Changelog

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

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

0.60.0 - 2026-10-09

Added

• Add decision-powered automatic model routing with /model router, persisted simple, standard, and complex provider/model/thinking selections, and authenticated-provider setup.
• Add /router to view routing configuration and /router config to reconfigure category models.
• Paginate per-model usage summaries with keyboard navigation.

Changed

• Replace /model auto with /model router; run classification only while routing is enabled, with standard-model fallback on errors or timeout.
• Keep classifier conversation history across routing activation changes and display classification results alongside the chosen response model.
• Require Go 1.26.0 for source builds and prefer the patched Go 1.26.9 toolchain; CI and release builds use Go 1.26.9.

Fixed

• Preserve partial assistant responses across automatic and manual model switches.
• Generate accurate line-based edit diffs for separated insertions and deletions.
• Upgrade Go and network dependencies to address reachable security vulnerabilities.

0.59.0 - 2026-10-08

Added

• Add the Liquid decision provider (d1, d1:free) to the decision registry.
• Classify task complexity with the Jev task classifier.
• Add latest provider model entries.
• Add a REPL widget for selecting decision provider and model.

Fixed

• Show only changed lines in edit diffs.
• Clean up thinking-effort metadata (glm-5.3-flash, hy3/hy4 efforts).

0.58.0 - 2026-10-01

Added

• Persist provider token usage to a global JSONL ledger at ~/.keen/usage/usage.jsonl, recording one entry per response from interactive turns, compaction, /btw, /adversary, subagents, and keen run.
• Add the /usage command with total and per-model input, output, cache read, and cache write bars for all time, 7 days, and 30 days, switchable with the arrow keys.
• Track cache read and write tokens separately across Anthropic, Bedrock, and the OpenAI clients, keeping CachedTokens as their combined total.

Changed

• Roll the usage ledger up into per-day totals at most once per UTC day, at startup (keen and keen run) and whenever /usage opens, so compaction no longer depends on the view being used; rollups older than one year are dropped.

Fixed

• Create the usage directory before locking and skip compaction when no ledger exists, so startup compaction neither fails on a fresh install nor creates an empty ledger.

0.57.1 - 2026-09-28

Changed

• Extract REPL stream handling into internal/cli/repl/stream and AskUser state and rendering into internal/cli/repl/askuser, replacing direct stream-handler field access with snapshot APIs.

0.57.0 - 2026-09-25

Added

• Add the Yolo-Auto provider (yolo-auto) with the yolo and yolo-small models.
• Add a worker subagent definition for bounded, self-contained tasks.
• Refresh the model registry with current Anthropic, OpenAI, Codex, Google AI, and OpenCode Go models: Claude Opus 5.5; GPT-6 Sol and GPT-6 Luna (OpenAI, Codex, and OpenCode Go); Gemini 3.8 and 3.7 Flash; Grok 4.7 and MiMo V2.6 Flash/Pro on OpenCode Go. Removes retired entries (Gemini 3.5 Flash, Grok 4.5, GLM-5.1, Qwen3.6 Plus) and routes newer OpenCode models through the Responses API.

Changed

• Route the REPL through a new agentcore agent facade: promote agentcore and appstate out of the CLI tree into importable internal packages and migrate the stream, headless, tooling, and ask-user paths onto it.
• Remove the in-request context reducer that pruned the oldest tool results, which invalidated provider prompt caches from the first pruned result onward and silently dropped tool output. Automatic compaction is now the only over-budget path; a request that still exceeds the provider window fails through the normal provider context-window error.
• Bump go.opentelemetry.io/otel/sdk from 1.44.0 to 1.45.0.

Fixed

• Make stream event delivery cancel-safe: cancelled consumers now close the stream instead of hanging provider goroutines, and tool execution is skipped when ToolStart cannot be delivered.
• Report context cancellation and deadlines on closed subagent streams instead of reading a closed event channel as success, and propagate parent session IDs into child runs.

0.56.2 - 2026-09-22

Fixed

• Preserve provider prompt caches across build/plan mode switches by keeping system prompts and tool definitions stable, persisting plan instructions on user messages, and blocking write_file/edit_file at execution time in plan mode.

0.56.1 - 2026-09-14

Fixed

+- Keep REPL input rules and prompt icon in the primary color across build, plan, and yolo modes. +- Move the mode badge to the end of the second input status line with the shared separator.

0.56.0 - 2026-09-14

Added

• Yolo agent mode (/mode yolo): skip all permission prompts, including dangerous bash, with a red input area as a reminder. Shift+Tab cycles through build → plan → yolo. Guard denials (system paths, .gitignore, home dotfiles) still block. Yolo has the full build tool registry (no plan-style tool stripping) and exits cleanly via /mode build.

