Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog and this project follows Semantic Versioning.

Unreleased

2.0.0-alpha.18 - 2026-05-26

Fixed

• Retried dropped Responses streams after assistant text deltas when no assistant message or tool side effect has been committed yet.
• Added retry decision metadata to no-retry backend logs so blocked retries explain whether they hit max attempts, non-replayable output, or a non-retryable error.

2.0.0-alpha.17 - 2026-05-25

Fixed

• Retried wrapped URL transport failures, including nested NSURLErrorDomain and URLError network connection loss errors, according to the configured retry policy.

2.0.0-alpha.16 - 2026-05-23

Fixed

• Fixed generated image turns by keeping assistant image attachments out of replayed Responses request content.

Changed

• Split the runtime agent model definitions into focused source files.
• Made SwiftPM target paths explicit and documented that the checked-in demo app is not part of published package products.

2.0.0-alpha.15 - 2026-05-22

Changed

• Updated the demo app's default Codex model to gpt-5.5.
• Added hosted Responses image generation support, generated-image metadata, flat-file attachment persistence, and demo rendering.

2.0.0-alpha.14 - 2026-05-13

Changed

• Removed prompt-only section labels from compiled persona and skill instructions to avoid sending debug metadata to the backend.
• Shortened request-visible context, options, and streamed structured-output helper prompts to reduce input token overhead.
• Coalesced redundant pending runtime store writes before persistence while preserving append/delete semantics.
• Added .verbose SDK logging for wire-level streaming payloads, keeping .debug focused on readable request/response and lifecycle diagnostics.
• Replaced the demo app's developer logging toggle with an Off / Debug / Verbose log-level picker.

2.0.0-alpha.13 - 2026-05-04

Changed

• Changed persona precedence so thread personas replace runtime/backend personality and request-level persona overrides replace inherited thread/runtime behavior for that turn instead of appending to it.
• Replaced request skill overrides with AgentSkillSelection so turns can explicitly use thread skills, replace them, or append request-local skills.

2.0.0-alpha.12 - 2026-04-29

Fixed

• Fixed a GitHub Actions loopback OAuth test race when an available localhost port is reused before the listener starts.

2.0.0-alpha.11 - 2026-04-29

Fixed

• Hardened ephemeral request tests for GitHub Actions runner differences.

2.0.0-alpha.10 - 2026-04-29

Added

• Added RequestExecutionMode.ephemeral for fast transient turns that skip prior thread history replay, context compaction, transcript/history persistence, pending state writes, and memory capture.
• Added demo app support for testing ephemeral turns from the Behavior Lab.

2.0.0-alpha.9 - 2026-04-29

Added

• Added AgentThreadConfiguration so threads can carry their own model and reasoning effort.
• Added runtime APIs for updating thread configuration after thread creation.

Changed

• Routed Codex responses requests and context compaction through thread-level model and reasoning configuration, with backend defaults as fallback.
• Updated the demo app and docs to create threads with model configuration and adjust reasoning per active thread.
• Restored the demo app's Xcode workspace metadata so local CodexKit and CodexKitUI package products resolve consistently.

2.0.0-alpha.5 - 2026-04-13

Added

• Added typed, structured request APIs through AgentMessageRequest<Input> and structured section support for machine-context turns.
• Added support for running runtime turns with mixed freeform text and structured input/section payloads in streaming and one-shot message paths.

Changed

• Renamed GRDBRuntimeStateStore to SQLiteRuntimeStateStore and aligned documentation/examples with the new naming.
• Improved runtime logging ergonomics to make request execution and turn lifecycle diagnostics easier to interpret.

2.0.0-alpha.1 - 2026-03-22

Added

• Schema-driven structured output support through AgentStructuredOutput, AgentStructuredOutputFormat, and the Swift-friendly JSONSchema DSL.
• Imported/share-friendly message construction through AgentImportedContent.
• Dedicated structured output demo tab plus App Intents / Shortcuts examples in the demo app.
• Thread detail navigation in the demo app so conversation views are separated from the main dashboard.

Changed

• Simplified the runtime messaging API so plain text uses sendMessage, typed replies use sendMessage(..., expecting:), and streaming uses streamMessage.
• Updated the demo app to better separate assistant controls, structured output demos, and thread views.
• Expanded README coverage for structured output, imported content, and App Intents integration.

1.1.0 - 2026-03-21

Added

• Runtime skill support with thread-pinned skills, per-turn skill overrides, and execution-policy enforcement for allowed tools, required tools, tool sequence, and max tool calls.
• Dynamic persona and skill loading from local files and remote URLs through AgentDefinitionSource.
• Resolved-instructions preview support so host apps can inspect the final compiled instructions for a turn.
• Transient request retry/backoff policy support in the Codex responses backend.
• Configurable reasoning effort (low, medium, high, xhigh) for gpt-5.4 style thinking control.
• Demo app UI for switching thinking level on future requests.

Changed

• Added CLI-style unauthorized-session recovery so runtime operations can refresh and retry once after auth expiry or invalidation.
• Improved the demo app’s skill and Health Coach flows to better show persona, skill, and tool orchestration together.
• Expanded README coverage for retries, skills, dynamic definition sources, and reasoning effort configuration.

1.0.0 - 2026-03-20

Added

• Stable CodexKit + CodexKitUI runtime surface for iOS agent integration.
• ChatGPT auth with .deviceCode and .oauth (localhost loopback callback flow).
• Threaded runtime state restore, streaming output handling, and approval-gated tool execution.
• Layered persona model (base instructions, thread persona stack, per-turn override).
• Text + image user input support and assistant image attachment hydration.
• Demo iOS app with:
  • dual auth flows
  • tool registration and logging
  • persona demos
  • Health Coach tab with HealthKit integration and proactive AI-generated coaching feedback
  • local reminder scheduling

Changed

• Refactored demo app into smaller Swift files for clearer ownership and readability.
• Updated README docs with production setup guidance and end-to-end examples.