Gemini CLI 0.47.0 — The Baseline Snapshot#
Published on July 14, 2026
Part of the Gemini CLI Version Tracker series. | Repository
Gemini CLI 0.47.0 is the first release in this tracker, so there is no prior version to compare against. Its source carries 288 environment-variable literals, 39 model-ID strings, and 25 experimental feature flags, recorded here as the baseline every later post diffs against. From the Gemini CLI source at tag v0.47.0.
The Baseline in Numbers#
| Metric | v0.47.0 |
|---|---|
| Environment-variable literals | 288 (270 GEMINI_*, 18 GOOGLE_*) |
| Model-ID strings | 39 |
| Experimental feature flags | 25 |
| Source files (packages) | 1,983 |
| Integration tests / eval files | 52 / 43 |
What the Raw Counts Include#
Most of the 288 literals are not settings you can change. Of the 220 that carry the GEMINI_CLI_ prefix, about 160 are telemetry attribute keys the tool attaches to spans and metrics to record what it did. GEMINI_CLI_API_RESPONSE_INPUT_TOKEN_COUNT, GEMINI_CLI_TOOL_CALL_DURATION_MS, and GEMINI_CLI_ROUTING_DECISION are examples. Another 13 are test fixtures or bare prefixes, such as GEMINI_CLI_FOO, GEMINI_PROMPT_, and GEMINI_CLI_TEST_VAR.
That leaves roughly 115 real configuration variables. Most of the environment surface exists to record what the agent did rather than change how it behaves.
Configuration Knobs by Area#
The variables you can set group into a handful of concerns.
| Area | What it controls | Representative variables |
|---|---|---|
| Authentication | Which credential the CLI presents | GEMINI_API_KEY, GOOGLE_API_KEY, GOOGLE_APPLICATION_CREDENTIALS, GOOGLE_CREDENTIALS |
| Backend | API key vs Vertex AI vs Cloud Code Assist | GOOGLE_GENAI_USE_VERTEXAI, GOOGLE_GENAI_USE_GCA, GOOGLE_CLOUD_PROJECT, GOOGLE_GEMINI_BASE_URL, GOOGLE_VERTEX_BASE_URL |
| Model and routing | Which model runs and how turns are routed | GEMINI_MODEL, GEMINI_MODEL_ALIAS_PRO, GEMINI_MODEL_ALIAS_FLASH, GEMINI_MODEL_ALIAS_FLASH_LITE, GEMINI_MODEL_ALIAS_AUTO |
| Sandbox and trust | Isolation and folder trust before execution | GEMINI_SANDBOX, GEMINI_SANDBOX_IMAGE, GEMINI_FOLDER_TRUST, GEMINI_CLI_APPROVAL_MODE, GEMINI_YOLO_MODE |
| Extensions | Installing extensions from a registry | GEMINI_CLI_EXTENSION_REGISTRY_URI, GEMINI_CLI_EXTENSION_SOURCE, GEMINI_CLI_COMPANION_EXTENSION_NAME |
| IDE integration | The editor companion server | GEMINI_CLI_IDE_SERVER_PORT, GEMINI_CLI_IDE_SERVER_STDIO_COMMAND, GEMINI_CLI_IDE_WORKSPACE_PATH |
| Telemetry | OpenTelemetry export and limits | GEMINI_TELEMETRY_ENABLED, GEMINI_TELEMETRY_OTLP_ENDPOINT, GEMINI_TELEMETRY_LOG_PROMPTS, GEMINI_STRICT_TELEMETRY_LIMITS |
| Storage and paths | Where config, credentials, and plans live | GEMINI_CONFIG_DIR, GEMINI_CLI_CREDENTIALS_PATH, GEMINI_PLANS_DIR, GEMINI_FORCE_ENCRYPTED_FILE_STORAGE |
| Debug and context | Tracing and runtime monitoring | GEMINI_DEBUG, GEMINI_CONTEXT_TRACE_ENABLED, GEMINI_EVENT_LOOP_MONITOR_ENABLED, GEMINI_MEMORY_MONITOR_INTERVAL |
Model routing and authentication carry the most machinery. The four GEMINI_MODEL_ALIAS_* variables map onto tiers named auto, pro, flash, and flash-lite. Telemetry keys such as GEMINI_CLI_ROUTING_CLASSIFIER_THRESHOLD, GEMINI_CLI_ROUTING_DECISION_SOURCE, and GEMINI_CLI_LLM_LOOP_CHECK_MAIN_MODEL describe a router: the CLI picks which model answers each turn, and a second model checks whether a run has fallen into a loop. For auth, a plain API key, Vertex AI, Cloud Code Assist, and Google One are all wired in, so the same CLI adapts to whatever Google account a user has.
