History cut

Gemini CLI

Google's open-source CLI for Gemini-powered agents. The moments that actually shaped how this agent behaves: model shifts, new tools, workflow rules, capability boundaries.

18 milestones 144 versions scanned Generated Jul 15, 2026
  1. 01
    2026
    of 18
    v0.10.0 major

    Shell token limits and examples removed

    A new "Shell tool output token efficiency" section was added, prescribing verbose-suppression flags and temp-file redirection to a hardcoded session path to curb excessive token consumption. The interactive-commands guidance was tightened to a firmer prohibition rather than a soft suggestion. The entire illustrative examples block (nine annotated dialogues) was deleted from the prompt.

    • promptsPrompt block / 0: -2,427 chars
    Promptstool-useworkflowenvironment
    Split-view diff2 fields changedprompts · 2
    FieldBeforeAfter
    promptstotal prompt size17,889 chars15,462 chars (-2,427)
    promptsblock 0previous texttext changed (-2,427 chars)
  2. 02
    2026
    of 18
    v0.13.0 notable

    Introduces codebase_investigator as primary tool

    The "Understand" step was renamed to "Understand & Strategize" and now mandates `codebase_investigator` as the first tool for complex refactoring, codebase exploration, or system-wide analysis, reserving `search_file_content`/`glob` for simple targeted lookups. The plan step was updated to require that `codebase_investigator` output form the foundation of any plan when it was used. Temp directory paths were also rotated to new session values, and minor blank-line and punctuation noise was cleaned up.

    • toolsTool added / codebase_investigator: new tool description
    • schemasSchema added / codebase_investigator: new input schema
    • promptsPrompt block / 0: +302 chars
    PromptsSchemasToolstool-useworkflowenvironment
    Split-view diff4 fields changedtools · 2prompts · 2
    FieldBeforeAfter
    toolscodebase_investigator—added
    toolstool count1112
    promptstotal prompt size15,462 chars15,764 chars (+302)
    promptsblock 0previous texttext changed (+302 chars)
  3. 03
    2026
    of 18
    v0.15.0 notable

    Todo tracking and Angular added to workflow

    The planning step now instructs the assistant to use `write_todos` for breaking complex tasks into tracked subtasks, formalizing a lightweight project-management loop. Angular was added alongside React as a recommended frontend and full-stack frontend option. Two smaller changes also landed: hardcoded session-specific temp directory paths were replaced with a generic `<temp_dir>` placeholder, and `read_many_files` was removed from the final reminder's list of approved file-reading tools.

    • toolsTool added / write_todos: new tool description
    • schemasSchema added / write_todos: new input schema
    • toolsTool description / replace: description +196 chars
    PromptsSchemasToolsworkflowtool-useenvironment
    Split-view diff5 fields changedtools · 3prompts · 2
    FieldBeforeAfter
    toolswrite_todos—added
    toolsreplace / descriptionprevious descriptiondescription changed (+196 chars)
    toolstool count1213
    promptstotal prompt size15,045 chars14,725 chars (-320)
    promptsblock 0previous texttext changed (-320 chars)
  4. 04
    2026
    of 18
    v0.20.0 major

    Agent recast as fully non-interactive

    The agent's self-description changed from "interactive" to "non-interactive," and all prompts to pause and confirm with the user were removed throughout the prompt. Ambiguity handling, user-approval gates for new-app plans, clarification questions, and feedback solicitation steps were all stripped out, replaced with an explicit directive to continue work autonomously using best judgment. Several guidelines (background processes, interactive commands, memory saving) were also shortened by removing their user-facing explanations and escape hatches.