0.55.0 - 2026-09-13

Added

• Reuse provider prompt caches during compaction: manual /compact and automatic compaction send the real system prompt plus full history so cached prefixes are reused instead of re-processing full history; compaction requests keep tool definitions but deny tool execution with an explicit error, and forced one-shot requests carry prompt-cache markers.
• Route incomplete compaction streams to clean errors so the REPL cannot hang waiting for a summary, and fail compaction application on empty summaries.
• Add gpt-6-astra to the OpenAI provider (1.05M context window, low/medium/high/xhigh/max thinking efforts) and to OpenCode Go with the Codex-mounted 272K context window.
• Add deepseek-v4.1-flash and low thinking-effort support.

Changed

• Extract dependency-safe leaf packages out of the internal/llm monolith.
• Bump google.golang.org/grpc from 1.83.1 to 1.83.2.

0.54.0 - 2026-09-09

Added

• Cache web_fetch results behind a checkCache flag: successful fetches up to 16 KiB are kept in a session-scoped in-memory LRU (512 entries) and larger ones on disk (~/.keen/web-fetch-artifacts/keen-web-fetch-<hash>.txt); cache hits skip re-fetching.
• Compress built-in tool outputs sent to LLM providers: read_file drops display-only bytes_read/lines_read metadata, and grep/glob factor shared directory prefixes (common_prefix with relative paths and grouped grep matches). Raw outputs still reach the UI and tool events, while compact results are replayed from historical tool memory.
• Add claude-fable-5-1 to the Anthropic provider plus grok-4.6, muse-spark-1.2-contributor, muse-spark-1.3-contributor, qwen3.8-flash, hy4-preview, and omen-alpha to OpenCode Go, with Responses-API routing for Grok 4.6 and Muse Spark models and reasoning-effort support for Hy4 preview and Omen Alpha.

Changed

• Pin govulncheck to x/vuln v1.7.0 in CI so the vulnerability scan works on Go 1.25 runners.

0.53.0 - 2026-09-07

Added

• Support multiline REPL suggestions and centralize layout constants.

Changed

• Refine Bash approval classification to distinguish secret exposure and destructive operations, reducing prompts for recoverable development commands.

Added

• Make call_mcp_tool cache successful responses behind checkCache: results up to 16 KiB in a session-scoped in-memory LRU (512 entries), larger ones on disk as ~/.keen/mcp-artifacts/keen-mcp-<hash>.txt shared with artifacts and /cleanup retention. Cache hits skip the permission prompt; large cached results keep the preview plus artifact_path shape.

Changed

• Centralize tool-name literals into shared constants in internal/tools.
• Refresh the README feature list (MCP, skill-driven MCP, subagents, memory, queuing, adversarial review, headless mode, hashline editing), rework the skills description, and remove the outdated telemetry section.
• Document headless mode (keen run) flags, progress output, and completion signals in docs/cli-usage.md.
• Bump google.golang.org/grpc from 1.82.1 to 1.83.1.

Added

• Add GLM-5.3 Flash to the Z.ai and OpenCode Go providers with a 1M context window and low/high/max thinking efforts.
• Fold consecutive reads of the same file in the REPL into one status line so multi-offset reads stay readable.

0.50.1 - 2026-08-29

Fixed

• Make manual compaction accept typed input, and make custom ask-user responses fully editable.
• Return only tool-result data discovered during execution and show read-file metadata for the returned slice.

0.50.0 - 2026-08-25

Added

• Add Z.ai GLM-5.3 and OpenCode Go LongCat model support with provider-specific thinking parameters.

0.49.0 - 2026-08-23

Added

• Let agents ask users interactive questions from the REPL.

Changed

• Update the homepage demo GIF.

0.48.1 - 2026-08-13

Fixed

• Bump Go to 1.25.13 to resolve stdlib vulnerabilities.

Added

• Add a demo GIF and move the demo recording guide to the docs.

0.48.0 - 2026-08-13

Added

• Add the session-scoped /tool-history full|none command to control whether future tool outputs are retained between turns.

Changed

• Document the tool-history setting and refresh CLI, architecture, provider, and tool references.

0.47.1 - 2026-08-13

Fixed

• Prevent recoverable headless tool failures from being rendered as REPL errors.

Changed

• Refine model tool-use guidance and expand coverage for REPL, OAuth, MCP, telemetry, and tool edge cases.

