Two independent bugs:
1. 429 "long context" — agentic sessions read enough files to exceed the
200K standard window; on the default [1m] model Synapse escalates into
the paid 1M tier. Pin claude_model to claude-opus-4-8 (200K window)
so Claude Code compacts context instead. Updated config.example.json
and the maintainer.py default / comment accordingly.
2. Bot silently ignores operator replies — the DM room was created with
is_direct:true, causing Element to auto-enable E2E encryption. This
stdlib-only bot cannot decrypt m.room.encrypted events, so all admin
messages were dropped without any log output. Fix:
- register-matrix-bot.sh now creates rooms with is_direct:false and
preset:private_chat (Synapse does not force encryption on non-DM
rooms).
- matrix_poll_admin() now logs a loud WARN if an encrypted event
arrives, instead of silently skipping it.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The homeserver authenticates humans via Keycloak OIDC and has password
login disabled, so register-matrix-bot.sh's password-login flow could
never obtain a token. Rewrite it to create the bot and mint a standalone
access token through Synapse's shared-secret admin API
(/_synapse/admin/v1/register), which works with registration + password
login turned off.
- Use the claude-code identity (@claude-code:semprini.me) as the bot.
- Avoid the admin "login as user" API: it returns a puppet token that
Synapse revokes when the issuing admin is deactivated.
- Make re-runs idempotent: reuse a still-valid token/room; only recreate
the account when no valid token is on hand.
- Read the registration shared secret from the running synapse container.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
systemd-deployed daemon that drives the semprini-core stack-support agent
headless via Claude Code, monitors health via Uptime Kuma, applies safe
upgrades, and escalates to the operator over Matrix.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
paul
