify a behavior without recovering original names or file topology.
Feature gating¶
Code can be dormant, experiment-gated, account-gated, platform-specific, or unreachable in the selected entrypoint. String presence is not activation. CLI presence is not account entitlement.
Configuration variability¶
Managed policy, user/project/local settings, environment, plugins, MCP servers, cloud credentials, and remote state can materially change effective behavior with identical executable bytes.
Confidence erosion over time¶
Documentation can become stale in three ways:
- a new executable changes local behavior;
- a server changes behavior used by the old executable;
- legal terms or public product documentation change.
Artifact facts remain historically valid, while deployment advice and legal links require periodic review.
Version-diff procedure¶
flowchart LR
accTitle: Versions and Limitations - Version-diff procedure
accDescr: Diagram showing version-diff procedure in the Versions and Limitations section.
Old["Old signed artifact"] --> OI["Inventory + anchors"]
New["New signed artifact"] --> NI["Inventory + anchors"]
OI --> Diff["Compare graph, hashes, CLI, anchors"]
NI --> Diff
Diff --> Class["Unchanged / changed / absent / new / unresolved"]
Class --> Review["Private security review"]
Review --> Publish["Behavioral summary after disclosure checks"]
Do not diff only pretty-printed bundle text. Stable layers include signature/build metadata, module graph, content hashes, help captures, semantic anchor resolution, and independently reconstructed contracts.
Research coverage and open gaps¶
| Surface | Current status | Remaining gap |
|---|---|---|
| Configuration precedence | Partially exercised | One scalar model selector was tested across persisted, explicit-settings, and CLI layers; per-key merge behavior and the complete schema remain open. |
| Runtime tool registry | Partially exercised | Read and Bash were exercised; schemas and mode-dependent catalogs for the remaining tools are unexercised. |
| Permission rules | Partially exercised | dontAsk denial and an explicit --allowedTools Bash allow were contrasted; full grammar and canonicalization remain open. |
| Hooks | Partially exercised | Sibling PreToolUse command hooks were observed dispatching concurrently; the full payload/result matrix and cross-event ordering remain open. |
| Stream JSON | Observed in named probes | Core wrapper and provider event ordering were captured; cancellation, malformed streams, retries, and compaction remain open. |
| MCP | Observed over fixture stdio | Initialization, discovery, and one tool call were captured; other transports, auth, elicitation, and failure paths remain open. |
| Sessions | Partially exercised | A mode-0600 JSONL transcript shape was captured; encryption, retention, migration, and external stores remain open. |
| IDE, Chrome, and remote control | Unexercised dynamically | IPC framing, peer authentication, approval flows, and lifecycle behavior. |
| Native add-ons | Unexercised dynamically | N-API interfaces and OS permission flows. |
| Provider resilience | Unexercised dynamically | Retry, fallback, cancellation, rate-limit, and error normalization behavior. |
| Telemetry | Unexercised dynamically | Effective payloads, redaction, batching, and disabled-mode guarantees. |
Each gap should be addressed with synthetic data and minimal authority. Security-relevant results must pass disclosure review before publication.