Model IDs#
The source lists 39 model strings, spanning the 1.0, 1.5, 2.0, 2.5, 3.0, 3.1, and 3.5 families. Several are clearly test values, including gemini-3.js, gemini-9001-super-duper, and gemini-3-test-only-model-string-for-testing.
| Model ID | Note |
|---|---|
gemini-3-pro, gemini-3.1-pro | Top reasoning tier the router maps to PRO |
gemini-3.5-flash | Newest fast tier; the release notes wire auto mode to it |
gemini-2.5-flash-lite, gemini-3.1-flash-lite | Cheapest tier the router falls back to |
gemini-2.5-computer-use-preview-10-2025 | Computer-use model for driving a screen |
gemini-2.5-flash-image | Image input and generation |
gemini-3.1-flash-live-preview | A live, streaming variant |
The source names gemini-3.5-flash and gemini-3.1-pro, versions ahead of what most people run today. Later posts will track which new model IDs appear here first.
Experimental Subsystems#
The 25 feature flags name the parts of the harness still marked experimental. Extensions get a registry and a config format (experimental.extensionRegistry, experimental.extensionConfig). Subagents show up as experimental.properties.enableAgents, experimental.properties.adk for the Agent Development Kit, and experimental.codebaseInvestigatorSettings for a search agent. Context handling has its own set: experimentalAgentHistorySummarization, experimentalAgentHistoryTruncation, and experimentalContextManagementConfig. The rest cover plan mode, automatic memory, git worktrees, voice input with a Whisper backend, ACP for editor clients, and experimental.gemmaModelRouter, which can route turns to Google's open Gemma models instead of Gemini.
The v0.47.0 Release Notes#
The published notes for the jump from v0.46.0 to v0.47.0 are mostly small fixes. Auto mode was switched to use gemini-3.5-flash, the policy engine gained an EBUSY fallback and TOML parse recovery, MCP tool discovery became atomic, and a Vertex model-mapping bug was fixed. Two entries point somewhere larger: the notes add migration commands and a transition banner for an "Antigravity CLI." What Antigravity is beyond a rename or successor path is unknown from the notes alone, and it is worth watching in the next release.
What the Shape Says About Google's Priorities#
Instrumentation comes first. More than half the environment literals exist to measure the agent: per-turn token counts, routing decisions, tool-call outcomes, agent recovery attempts, hook timings, and billing credit balance. That is the telemetry footprint of a measured product with a full analytics pipeline.
Model routing runs deep in this build. Four named tiers, a routing classifier with a tunable threshold, a loop-check model, and a Gemma fallback all ship. The pieces describe a router that picks a model per turn and can fall back to a cheaper tier.
Auth spans an API key, Vertex AI, Cloud Code Assist, and Google One, so the same binary serves a solo user and a Google Cloud enterprise. An extension registry, a companion IDE server, and GitHub Actions variables (GEMINI_CLI_GH_PR_NUMBER and others) point the CLI at editors and CI alongside the terminal.
This is the reference point. The next tracked Gemini CLI release will be reported as a diff against these numbers: new environment variables, new model IDs, and any feature flag that graduates out of experimental.
This analysis is based on publicly available open-source code and release metadata, conducted for educational and research purposes. All trademarks and software referenced belong to their respective owners.