    • promptsPrompt block / 0: -753 chars
    Promptspersonaworkflowpermissions
    Split-view diff2 fields changedprompts · 2
    FieldBeforeAfter
    promptstotal prompt size14,725 chars13,972 chars (-753)
    promptsblock 0previous texttext changed (-753 chars)
  5. 05
    2026
    of 18
    v0.22.0 notable

    Adds sub-agent delegation section

    A new "Available Sub-Agents" section was introduced, documenting the `delegate_to_agent` tool and the `codebase_investigator` sub-agent with its purpose and return format. The Step 1 of the Software Engineering workflow was reworked: the strong mandate to use `codebase_investigator` as the "first and primary tool" for complex tasks was replaced with a broader directive to use `search_file_content`, `glob`, and `read_file` extensively and in parallel. This shifts the default investigation strategy away from the specialized sub-agent toward direct file-exploration tools, while still advertising `codebase_investigator` as an available resource via `delegate_to_agent`.

    • toolsTool added / delegate_to_agent: new tool description
    • schemasSchema added / delegate_to_agent: new input schema
    • toolsTool removed / codebase_investigator: tool removed
    PromptsSchemasToolsTransporttool-useworkflow
    Split-view diff5 fields changedtools · 2prompts · 2transport · 1
    FieldBeforeAfter
    toolsdelegate_to_agent—added
    toolscodebase_investigatorpresent—
    promptstotal prompt size13,971 chars14,288 chars (+317)
    promptsblock 0previous texttext changed (+317 chars)
    transportheader / user_agentGeminiCLI/v22.22.2 (linux; x64)GeminiCLI/0.22.0/gemini-2.5-pro (linux; x64)
  6. 06
    2026
    of 18
    v0.22.1 notable

    codebase_investigator promoted to primary tool

    The separate "Available Sub-Agents" section describing `codebase_investigator` was removed, and its guidance was folded directly into Step 1 of the Software Engineering workflow. The step was renamed "Understand & Strategize" and now mandates `codebase_investigator` as the first and primary tool for complex refactoring or system-wide analysis, while reserving simpler search tools for targeted lookups. Step 2 was also updated to require that any `codebase_investigator` output serve as the foundation of the plan.

    • toolsTool added / codebase_investigator: new tool description
    • schemasSchema added / codebase_investigator: new input schema
    • toolsTool removed / delegate_to_agent: tool removed
    PromptsSchemasToolsTransportworkflowtool-usereorganization
    Split-view diff5 fields changedtools · 2prompts · 2transport · 1
    FieldBeforeAfter
    toolscodebase_investigator—added
    toolsdelegate_to_agentpresent—
    promptstotal prompt size14,288 chars13,971 chars (-317)
    promptsblock 0previous texttext changed (-317 chars)
    transportheader / user_agentGeminiCLI/0.22.0/gemini-2.5-pro (linux; x64)GeminiCLI/v22.22.2 (linux; x64)
  7. 07
    2026
    of 18
    v0.22.2 notable

    Sub-agent registry and revised workflow step

    A new "Available Sub-Agents" section was added, formally documenting the `delegate_to_agent` tool and the `codebase_investigator` sub-agent with a description of when and how to invoke it. The "Understand & Strategize" workflow step was simultaneously rewritten to de-emphasize `codebase_investigator` as the mandatory first tool for complex tasks, replacing it with a broader directive to use `search_file_content`, `glob`, and `read_file` extensively and in parallel. This shifts the agent's default investigation strategy toward direct file exploration while the sub-agent capability is surfaced in its own dedicated section.

    • toolsTool added / delegate_to_agent: new tool description
    • schemasSchema added / delegate_to_agent: new input schema
    • toolsTool removed / codebase_investigator: tool removed
    PromptsSchemasToolsTransporttool-useworkflow
    Split-view diff5 fields changedtools · 2prompts · 2transport · 1
    FieldBeforeAfter
    toolsdelegate_to_agent—added
    toolscodebase_investigatorpresent—
    promptstotal prompt size13,971 chars14,288 chars (+317)
    promptsblock 0previous texttext changed (+317 chars)
    transportheader / user_agentGeminiCLI/v22.22.2 (linux; x64)GeminiCLI/0.22.2/gemini-2.5-pro (linux; x64)
  8. 08
    2026
    of 18
    v0.26.0 major

    Adds skill system and stronger sub-agent guidance

    A new `activate_skill` / `available_skills` framework was introduced, giving the assistant a mechanism to load expert procedural instructions at runtime via `<activated_skill>` tags. The sub-agent section was substantially expanded with concrete delegation examples and a standing directive to *always* use `delegate_to_agent` when a relevant agent exists. A minor clarification was also added requiring `git --no-pager` as an example of non-interactive command usage.

