History cut

Codex CLI

OpenAI's terminal coding assistant. The moments that actually shaped how this agent behaves: model shifts, new tools, workflow rules, capability boundaries.

10 milestones 55 versions scanned Generated Jul 15, 2026
  1. 01
    2026
    of 10
    v0.99.0 major

    Major persona and formatting overhaul

    Version 0.99.0 introduced a detailed Personality section defining Codex's values (clarity, pragmatism, rigor), interaction style, and escalation behavior. Formatting and presentation rules were reorganized from a single trailing section into structured guidance embedded earlier in the prompt, with several rules tightened or rewritten. A new constraint on non-interactive git commands was also added.

    • promptsPrompt block / 0: +1,576 chars
    • promptsPrompt block / 1: -598 chars
    • transportHeader / user_agent: codex_exec/0.98.0 (Debian 12.0.0; x86_64) vscode/1.115.0 -> codex_exec/0.99.0 (Debian 12.0.0; x86_64) vscode/1.115.0
    PromptsTransportpersonaformattingworkflow
    Split-view diff4 fields changedprompts · 3transport · 1
    FieldBeforeAfter
    promptstotal prompt size8,502 chars9,480 chars (+978)
    promptsblock 0previous texttext changed (+1,576 chars)
    promptsblock 1previous texttext changed (-598 chars)
    transportheader / user_agentcodex_exec/0.98.0 (Debian 12.0.0; x86_64) vscode/1.115.0codex_exec/0.99.0 (Debian 12.0.0; x86_64) vscode/1.115.0
  2. 02
    2026
    of 10
    v0.106.0 major

    Major persona, workflow, and formatting overhaul

    The persona was stripped of its "quiet joy/enthusiasm" framing in favor of purely direct, factual communication, and cheerleading-acknowledgment language was removed entirely. The "Working with the user" section was substantially rewritten to introduce a two-channel model (commentary vs. final), explicit autonomy/persistence directives, detailed intermediary-update cadence rules, and revised file-reference conventions (markdown links instead of inline code). Several sections were removed or restructured, including the plan-tool guidance block, while tool-use gained a new parallel-calls directive and an instruction to avoid Python for simple file I/O.

    • promptsPrompt block / 0: +3,164 chars
    • modelModel setting / model: gpt-5.2-codex -> gpt-5.3-codex
    • transportHeader / user_agent: codex_exec/0.105.0 (Debian 12.0.0; x86_64) vscode/1.115.0 -> codex_exec/0.106.0 (Debian 12.0.0; x86_64) vscode/1.115.0
    ModelPromptsTransportpersonaworkflowformatting
    Split-view diff4 fields changedprompts · 2model · 1transport · 1
    FieldBeforeAfter
    promptstotal prompt size9,480 chars12,644 chars (+3,164)
    promptsblock 0previous texttext changed (+3,164 chars)
    modelmodelgpt-5.2-codexgpt-5.3-codex
    transportheader / user_agentcodex_exec/0.105.0 (Debian 12.0.0; x86_64) vscode/1.115.0codex_exec/0.106.0 (Debian 12.0.0; x86_64) vscode/1.115.0
  3. 03
    2026
    of 10
    v0.115.0 major

    Adds skill-discovery and execution system

    A new `<skills_instructions>` block introduces a full "Skills" subsystem, pre-loading three built-in skills (openai-docs, skill-creator, skill-installer) and defining detailed rules for triggering, loading, coordinating, and safely falling back on SKILL.md files. This is a significant new capability: the assistant can now discover and follow locally-stored instruction sets on demand, scoped per-turn. Engineers relying on consistent, stateless behavior should review the trigger and context-hygiene rules carefully.

    • toolsTool added / close_agent: new tool description
    • schemasSchema added / close_agent: new input schema
    • toolsTool added / resume_agent: new tool description
    PromptsSchemasToolsTransportworkflowtool-usecontext
    Split-view diff9 fields changedtools · 6prompts · 2transport · 1
    FieldBeforeAfter
    toolsclose_agentadded
    toolsresume_agentadded
    toolssend_inputadded
    toolsspawn_agentadded
    toolswait_agentadded
    toolstool count712
    promptstotal prompt size12,684 chars16,120 chars (+3,436)
    promptsblock 1previous texttext changed (+3,436 chars)
    transportheader / user_agentcodex_exec/0.114.0 (Debian 12.0.0; x86_64) vscode/1.115.0 (codex-exec; 0.114.0)codex_exec/0.115.0 (Debian 12.0.0; x86_64) vscode/1.115.0 (codex-exec; 0.115.0)
  4. 04
    2026
    of 10
    v0.117.0 notable

    Two new system skills added

    Two new built-in skills were registered in the available skills list: `imagegen`, for AI-generated or AI-edited raster/bitmap images, and `plugin-creator`, for scaffolding Codex plugin directories and marketplace metadata. These expand the assistant's out-of-the-box capability surface without changing any existing skills or workflow logic.