0.47.0 - 2026-08-11

Changed

• Expand deterministic coverage across CLI, OAuth, REPL, LLM, MCP, telemetry, filesystem, and utility behavior.
• Publish test coverage reports to Codecov.
• Refresh the project introduction with current terminal screenshots.
• Use the canonical GitHub module path for Go tooling and pkg.go.dev.
• Remove the obsolete benchmark harness.
• Recommend semantic-version bumps in the release workflow.

Fixed

• Hide expected edit failures from REPL error output.

0.46.0 - 2026-08-10

Added

• Hashline-style editing (issue #80): read_file emits N:HASH| line anchors (line number + 3-char FNV-1a hash), edit_file takes a multi-op ops[] array validated against one file snapshot and applied atomically, and structured grep matches include line_hash. No file-level hash; anchors never relocate; re-read before a later same-file edit.

0.45.0 - 2025-08-07

Added

• Show context usage breakdown in the REPL.

0.44.0 - 2026-08-05

Changed

• Migrate OpenAI, OpenAI Responses, and Codex clients to the openai-go v3 SDK, converting tools and tool calls to v3 union types.
• Replay Responses reasoning items across tool continuations so reasoning models retain their full reasoning state.
• Surface response.failed and response.incomplete terminal events as stream errors instead of silent success.
• Handle array-backed Responses function-call outputs in the context reducer.

Fixed

• Handle the v3 response.reasoning_text.delta streaming event in Responses and Codex clients.
• Filter Chat Completions tool calls by non-empty function name so OpenAI-compatible providers without a union type still execute tools.

0.43.0 - 2026-08-06

Added

• keen run now streams live text chunks and tool completion lines to stderr while a headless turn is in flight.
• keen run --completion-signal lets Ralph-style loops know when to stop: exit 0 when the marker is present, exit 2 (without the Error: prefix) when it is missing so the loop can continue.

0.42.0 - 2026-08-04

Added

• Add thinking support and refreshed model entries across supported providers.
• Add provider model configuration loading and validation coverage.

Changed

• Refresh provider documentation and model registry entries.
• Improve provider-specific thinking configuration across Anthropic, Bedrock, Genkit, OpenAI, and Responses clients.
• Streamline REPL model selection and improve input metadata and suggestions.

0.41.0 - 2026-08-03

Added

• Automatically compact long-running agent context across all LLM providers while preserving active tool progress.
• Persist and replay automatic compaction checkpoints in interactive and headless sessions.
• Document manual and automatic context compaction flows.

Changed

• Refine agent and compaction prompts for concise continuation summaries.
• Preserve partial headless output when a post-compaction stream fails.

Fixed

• Stop Genkit test iterators when stream consumers stop early.

0.40.0 - 2026-08-05

Added

• Prune stale Keen data from the local application directory.
• Spill large web_fetch results into readable artifacts.

Fixed

• Restrict delegated subagent tasks to configured profiles.

0.39.0 - 2026-07-30

Changed

• Refine the initial REPL screen and make regex search patterns easier to read in tool status output.
• Continue queued prompts after interrupting a stream or redirecting a permission request.
• Remove obsolete subagent and MCP guidance.

0.38.1 - 2026-07-29

Changed

• Prompt before updating an already configured provider while changing its model.
• Shorten REPL tool-input previews and render argument values without quotes.

0.38.0 - 2026-07-29

Added

• Configurable subagent profiles with custom instructions, model overrides, and permission-scoped tool access.
• Isolated child-agent execution with auto-approved permitted operations and sanitized activity in the REPL.
• Provider-specific model configuration resolution for profile overrides.
• Documentation for configuring and using subagents.

Changed

• Replace the bundled read-only explorer profile with discoverable, user-configurable profiles.

Fixed

• Hide OpenAI-compatible providers from interactive model selection.
• Suppress expected tool failures from REPL error output.

0.37.0 - 2026-07-26

Added

• Anonymous GA4 session start and end telemetry with coarse country reporting, privacy-focused fields, and standard opt-out controls.

0.36.3 - 2026-07-26

Fixed

• Retain bounded write_file and edit_file inputs in tool history so the model no longer learns from empty historical tool calls.

0.36.2 - 2026-07-25

Changed

• Align system prompt guidance with retained historical tool inputs and tool-level Git enforcement.
• Remove dead turn-memory rebuilds from retry handling.

Fixed

• Omit obsolete read-file truncation metadata from turn memory.

0.36.1 - 2026-07-24

Changed

• Centralize delegated-subagent timeout configuration in agent profiles.

0.36.0 - 2026-07-23

Added

• Git branch display in the REPL location line, refreshed on startup and after each completed turn.
• Provider prefix on the model name in the REPL meta view.