    • toolsTool description / activate_skill: description +29 chars
    • promptsPrompt block / 0: +2,068 chars
    • transportHeader / user_agent: GeminiCLI/0.25.2/gemini-2.5-pro (linux; x64) -> GeminiCLI/0.26.0/gemini-2.5-pro (linux; x64)
    PromptsToolsTransporttool-useworkflowpermissions
    Split-view diff4 fields changedprompts · 2tools · 1transport · 1
    FieldBeforeAfter
    toolsactivate_skill / descriptionprevious descriptiondescription changed (+29 chars)
    promptstotal prompt size14,549 chars16,617 chars (+2,068)
    promptsblock 0previous texttext changed (+2,068 chars)
    transportheader / user_agentGeminiCLI/0.25.2/gemini-2.5-pro (linux; x64)GeminiCLI/0.26.0/gemini-2.5-pro (linux; x64)
  9. 09
    2026
    of 18
    v0.27.0 notable

    Stricter implicit-change confirmation required

    The ambiguity-handling rule was extended to explicitly prohibit acting on implied changes (e.g., a reported bug) without user confirmation, and the planning step now mandates asking before modifying code when a request only implies a change. Sub-agent delegation instructions were restructured from a `delegate_to_agent` call pattern to direct tool invocation, with simplified agent listings and stronger guidance to use the closest relevant agent even for broader tasks. A new Hook Context section was added, instructing the model to treat `<hook_context>` content as read-only data and never as overriding instructions.

    • toolsTool added / cli_help: new tool description
    • schemasSchema added / cli_help: new input schema
    • toolsTool added / codebase_investigator: new tool description
    PromptsSchemasToolsTransportworkflowsafetytool-use
    Split-view diff8 fields changedtools · 5prompts · 2transport · 1
    FieldBeforeAfter
    toolscli_help—added
    toolscodebase_investigator—added
    toolsdelegate_to_agentpresent—
    toolsrun_shell_command / descriptionprevious descriptiondescription changed (-220 chars, -3 lines)
    toolstool count1314
    promptstotal prompt size16,617 chars16,548 chars (-69)
    promptsblock 0previous texttext changed (-69 chars)
    transportheader / user_agentGeminiCLI/0.26.0/gemini-2.5-pro (linux; x64)GeminiCLI/0.27.0/gemini-2.5-pro (linux; x64)
  10. 10
    2026
    of 18
    v0.28.0 notable

    Rename search tool to grep_search

    The workflow instruction for the "Understand" step was updated to reference `grep_search` instead of `search_file_content` as the recommended search tool. The skill-creator path was also bumped from version 0.27.4 to 0.28.0, reflecting the package version increment.

    • toolsTool added / ask_user: new tool description
    • schemasSchema added / ask_user: new input schema
    • toolsTool added / grep_search: new tool description
    PromptsSchemasToolsTransporttool-useworkflowbugfix
    Split-view diff8 fields changedtools · 5prompts · 2transport · 1
    FieldBeforeAfter
    toolsask_user—added
    toolsgrep_search—added
    toolssearch_file_contentpresent—
    toolsrun_shell_command / descriptionprevious descriptiondescription changed (+68 chars)
    toolstool count1415
    promptstotal prompt size16,548 chars16,540 chars (-8)
    promptsblock 0previous texttext changed (-8 chars)
    transportheader / user_agentGeminiCLI/0.27.4/gemini-2.5-pro (linux; x64)GeminiCLI/0.28.0/gemini-2.5-pro (linux; x64)
  11. 11
    2026
    of 18
    v0.29.0 major

    Major persona, workflow, and safety overhaul

    The agent was formally named "Gemini CLI" and given an explicit senior-engineer persona. The primary workflow was restructured from a linear 6-step process into a Research → Strategy → Execution lifecycle with iterative Plan/Act/Validate cycles, and new Directive vs. Inquiry intent-classification semantics were introduced. Security mandates (credential protection, source-control guardrails), sub-agent XML descriptors, stricter memory-tool scoping, and a mandatory "Explain Before Acting" transparency rule were all added, while the "Outside of Sandbox" section and verbose shell-output guidelines were removed.

