This release ships a unified human-in-the-loop /resume mechanism across every channel, a slash composer for invoking skills and other agents directly, typed file tools that cut token usage and hallucinations, and per-channel message splitters so long replies arrive cleanly instead of getting truncated. Approvals (app deploys, multiple-choice questions, credential picks) now render as buttons on Telegram, Discord, Slack, and WhatsApp Cloud, and as Reply YES or NO
prompts on Signal, SMS, and personal WhatsApp, with the inbound parser accepting the obvious variants. Behind the scenes, a new Harness
struct consolidates the agent runtime view of AppState
, sandboxing moves to a single SandboxManager
per principal kind, and the channel adapter framework now classifies failures with a typed ChannelError
.
Supported channels: Slack, Discord, Telegram, Signal, SMS, WhatsApp Cloud, WhatsApp Personal.
Human in the loop
- Replace opaque
tool_data
blobs on tool calls with typedHitl
values; build typed HITL tool components and consume new SSE event names. - Persist
d
message status and resume with an atomic compare-and-swap; translate the newInferenceResponse
outcomes into /resume persistence. - Add a
POST /api/chats/{chat_id}/tool-calls/resolve
endpoint, lift the HITL callback parser and response label into a sharedhitl
module, and drop the legacy app and vault approval routes superseded by the resolve endpoint. - Re-seed pending HITLs from
msg.tool_calls
oninference_done
so d turns continue to surface new questions. - Share an inbound reply HITL resolver with YES/NO variant parsing (
y
,yeah
,ok
,nope
, thumbs-up emoji, and the usual variants). - Resume HITL delivery on resolve so sequential adapters render the next prompt; use the absolute
public_base_url
for HITL fallback URLs on channels that can't render the picker. - Render HITL prompts and resolve responses on Telegram (inline keyboard), Slack (Block Kit via Socket Mode interactions), Discord (buttons), WhatsApp Cloud (interactive messages and button taps), WhatsApp personal account (quote replies), Signal (quote replies), and SMS (
Reply YES or NO
hint on App-deploy approval prompts). - Hold the Discord and WhatsApp typing indicator across long inferences so the user sees activity while the agent is generating.
- Rename the
manage_service
tool tomanage_app
and adopt typed HITL hooks; exit tasks on HITL and respawn viarun_task
. - Test HITL persistence, rendering, batched resolve, and the resume race.
Slash composer
- Add a
chat::slash
invocation parser for/
and@
prefixes; add aMessageCommand
side field onMessage
for parsed slash invocations. - Wire up
/commands
end to end: dispatch, terminal-write refactor, cross-agent attribution, slash parsing. - Add a
GET /api/chats/{id}/commands
discovery endpoint and alistCommands
API client. - Frontend: slash composer with Lexical triggers, directive chips, and a
/new
builtin that opens a fresh chat with the current agent. @<agent>
per-turn override: the reply is attributed to the target agent and the next message reverts to the chat's default.- Add
save_chat
,save_updated_message
, anddelete_messages_for_chat
helpers. - Skills: new SKILL.md frontmatter fields
disable-model-invocation
,argument-hint
, andarguments
; hidedisable-model-invocation
skills from the model's available-skills block while keeping them in the/
menu for users to invoke directly. - Add e2e tests for
/agent
dispatch and commands discovery.
Channels and message splitter
- Add a per-format message splitter with shared boundary primitives that break at paragraph, then line, then word, then UTF-8 character, and never inside a code fence or escape sequence.
- Wire
TelegramMarkdownV2Splitter
into the Telegram adapter (4,096-char per-message limit) and render markdown tables as monospace code blocks since Telegram has no native table support. - Wire
MarkdownSplitter
into the Discord adapter (2,000-char per-message limit), remove the oldchunk_for_discord
helper, and render markdown tables as monospace code blocks. - Wire
PlainSplitter
into the Slack adapter (silent multi-chunk). - Wire
PlainSplitter
into the SMS adapter with a 1,600-char hard cap andFull reply: {short-link}
overflow that points at aChat
-kind share URL (reused across overflow events in the same chat). - Wire
MarkdownSplitter
into the WhatsApp Cloud and WhatsApp Personal adapters. - Wire
SignalSplitter
into the Signal adapter with per-chunkSignalText { body, ranges }
. - Add
ChannelError
type andMessageDelivery.failure_kind
field; migrate channel adapters to typedChannelError
classification (Transient
,Forbidden
,NotFound
,PayloadInvalid
,PayloadTooLarge
,Unauthorized
,Other
) with an optionalretry_hint
. - Make
start_with_retry
idempotent so a single start no longer double-spawns the adapter. - Replace the bounded broadcast bus with unbounded fan-out so channel inbound stops losing events.
