chore: release

[github-actions]

🤖 New release

• rara-api: 0.0.1
• base: 0.0.1 -> 0.0.2
• common-runtime: 0.0.1
• common-worker: 0.0.1
• rara-browser: 0.0.1
• rara-domain-shared: 0.0.1
• rara-paths: 0.0.1
• rara-soul: 0.0.1
• rara-stt: 0.0.1
• rara-tool-macro: 0.0.1
• rara-kernel: 0.0.1
• rara-agents: 0.0.1
• rara-keyring-store: 0.0.1
• rara-mcp: 0.0.1
• rara-tts: 0.0.1
• rara-channels: 0.0.1
• rara-composio: 0.0.1
• rara-git: 0.0.1
• rara-sessions: 0.0.1
• rara-skills: 0.0.1
• yunara-store: 0.0.1
• rara-backend-admin: 0.0.1
• rara-codex-oauth: 0.0.1
• rara-pg-credential-store: 0.0.1
• rara-error: 0.0.1
• rara-server: 0.0.1
• rara-symphony: 0.0.1
• rara-app: 0.0.1
• common-telemetry: 0.0.1
• rara-cli: 0.0.1
• crawl4ai: 0.0.1
• rara-model: 0.0.1

Changelog

rara-api

[0.0.1] - 2026-04-13

Bug Fixes

• Resolve all clippy warnings across codebase (#313)

Documentation

• api: Add proto comments for buf lint
• Update README to reflect tape-based architecture (#783) (#784)
• Simplify readme with logo (#1088) (#1089)
• Add inspired-by credits (#1091) (#1092)

Features

• execution: Add rara py worker grpc runtime skeleton
• execution: Add jobspy capability and capability discovery
• Integrate mem0 via execution worker

Miscellaneous Tasks

• Establish job backend baseline
• Rename to rara
• Change default HTTP port from 3000 to 25555
• Format
• Format

Refactor

• Rename rsketch crates to job
• Rename remaining rsketch references to job
• Decouple telegram bot with grpc/http boundary (#94)
• Unify runtime state into AppState with init() and routes()
• kernel: Consolidate queue/KV/LLM subsystems, remove rara-queue crate
• Fix naming conventions and convention drift (#1032) (#1035)

base

[0.0.2] - 2026-04-13

Documentation

• Update README to reflect tape-based architecture (#783) (#784)
• Simplify readme with logo (#1088) (#1089)
• Add inspired-by credits (#1091) (#1092)

Features

• base: Add ArcCow and SharedString utility types
• base: Add deterministic UUID constructor to define_id (#164)

Miscellaneous Tasks

• Establish job backend baseline
• Rename to rara
• Change default HTTP port from 3000 to 25555
• Make lint pass across workspace
• Format
• Add missing AGENT.md files for all crates (#535) (#539)

Performance

• base: Use NonZeroU128 in define_id! for Option niche optimization (#1149) (#1160)

Refactor

• Rename rsketch crates to job
• Decouple telegram bot with grpc/http boundary (#94)
• Extract crawl4ai into common crate and remove unused downloader
• Add keyring-store crate, process group utils, layer READMEs, and dep upgrades
• kernel: Consolidate queue/KV/LLM subsystems, remove rara-queue crate
• Make it compile
• Remove dead code identified by desloppify scan (#1025) (#1029)
• Fix naming conventions and convention drift (#1032) (#1035)

common-runtime

[0.0.1] - 2026-04-13

Documentation

• Update README to reflect tape-based architecture (#783) (#784)
• Simplify readme with logo (#1088) (#1089)
• Add inspired-by credits (#1091) (#1092)

Miscellaneous Tasks

• Establish job backend baseline
• Rename to rara
• Change default HTTP port from 3000 to 25555
• Format
• Add missing AGENT.md files for all crates (#535) (#539)

Refactor

• Rename rsketch crates to job
• Decouple telegram bot with grpc/http boundary (#94)
• Extract crawl4ai into common crate and remove unused downloader
• kernel: Consolidate queue/KV/LLM subsystems, remove rara-queue crate
• Fix naming conventions and convention drift (#1032) (#1035)

common-worker

[0.0.1] - 2026-04-13

Bug Fixes

• app: Connection pool timeout and Ctrl+C shutdown panic (#296)
• telegram: Harden tool-call XML stripping for streaming edge cases (#314)

Documentation

• Update README to reflect tape-based architecture (#783) (#784)
• Simplify readme with logo (#1088) (#1089)
• Add inspired-by credits (#1091) (#1092)

Features

• Integrate mem0 via execution worker

Miscellaneous Tasks

• Establish job backend baseline
• Clean
• Rename to rara
• Change default HTTP port from 3000 to 25555
• Format
• Format
• Format
• Add missing AGENT.md files for all crates (#535) (#539)