Changed

• Split the REPL meta view into location and context status lines with consistently styled separators.
• Group REPL model state into sub-structs.

Fixed

• Bump golang.org/x/text to v0.39.0 to address GO-2026-5970 (infinite loop on invalid input).
• Bump google.golang.org/grpc to v1.82.1.

0.35.0 - 2026-07-22

Changed

• Redesign REPL tool status display with friendly labels, structured metadata (line counts, byte sizes, match counts), ellipsis truncation, and duration formatting on all tool results.
• Combine adjacent tool start/end into a single done line.
• Bump subagent default timeout to 30 minutes.

0.34.0 - 2026-07-21

Added

• Run up to ten delegated subagent tasks in parallel with per-agent progress and completion counts.

0.33.0 - 2026-07-20

Added

• Retain bounded inputs for read-only, shell, delegation, and MCP tools in turn memory and provider history.

Changed

• Allow explorer subagents more time to complete scoped investigations.

0.32.0 - 2026-07-20

Added

• Expose file-change and failed-command outcomes in tool results and turn memory.

Changed

• Prefer activity already recorded on stream segments before building turn memory.
• Record turn memory at turn end instead of during streaming.

0.31.2 - 2026-07-14

Fixed

• Upgrade Go dependencies to address known security vulnerabilities.

Changed

• Scan reachable Go vulnerabilities and run CodeQL security analysis in CI.

0.31.1 - 2026-07-14

Fixed

• Improve model-facing tool validation errors and preserve literal source characters in serialized tool outputs.
• Reduce historical tool-result placeholders to concise status metadata.

0.31.0 - 2026-07-13

Added

• InputValidator interface and ValidateInput helper for pre-execution tool-input validation.
• Centralized executeValidatedTool path so validation failures are handled consistently.

Changed

• ToolStart event is now emitted only after input validation succeeds, keeping malformed tool calls out of the REPL transcript.

0.30.0 - 2026-07-12

Added

• OpenAI-compatible provider support for custom endpoints without registry-defined models.

Changed

• Replay historical tool activity as provider-native tool calls and results, preserving invocation order while keeping discarded tool data out of assistant text.
• Updated provider and model documentation to match the registry.

0.29.0 - 2026-07-11

Added

• Adversary stream tool start/end event handling so tool calls are flushed and rendered like normal assistant tool calls.

Fixed

• Install script now resolves the latest release from GitHub's public redirect instead of the REST API, avoiding the 60 req/hour unauthenticated rate limit on shared/CGNAT IPs.

0.28.0 - 2026-07-10

Added

• Global and project memory with managed updates, secret detection, and /memory commands.
• Historical tool activity annotations in Turn Memory to preserve execution context across long-running sessions.

Changed

• Anthropic debug logging now shows the ordered provider-facing conversation content.

0.27.0 - 2026-07-10

Added

• GPT-5.6 model entries in the provider registry.
• Debug logging for unhandled LLM tool tokens and malformed tool JSON in Anthropic and OpenAI clients.

0.26.1 - 2026-07-08

Changed

• Tightened tool-memory guidance in the system prompt to prevent hallucinated findings from stale memory, requiring a fresh tool call before relying on external evidence.

0.26.0 - 2026-07-07

Added

• Claude Fable 5 and Sonnet 5 model entries in the provider registry.

0.25.3 - 2026-07-03

Fixed

• Clear REPL input on Esc instead of quitting the application.

0.25.2 - 2026-07-02

Fixed

• Require follow-through in tool-use claims to prevent unsupported completion summaries.

0.25.1 - 2026-07-02

Added

• Batched REPL stream viewport redraws to reduce render churn while preserving immediate diff and permission updates.

0.25.0 - 2026-06-29

Added

• Distinct input styles for /btw and /adversary commands in the REPL.
• Spill large MCP tool results to artifact files.

Fixed

• Send required metadata in OAuth dynamic client registration for MCP servers.

0.24.12 - 2026-06-29

Fixed

• Clear turn elapsed display on /clear command.
• Remove delegate_task from adversary tool registry.

0.24.11 - 2025-07-22

Added

• Turn elapsed time display after agent completes in the REPL.

Changed

• Use dim faint style for turn elapsed display in the REPL.

0.24.10 - 2025-07-21

Fixed

• Remove false-positive dangerous classification for 2>&1 and 2>/dev/null redirects in bash classifier.

0.24.9 - 2026-06-27

Added

• Automatic dangerous bash command classification in the bash tool, gating risky commands even when the LLM omits isDangerous.
• File tagging (@) suggestions inside slash commands and skill invocations.