    • toolsTool description / close_agent: description +44 chars
    • schemasTool schema / close_agent: properties +target / properties -id / required changed
    • schemasTool schema / send_input: properties +target / properties -id / required changed
    PromptsSchemasToolstool-useworkflow
    Split-view diff7 fields changedschemas · 3tools · 2prompts · 2
    FieldBeforeAfter
    toolsclose_agent / descriptionprevious descriptiondescription changed (+44 chars)
    schemasclose_agenthad: id · required list changednow has: target · required list changed
    schemassend_inputhad: id · required list changednow has: target · required list changed
    toolsspawn_agent / descriptionprevious descriptiondescription changed (-27 chars)
    schemaswait_agenthad: ids · required list changednow has: targets · required list changed
    promptstotal prompt size16,120 chars17,261 chars (+1,141)
    promptsblock 1previous texttext changed (+1,141 chars)
  5. 05
    2026
    of 10
    v0.123.0 major

    Major persona, workflow, and formatting overhaul

    This version introduces a top-level expert-coding-agent persona statement, promotes `apply_patch` from preferred to mandatory for manual edits, and significantly rewrites the final-answer formatting rules to favor concise prose over bullets and long explanations. The autonomy/persistence section was moved earlier in the document, unexpected-change handling was softened from "stop immediately" to "only stop if it conflicts with your task," and bash command chaining (e.g. `echo "===="`) was explicitly banned. User-update frequency shifted from every 20s to every 30s, and React best-practice guidance was added to the frontend section.

    • toolsTool description / spawn_agent: description -381 chars
    • schemasTool schema / spawn_agent: schema changed
    • promptsPrompt block / 0: +2,389 chars
    ModelPromptsSchemasToolspersonaworkflowformatting
    Split-view diff6 fields changedprompts · 3tools · 1schemas · 1model · 1
    FieldBeforeAfter
    toolsspawn_agent / descriptionprevious descriptiondescription changed (-381 chars, -1 lines)
    schemasspawn_agentprevious schemaupdated schema
    promptstotal prompt size17,261 chars19,650 chars (+2,389)
    promptsblock 0previous texttext changed (+2,389 chars)
    promptsblock 1previous texttext changed (0 chars net)
    modelmodelgpt-5.3-codexgpt-5.4
  6. 06
    2026
    of 10
    v0.125.0 major

    Adds extensive frontend and engineering-judgment guidance

    The update adds two large new sections: a detailed "Engineering judgment" block prescribing conservative, codebase-sympathetic implementation choices, and a comprehensive "Frontend guidance" block covering design empathy, UI component rules, typography, layout, 3D/Three.js, and responsive constraints. The old "Frontend tasks" section (which pushed bold, anti-"AI slop" aesthetics) was removed and replaced by the more structured and prescriptive frontend guidance. Persona, workflow, and intermediary-update instructions were also extensively reworded in a more conversational first-person voice, and several new commentary rules were added (e.g., checklist incremental updates, anti-goblin/creature clause, and ghost-request sanity check on resume).

    • toolsTool description / spawn_agent: description +369 chars
    • promptsPrompt block / 0: +6,603 chars
    • promptsPrompt block / 1: +0 chars
    ModelPromptsToolsworkflowpersonaformatting
    Split-view diff5 fields changedprompts · 3tools · 1model · 1
    FieldBeforeAfter
    toolsspawn_agent / descriptionprevious descriptiondescription changed (+369 chars, +1 lines)
    promptstotal prompt size19,639 chars26,242 chars (+6,603)
    promptsblock 0previous texttext changed (+6,603 chars)
    promptsblock 1previous texttext changed (0 chars net)
    modelmodelgpt-5.4gpt-5.5
  7. 07
    2026
    of 10
    v0.133.0 notable

    Plugin-creator skill gains update flow

    The `plugin-creator` skill description was expanded to include a new use case: updating an existing local plugin during development using a CLI-driven cachebuster and reinstall flow. This broadens the skill's declared scope beyond initial creation and marketplace entry management to cover iterative plugin development.

    • toolsTool added / create_goal: new tool description
    • schemasSchema added / create_goal: new input schema
    • toolsTool added / get_goal: new tool description
    PromptsSchemasToolsTransporttool-useworkflow
    Split-view diff12 fields changedtools · 9prompts · 2transport · 1
    FieldBeforeAfter
    toolscreate_goaladded
    toolsget_goaladded
    toolsupdate_goaladded
    toolsclose_agentpresent
    toolsresume_agentpresent
    toolssend_inputpresent
    toolsspawn_agentpresent
    toolswait_agentpresent
    toolstool count1211
    promptstotal prompt size26,213 chars26,316 chars (+103)
    promptsblock 1previous texttext changed (+103 chars)
    transportheader / user_agentcodex_exec/0.132.0 (Debian 12.0.0; x86_64) vscode/1.121.0 (codex_exec; 0.132.0)codex_exec/0.133.0 (Debian 12.0.0; x86_64) vscode/1.121.0 (codex_exec; 0.133.0)
  8. 08
    2026
    of 10
    v0.136.0 notable

    openai-docs skill gains Codex self-knowledge

    The `openai-docs` skill description was updated to explicitly cover questions about Codex itself and choosing Codex surfaces, and its routing logic was refined to prefer a "Codex manual helper" for broad Codex self-knowledge questions before falling back to general OpenAI docs MCP tools. All other skill entries were unchanged in substance; only the temporary working-directory path segment changed across all entries.