    • toolsTool removed / ask_user: tool removed
    • schemasSchema removed / ask_user: input schema removed
    • toolsTool removed / write_todos: tool removed
    ModelPromptsSchemasToolsTransportpersonaworkflowsafety
    Split-view diff12 fields changedtools · 5transport · 3prompts · 2model · 2
    FieldBeforeAfter
    toolsask_userpresent—
    toolswrite_todospresent—
    toolsrun_shell_command / descriptionprevious descriptiondescription changed (+363 chars, +4 lines)
    toolssave_memory / descriptionprevious descriptiondescription changed (-530 chars, -9 lines)
    toolstool count1513
    promptstotal prompt size16,540 chars15,934 chars (-606)
    promptsblock 0previous texttext changed (-606 chars)
    modelmodelgemini-2.5-flashgemini-3-flash-preview
    modelthinking{"includeThoughts": true, "thinkingBudget": 8192}{"includeThoughts": true, "thinkingLevel": "HIGH"}
    transportheader / user_agentGeminiCLI/0.28.2/gemini-2.5-pro (linux; x64)GeminiCLI/0.29.0/gemini-3-pro-preview (linux; x64)
    transportendpoint / path/v1beta/models/gemini-2.5-flash:streamGenerateContent/v1beta/models/gemini-3-flash-preview:streamGenerateContent
    transportendpoint / queryalt=ssealt=sse
  12. 12
    2026
    of 18
    v0.30.0 notable

    Context limits and scaffolding guardrails added

    A new "Context Efficiency" section was added prescribing search-scoping strategies (e.g., `names_only`, `total_max_matches`) to prevent context window exhaustion. The prototype-build workflow gained a mandatory requirement to pass non-interactive flags when running scaffolding tools, preventing hanging environments. Minor wording and formatting cleanups (plan step rewording, blank-line removal, version path bump) round out the release.

    • toolsTool description / grep_search: description -17 chars
    • toolsTool description / save_memory: description -90 chars
    • promptsPrompt block / 0: +584 chars
    PromptsToolsTransportcontextworkflowenvironment
    Split-view diff5 fields changedtools · 2prompts · 2transport · 1
    FieldBeforeAfter
    toolsgrep_search / descriptionprevious descriptiondescription changed (-17 chars)
    toolssave_memory / descriptionprevious descriptiondescription changed (-90 chars)
    promptstotal prompt size15,934 chars16,518 chars (+584)
    promptsblock 0previous texttext changed (+584 chars)
    transportheader / user_agentGeminiCLI/0.29.7/gemini-3.1-pro-preview (linux; x64)GeminiCLI/0.30.0/gemini-3.1-pro-preview (linux; x64)
  13. 13
    2026
    of 18
    v0.31.0 major

    Major context efficiency guidance expansion

    The context efficiency section was substantially rewritten, replacing two terse bullet points with a detailed structured guide covering cost estimation, search/read optimization strategies, and concrete usage examples. A new "User Hints" directive was also added under Engineering Standards, instructing the assistant to treat real-time user hints during execution as high-priority, scope-preserving course corrections. The skill-creator path was bumped from version 0.30.1 to 0.31.0.

