What's Changed

Features

• Bedrock API Key support — AWS Bedrock now accepts the simpler bearer-token API keys (AWS 2024-Q4 feature) as an alternative to traditional SigV4. Both CLI setup wizard and Web UI offer a toggle between API Key (default) and AWS Access Key modes.

Fixes

• CI's fmt and clippy jobs were failing because the nightly toolchain pinned via rust-toolchain.toml didn't install those components. Now declared explicitly.

Full Changelog: v0.1.0-alpha.41...v0.1.0-alpha.42

What's Changed

Features

• Amazon Bedrock provider — Converse API support for Claude, Llama 3.1, Nova, Mistral, and Cohere. Both streaming (via hand-written AWS event-stream decoder) and non-streaming, with tool use. First-class option in clawhive setup and the Web UI.
• Skill management — install skills from GitHub shorthand (user/repo) or a single SKILL.md file; new /skill list and /skill remove chat commands; install metadata (.metadata.json) recorded for future skill update.
• Web search — multi-provider fallback with circuit breaker.
• Memory — detect time-range keywords in queries for better recall.

Improvements

• Typed errors at crate boundaries (AuthError, MemoryError, BusError, ProviderError) replace ad-hoc anyhow in public APIs.
• Centralized command registry shared by all channel adapters (Discord, Telegram, web).

Fixes

• Telegram: fix UTF-8 boundary panic when splitting long messages.
• Discord: include subcommand name for no-arg skill commands.
• Channel autocomplete now lists new skill commands.

Full Changelog: v0.1.0-alpha.40...v0.1.0-alpha.41

What's Changed

Fixes

• WhatsApp and iMessage channels no longer silently fail during setup wizard — validation and error feedback now work correctly

Full Changelog: v0.1.0-alpha.39...v0.1.0-alpha.40

What's Changed

Prompt System

• PromptMode enum (Full/Minimal/Scheduled/Heartbeat) — context files loaded per mode, reducing token waste
• New identity_parser module for structured agent identity extraction

Memory

• Per-session compaction mutex with 60s timeout
• Patch retry before full consolidation fallback
• Progressive disclosure with snippets and score breakdown in search results
• memory audit, rebuild-index, stats CLI commands
• Per-item dirty source processing to prevent queue blocking

CLI

• auth reset-password command

WhatsApp

• LID→phone retry (3s delay) before falling back to raw JID
• Adapted to upstream media key type change (Vec<u8> → [u8; 32])

Fixes

• JSON fence stripping handles arbitrary language tags
• Consolidation idle_minutes / daily_at_hour default to None

Full Changelog: v0.1.0-alpha.38...v0.1.0-alpha.39

What's Changed

Memory System Rewrite

• Episode-first pipeline with memory lineage tracking and file audit
• Upgraded SQLite hybrid search (vector + FTS5)
• New clawhive memory CLI commands

Approval Improvements

• LLM-powered human-readable approval summaries across all channels
• 10-minute approval timeout (prevents permanent session blocking)
• WhatsApp & WeChat approval listeners with quick reply (yes/no/always <id>)

WhatsApp QR Pairing

• QR pairing API + web console UI (Pair Device, status polling, Re-login)
• Merged ~50 upstream whatsapp-rust fixes
• LID-to-phone resolution for access policy matching

Fixes

• Config change detection: channel changes now trigger reload correctly
• Fix UTF-8 panic in shell_tool output truncation
• Auto-create routing binding when adding connector
• Update provider presets (gpt-5.3, gpt-5.4)

Full Changelog: v0.1.0-alpha.37...v0.1.0-alpha.38

What's Changed

Fix: Agent Stalling on Long Tasks (Empty Promise Detection)

• fix(core): decouple empty-promise retry budget from other nudge mechanisms
  • Empty-promise detection (detect_empty_promise_structural) now uses its own independent retry counter instead of sharing scheduled_task_retries with incomplete-thought and fabricated-response checks
  • Previously, if should_retry_incomplete_scheduled_thought consumed the shared retry budget, the structural empty-promise detector (colon-ending + short text + no tool_use) would be permanently bypassed, causing the agent to stall mid-task
  • Empty-promise retry cap raised from 1 to 2