    • schemasTool schema / create_goal: schema changed
    • schemasTool schema / exec_command: schema changed
    • schemasTool schema / update_plan: schema changed
    PromptsSchemasTransportworkflowcontextbugfix
    Split-view diff8 fields changedschemas · 5prompts · 2transport · 1
    FieldBeforeAfter
    schemascreate_goalprevious schemaupdated schema
    schemasexec_commandprevious schemaupdated schema
    schemasupdate_planprevious schemaupdated schema
    schemasview_imageprevious schemaupdated schema
    schemaswrite_stdinprevious schemaupdated schema
    promptstotal prompt size26,376 chars26,462 chars (+86)
    promptsblock 1previous texttext changed (+86 chars)
    transportheader / user_agentcodex_exec/0.135.0 (Debian 12.0.0; x86_64) vscode/1.122.1 (codex_exec; 0.135.0)codex_exec/0.136.0 (Debian 12.0.0; x86_64) vscode/1.122.1 (codex_exec; 0.136.0)
  9. 09
    2026
    of 10
    v0.140.0 major

    Skills system gains multi-locator support

    The skills framework was expanded beyond filesystem-only `SKILL.md` paths to support multiple resource locator types: `file`, `environment resource`, `orchestrator resource`, and `custom resource`. Reading instructions were substantially rewritten to handle each locator type distinctly, including using `skills.list`/`skills.read` for orchestrator-backed skills and explicit rules against treating `skill://` URIs as filesystem paths. Temporary file paths in skill entries were also scrubbed to a generic `<scrub>` placeholder, and the context-hygiene rule was clarified to distinguish progressive disclosure of files from partial reading of a selected instruction file.

    • toolsTool description / create_goal: description +12 chars
    • schemasTool schema / create_goal: schema changed
    • toolsTool description / request_user_input: description +230 chars
    PromptsSchemasToolsTransporttool-useworkflowenvironment
    Split-view diff7 fields changedtools · 2schemas · 2prompts · 2transport · 1
    FieldBeforeAfter
    toolscreate_goal / descriptionprevious descriptiondescription changed (+12 chars)
    schemascreate_goalprevious schemaupdated schema
    toolsrequest_user_input / descriptionprevious descriptiondescription changed (+230 chars)
    schemasrequest_user_inputprevious schemanow has: autoResolutionMs
    promptstotal prompt size26,462 chars27,781 chars (+1,319)
    promptsblock 1previous texttext changed (+1,319 chars)
    transportheader / user_agentcodex_exec/0.137.0 (Debian 12.0.0; x86_64) vscode/1.123.0 (codex_exec; 0.137.0)codex_exec/0.140.0 (Debian 12.0.0; x86_64) vscode/1.124.2 (codex_exec; 0.140.0)
  10. 10
    2026
    of 10
    v0.144.0 major

    Major persona and workflow overhaul

    The persona was substantially rewritten from a terse, pragmatic senior-engineer voice to a warmer, curiosity-driven communicator that adapts to user tone. Large swaths of frontend design rules, engineering judgment guidelines, and git/editing constraints were condensed or removed, replaced with a cleaner autonomy model and more explicit commentary/final-channel guidance. Skill-use instructions were restructured with alias resolution, clearer trigger rules, and new transparency requirements around skill announcements.

    • toolsTool removed / apply_patch: tool removed
    • schemasSchema removed / apply_patch: input schema removed
    • toolsTool removed / create_goal: tool removed
    ModelPromptsSchemasToolsTransportpersonaworkflowtool-use
    Split-view diff16 fields changedtools · 10prompts · 3model · 2transport · 1
    FieldBeforeAfter
    toolsapply_patchpresent
    toolscreate_goalpresent
    toolsexec_commandpresent
    toolsget_goalpresent
    toolsrequest_user_inputpresent
    toolsupdate_goalpresent
    toolsupdate_planpresent
    toolsview_imagepresent
    toolswrite_stdinpresent
    toolstool count110
    promptstotal prompt size27,781 chars16,299 chars (-11,482)
    promptsblock 0previous texttext changed (-5,036 chars)
    promptsblock 1prompt block present
    modelmodelgpt-5.5gpt-5.6-sol
    modelthinking{"effort": "medium"}{"context": "all_turns", "effort": "low"}
    transportheader / user_agentcodex_exec/0.143.0 (Debian 12.0.0; x86_64) vscode/1.128.1 (codex_exec; 0.143.0)codex_exec/0.144.0 (Debian 12.0.0; x86_64) vscode/1.128.1 (codex_exec; 0.144.0)