WURZEL: Ein Agent empfängt SSE-Events ausschließlich, solange er in
agenthub_work blockiert. Während er einen Task AUSFÜHRT, ist er vollständig
taub — kein Reopen, kein Cancel, keine Nachricht erreicht ihn. Ein echter
Interrupt in einen laufenden Agenten-Turn existiert nicht. Heartbeat,
Watchdog, Polling-Fallback und Presence-Tracking waren allesamt Umgehungen
dieser einen Tatsache; deshalb kam das Problem jede Session zurück.
LÖSUNG — der einzige reale Kanal: die Momente, in denen der Agent von sich
aus mit dem Hub spricht.
- checkinService.resolvePending(): zustandslos aus Tasks + Messages
abgeleitet, nichts, was ein Neustart verliert. Erkennt, dass dem Agenten
die Arbeit entzogen wurde (reopen/cancel/fremder Claim), und liefert
ungelesene Nachrichten + offene Zuweisungen mit.
- agenthub_task_log gibt den pending-Block zurück; der Tool-Text macht ihn
verbindlich (pending.interrupted ⇒ sofort aufhören, nicht einreichen).
- Neu: agenthub_checkin(agent, taskId) für lange Strecken ohne Log-Zeile,
plus GET /agents/:agent/pending.
- resolvePending degradiert auf reinen Namensvergleich, wenn keine Config da
ist — ein Check-in darf nie an Konfiguration scheitern.
DEC-0035: `review` bindet den Agenten wie `in_progress`. Vorher galt er in
der Sekunde des Einreichens als frei, griff sich die nächste Task, und ein
Reopen prallte danach am Ein-Task-Guard ab (so ging der Reopen von TSK-0218
verloren). Der Loop bleibt aktiv — wach warten, nichts Neues anfangen.
Außerdem: in_progress → open erlaubt, damit der Architekt eine festhängende
Arbeit überhaupt entziehen kann (vorher 400, Agent blieb dauerhaft blockiert).
SICHTBARKEIT: /health und /agent-health zeigen pro Agent pendingCount
(wartende Tasks) und deafForSec (wie lange ohne Check-in bei laufender
Arbeit). Ungelesene Nachrichten stehen separat — der dreistellige
Altbestand einzelner Agenten hätte das Signal sonst erschlagen.
Tests: 279 → 290. tests/singleClaim (b) auf den neuen Vertrag umgeschrieben
(+ (b2): ein Reopen gewinnt gegen die nächste Task).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Drei vom Architekten abgenommene Tasks, gebündelt als Checkpoint:
- TSK-0242: Agent-Alias-Mapping (kimi-ah → kimi kanonisiert, Rollen → preferredAgent),
reopenTask räumt claimedBy ab, fsWatch reindiziert direkte Datei-Edits,
work-Default 300s → 50s, task_list mit Limit.
- TSK-0245: zwei Agent-Klassen (dispatch loop|architect). Watchdog mahnt
architekt-getriebene Agenten nur noch EINMAL statt im Minutentakt;
`task dispatch` startet sie explizit, `task record` trägt extern
erledigte Arbeit mit origin=external nach.
- TSK-0249: Lifecycle wird serverseitig erzwungen (open→review scheitert mit
klarer Meldung), claimedBy/doneBy überleben bis done, Presence pro Agent,
Review-Watchdog, GET /architect/pulse (1.4 kB statt 34 kB, since-Cursor,
omitted statt stillem Abschneiden), unbekannter Agent → 400 statt 500.
Alle Punkte live am laufenden Hub nachgemessen, nicht aus Agenten-Logs übernommen.
Tests: 242 → 279 grün, tsc sauber.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
- probe configured server URL (~1s) before remote commands; on failure
re-discover via LAN discovery, persist and use the resolved URL
- poll tryClaim/trySurfaceMessages every ~4s during the SSE wait in CLI
work and MCP waitForTask — a lost SSE frame costs seconds, never sleep
- work/MCP no longer auto-mark messages read: surfacing sets delivered,
explicit ack/read required; loops still wake only on unread (no spin)
- parseSSEBuffer accepts CRLF frame separators and joins multi data: lines
- bump version to 0.10.1 (package.json, CLI --version, MCP server)
- work.ts: the CLI work loop now wakes on architect follow-up MESSAGES, not just
tasks. drainAgentMessages() surfaces + marks-read the agent's inbox; wired into
waitAndClaim (SSE message events) and as an upfront check in workAgent. reopen +
next-assign already wake via task events; a plain message no longer leaves the
re-armed loop dormant.
- mcp/server.ts: implementer agenthub_work LOOP reminder now explicitly instructs
re-arming after agenthub_task_review so the architect's approval/reopen/message
wakes it (waitForTask already wakes on task+message).
- tests: work.test.ts +message-wake +reopen-wake
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- routes PATCH /tasks/🆔 uniformly appendTaskLog + publishLog for every status
transition (claim/review/done/cancel/reopen) and on assign; best-effort, never
fails the mutation. No double-emission (log rides the 'log' channel; fsWatch
only watches .md so the .log append is not re-emitted).
