Codex and Claude Code OAuth tokens with refresh_token can auto-renew,
so they should not be considered expired based on last_refresh time or
file mtime. Previously, credentials were hardcoded to expire after 1
hour, causing valid Codex sessions to show as unconfigured.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move channel plugin initialization (initChannels + ChannelManager
constructor) before restoreAgents() to prevent TypeError when
createAgent() calls channelManager.listChannelInfos() during restore.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Create onChannelSendFile callback in Hub.createAgent() that tries the
channel plugin path first (local file), then falls back to the gateway
path (base64 over RoutedMessage). Also pass channel info to agent.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add ChannelInfo type and buildChannelsSection() that informs the LLM
about connected messaging channels and their capabilities (e.g. send
files). Wire through SystemPromptOptions and runner.ts.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Prepend [ChannelName · private/group] prefix to debounced messages so
the LLM knows the message source. Add sendFile() for outbound media
routing and listChannelInfos() for system prompt channel awareness.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New send_file tool with TypeBox schema, auto-detect media type from
file extension, and file validation. Wired through AgentOptions and
resolveTools with conditional registration when callback is provided.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add OutboundMedia interface and OutboundMediaType to the channel type
system. Implement sendMedia in the Telegram plugin using grammy's
InputFile API with HTML caption formatting and plain-text fallback.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Guide the LLM to evaluate snippet quality after web_search and
follow up with web_fetch on the most relevant URLs when deeper
content is needed for accurate answers.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Hidden directory ~/.super-multica is not user-friendly for a working
directory. Move default workspace base to ~/Documents/Multica so
users can easily find agent-created files in Finder.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add MessageSource type to track where user messages originate (local, gateway, channel)
- Broadcast inbound messages from all channels to local Desktop UI via Hub.onInboundMessage()
- Persist source field in JSONL session storage so it survives page refresh
- Display source icon (Monitor/Smartphone/Send) with tooltip for non-local user messages
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Each agent profile now gets a dedicated workspace directory
(~/.super-multica/workspace/{profileId}) used as the default CWD
for tool operations. Supports override via MULTICA_WORKSPACE_DIR
env var or config.json workspaceDir field. The workspace path is
injected into system prompt and runtime info.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
A subagent that times out loses all its work, so a generous default
reduces wasted compute. Update the constant, tool description, system
prompt guidelines, and documentation to reflect the new 1800s default.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When writeInternal fails with a transient error (e.g. undici
"terminated", connection reset, 502/503), wait 5 seconds and retry
once. This prevents continuation tasks from silently failing when
the LLM streaming connection is interrupted mid-generation.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
sessions_spawn now accepts groupId and next parameters. First spawn with
next auto-creates a group; subsequent spawns join via groupId. sessions_list
groups runs by groupId with completion progress display.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When all runs in a group complete, deliver combined findings plus the
`next` continuation prompt to the parent agent via writeInternal. The
parent can then act on the collected data (summarize, write files, etc).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add SubagentGroup for "collect all, then act" workflows where multiple
subagents complete before a continuation task runs. Groups are persisted
alongside run records and support a `next` continuation prompt.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Root cause: compaction only ran after successful assistant response,
never before calling the LLM API. Internal runs skipped it entirely,
and the fire-and-forget void swallowed errors with race conditions.
Layer 1 - Pre-flight compaction in transformContext:
Runs before EVERY LLM call (including internal runs). Prunes tool
results and drops oldest messages when utilization exceeds 80%.
Pure in-memory, no disk writes.
Layer 2 - Fix post-hoc compaction:
Store compaction promise instead of voiding it, await at _run() start
to prevent race conditions. Emit compaction_start before the actual
compaction. Remove useless catch-rethrow block.
Layer 3 - Context overflow recovery:
Catch context overflow 400 errors → auto-compact → retry (up to 2
attempts). Runs before classifyError/auth rotation so overflow errors
are handled by compaction, not profile switching.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- ESTIMATION_SAFETY_MARGIN: 1.2 → 1.5 (50% buffer covers CJK text)
- estimateSystemPromptTokens: /3 → /2 (conservative for mixed content)
This makes the 80% compaction trigger fire earlier, reducing the gap
between estimated and actual token counts that caused overflow errors.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Detects context overflow 400 errors from various LLM providers
(prompt too long, context length exceeded, request too large, etc.)
for use in auto-compaction recovery.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add stop button to interrupt agent generation mid-stream. The send button
toggles to a stop icon during loading. Abort propagates from UI through
IPC to the Agent layer (PiAgentCore.abort()), preserving all partial
content in the agent's context so users can follow up immediately.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add AppState module in core for managing app state persistence
- Add app-state IPC handlers for reading/writing onboarding state
- Hydrate onboarding state from file system on app startup
- Prevent flash by showing blank screen during hydration
- Update onboarding store to sync with file system
- Improve MulticaIcon with enhanced animation states
- Minor UI fixes in chat and device list components
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Web tools are always available, so the conditional branching was
unnecessary. Data section now always includes the dynamic evidence
decision guidance.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The finance-decisioner used hardcoded keyword matching and arbitrary
scoring weights to decide evidence plans — the LLM reading the
improved SKILL.md already makes this decision more intelligently.