Refactor

• Rename rsketch crates to job
• Decouple telegram bot with grpc/http boundary (#94)
• Move TelegramService into telegram-bot crate, add outbox pattern
• Extract crawl4ai into common crate and remove unused downloader
• metrics: Migrate to lazy_static grouped pattern with eager init
• kernel: Consolidate queue/KV/LLM subsystems, remove rara-queue crate
• telemetry: Migrate metrics from Prometheus pull to OpenTelemetry push (#568)
• Fix naming conventions and convention drift (#1032) (#1035)

rara-browser

[0.0.1] - 2026-04-13

Documentation

• Update README to reflect tape-based architecture (#783) (#784)
• Simplify readme with logo (#1088) (#1089)
• Add inspired-by credits (#1091) (#1092)

Miscellaneous Tasks

• Establish job backend baseline
• Change default HTTP port from 3000 to 25555
• Format

Refactor

• Decouple telegram bot with grpc/http boundary (#94)
• kernel: Consolidate queue/KV/LLM subsystems, remove rara-queue crate
• workspace: Extract browser/stt from kernel into driver crates (#1146) (#1154)

rara-domain-shared

[0.0.1] - 2026-04-13

Bug Fixes

• notifications: Avoid 500 when pgmq archive table is absent
• settings: Add pipeline_cron field to test initializers
• shared: Add missing recall_every_turn field in test initializer (#322)
• kernel: Use ModelRepo for runtime model resolution, read Telegram token from settings

Documentation

• Update README to reflect tape-based architecture (#783) (#784)
• Simplify readme with logo (#1088) (#1089)
• Add inspired-by credits (#1091) (#1092)

Features

• Add runtime settings with hot reload for ai and telegram
• settings: Add updated timestamp and toast-based feedback
• notifications: Rebuild observability around pgmq queue semantics
• Improve settings UX and markdown preview flow
• settings: Per-scenario model configuration
• agents: Add proactive agent with personality (Agent Soul)
• chat: Make default system prompt configurable via settings (#121)
• memory: Integrate agent memory with pg/sqlite backends
• telegram: Add group chat support with mention-based triggering
• chat: Dynamic OpenRouter model list + favorites (#151)
• tools: Add screenshot tool with Playwright + Telegram photo sending (#157)
• openapi: Add OpenAPI support with utoipa + Swagger UI (#159)
• openapi: Annotate chat, job, typst, resume, settings routes (#178)
• agents: Model fallback chain (#193)
• Integrate composio
• email: Integrate lettre for Gmail sending (#216)
• pipeline: Implement job pipeline agent and service (#217)
• pipeline: Add cron scheduling for automatic pipeline runs (#220)
• pipeline: Add MCP tool support to pipeline agent
• pipeline: Send notifications to dedicated Telegram channel (#227)
• pipeline: Add report_pipeline_stats tool and recipient-based notify
• contacts: Add telegram contacts allowlist (#232)
• ai: Add Ollama provider support for local LLM inference (#240)
• settings: Add SSH public key API endpoint
• settings: Add llmfit model recommendations endpoint (#256)
• settings: Integrate Ollama model management endpoints (#268)
• web: Add capability-based filtering to Ollama model selector (#274)
• Modularize settings admin routes
• memory: Add post-compaction recall and per-turn recall config (#319)
• kernel: Settings-driven SandboxConfig + hot reload (#453)
• kernel: Wire IngressRateLimiter into IOSubsystem resolve path
• channels: Add WeChat iLink Bot channel adapter (#827) (#830)

Miscellaneous Tasks

• Establish job backend baseline
• Fmt code
• Format
• Rename to rara
• Format & some improvement & prompt markdown
• Change default HTTP port from 3000 to 25555
• Remove unused ollama-rs dependency
• Remove legacy domain settings router file
• Format
• Format
• Rustfmt formatting pass, fix Helm replicas/workers from true to 1
• Add missing AGENT.md files for all crates (#535) (#539)