Improvement: Proactive Context Compaction at 75%

• fix(core): trigger context compaction at 75% of available tokens (was 100%)
  • Prevents the LLM from slowing down and losing focus as context approaches the limit
  • Memory flush threshold adjusted accordingly — now triggers before the 75% compaction threshold instead of at ~93%
  • With 128K context: compaction now at ~93K tokens (was ~124K), memory flush at ~85K tokens (was ~116K)

Full Changelog: v0.1.0-alpha.36...v0.1.0-alpha.37

What's Changed

New: /model Command — Switch Agent Model at Runtime

• feat(core): extend /model slash command to accept provider/model argument
  • /model (no args) still shows current model info
  • /model provider/model validates provider & model, persists to agent YAML, and hot-swaps in-memory config
  • Strict validation: provider must be configured, model must exist in presets (custom providers skip model check)
  • Error messages list available providers/models when validation fails
• feat(channels): register /model command with parameter support in Discord, Telegram, and Web UI
  • Discord: interactive parameter input via slash command option
  • Telegram: updated bot command description with usage hint
  • Web: autocomplete shows [provider/model] argument

Improvements

• feat(setup): include all preset models in provider YAML during clawhive setup (was only writing default model)
• feat(setup): add Moonshot API region selection (China / International) during provider setup

Bug Fixes

• fix(gateway): drop unrouted messages instead of falling back to default agent
• fix(feishu): return proper card action ACK with toast for approval cards (fixes error 200340 on button click)
• fix(web): hide slash command autocomplete dropdown when input exactly matches a command (was blocking Enter to send)

Full Changelog: v0.1.0-alpha.35...v0.1.0-alpha.36

What's Changed

New: WeChat Channel via iLink Protocol

• feat(channels): add WeChat personal account as a messaging channel via Tencent's official iLink Bot protocol
  • Long-polling message receiver (similar to Telegram Bot API)
  • Text + image support with AES-128-ECB encrypted CDN media
  • QR code login with session persistence (no re-scan on restart)
• feat(cli): add WeChat to clawhive setup interactive wizard with terminal QR code scanning
• feat(server,web): add WeChat to web console Channels page with in-browser QR code login
  • Backend API endpoints for QR login flow (/api/channels/weixin/qr-login, /api/channels/weixin/qr-status)
  • QR code rendered client-side via qrcode.react

Bug Fixes

• fix(memory): isolate search index by agent_id to prevent cross-agent memory leakage — agents no longer see each other's sessions/memory in search results
• fix(channels): use compact Dense1x2 QR renderer for WeChat terminal login (was too large to scan)
• fix(channels): suppress tracing logs during interactive QR login flow
• fix(server): report WeChat channel status correctly (no token credential check needed)
• fix(web): handle WeChat in add-connector dialog without token field

Full Changelog: v0.1.0-alpha.34...v0.1.0-alpha.35

What's Changed

• fix(scheduler): remap POSIX day-of-week values (0=Sun) to cron crate convention (1=Sun) in 5-field cron expressions — fixes scheduled tasks firing on wrong days
• fix(server): web console chat now routes to the explicitly selected agent instead of falling back to default routing
• fix(core): auto-inject all ~/.clawhive/.env variables into tool execution sandbox — no longer requires explicit sandbox.env_inherit declaration
• fix(core): add configurable max_iterations per agent (default 50, was hardcoded 25)
• fix(provider): increase API request timeout from 60s to 180s for long-running agent tasks
• fix(provider): rename "openai api error" to "api error" for provider-agnostic error messages
• fix(channels): fallback from send_photo to send_document when Telegram returns IMAGE_PROCESS_FAILED

Full Changelog: v0.1.0-alpha.33...v0.1.0-alpha.34

What's Changed

• fix(core): increase default max_response_tokens from 2048 to 4096 for conversations, add per-agent YAML configuration
• fix(core): increase conversation truncation retries from 1 to 2 with improved "summarize concisely" nudge
• fix(core): reduce MAX_OUTPUT_BYTES from 50KB to 20KB to prevent tool output from overwhelming context
• fix(core): provide friendly error message when stop_reason=length causes empty reply
• fix(schema): improve ApprovalDisplay program extraction to skip shell variable assignments and subshells

Full Changelog: v0.1.0-alpha.32...v0.1.0-alpha.33