    • toolsTool description / google_web_search: description +218 chars
    • toolsTool description / grep_search: description +238 chars
    • toolsTool description / read_file: description +7 chars
    PromptsToolsTransportcontextworkflowbugfix
    Split-view diff10 fields changedtools · 7prompts · 2transport · 1
    FieldBeforeAfter
    toolsgoogle_web_search / descriptionprevious descriptiondescription changed (+218 chars)
    toolsgrep_search / descriptionprevious descriptiondescription changed (+238 chars)
    toolsread_file / descriptionprevious descriptiondescription changed (+7 chars)
    toolsreplace / descriptionprevious descriptiondescription changed (-1,678 chars, -10 lines)
    toolssave_memory / descriptionprevious descriptiondescription changed (-60 chars, -7 lines)
    toolsweb_fetch / descriptionprevious descriptiondescription changed (+170 chars)
    toolswrite_file / descriptionprevious descriptiondescription changed (+97 chars, -2 lines)
    promptstotal prompt size16,518 chars19,473 chars (+2,955)
    promptsblock 0previous texttext changed (+2,955 chars)
    transportheader / user_agentGeminiCLI/0.30.1/gemini-3.1-pro-preview (linux; x64)GeminiCLI/0.31.0/gemini-3.1-pro-preview (linux; x64)
  14. 14
    2026
    of 18
    v0.32.0 major

    New generalist sub-agent and orchestration guidance

    A new `generalist` sub-agent was added for turn-intensive or high-volume tasks, paired with an expanded "Strategic Orchestration & Delegation" section that instructs the model to treat its context window as a scarce resource and delegate batch, verbose, or speculative work to sub-agents. A minor safety rule was appended prohibiting use of `ask_user` to seek permission before running shell commands. A parameter name fix (`include`/`exclude` → `include_pattern`/`exclude_pattern`) and a version-bumped skill path were also included.

    • toolsTool added / generalist: new tool description
    • schemasSchema added / generalist: new input schema
    • promptsPrompt block / 0: +1,672 chars
    PromptsSchemasToolsTransportworkflowtool-usesafety
    Split-view diff5 fields changedtools · 2prompts · 2transport · 1
    FieldBeforeAfter
    toolsgeneralist—added
    toolstool count1314
    promptstotal prompt size19,473 chars21,145 chars (+1,672)
    promptsblock 0previous texttext changed (+1,672 chars)
    transportheader / user_agentGeminiCLI/0.31.0/gemini-3.1-pro-preview (linux; x64)GeminiCLI/0.32.0/gemini-3.1-pro-preview (linux; x64)
  15. 15
    2026
    of 18
    v0.36.0 notable

    Plan-mode required for new app builds

    The app-scaffolding workflow now mandates use of `enter_plan_mode` before writing any code, replacing the prior implicit planning step. Research phase guidance was also extended to recommend `enter_plan_mode` for ambiguous or architecturally broad requests. Separately, a new engineering standard explicitly prohibits suppressing type-system warnings or using unsafe casts, directing use of idiomatic alternatives instead.

    • toolsTool added / enter_plan_mode: new tool description
    • schemasSchema added / enter_plan_mode: new input schema
    • toolsTool added / exit_plan_mode: new tool description
    PromptsSchemasToolsTransportworkflowtool-usesafety
    Split-view diff8 fields changedtools · 5prompts · 2transport · 1
    FieldBeforeAfter
    toolsenter_plan_mode—added
    toolsexit_plan_mode—added
    toolscli_help / descriptionprevious descriptiondescription changed (+159 chars)
    toolscodebase_investigator / descriptionprevious descriptiondescription changed (-8 chars, -2 lines)
    toolstool count1416
    promptstotal prompt size22,525 chars23,011 chars (+486)
    promptsblock 0previous texttext changed (+486 chars)
    transportheader / user_agentGeminiCLI/0.35.3/gemini-3.1-pro-preview (linux; x64; codespaces)GeminiCLI/0.36.0/gemini-3.1-pro-preview (linux; x64; codespaces)
  16. 16
    2026
    of 18
    v0.39.0 notable

    Adds retry cap and re-evaluation rule

    A new "Strategic Re-evaluation" rule mandates that after three failed fix attempts the agent must stop, audit its assumptions, and pivot to a different architectural approach rather than continuing to patch. The sub-agent invocation description was clarified to specify the `invoke_agent` tool and `agent_name` parameter explicitly. A skill-creator path was also updated to reflect a new local cache location.