Refactor

• Rename job-domain-core to job-domain-shared and clean up unused types
• Merge convert.rs into types.rs and remove redundant tests
• Move TelegramService to shared crate, add notify-driven workers
• Decouple telegram bot with grpc/http boundary (#94)
• Move TelegramService into telegram-bot crate, add outbox pattern
• Remove notify domain crate, replace with lightweight shared notify client
• Realign domain boundaries and rename domain crates
• notify: Route observability through NotifyClient
• Unify runtime state into AppState with init() and routes()
• Extract crawl4ai into common crate and remove unused downloader
• message-bus: Rewrite as Coordinator + Command trait
• memory: Use Chroma server-side embeddings, remove HashEmbedder
• memory: Simplify to PG-only backend with required Chroma
• Add keyring-store crate, process group utils, layer READMEs, and dep upgrades
• pipeline: Move agent prompt into extension crate
• settings: Replace llmfit subprocess with llmfit-core git dependency (#258)
• agent-core: Centralize model configuration behind ModelRepo trait (#279)
• Remove compose_with_soul/resolve_soul and settings prompt fields (#281)
• Split settings admin domains and reorganize settings UI
• backend-admin: Move all domain routers into backend-admin (#295)
• Move contacts to telegram-bot, add ContactLookup trait (#307)
• settings: Move SettingsSvc + ollama from domain/shared to backend-admin (#310)
• memory: Integrate new MemoryManager into tools, orchestrator, and settings (#313)
• codex: Move oauth core logic out of backend-admin
• codex: Move oauth core into integrations crate
• Remove legacy proactive agent and agent scheduler
• Remove job pipeline module and related code
• settings: Unify runtime settings into flat KV store (#401)
• kernel: Consolidate queue/KV/LLM subsystems, remove rara-queue crate
• Remove PGMQ NotifyClient — replaced by kernel egress
• llm: Per-provider default_model and fallback_models (#47)
• app: Align knowledge config with settings-first architecture

rara-paths

[0.0.1] - 2026-02-06

Documentation

• Add features section to README

Features

• Integrate Buf for multi-language gRPC code generation

Security

• Fix GitHub Actions security issues and update dependencies

Chore

• Update remade

rara-soul

[0.0.1] - 2026-04-13

Bug Fixes

• Resolve all clippy warnings across codebase (#313)
• kernel: Disable auto-fold, remove context warnings, align with bub (#1315) (#1316)

Documentation

• Update README to reflect tape-based architecture (#783) (#784)
• Simplify readme with logo (#1088) (#1089)
• Add inspired-by credits (#1091) (#1092)

Features

• soul: Add rara-soul crate with structured persona framework (#172)
• soul: Add default soul definitions for rara and nana (#172)
• soul: Rewrite rara soul with expressive personality (#175)
• soul: Implement evolve-soul tool and auto-notifications for Mita tools
• soul: Add self-identity to rara's soul prompt (#684) (#685)
• soul: Add user autonomy boundary to disable content censorship (#851) (#852)
• soul: Add working habits for tape usage to rara soul prompt (#1311) (#1312)

Miscellaneous Tasks

• Establish job backend baseline
• Change default HTTP port from 3000 to 25555
• Format
• Format
• Add missing AGENT.md files for all crates (#535) (#539)

Refactor

• Decouple telegram bot with grpc/http boundary (#94)
• kernel: Consolidate queue/KV/LLM subsystems, remove rara-queue crate
• soul: Remove all fallback logic, use built-in defaults directly
• soul: Redesign rara personality to tsundere style
• soul: Redesign rara personality to yamato nadeshiko style
• agents: Prompt diet — cut tokens ~49% (#823) (#824)
• agents: Optimize agent prompts with prompt-refinery framework (#1105) (#1106)
• soul: Introduce StyleScore newtype for bounded 1-10 values (#1126) (#1128)

rara-stt

[0.0.1] - 2026-04-13

Bug Fixes

• drivers: Structured STT errors with retry for transient failures (#1164) (#1168)

Documentation

• Update README to reflect tape-based architecture (#783) (#784)
• Simplify readme with logo (#1088) (#1089)
• Add inspired-by credits (#1091) (#1092)

Miscellaneous Tasks

• Establish job backend baseline
• Change default HTTP port from 3000 to 25555
• Format

Refactor

• Decouple telegram bot with grpc/http boundary (#94)
• kernel: Consolidate queue/KV/LLM subsystems, remove rara-queue crate
• workspace: Extract browser/stt from kernel into driver crates (#1146) (#1154)

Testing

• drivers: Wiremock-based integration tests for STT and TTS services (#1176) (#1181)

rara-tool-macro

[0.0.1] - 2026-04-13

Bug Fixes

• app: Context-mode interceptor whitelist + summary + system prompt (#722) (#732)

Documentation

• Update README to reflect tape-based architecture (#783) (#784)
• Simplify readme with logo (#1088) (#1089)
• Add inspired-by credits (#1091) (#1092)

Features

• tool-macro: Add ToolDef derive macro infrastructure (#514)
• kernel: Deferred tool loading — reduce per-turn token overhead (#756) (#768)
• kernel: Per-tool execution timeout granularity (#778) (#782)
• kernel: Add AgentTool::validate semantic validation step (#1183) (#1185)
• kernel: Add safety axes + concurrency partitioning (#1186) (#1192)

Miscellaneous Tasks

• Establish job backend baseline
• Change default HTTP port from 3000 to 25555
• Format