Removes: finance-decisioner.ts, its tests, runner.ts integration
(applyFinanceResearchGuidance, saveFinanceDecisionMeta,
rebuildSystemPromptWithExtra shim), and researchDecision from
session meta.
Keeps: SKILL.md improvements, sections.ts dynamic data/web guidance.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The heartbeat runner used agent.write() (normal write), which persisted
both the heartbeat prompt and HEARTBEAT_OK response as regular messages
visible in the UI chat history.
Switch to runInternalForResult() — a new AsyncAgent method that runs
via runInternal() (messages marked internal: true, rolled back from
in-memory state). This hides both the heartbeat prompt and response
from the UI entirely, while still persisting to JSONL for diagnostics.
The previous commit's event-stream heartbeat ACK filter remains as a
defense-in-depth layer for edge cases.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The heartbeat runner uses agent.write() (normal write), so heartbeat
ACK responses like "HEARTBEAT_OK" were not suppressed by the internal
run filter and leaked into the desktop UI chat. The Gateway path was
already fixed (Hub has delayed-start + isHeartbeatAckEvent filtering),
but the local Desktop path through AsyncAgent had no such filtering.
Add createFilteredHandler() to AsyncAgent that buffers message_start
for assistant messages and checks subsequent events with
isHeartbeatAckEvent(). Pure heartbeat ACKs are suppressed end-to-end;
all other messages are forwarded normally.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Changed forwardAssistant from false to true in sendAnnounceDirect() so
the assistant's summary response is streamed to the desktop UI in
real-time. The announcement prompt stays internal (hidden from UI), but
the user now sees the completion notification.
Previously, persistAssistantSummary saved the response to JSONL but
never emitted events to the UI subscriber, leaving users with no
visible feedback after subagent tasks completed.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
setProvider() updated toolsOptions.provider but didn't reload tool
instances. The sessions_spawn tool captured the old provider in its
closure at creation time, causing subagents to inherit a stale provider.
Now calls resolveTools() + setTools() after updating toolsOptions so
sessions_spawn gets a fresh closure with the correct provider.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add an `announce` parameter to sessions_spawn that controls when findings
are delivered to the parent agent:
- "immediate" (default): announce per-completion (existing behavior)
- "silent": defer until ALL silent runs from the same requester complete,
then deliver ONE coalesced announcement with all findings
This enables workflows like "spawn 10 parallel subagents to collect data,
then summarize everything at once" without intermediate results.
Changes:
- types.ts: add `announce` field to SubagentRunRecord & RegisterSubagentRunParams
- sessions-spawn.ts: add `announce` parameter to tool schema
- registry.ts: split checkAndAnnounce into immediate/silent groups,
extract announceGroup helper, use count-match guard for silent readiness
- sections.ts: add announce mode guidance to system prompt
- registry.test.ts: add silent mode tests (field storage, group isolation)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Extract SILENT_REPLY_TOKEN and isSilentReplyText() into a shared
module. Detects NO_REPLY at the start or end of text (with optional
whitespace/punctuation) to filter out silent announcement responses
that should not be forwarded to the user.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Document the full subagent lifecycle: spawn, concurrency queue,
execution, completion handling, two-tier announcement delivery,
and record archival. Include provider inheritance chain and
error propagation diagrams.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add error reporting rule to subagent system prompt: subagents must
explicitly report tool failures and missing credentials in their
final message. Add timeout guidelines to parent system prompt with
recommended values by task complexity (10-30 min).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When subagents are still running, sessions_list now returns an
explicit instruction telling the LLM not to poll again and wait
for automatic result delivery. Normalizes status display to
uppercase ([RUNNING], [OK], [ERROR]).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Pass parent's resolvedProvider through the tool chain (tools.ts →
sessions-spawn.ts → hub.createSubagent) so subagents use the same
LLM provider as the parent. Previously subagents fell back to the
hardcoded default provider, causing API key errors.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
