v0.21.0

✨ New Features

🖥️ A2A Embedded Chat UI

• Added a browser-based chat UI served directly by the A2A server (root /,
  reserved /_a2a-ui/ prefix) so you can chat with your agents without a
  separate client.
• Streaming responses with append/replace artifact semantics, thinking blocks,
  status pills, and live task recovery after a page refresh (GetTask →
  resubscribe).
• Conversation management: create/rename/delete, persisted in IndexedDB;
  full A2A contextId/taskId continuity.
• Agent selector per turn, input-required continuation locking, inline file
  upload (up to 5 files, 10 MB each) as A2A raw parts, and retry support.
• Optional API-key auth: the key is kept in sessionStorage only (never
  persisted); public UI routes stay public while protected routes require a
  Bearer token. The public /_a2a-ui/config endpoint never leaks the key.
• New public /agents route for client-side agent listing.

🧠 Reasoning Effort Control (--reason-effort)

• New CLI argument --reason-effort (none, minimal, low, medium,
  high, xhigh) on all commands (chat / console / GUI, A2A server, ACP
  agent, job).
• New per-agent reason_effort field in agent configs (agents.toml),
  editable from the GUI agent editor.
• /think slash command now accepts reasoning levels; the legacy numeric
  thinking budget still works.
• Documented reasoning precedence: user switch (/think) > CLI > agent config

  model default. New reasoning_selection module shares canonical levels and
  an Anthropic-family token-budget adapter with the ACP model controller.

📇 Reworked Model Selection

• Typed value objects (RuntimeModelInput, ModelSelection,
  ModelSelectionSource) with a single documented precedence chain:
  user switch (/model) > explicit runtime args > agent config > detected
  AGENTCREW_MODEL_ID > persisted last-used > provider default.
• /model force switches now survive config reloads; is_forced / is_pinned
  are explicit per-selection properties.
• Detected values are never masked as explicit arguments, so precedence stays
  honest across all startup routes (console, GUI, A2A server, ACP, job).
• Every local agent now owns a dedicated, uncached LLM service so per-agent
  reasoning mutations never leak to another agent.

🔍 Full-text Conversation Search (PR #48 by @fy17ohhh)

• Console conversation browser now searches both titles and message
  content with highlighted match previews (role context + ellipses).
• Session-local lazy index: histories are normalized once (Unicode casefold,
  whitespace collapse) and cached, so later queries skip filesystem reads.
• Internal-only content (memory preamble, injected file blocks, non-chat tool
  blocks) is excluded from the index.

🤖 New Models

• Qwen 3.8 2.4T A95B (DeepInfra, Fireworks, Together): 262k context,
  tool use + thinking + vision, default reasoning high.
• Meta Muse Glimmer 30B (Fireworks, Together): multimodal agent model,
  131k context, schema-based tool calling, default reasoning high.
• DeepSeek V4 Pro Latest and Qwen3.8 27B (CrofAI).
• Model default_reasoning extended with xhigh, max, ultra.

🧾 Documentation (PR #47 by @fy17ohhh)

• Custom LLM provider guide with Ollama / llama.cpp / LM Studio schema,
  remote SSE MCP example, and background MCP discovery notes.

---

⚡ Performance & Startup

• Startup time optimized: version-check wait reduced from 5s → 2s, and
  httpcore, httpcore2, httpx, httpx2 loggers silenced to ERROR.
• RemoteAgent is now lazily imported via module __getattr__, dropping
  heavy A2A imports from the startup path.
• MCP discovery moved to durable, manager-owned background threads (with their
  own event loop) that survive command-loop closure; built-in tools sync
  immediately, MCP tools final-sync on the first message with a bounded
  30-second fail-open wait. A timeout never cancels the underlying discovery.
• Per-agent discovery deduplication (concurrent.futures.Future) and a stale
  completion guard so a completed background worker can never repopulate a
  deactivated agent.

🐛 Bug Fixes

• Token stats: ConversationUsage.add_turn no longer accumulates
  per-turn input/output/cached tokens into a running counter — turns are now
  recorded correctly while totals/cost still accrue.
• /debug no longer mutates messages: debug output deep-copies chat and
  streamed histories so debug can't affect real message state.
• Debug JSON rendering: markup=False prevents Rich markup from hiding
  message content.
• Agent history behavior: switching agents only inherits the previous
  agent's history when the new agent has none of its own, and no longer
  resets an agent already involved in a conversation.
• Remote agent state reset is now only applied when the target agent is
  actually a RemoteAgent (lazy imported), preserving conversation manager
  behavior for local agents.
• MCP session lifecycle: fixed rework replaces brittle setup timeouts
  with a strict reverse-order scope teardown (session → transport → HTTP
  client) that closes every resource exactly once; teardown failures and
  cancellation never mask the primary error (fixes a historical OAuth/AnyIO
  hang).
• MCP v2 field compat: pagination uses next_cursor, resources use
  mime_type, tools use input_schema, resource links use plain string URIs;
  Error tool results now wrap MCPError messages.
• OAuth: callback captures RFC 9207 iss issuer and returns
  AuthorizationCodeResult, while legacy stored client info (no issuer)
  remains loadable.

---

⚙️ Upgrades & Breaking Changes

• mcp 1.27.1 → 2.0.0 (MCP v2 spec): SDK split into mcp + mcp-types,
  API renames above, OAuth AuthorizationCodeResult semantics.
• httpx → httpx2 (+ httpcore2): all internal HTTP clients (A2A
  downloads, remote agent streaming, OAuth, usage APIs, image fetch, web
  extraction) migrated. ⚠️ Accepted residual risk: a2a-sdk still types and
  catches httpx exceptions, so httpx2 transport errors may bypass the SDK's
  A2AClientError conversion until a2a-sdk adopts httpx2 (documented in
  TestRemoteAgentHttpClientBoundary).
• streaming_server semantics documented precisely: URLs ending in /sse
  use the legacy SSE transport; any other URL uses Streamable HTTP.

---

🧰 Internal Refactors

• LLM ownership lifecycle extracted into AgentLLMLifecycle
  (dedicated_llm, is_service_owned, close_superseded, release_llm,
  ensure_reasoning_isolated, apply_reasoning, sync + async replacements).
  ServiceManager gained clone_service, get_service_for_selection,
  close_service, and draining of pending close tasks at shutdown with
  exactly-once per-service dedup.
• Collaborator extraction from LocalAgent into
  AgentMemoryCoordinator (memory parsing/storage) and
  AgentMessageFormatter (assistant/tool formatting), mirroring the existing
  AgentContextManager / AgentToolRegistrar pattern; public wrappers keep
  every caller unchanged.
• AgentManager.deregister_agent now releases a removed LocalAgent's owned
  dedicated LLM service (exactly once, when not cached and not referenced by
  a remaining agent); shutdown drains all dedicated services.
• Debug message arg preview now shows both start and end of long values
  (...original[-25:]) for more useful traces.

---

👏 Contributors

We'd like to thank everyone who contributed to this release:

• @daltonnyx (Quy Truong) — author of most of v0.21.0: A2A chat UI, reasoning-effort controls, model selection rework, MCP v2 + httpx2 upgrades, performance work, and bug fixes.
• @fy17ohhh — full-text conversation search with match previews (#48), and custom provider documentation (#47).

Your time and effort make AgentCrew better for everyone. 💙

---

Full changelog: see the commit history