Refactor

• Decouple telegram bot with grpc/http boundary (#94)
• kernel: Consolidate queue/KV/LLM subsystems, remove rara-queue crate
• tool: Typed Output associated type for ToolExecute (#524) (#533)
• kernel: Drop output interceptor (#809) (#811)

rara-kernel

[0.0.1] - 2026-04-13

Bug Fixes

• egress: Fallback to persistent platform identities for stateless channels (#25)
• kernel: Unify agent_turn error type from String to KernelError
• kernel,channels: Prevent duplicate telegram messages and stuck Running processes
• kernel,boot: Address user tape code review issues (#75)
• kernel: Wire spawn_child result_tx through Session lifecycle (#76)
• kernel: Cleanup spawn_child agents after first turn completion (#76)
• kernel: Catch panics in turn task and log actual panic message (#88)
• kernel: Store origin_endpoint in Session to prevent cross-channel reply leaks (#96)
• memory: Deduplicate user message in LLM context assembly (#101)
• security: Enforce tool permissions in agent loop
• kernel: User-friendly context window error message (#129)
• tools: Rename all tool names to match OpenAI ^[a-zA-Z0-9-]+$ pattern
• kernel: Schedule-add parameter validation for LLM compatibility (#132)
• kernel,symphony: Offload blocking I/O to spawn_blocking to prevent tokio starvation
• kernel: Scheduled task isolation — independent tape and silent delivery (#140)
• kernel: Build meaningful summary for auto-handoff anchor
• Tape memory
• memory: Improve tape search relevance
• mita: Persist MitaDirective as Event entry in session tape (#173)
• kernel: Move syscall job_wheel persist to spawn_blocking (#184)
• kernel: Prevent orphan tape in checkout and add rollback (#188)
• channels: Emit TextClear to fix tool progress notifications (#207)
• kernel: Emit ToolCallStart before argument parsing (#207)
• kernel: Normalize empty tool call arguments to valid JSON
• kernel: Log successful tool calls
• kernel: Log tool call arguments at start
• kernel: Include request args in tool error log
• kernel: Use info level for LLM request/response logs
• kernel: Ensure ToolCallStart is emitted before ToolCallArgumentsDelta
• kernel: Address review feedback — configurable rate limit, memory eviction, serde parse
• kernel: Address PR review — gc wiring, clock-testable rate limiter, strum parsing (#223)
• kernel: Preserve original message in NonRetryable/RetryableServer errors (#227)
• kernel: 强制执行 max_concurrency 和 child_semaphore 并发限制 (fix(kernel): spawn_parallel 未消费 max_concurrency 参数 #242, fix(kernel): child_semaphore 已初始化但未 acquire #243)
• kernel: Add default_execution_mode to worker manifest
• telegram: Harden tool-call XML stripping for streaming edge cases (#314)
• kernel: Mark TurnTrace as failed and emit warning when max_iterations exhausted (#319) (#326)
• llm: Add frequency_penalty to prevent repetition loops (#317) (#318)
• kernel: Skip empty notifications instead of sending placeholder string (#334) (#336)
• agents: Add marketplace tool to rara agent manifest (#347)
• telegram: Pre-render trace HTML for instant callback response (#343)
• kernel: Drop PublishEvent with missing/blank payload.message (#350)
• Resolve all clippy warnings across codebase (#313)
• kernel: Suppress duplicate error message on user interrupt (#355)
• kernel: Treat plan step iteration exhaustion as replan trigger (#477)
• kernel: Use origin_endpoint directly for outbound routing (#482)
• kernel: Add logging and thinking model fallback for session title generation (#493) (#494)
• syscall-scheduler: Move scheduler to workspace dir
• scheduler: List scheduler jobs across sessions
• scheduler: Update scheduler-tool's prompt
• kernel: Default execution mode to Plan instead of Reactive (#567) (#570)
• kernel: Inject agent context into plan-mode planner (#567) (#576)
• kernel: Reduce plan step iterations and add early-exit guidance (#586) (#588)
• kernel: Replace byte truncation with self-summarizing child agents (#578) (#581)
• kernel: Persist intermediate assistant messages to tape for cascade tick detection (#606) (#608)
• guard: Expand whitelist and fix approval timeout race (fix(guard): whitelist should include data_dir and temp_dir #604, fix(guard): approval resolve race shows misleading error on timeout #605) (#609)
• kernel: Repetition guard fixes (#616) (#619)
• kernel: Pass parsed arguments to ToolCallStart for trace summaries (#621) (#623)
• kernel: Register notified() before emptiness check in ShardQueue::wait() (#687) (#692)