Share links and preview pages
- Add a
Share
entity withFile
andChat
kinds plus aShareService
for issuing short links; wire the service intoAppState
and schedule periodic cleanup. - Add
share.ttl_secs
(default 30 days) andshare.cleanup_interval_secs
(default 6 hours) config knobs. - Add
`GET /s/{id}`
short-link resolver with a 303 redirect to the canonical or presigned URL. - Add a
/p
preview page that renders markdown and code with syntax highlighting; add aGET /p/{slug}
redirector to the query-param form. - Route channel attachments through
outbound_url
in all adapters via a new`chat::channel::attachment`
helper: inline-channel previewable files get`/p/{8-char-id}`
, inline non-previewable files get`/s/{8-char-id}`
, button channels emit long-form`/p/{owner}/{handle}/{path}`
or/api/files/...
URLs. - Add an
anonymous_not_found
helper so unknown / expired / unauthorized share IDs return byte-identical 404s. - Extract
FilePreviewContent
into a shared module; extractapiFetch
fromrequest<T>
for non-JSON authenticated requests. - Add
`ShareKind::Chat`
variant with lazylookup_or_issue_chat
for SMS chat-share reuse.
Tasks
- Accept
result_description
oncreate_task
andcreate_recurring_task
so prose answers skip schema authoring; reject when bothresult_description
andresult_schema
are passed (or neither). - Require a top-level
summary: string
on complex task result schemas; reject complex schemas without it at task creation. - Carry
result_schema
onTaskCompletion
events and persist the JSON content. - Render the
TaskCompletion
bubble with a schema-driven body and a link to the task chat; renderTaskCompletion
via a shared markdown helper across channel adapters. - Researcher now publishes the full research as a markdown attachment, named after the topic (e.g.
h100-used-prices.md
) so successive research tasks don't overwrite each other.
Typed file tools
- Add typed
read
,write
,edit
,glob
, andgrep
tools. Workspace-scoped by default; absolute and policy-authorized sibling paths allowed; virtual-path URIs rejected. edit
uses Unicode normalization (NFKC, ASCII quotes/dashes/spaces, collapsed whitespace) so a slightly-misremembered snippet still matches its target. The motivation is token usage and hallucinations: a typededit
returns a structured diff in a fraction of the tokens a shell round-trip eats, without the model free-styling shell escapes against a half-remembered path.- Add the
frona-text
crate with shared text and search primitives (NormalizedString
,LineEnding
,walk_with_ignore
for.gitignore
-aware traversal). - Refactor sandbox:
SandboxManager
is now the single entry point per principal kind.
### Tool views (frontend)
- Extract a per-tool view registry:
ToolRow
primitives,DefaultView
, slim chrome. - Add views for shell, Python, and Node (with sandbox-deny extraction); for the typed file tools (
read
,write
,edit
,glob
,grep
) with syntax highlighting and line numbers; forproduce_file
(filename / size / content-type card); for`web_search`
(result list) and`web_fetch`
(markdown + clickable URL); for task tools (`create_task`
,create_recurring_task
withcronstrue
,delete_task
); for memory tools (auto-expand on large text); for`update_identity`
(set/remove attribute rendering); and for`set_heartbeat`
(humanized interval and next-tick time). - Render the shell tool command as a bash-highlighted block with wrap; render
produce_file
as a filename card with size and content type. - Add
vitest
test runner andcronstrue
/unbash
dependencies.
Harness / runtime
- Add a
Harness
struct as the agent runtime view ofAppState
; route session build, voice WS, message stream, and resume-all through it. - Narrow
TaskExecutor
to holdArc<Harness>
and own its own resolution-notifier map. - Move HITL resolution onto Harness and drop
ChannelCtx.app_state
. - Move task-tool registration into
ToolManager
; extractdeliver_event_to_source
soReportSignalTool
no longer needsTaskExecutor
. - Delete
agent/execution.rs
and update test fixtures for Harness. - Inject heartbeats as a transient turn so the agent stops replying to them like a user message.
- Fold per-turn lifecycle events into
InferenceEventKind
.
Browser
- Keep browser sessions alive past Browserless's hard 408-second timeout: pass a long timeout at connect time, self-evict and recycle sessions with a 60-second margin, and evict dead WebSockets immediately.
Chat / Frontend
- Stop dropping the newest message when a chat crosses the page limit.
- Move the reasoning toggle from inline content to a sparkle icon in the message header.
- Disable single-tilde strikethrough in assistant markdown.