0.24.8 - 2026-07-18

Added

• api_key_helper credential refresh on 401 responses.
• REPL loading step shown while fetching credentials for api_key_helper providers.

0.24.7 - 2026-06-24

Added

• api_key_helper for shell-based API key resolution.
• Block-level prompt caching for Anthropic.

0.24.6 - 2026-06-23

Added

• --resume flag to restore previous REPL sessions.

0.24.5 - 2026-06-23

Fixed

• Ignore whitespace-only input submissions in the REPL.

Changed

• Document KEEN_LOG_LEVEL debugging instructions in CONTRIBUTING.md.

0.24.4 - 2026-06-22

Added

• Surface MCP startup errors in the REPL UI.
• Per-provider custom HTTP headers configuration.

Fixed

• Unblock bang (!) output readers when context is cancelled.

0.24.3 - 2026-06-21

Fixed

• Prefer native clipboard writes before OSC52 fallback in the REPL and log native clipboard failures in debug mode.

0.24.2 - 2026-06-21

Fixed

• Adjusted viewport height to account for notifications and queued inputs in the REPL.

0.24.1 - 2026-07-18

Changed

• Removed vim-style j/k keybindings from permission prompts, session picker, and model selection, keeping only arrow keys for navigation.
• Applied faint styling to loading and queue item text in the REPL.

0.24.0 - 2026-07-17

Added

• Queue user inputs while the agent is streaming, preventing input loss during active LLM turns.
• Loading spinner for bash and bang (!) shell commands, providing visual feedback during shell execution.

Changed

• Reduced hallucinated MCP tool names with exact-name guidance in the system prompt.

0.23.7 - 2026-06-20

Added

• OpenAI-compatible and Responses clients now reuse prompt cache context per session to improve cache hits across turns.

0.23.6 - 2026-06-18

Fixed

• Append a config fix hint pointing to ~/.keen/configs.json to LLM config errors raised during resolve, load, and client construction.
• Clear pendingState on context overflow after tool-result reduction so stale tool traces and thinking tokens are not re-injected on the next call.

0.23.5 - 2026-06-18

Added

• Atom One Dark color theme for markdown code highlighting in the REPL.
• First Ctrl+C now interrupts active REPL work before a second Ctrl+C exits.

Removed

• Bundled cleanup and plan skills.

0.23.4 - 2026-06-18

Added

• Asynchronous ! shell command execution with streaming output and cancellation support.
• Bash tool now writes truncated output artifacts to ~/.keen/bash and grants read access to them.

Changed

• Refreshed GLM model entries in the provider registry.

Fixed

• Corrected spacing in the REPL input meta display.
• Skip context-reducer targets smaller than the placeholder threshold.

0.23.3 - 2026-06-17

Changed

• Relocated provider registry and loader from providers/ to internal/providers/ to keep the provider API internal.

0.23.2 - 2026-06-15

Added

• Spinner animations for REPL loading states.
• Qwen3.7 Plus model entry in the provider registry.

Changed

• Shell output now renders with a distinct muted background style in the REPL.
• Refined input meta display in the REPL.
• Pruned superseded models from the provider registry.

Fixed

• Updated test expectations for context status rendering and model count.

0.23.1 - 2026-06-15

Added

• Kimi K2.7 Code model entry in the provider registry.

Changed

• Extracted REPL input selection into a dedicated type.

0.23.0 - 2026-06-14

Added

• Read-only subagent delegation via delegate_task with the bundled explorer subagent for scoped codebase investigations.

Changed

• Simplified REPL tool invocation display.
• Simplified displayed tool input fields in the REPL output.

0.22.3 - 2026-06-14

Added

• Google Analytics tracking for the documentation site.
• Automatic clipboard copy when mouse text selection is released in the REPL, with a transient copy notification.

Changed

• Reduced tool context sent to the LLM before requests.
• Removed Ctrl+C and Cmd+C as REPL selection copy shortcuts in favor of copy-on-release.

0.22.2 - 2026-06-15

Added

• Input/output token accumulation and display in the REPL status bar.
• Session-level token totals that reset on /clear, /new, and /resume.
• Compact token formatting (e.g. 1.2k, 3.4M) in the meta status view.

0.22.1 - 2026-06-14

Added

• Optional API key support for the Amazon Bedrock provider, enabling usage with temporary credentials or cross-account authentication.
• Bearer token authentication for Bedrock models.
• Config helpers to detect optional vs required API key configuration.

Changed

• Updated model selection UI to show hints for optional vs required API keys.