• kernel: Close zombie streams when opening new stream (#688) (#693)
• kernel: Log and recover from job_wheel mutex poisoning (#689) (#694)
• kernel: Prefer accumulated_text over reasoning for TurnRationale (#713) (#715)
• kernel: Allow ancestor paths in path_scope guard for directory tools (#727) (#733)
• app: Context-mode interceptor whitelist + summary + system prompt (#722) (#732)
• kernel: Align OpenAI driver wire format with API spec (#743) (#745)
• kernel: Default ExecutionMode to Reactive and cap plan-mode iterations (#752) (#754)
• kernel: Openai wire format alignment (#747) (#750)
• kernel: Make interceptor prompt fragment dynamic to track MCP state (#763) (#769)
• kernel: Tune agent loop timeouts for responsiveness (#770) (#772)
• kernel: Add logs_dir and ~/.claude to path-scope whitelist (#780) (#781)
• kernel: Set max_tokens to 1024 in agent loop CompletionRequest (#802) (#803)
• kernel: Increase max_tokens to 2048 and optimize agent prompt (#815) (#816)
• kernel: Nudge LLM when it returns empty text after tool calls (#855) (#856)
• kernel: Send fallback reply on empty LLM response after error recovery (#894) (#896)
• kernel: Make typing_refresh respect turn cancellation (#900) (#903)
• kernel: Include descriptions in discoverable tools system prompt (#914) (#915)
• kernel: Inject system paths and improve discover-tools guidance in agent system prompt (#919) (#920)
• kernel: Reword deferred tools prompt to stop agent claiming tools unavailable (#935) (#937)
• kernel: Align deferred tool catalog with executable tool registry (#941) (#942)
• kernel: Add timeout enforcement for background agent watchers (#928) (#931)
• kernel: Allow multiple LLM error recoveries per agent turn (#930) (#932)
• kernel: Use parking_lot mutex (#927) (#933)
• kernel: Improve context estimation (#929) (#934)
• kernel: Expose full error chain in LLM provider errors (#1017) (#1018)
• kernel: Bypass proxy for local/private-network LLM providers (#1020) (#1021)
• Resolve contract violations found by code quality scan (#1026) (#1030)
• kernel: Merge system messages for strict chat template providers (#1034) (#1038)
• kernel: Strip tags from content field (#1054) (#1055)
• kernel,web: Preserve multimodal image content through LLM pipeline (#1063) (#1064)
• chat: Preserve multimodal in api (#1066) (#1067)
• kernel: Preserve multimodal content on first-turn spawn (#1068) (#1070)
• kernel: Treat empty LLM stream as error with auto-fold recovery (#1116) (#1117)
• kernel: Use SessionKey/ChannelType types in ChannelMessage and ChannelBinding (#1120) (#1132)
• cmd: Index 'rara debug' lookup via execution_traces SQL (#1138) (#1139)
• kernel: Explicit backpressure retry for IOError::Full in ingress pipeline (#1148) (#1158)
• cmd: Render full ExecutionTrace from SQL in 'rara debug' (#1156) (#1161)
• kernel: Intercept XML tool calls from MiniMax in content stream (#1166) (#1170)
• kernel: Tool_xml safe_emit_len panics on multi-byte UTF-8 (#1173) (#1174)
• kernel: Handle streamed tool deltas even when stop_reason mismatches (#1201)
• channels: /debug shows all entry kinds via entries_by_message_id (#1207) (#1209)
• kernel: Detect vision support by model name, stop stripping images (#1216) (#1218)
• codex-oauth: Add ChatGPT-Account-Id header for API access (#1241) (#1244)
• kernel: CodexDriver reads HTTPS_PROXY for chatgpt.com connectivity (#1248)
• kernel: CodexDriver proxy support + required instructions field (#1249)
• kernel: Codex driver — extract system messages into top-level instructions field (#1252)
• kernel: Codex driver — set store=false (endpoint rejects store=true) (#1253)
• kernel: Codex driver — remove unsupported max_output_tokens parameter (#1254)
• kernel: Codex driver — strip all unsupported parameters from request body (#1255)
• kernel: Codex driver — explicitly send store=false (API defaults to true) (#1256)
• kernel: Ensure object schemas always have properties key (#1257)
• kernel: Codex driver — add 5-minute absolute stream timeout (#1260)
• kernel: Simplify context contract following bub's approach (#1309) (#1310)
• kernel: Disable auto-fold, remove context warnings, align with bub (#1315) (#1316)
• channels: Route approval requests to originating chat, not primary_chat_id (#1319) (#1322)
• kernel: Ack detector now catches mid-turn planning responses (#1332) (#1333)
• kernel: Sanitize tool names for OpenAI API compatibility (#1342) (#1343)