    • toolsTool added / invoke_agent: new tool description
    • schemasSchema added / invoke_agent: new input schema
    • toolsTool removed / cli_help: tool removed
    PromptsSchemasToolsTransportworkflowtool-useenvironment
    Split-view diff8 fields changedtools · 5prompts · 2transport · 1
    FieldBeforeAfter
    toolsinvoke_agent—added
    toolscli_helppresent—
    toolscodebase_investigatorpresent—
    toolsgeneralistpresent—
    toolstool count1715
    promptstotal prompt size23,615 chars24,051 chars (+436)
    promptsblock 0previous texttext changed (+436 chars)
    transportheader / user_agentGeminiCLI/0.38.2/gemini-3.1-pro-preview (linux; x64; codespaces)GeminiCLI/0.39.0/gemini-3.1-pro-preview (linux; x64; codespaces)
  17. 17
    2026
    of 18
    v0.40.0 major

    Adds topic model, overhauled memory system

    The "Explain Before Acting" and post-change summary rules were replaced with a structured `update_topic` workflow that prescribes when and how to publish progress updates across multi-step tasks. The `save_memory` tool was removed entirely and replaced with a four-tier file-based memory system (project GEMINI.md, subdirectory GEMINI.md, private project MEMORY.md, and global personal GEMINI.md), complete with explicit routing rules for which facts belong in which tier. Tone/style references to the old "Explain Before Acting" mandate were updated to reference the new "Topic Model".

    • toolsTool added / update_topic: new tool description
    • schemasSchema added / update_topic: new input schema
    • toolsTool removed / save_memory: tool removed
    PromptsSchemasToolsTransportworkflowtool-usecontext
    Split-view diff6 fields changedtools · 3prompts · 2transport · 1
    FieldBeforeAfter
    toolsupdate_topic—added
    toolssave_memorypresent—
    toolsgrep_search / descriptionprevious descriptiondescription changed (-221 chars)
    promptstotal prompt size24,021 chars28,295 chars (+4,274)
    promptsblock 0previous texttext changed (+4,274 chars)
    transportheader / user_agentGeminiCLI/0.39.1/gemini-3.1-pro-preview (linux; x64; codespaces)GeminiCLI/0.40.0/gemini-3.1-pro-preview (linux; x64; codespaces)
  18. 18
    2026
    of 18
    v0.44.0 major

    Major prompt restructure and persona reset

    The agent's self-description was stripped of "Gemini CLI" branding and YOLO-mode framing, replaced with a generic non-interactive CLI agent identity. Extensive sections on context-efficiency strategy, orchestration heuristics, GEMINI.md precedence rules, and the Research→Strategy→Execution lifecycle were removed or heavily condensed. New sections were added covering shell output token efficiency, an outside-sandbox warning, a final reminder, and inline guidance on comment discipline, change summaries, and revert behavior.

    • toolsTool description / run_shell_command: description -68 chars
    • promptsPrompt block / 0: -6,235 chars
    • transportHeader / user_agent: GeminiCLI-tui/0.43.0/gemini-3.1-pro-preview (linux; x64; codespaces) -> GeminiCLI-tui/0.44.0/gemini-2.5-pro (linux; x64; codespaces)
    PromptsToolsTransportpersonaworkflowreorganization
    Split-view diff4 fields changedprompts · 2tools · 1transport · 1
    FieldBeforeAfter
    toolsrun_shell_command / descriptionprevious descriptiondescription changed (-68 chars)
    promptstotal prompt size28,511 chars22,276 chars (-6,235)
    promptsblock 0previous texttext changed (-6,235 chars)
    transportheader / user_agentGeminiCLI-tui/0.43.0/gemini-3.1-pro-preview (linux; x64; codespaces)GeminiCLI-tui/0.44.0/gemini-2.5-pro (linux; x64; codespaces)