0.22.0 - 2026-06-13

Added

• Amazon Bedrock provider with Claude models (Opus, Sonnet, Haiku) via the AWS SDK, supporting streaming, tool use, and reasoning.
• AWS credential-based authentication with automatic region resolution and optional base URL override.
• MCP tool call validation: reject missing required arguments and normalize server/tool names before invoking the server.
• Debug logging for token usage (prompt, completion, cached, reasoning) in OpenAI-compatible, Codex, and Responses clients.

Changed

• Updated Amazon Bedrock documentation and MCP tool description in README and docs.

0.21.2 - 2026-06-11

Added

• Clickable URLs in REPL output: bare URLs render as underlined OSC 8 hyperlinks and open on Alt/Option or Ctrl+click.

0.21.1 - 2026-06-09

Changed

• Refreshed status bar glyphs and layout.
• Moved mode chip to input border with plan mode styling.
• Consolidated demo GIF and removed unused assets.
• Added cast to GIF conversion command to the agent release skill.

0.21.0 - 2026-06-08

Added

• ! prefix in the REPL input for direct shell command execution, bypassing the LLM.
• Shell mode input styling: accent-colored rules with a shell chip and recolored prompt when ! is typed.
• Syntax-highlighted command output via the built-in markdown renderer with automatic language detection.

0.20.5 - 2026-06-07

Changed

• Enabled prompt caching for all providers using the Anthropic client, no longer restricted to the Anthropic provider.
• Cached the last system block to improve hit rates on multi-turn conversations.

0.20.4 - 2026-06-05

Added

• Product Hunt badge and launch highlight to documentation homepage.

Changed

• Reduce long tool contexts in LLM calls.

Fixed

• Tighten Product Hunt badge spacing on documentation homepage.

0.20.3 - 2026-06-03

Added

• Styled inline code spans in REPL startup tips.

Changed

• Updated demo assets.

0.20.2 - 2026-06-02

Added

• Shimmer loading effect and did-you-know tips replacing the previous loading spells in the REPL.

0.20.1 - 2026-06-02

Added

• REPL startup screen hints showing the last session and rotating usage tips.
• Support for the MiniMax m3 model in the provider registry.

Changed

• Polished REPL input UX with dynamic textarea height and up-arrow history navigation.
• Renamed MiniMax model identifiers to lowercase in the provider registry.

0.20.0 - 2026-06-01

Added

• /adversary REPL command to trigger an adversarial review of the current conversation for gaps, risks, and missed alternatives.

0.19.4 - 2026-05-31

0.19.3 - 2026-05-26

Added

• Inline /btw side-question rendering with full conversation context instead of an overlay popup.
• Styled chip mode indicators in the REPL status bar (plan/build) replacing glyph+text labels.

0.19.2 - 2026-05-25

Added

• Support Claude skill directories in skill discovery and filesystem permissions.
• Expanded tool reference documentation and roadmap updates.

Changed

• Updated demo assets in the README and documentation site.

0.19.1 - 2026-05-23

Changed

• Preserved MCP skill enablement preferences during server sync while removing stale generated skills for deleted servers.
• Simplified MCP tool display and capped skill descriptions in REPL list output.
• Used MCP server instructions for generated skill descriptions.
• Store logs in ~/.keen/logs instead of ~/.keen-code/logs.
• Refreshed documentation landing page styling and clarified skill activation persistence.

0.19.0 - 2026-05-21

Added

• web_fetch tool to fetch URL content and convert HTML pages to Markdown for LLM consumption.
• MCP server support with configurable transports, authentication, connection management, and tool discovery.
• MCP tool-calling support through generated MCP skills and the call_mcp_tool tool.
• Documentation for MCP servers, skill-driven MCP integration, and OAuth-authenticated MCP servers.
• GitHub Pages documentation site powered by MkDocs Material.
• Suggested subcommands for /mcp connect, /mcp status, /skills list, /skills enable, and /skills disable.

Changed

• Streamlined README intro section.
• Updated REPL mode glyphs and removed mode-change confirmation messages.
• MCP skills now enable or disable based on connection status while preserving generated skill files.
• Improved docs site styling, navigation, badges, fonts, and local preview support.

Fixed

• Render markdown table row rules safely.
• Fixed broken or misleading documentation links and labels.

0.18.0 - 2026-05-16

Added

• Plan and build modes for structured REPL interaction workflows.

0.17.0 - 2026-05-15

Added

• Project-level tool allow lists for pre-approved permission checks.
• Anthropic prompt caching support and improved token usage tracking.
• Benchmark runner with updated benchmark documentation and demo assets.

Changed

• Improved REPL markdown rendering with width-aware horizontal rules, wrapped tables, connected table borders, and outer table frames.
• Refined assistant formatting guidance to prefer semantic GitHub-flavored markdown.
• Updated CLI usage and permission system documentation.