Documentation

• memory: Add detailed what/how/why comments to tape memory module (#64)
• kernel: Add detailed comments to start_llm_turn explaining lifecycle phases
• telegram: Add implementation comments for anchor tree flows
• kernel: Enrich checkout action description in TapeTool (#202)
• kernel: Add AGENT.md guidelines for IngressRateLimiter and GroupPolicy
• kernel: Add 'why' reasoning to AGENT.md guidelines
• kernel: Add AGENT.md section for tape-driven message rebuild + context budget (#229)
• Update README to reflect tape-based architecture (#783) (#784)
• Simplify readme with logo (#1088) (#1089)
• Add inspired-by credits (#1091) (#1092)
• kernel: Add ARCHITECTURE.md mapping kernel data flow and module roles (#1140) (#1150)

Features

• kernel: Implement KernelHandle as event-queue-based public API (#23)
• session: 添加 SessionIndex trait 和 FileSessionIndex 实现 (#43)
• cmd: Improve TUI session Gantt chart with metrics overlay and time axis
• memory: Add user tape for cross-session user memory (#70)
• kernel: Add group chat proactive reply with two-step LLM judgment (#71)
• kernel: Add KnowledgeConfig struct (#81)
• kernel: Add knowledge layer — items, categories, embedding, extractor, tool (#81)
• kernel: Wire knowledge layer into kernel event loop and boot sequence (#81)
• channels: Support sending images to users in Telegram (#91)
• memory: Add source_ids to compaction anchor and entry lookup by ID (#104)
• memory: Expose source references in knowledge search and add resolve_sources (#105)
• memory: Support fork from specific entry ID (#107)
• agent: Emit intent/progress during long multi-step tool executions (#116)
• kernel: Dynamic MCP tool injection into agent loop (#126)
• kernel: Replace oneshot result channel with mpsc AgentEvent channel (#127)
• kernel: Run_agent_loop emits milestones via mpsc channel (#127)
• kernel: Exec_spawn collects milestones into tool result (#127)
• kernel: Usage collection, tape tools, and context contract (#130)
• llm: Image compression pipeline for vision input (#131)
• kernel: Add scheduled task system (#132)
• kernel: Auto-handoff on context window overflow (#134)
• kernel: ScheduledJobAgent + enriched task notifications (#135)
• kernel: KernelEvent::SendNotification + fix PublishEvent syscall (#137)
• kernel: Runtime context guard with token feedback (#149)
• llm: Add Message::tool_result_multimodal() constructor
• kernel: Add ToolOutput type and update AgentTool::execute() signature
• kernel: Add desired_session_key to spawn_with_input (#164)
• kernel: Store LLM usage metadata on assistant tape entries (#165)
• memory: Add estimated_context_tokens to TapeInfo (#165)
• kernel: Expose estimated_context_tokens in tape.info tool (#165)
• kernel: Use estimated_context_tokens in context pressure warnings (#165)
• memory: User tape knowledge distillation via anchor (#170)
• kernel: Render soul prompt with runtime state via SoulRenderer (#174)
• kernel: Add mood inference hook at end of agent loop (#176)
• kernel: Add rate limit retry with exponential backoff for LLM calls
• telegram: Add /anchors and /checkout commands
• kernel: Add checkout action to tape tool (#188)
• kernel: Teach LLM about anchor navigation in runtime prompt (#188)
• kernel: Inject SessionIndex into TapeTool for real checkout (#193)
• kernel: Add checkout_root action to TapeTool (#204)
• channels: Include raw args in tool parse failure error (#207)
• kernel: Log LLM request and response at debug level
• kernel: Add GroupPolicy enum to channel types (security(channels): Zeroize bot token in config and adapter path #219-adjacent)
• kernel: Add IngressRateLimiter with sliding-window per-key limiting
• kernel: Wire IngressRateLimiter into IOSubsystem resolve path
• kernel: Add context budget for tool result truncation (#228)
• kernel: Add /msg_version command and session/manifest routing (#257)
• channels: Plan-execute TG 三级显示策略 + 单消息编辑流 (#267)
• telegram: Show input/output token counts in progress UX (#304)
• kernel,telegram: Rara_message_id end-to-end tracing and debug_trace tool (#337)
• kernel: Background agent spawning with proactive result delivery (#340)
• kernel: Context folding — auto-anchor with pressure-driven summarization (#357)
• kernel: Implement ContextFolder and FoldBranchTool (#376)
• kernel,telegram: Auto-generate session title & redesign /sessions UI (#434)
• kernel: External agent.md prompt (#451)
• kernel: Centralize loading hints with random selection (#455)
• channels: Add /status command with session info and scheduled jobs (#450) (#453)
• kernel: Implement TaskReport type system and notification subscription bus (#438) (#439)
• dock: Generative UI canvas workbench (#424)
