T3 Code

Permission-mode docs clearly distinguish Supervised/Auto-accept/Auto/Full access, note the default is Full access, and explain provider-specific enforcement, giving real substance on least privilege and confirmation (2). Deductions: default full access, projects explicitly not filesystem sandboxes, and filesystem:read can read absolute paths. Data-flow transparency is strong in environment-auth (scopes, DPoP, separate MCP audience, signed asset URLs) (2), but signed URLs survive revocation until expiry and are not session-bound. Sensitive data handling notes secrets marked sensitive are not redisplayed and Keychain access is opt-in (2), yet storage-encryption detail is absent. Dependency security shows only pinned versions and CI, no audit/SBOM evidence (1). External effects: install via curl|sh and irm|iex remote execution is a high-risk default (1). Rollback is limited to t3 uninstall and git pull rebuild, with no data/state rollback (1). Source attribution is only the MIT copyright line, no third-party notices (1).

Docs are internally consistent across permission modes, provider differences, install and update guidance (2). Dependency availability is demanding: Node 24, pnpm 11, global Vite+ vp, multiple provider CLIs, and no Intel Mac binary requiring source build (1). Failure messages are good: staged T3 Connect troubleshooting, error-code table, and diagnostics copy (2).

Audience and scenarios cover desktop/web/mobile/remote/background service clearly (2). Capability boundaries are well stated: non-sandbox projects, provider differences, Muse/Grok mode limits, Intel Mac limits (2). Trigger precision is weak: per-thread modes, full-access default, and complex provider fallbacks make it hard to predict when approval is required (1). Environment fit covers WSL, Tailscale, proxies, and port 7844 (2).

Information architecture is clear with README plus docs/user and docs/internals (2). Install notes are thorough across package managers, CLI, desktop, mobile, and Intel Mac source build (3). Naming stability is low: self-described very early, expect bugs, mostly not accepting contributions (1). Examples and FAQ are sparse, mostly procedural steps (1). Known limitations have dedicated sections and provider caveats (2). License is full MIT text (3). Versioning/changelog is absent from provided files beyond nightly/version env vars (1). Maintenance responsibility is unclear: unverified publisher and closed contribution policy (1).

Output usability is solid: CLI, desktop, mobile, and remote control with actionable docs (2). Marginal value is a unified multi-provider agent control surface with usage aggregation (2). Cost-benefit is weakened by full-access default, remote tunnel, heavy dependency install, and early-stage stability (1).

claim_traceability: docs cite concrete source paths (auth/http.ts, McpOAuth.ts) but those sources are not in the provided files, so claims cannot be checked (1). cross_source_corroboration: README and docs corroborate install and permission descriptions, but no test or run evidence is provided (1). fact_inference_separation: docs are largely design statements without separating verified fact from inference (1).