0.16.3 - 2026-05-13

Added

• Paginate read_file output and add line number prefixes.
• OpenCode usage scripts and restructured benchmark output with usage timestamp filtering.
• Refined system prompt exploration guidance for efficient tool use.

Changed

• Restructured benchmark layout.

0.16.2 - 2026-05-12

Added

• Toggle focus between input and viewport via Tab and mouse clicks.
• Route up/down keys based on focused region.
• Dim input chrome and prompt glyph when focus is in the viewport.

Changed

• Merged PR #41: add basic benchmark.

0.16.1 - 2026-05-12

Added

• keen run headless command for non-interactive task execution.
• --provider and --model flags to override LLM configuration in keen run.

0.16.0 - 2026-05-11

Added

• Bundled workflow skills for common agent tasks.
• /btw side questions for asking context-aware questions without interrupting the main conversation.
• Documentation for /btw side questions.

Changed

• Constrained REPL suggestion list height to fit the available viewport.

0.15.3 - 2026-05-08

Changed

• Moved release guide from README into a local skill at .agents/skills/release/.

Added

• Documentation for turn memory KV cache and token cost analysis.

0.15.2 - 2026-05-08

Changed

• Avoid repeated file suggestion cache rebuilds in large repositories.

0.15.1 - 2026-05-07

Added

• Horizontal padding for submitted user input blocks.

Changed

• Use git ls-files for faster cached file suggestions.
• Improved REPL status display and usage documentation.

Fixed

• REPL submitted input wrapping test expectation.

0.15.0 - 2026-05-07

Added

• MiniMax provider support for M2.7 and M2.5 via the Anthropic-compatible API.

0.14.0 - 2026-05-06

Added

• OpenCode Go provider support routed through OpenAI-compatible or Anthropic clients, including provider registry entries and thinking parameter handling
• REPL session IDs propagated through LLM stream calls and attached as hyphenless OpenCode Go request headers (Anthropic and OpenAI-compatible)
• Architecture and system documentation covering AI providers, permission system, session management, skills system, tools, and turn memory

Changed

• Preserve Anthropic thinking blocks across tool continuations
• Simplified LLM test coverage by removing redundant provider, thinking effort, and system prompt tests

0.13.0 - 2026-05-06

Added

• Agent skills discovery, slash-command invocation, frontmatter validation, and argument substitution
• Bundled commit skill embedded in the binary and extracted to the user skills directory at runtime
• Additional model registry entries

Changed

• Reset LLM provider state when starting new REPL sessions

0.12.2 - 2026-05-03

Added

• Permission option to ask what the agent should do instead, interrupting the current stream while preserving partial state

0.12.1 - 2026-05-02

Changed

• Improved REPL loading status display with elapsed time

Fixed

• Indent wrapped submitted user input lines in the REPL transcript

0.12.0 - 2026-05-01

Added

• In-app text selection for REPL output and input, with copy support for active selections via Ctrl+C or forwarded Cmd+C

0.11.2 - 2026-04-30

Changed

• Split REPL command handling into dedicated command handler components

0.11.1 - 2026-04-30

Added

• Shaded background block for the echoed user input that grows with line count and resizes responsively with the viewport

Changed

• Refreshed the prompt glyph from > to ▶ across the textarea, echoed input, model selection inputs, session picker, and permission card cursor

0.11.0 - 2026-04-30

Added

• Retry support across streaming clients for improved LLM reliability
• Pending tool state preservation across turns
• Pending state recovery for all LLM clients

0.10.0 - 2026-04-27

Added

• ChatGPT OAuth support for the Codex provider

0.9.0 - 2026-04-27

0.8.0 - 2026-04-25

Added

• Retry transient LLM stream errors with backoff for OpenAI-compatible clients

0.7.0 - 2026-04-25

Added

• Startup update checker that notifies REPL users when a newer version is available

0.6.1 - 2026-04-25

Fixed

• Improved assistant markdown colors on light terminals while preserving inline code color and syntax highlighting

0.6.0 - 2025-07-18

Added

• Z.ai (GLM) as an OpenAI-compatible provider (#3)
• File suggestion with @ prefix in the input textarea
• filesearch package with gitignore-aware file indexing and glob-escaped query matching

Fixed

• Materialize partial message and TurnMemory on LLM error

Changed

• Bumped Anthropic max output tokens

0.5.0 - 2026-04-23

Added

• Configurable base URL per provider
• CONTRIBUTING.md for contributors
• Public-facing ROADMAP.md
• Turn-memory documentation
• Project tour, issue templates, and pull request template
• Demo GIF in README

