v0.18.1 - 2026-04-04

• Documentation refresh and restructuring: Made docs/WALKTHROUGH.md the canonical onboarding guide, added docs/README.md as a docs index, and aligned README.md to point users at the new walkthrough-first flow.
• Docs cleanup: Removed the redundant docs/USERGUIDE.md, moved implementation/history material into docs/impl/, and refreshed cross-links so user-facing docs and design-history docs are clearly separated.
• Workflow docs: Refreshed Mail Mode, mail-loop, OAuth, contributing, and subquery documentation to reflect current tool names, current mail-loop behavior, and the latest authentication/configuration guidance.

Full Changelog: v0.18...v0.18.1

OpenAI OAuth bootstrap + CLI cleanup: Added built-in OpenAI OAuth bootstrap flows for both browser/manual and headless device-code login, improved device-flow retry and backoff behavior, and cleaned up the CLI by renaming the bootstrap flags to --fetch_openai_oauth_token and --fetch_openai_oauth_device_token. --help and --helpfull now focus on command-line options only, without Hotword or Slash Command sections. This release also updates related docs/helper guidance, moves PKCE SHA-256 generation into a dedicated helper with tests, and includes a small input-parsing fuzz fix.

Full Changelog: v0.17.3...v0.18

Full Changelog: v0.17.3...v0.18

• Fix CI-CD workflow: update yaml file

Full Changelog: v0.17.2...v0.17.3

Full Changelog: v0.17.2...v0.17.3

• Documentation: Updated architecture/config docs to reference app/main.cpp instead of the old root-level main.cpp path.

v0.17.1 - 2026-04-04

• Subquery infinite-context support (context_window = 0): For context window configuration of llm_query subagents, 0 now means infinite, similar to how /context window 0 means infinite. This is useful for certain workflows. 0 now explicitly means infinite history for delegated subqueries; added unit/fuzz coverage and updated subquery docs/examples accordingly.

Full Changelog: v0.17...v0.17.1

Full Changelog: v0.17...v0.17.1

v0.17 - 2026-04-04

New Feature

Config defined sub-agents with persistent context. See README.md and docs/USERGUIDE.md for details

• Config-defined llm_query sub-agents: Config now supports specialization loading for llm_query via [llm_tool_<name>] sections with required field validation (system_prompt_patch, session_id, skill) and optional context_window. These allow sub agetns to have specialized and persistent context windows that can be delegated into.
• Startup registration for specializations: Added startup wiring that registers each specialization as a first-class tool and keeps stale specialization tools in sync with active config.
• Delegation API and policy hardening: Generalized llm_query execution options and centralized subquery policy checks to block recursion paths (llm_query/llm_tool_* in subquery scope) with fixed depth constraints.
• Fuzz + test coverage for subquery paths: Added dedicated fuzz targets for specialization config parsing, subquery policy boundaries, and query options handling; expanded unit coverage around registration/validation behavior.
• Naming alignment to valid tool identifiers: Standardized specialization naming and docs/tests from dot form to underscore form (llm_tool_...) to match tool-name constraints.
• Docs discoverability improvements: Added/expanded docs/subqueries.md, docs/example_subqueries.ini, and prominent README/USERGUIDE guidance for configuring and invoking specialization tools.

Full Changelog: v0.16...v0.17

Full Changelog: v0.16...v0.17

• Tooling split and API cleanup: Moved the tool executor/dispatcher and built-in tool handlers out of core/ into top-level tools/
• Fuzzing coverage expansion: Added FuzzTest targets for core JSON helpers, dispatcher validation, orchestrator normalization, tool-argument validation, and interface input parsing to harden boundary code against malformed input. See docs/fuzzing.md
• OpenAI default model update: Bumped the default OpenAI model to gpt-5.4-mini:high for both standard OpenAI usage and Responses API routing. Either that or gpt-5.3-codex:low is a good default for most tasks.
• Docs and contributor guidance: Added a dedicated fuzzing guide, and improved AGENTS.md with info on validation boundaries, and test selection.

Full Changelog: v0.15.4...v0.16

Full Changelog: v0.15.4...v0.16

See CHANGELOG.md for release notes.

Full Changelog: v0.15.3...v0.15.4

Full Changelog: v0.15.3...v0.15.4

v0.15.3 - 2026-03-12

• OpenAI Responses duplication fix: Prevented duplicate final assistant messages when SSE streams include both response.output_text.delta and response.output_item.done output text; added a regression test for this flow.
• execute_bash timeout plumbing: Added explicit timeout_seconds wiring with a 180s default and structured timeout handling to improve reliability and reporting.
• Diff rendering polish: Improved unified-diff markdown highlighting and UI diff rendering behavior. Uses https://github.com/hsaliak/tree-sitter-unified-diff

Full Changelog: v0.15.2...v0.15.3

Full Changelog: v0.15.2...v0.15.3

• Mail mode UX and flow improvements: Added /mode mail <branchname> auto-staging with validation, then refined mail model smoothness and formatting through follow-up rerolls.
• Scratchpad support restored: Reintroduced the session scratchpad table plus commands/tools, and followed up with scratchpad-focused fixes and README guidance.
• Patch tool output rendering: Added markdown diff rendering for patch_tool unified diffs in the UI while preserving existing truncation behavior.
• Tooling/docs cleanup: Removed an unnecessary tool, updated /mode help text, and renamed mail model docs to mail_mode for consistency.
• Maintenance: Applied style and clang-format cleanup commits.

Full Changelog: v0.15.1...v0.15.2

Full Changelog: v0.15.1...v0.15.2