• kernel: Agent knowledge directory with index + on-demand loading (#466) (#469)
• kernel: Read-file adaptive paging based on context window (#468) (#471)
• kernel: Default execution mode to plan (v2) (#479) (#481)
• kernel: Add browser automation subsystem via Lightpanda + CDP (#473)
• kernel: Inject installed skills into agent system prompt (#487)
• channels: Add session delete buttons and relative time in /sessions (#492)
• tool-macro: Add ToolDef derive macro infrastructure (#514)
• kernel: Implement pause_turn circuit breaker for agent loop (#506) (#508)
• tool-macro: Migrate kernel tools to ToolDef (#510) (#517)
• web: Cascade viewer — agent execution trace side panel (#513)
• acp: Add native acp client crate (#504)
• kernel: Enrich tape metadata with typed structs for latency, model, and stop_reason (#548) (#551)
• kernel: Add PathScopeGuard for file-access scope enforcement (#579) (#582)
• kernel: Session-level dynamic whitelist for PathScopeGuard approvals (#596) (#597)
• kernel: Streaming repetition guard for LLM output loops (#602) (#614)
• kernel: Add session-length handoff reminder and strengthen topic-switch prompt (#615)
• kernel: Show LLM reasoning for tool calls in progress display (#661) (#664)
• channels: Enhance guard approval UI with context, timing, and expiration (#653) (#665)
• kernel: Add LlmModelLister and LlmEmbedder extension traits (#762) (#766)
• kernel: Add tool call loop breaker (#773) (#775)
• kernel: Deferred tool loading — reduce per-turn token overhead (#756) (#768)
• kernel: Per-tool execution timeout granularity (#778) (#782)
• kernel: Stream bash stdout in real-time during tool execution (#777) (#788)
• app: Port kota file tools — in-process grep/find, delete-file, create-directory (#808) (#810)
• channels: Add WeChat iLink Bot channel adapter (#827) (#830)
• kernel: Discover-tools finds skills (#833) (#835)
• kernel: Task tool — preset-based background agent delegation (#845) (#847)
• kernel: Add explore task preset and delegation routing (#873) (#875)
• kernel: Inject deferred tool names into agent system prompt (#902) (#906)
• tools: Include parameter summaries in discover-tools results (#925) (#926)
• channels: Improve Telegram progress UX (feat(channels): collapse tool command details into cascade instead of truncating #947, feat(channels): fix thinking content positioning in message timeline #948, feat(channels): collapse guard approval message after decision #949) (#951)
• kernel: Add ToolHint, UserQuestionManager, and ask-user tool (#945) (#952)
• kernel: Codex OAuth as first-class LLM provider (#950) (#953)
• kernel: Telegram voice message STT via local whisper-server (#998) (#1003)
• cmd: Rara setup interactive configuration wizard (#1008) (#1011)
• kernel: Detect model vision capability and strip image blocks (#1076) (#1078)
• app: Auto-manage whisper-server lifecycle (#1081) (#1082)
• web: Voice message input via microphone recording (#1084) (#1085)
• browser: Enforce navigation timeout and make wait_for timeout configurable (#1107) (#1108)
• kernel: Batch file reads and 429 rate-limit recovery (#1118) (#1119)
• cmd: Add 'rara debug <message_id>' CLI command (#1135) (#1136)
• kernel: Per-session child concurrency semaphore (#1144) (#1155)
• kernel: Add AgentTool::validate semantic validation step (#1183) (#1185)
• kernel: GLM-5.1 provider-specific optimizations (#1196) (#1197)
• kernel: Add safety axes + concurrency partitioning (#1186) (#1192)
• kernel: Tool execution hooks — PreToolUse / PostToolUse / PostToolUseFailure (#1204) (#1208)
• kernel: Scheduling architecture hardening — 10 improvements from matklad review (#1222) (#1245)
• kernel: Add CodexDriver for ChatGPT backend Responses API (#1246) (#1247)
• kernel: Add LifecycleHook trait and registry for turn/fold events (#1286) (#1288)
• kernel: Implement SkillNudge + MemoryNudge lifecycle hooks (#1290) (#1292)
• kernel: Self-continuation signal to prevent GPT stop-and-ask (#1301) (#1306)
• kernel: Interrupt agent turn on new user message (#1317) (#1318)
• kernel: Add task delegation guidance to context_contract (#1321) (#1323)
• agents: Make safety fragment act-by-default to improve proactivity (#1320) (#1324)
• kernel: Runtime ack detection to prevent lazy LLM responses (#1329) (#1330)
• kernel: GPT-specific anti-narration prompt + raise ack length to 2000 (#1340) (#1341)
• kernel: Hermes-aligned tool-use enforcement for GPT/Gemini/Grok (#1340) (#1344)
• kernel: Add WHEN TO USE guidance to task tool and promote spawn-background to core (#1348) (#1352)
• kernel: Three-tier ack detection for verbose GPT narration (#1349) (#1353)
• kernel: Integrate talk-normal anti-slop output style rules (#1350) (#1354)