Changed

• Wrapped REPL diff output in a viewport
• Updated LLM configuration
• Refreshed the demo GIF rendering with Monaspace Argon NF

0.4.1 - 2026-04-22

Fixed

• Retain tool memory on stream interrupt

0.4.0 - 2026-04-22

Added

• Provider-backed context status replacing the local word-count heuristic
• Token usage events emitted from all provider clients (OpenAI Responses, Anthropic, Genkit/Google AI, DeepSeek/Moonshot)
• Cache-aware token accounting for Anthropic (includes cache creation and read tokens)
• Anthropic adaptive effort display in the status bar
• N/A display when context window is unknown instead of a misleading percentage

Changed

• Context status now reports actual provider-counted token usage against the model context window
• Compaction suggestions are grounded in real tokenization rather than local estimates
• /clear and /new reset context metrics for new sessions
• Updated provider registry with new models and context windows

Removed

• Local word-count token estimation helpers (estimateTokensFromWordCount, countWords, estimateToolDefinitionTokens, buildConversationForEstimation)

0.3.0 - 2026-04-21

Added

• Configurable thinking effort selection in model setup and via the /thinking runtime command
• Direct Anthropic SDK client with expanded Anthropic streaming and tool-loop test coverage
• Refactored REPL helper packages for app state, output, permissions, theme, tooling, widgets, and streaming

Changed

• Thread thinking effort configuration through OpenAI Responses, Anthropic, and Genkit clients
• Added Anthropic prompt caching support in the REPL streaming path
• Refreshed phase-5 design notes and removed stale scratch artifacts

0.2.3 - 2026-04-19

Added

• Structured TurnMemory system to replace in-band XML tags for tracking durable tool outcomes
• turnMemoryAccumulator in REPL to automatically capture file changes and failed bash commands

Changed

• Refactored LLM provider interface to deterministically append tool memory metadata
• Simplified system and compaction prompts by removing manual memory tag instructions
• Improved session persistence to support structured tool outcomes

Removed

• Legacy <keen_memory> tag parsing and stripping logic

0.2.2 - 2026-04-17

Added

• Hidden keen_memory blocks to preserve durable tool outcomes across turns without showing them in the REPL transcript

Changed

• Session picker now constrains its visible list to the current viewport height

Fixed

• Only extract trailing dedicated keen_memory blocks for logging and compaction-aware handling

0.2.1 - 2026-04-16

Changed

• Simplified REPL context status display and metadata emphasis

0.2.0 - 2026-04-16

Added

• Conversation session management with transcript persistence
• /sessions command to list recent sessions with metadata
• /resume command with interactive picker to restore conversations
• /compact command to summarize conversation history via LLM
• Event-sourced storage (session_started, user_message, assistant_turn, compaction_applied)
• Store tool outputs, bash results, and file diffs in transcript for full replay

0.1.7 - 2026-03-24

Added

• REPL context status indicator with progress bar and percentage based on model context window
• Slash command autosuggestion dropdown for /help, /model, and /exit

Changed

• Consolidated REPL styling for context status and suggestion UI

0.1.6 - 2026-03-22

Changed

• Improved spinner UX with smoother feedback during LLM streaming
• Refined tool descriptions for better LLM tool selection
• Improved Genkit streaming reliability

0.1.5 - 2026-03-22

Added

• Install script for easier local setup
• npm wrapper package documentation

0.1.4 - 2026-03-22

Changed

• Switched npm publishing to trusted publishing (removes need for legacy token)

0.1.3 - 2026-03-22

Fixed

• Release pipeline corrections from v0.1.2

0.1.2 - 2026-03-22

Fixed

• Improved release flow and startup behavior

0.1.1 - 2026-03-22

Fixed

• npm wrapper publish and install flow

0.1.0 - 2026-03-22

Added

• Interactive REPL powered by Bubble Tea with streaming LLM responses
• Multi-turn tool calling with Genkit integration
• read_file tool with interactive permission system
• write_file tool with inline diff rendering
• edit_file tool with inline diff rendering
• bash tool with permission gating
• glob tool for file pattern searching
• grep tool for content search
• File guard with .gitignore awareness and permission levels (granted/pending/denied)
• Inline permission card UI (replaces full-screen modal)
• Dynamic system prompt generation with project context
• OpenAI-compatible client supporting DeepSeek (including reasoning/chain-of-thought)
• MoonshotAI provider via OpenAI-compatible client
• Dedicated OpenAI Responses API client
• GoReleaser config for cross-platform binary distribution
• npm wrapper package for installation via npm install -g keen-code