- MCP agenthub_task_log tool + remoteClient.appendTaskLog + CLI 'task log <id>
--text [--agent][--level]'
- agenthub_work LOOP reminder: 'Report meaningful progress with agenthub_task_log'
- taskDetail: Live Console panel — historic lines via readTaskLog + live tail via
the named task-log SSE event filtered to this task id
- tests: +taskLog-live.test.ts (PATCH->one log event, no double change, historic
render, POST /log, remoteClient)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Board (/board):
- Drag an AGENT chip onto a task card to (re)assign it (realtime-notified).
- Drag a task card to a column to change status; open→in_progress auto-assigns
from the title's "<agent>:" prefix — no manual agent picking.
- "+ New task" composer (agent dropdown removed; agent comes from the title).
- Live Cost & Budget panel.
Team (/team): live SSE sync of busy state + always-on ambient animation
(connector shimmer, idle glow) that brightens to a busy pulse when an agent
works. Org-chart hierarchy stays.
Token accounting: new budgetService/rosterService + GET /budget and /agents.
Real doneTokens + time-on-task estimate capped at 45 min/task (avoids the
wall-clock overcount that produced multi-million-token totals), blended
per-model EUR cost + optional budget bars. All estimates flagged "~".
Autostart: `agent setup` writes deterministic SessionStart hooks for Codex
(~/.codex/config.toml) and Kimi (~/.kimi-code/config.toml), not just Claude
Code. Verified: both auto-enter the agenthub_work loop.
Realtime architect review: agenthub_work is role-aware — architect/reviewer
blocks on SSE and wakes when a task hits review (no manual watcher re-arm).
mDNS: server advertises agenthub.local (bonjour-service) so the hub is
reachable in a browser on the LAN without an IP.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The recurring "agent doesn't read messages in realtime / isn't triggered for the
next task" problem: agents weren't STAYING in the agenthub_work loop, so they went
dormant after each task. Two reinforcements:
- agenthub_work now ends every response with a LOOP directive: "the instant you
finish, call agenthub_work again (in the background) so the next task/message
reaches you in realtime — never end your turn without relaunching it."
- The auto-start hook-context now stresses the BACKGROUND loop + relaunch pattern,
so a Claude Code session is woken on each event instead of blocking.
Layer 1 (auto-start) gets agents INTO the loop; this keeps them IN it. Honest
limit: turn-based agents still rely on obeying the relaunch nudge — there's no
100% always-on without a supervisor. Bump 0.9.0 -> 0.9.1.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Direct messages between agents through the hub — architect↔agent and agent↔agent
questions, clarifications and pings that handoffs/memories don't cover. Built as
MCP tools so kimi + codex get it too (not Claude-only), plus REST + CLI.
- Message entity (MSG-NNNN): schema, counter, messages/ dir, messageService
(createMessage / listInbox / listMessages / markMessageRead).
- REST: POST /messages, GET /messages[?agent&unread], POST /messages/:id/read,
with SSE emit. fsWatch + statusRefresh wired ('message' events; status skips them).
- MCP: agenthub_message {from,to,text,taskId?} + agenthub_inbox {agent,unreadOnly?}.
agenthub_work now surfaces + drains unread messages and wakes on message events,
so an agent sees messages inside its work loop. (18 tools total.)
- CLI: `agenthub message <to> <text> --from <agent>` + `agenthub inbox --agent`.
watch stream shows "Message" events (architect-visible).
- Architect-visible by design: the hub stays the coordination point.
Bump 0.7.5 -> 0.8.0.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
One command, no hand-editing: `agenthub mcp install` writes/merges a project
.mcp.json (the standard config Claude Code and other .mcp.json-aware clients
auto-discover) with the hub URL filled in — so those agents register
automatically. Prints the global-config snippet for Codex/Kimi as fallback.
`--print` is a dry run. Bump 0.7.3 -> 0.7.4.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Removes the registration friction: `agenthub mcp install` prints ready-to-paste
MCP config for Claude Code, Codex and Kimi, with the hub URL filled in from the
project config. `agenthub mcp` (no action) still starts the stdio server, so the
MCP-client spawn command is unchanged. Bump 0.7.2 -> 0.7.3.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Solid-collaboration foundation: agents invoke STRUCTURED tools instead of
shelling out to the CLI — so a tool can't be narrated away, hallucinated as
"doesn't exist", or mistyped. Confirmed working for Claude / Codex / Kimi
(all speak MCP).
- src/mcp/server.ts: `agenthub mcp` starts a stdio MCP server exposing 16
tools (work, task_list/show/create/assign/claim/review/reopen/done,
memory_add/search, handoff_read/create, decision_create, hello, status).
Thin layer over the SAME core: proxies through the REST API (remoteClient)
when a hub is configured — so board, SSE, status auto-refresh and CLI keep
working unchanged — else local services. agenthub_work blocks on SSE until
a task addressed to the agent is claimable, then returns it + its handoff.
- `agenthub mcp` command; excluded from the update-notify (stdio purity).
- Verified: initialize + tools/list (16) + tools/call (status, task_list)
proxy to the live hub.
Does NOT change the dormant-agent reality (agents still loop agenthub_work);
it makes every interaction reliable. Bump 0.6.1 -> 0.7.0.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