Miscellaneous Tasks

• Establish job backend baseline
• Change default HTTP port from 3000 to 25555
• Format
• Tidy project
• Format code
• kernel: Replace From<(&str, Option)> with explicit classify_provider_error, add stage constants
• Rustfmt formatting pass, fix Helm replicas/workers from true to 1
• Format
• Clean
• kernel: Add usearch dependency for knowledge layer (#81)
• Change jobs_path
• Format
• Add tool timeout
• Support composio config
• kernel: Add perf TODO for anchor tree session loading (#188)
• Format
• Format
• Auto tool
• channels: Add delivery debug logging for wechat send path (#881) (#882)
• kernel: Increase default_max_iterations from 12 to 60 (#1313) (#1313)

Performance

• cascade: Build CascadeTrace incrementally during agent loop (#625) (#632)
• kernel: Batch sequential event handling in run_processor (#1141) (#1152)
• kernel: In-memory tape index for anchor and kind lookups (#1147) (#1157)
• base: Use NonZeroU128 in define_id! for Option niche optimization (#1149) (#1160)

Refactor

• Decouple telegram bot with grpc/http boundary (#94)
• kernel: Extract SecuritySubsystem
• kernel: Extract AuditSubsystem
• kernel: Flatten KernelInner, add strum derives, instrument macros, and Arc type aliases
• kernel: Use join_all for concurrent event batch processing (#20)
• kernel: Migrate external callers to KernelHandle, demote Kernel methods (#24)
• kernel: Remove redundant spawn methods from Kernel
• kernel: Remove async-openai and legacy LLM provider layer
• kernel: Add OutboundEnvelope constructors, eliminate duplicate struct literals
• kernel: Extract routing helpers from handle_user_message
• kernel: Reduce ProcessHandle boilerplate with helper methods
• kernel: Consolidate queue/KV/LLM subsystems, remove rara-queue crate
• kernel: Replace manual map_err with snafu ResultExt (#33)
• kernel: Replace manual enum match with strum derives (#34)
• kernel: Replace manual Debug impls with derive_more::Debug (#35)
• kernel: Dissolve defaults/ module into domain modules (#36)
• kernel: Split session.rs into directory module, fix external import paths (#36)
• kernel: 将 RuntimeTable 从类型别名提升为领域结构体 (#39)
• kernel: 提取 DeliverySubsystem 子组件 (#39)
• kernel: 提取 SyscallDispatcher 子组件 (#39)
• kernel: 平铺过度拆分的子模块 (#40)
• Replace Consul with YAML config, continue SQLite migration, add session resumption
• kernel: Introduce EventBase for unified event metadata
• llm: Per-provider default_model and fallback_models (#47)
• kernel: Session-centric runtime (#48)
• kernel: Migrate from Memory+SessionRepository to tape (#51)
• Make it compile
• kernel: Remove SessionResolver, simplify ChannelBinding (#63)
• kernel: Extract TapeTool from SyscallTool into dedicated tool
• kernel: Improve TapeTool error handling, add between_anchors action (#68)
• kernel: Decouple proactive judgment into GroupMessage event (#79)
• kernel: Make knowledge layer a required component, not optional (#81)
• kernel: Remove enabled flag from KnowledgeConfig (#81)
• app: Align knowledge config with settings-first architecture
• memory: Compaction shrinks read set instead of deleting history (#103)
• memory: Define typed HandoffState contract for anchor state (#106)
• kernel: ScheduledTask as dedicated KernelEvent + notifications (#133)
• kernel: Split schedule-add into three LLM-friendly tools
• kernel: Rewrite tape tool description with topic-driven anchor and recall guidance
• mita: Replace submit_message with typed MitaDirective (#171)
• soul: Remove all fallback logic, use built-in defaults directly
• kernel: Spawn per event instead of join_all to prevent batch blocking (#185)
• kernel: Extract InMemorySessionIndex to shared test utility (#188)
• kernel: Move checkout_anchor to TapeService (#188)
• kernel: 每次迭代从 tape 重建 LLM messages，消除双写冗余 (#229)
• kernel: Replace Chinese prompts with English in agent loop (#229)
• kernel: Consolidate tool impls into tool/ module (#264)
• memory: Remove MAX_USER_NOTES truncation, trust anchor boundary (#407)
• kernel: Migrate browser tools to ToolDef derive macro (#511) (#518)
• app: Migrate app, dock, and knowledge tools to ToolDef derive macro (#512) (#519)
• tool: Typed Output associated type for ToolExecute (#524) (#533)
• telemetry: Migrate metrics from Prometheus pull to OpenTelemetry push (#568)
• tg: Unify plan and progress into single in-progress message (#580) (#583)
• kernel: Plan mode agent loop fixes (#648 #649 #650) (#667)
• kernel: Prompt review — fix 12 findings (#755) (#758)
• kernel: Flatten spawn-background params — remove nested manifest (#764) (#767)
• tools: Token diet — aggressive tool tiering + new file tools + browser prompt (#805) (#806)
• kernel: Drop output interceptor (#809) (#811)
• agents: Prompt diet — cut tokens ~49% (#823) (#824)
• kernel: Tool schema diet — split tape + compress descriptions (#825) (#826)
• kernel: Prompt diet tool tiering (#831) (#832)
• agents: Align system prompts with Claude Code patterns ([#841](https://github.com/rararulab/rara/