fix(ops): make telegram-claude-bridge daemon resilient #21

Merged

navigator merged 1 commit from feature/bridge-resilience-fix into main

2026-05-24 13:31:11 -03:00

navigator commented

2026-05-24 13:30:02 -03:00

Owner

Three fixes after PR #20 daemon crashed in first production deploy on agent host.

Symptoms (from agent journalctl):

Service crashed exit 1 ~100s after each msg arrived
systemd restarted, daemon picked same msg again (offset never persisted)
Logs showed only "msg from..." and two HTTP 200s (ack + maybe response), no "reply sent"

Root causes:

set -e killed the daemon on any non-zero exit anywhere in the per-message path. Removed (kept -u, pipefail).
save_offset ran AFTER process_message; if process crashed first, offset lost, infinite loop on the same msg.
The main loop iterated updates via ... | jq | while read, creating a subshell that lost state (offset writes from subshell visible but parent died first).

Fix:

Drop -e, keep -u -o pipefail
Move save_offset BEFORE process_message
Iterate via tempfile + redirected while-read (no subshell)
Add || fallback on jq calls so malformed Telegram responses don't kill the loop
Per-message errors get logged and skipped, daemon stays alive

Verified locally: claude -p reply ok returns in ~3s on agent (Xeon E5-2630 v2 with AVX). The failing component is the daemon shell, not claude or systemd.

Tests: shell only. shell-lint validates the daemon. Auto-merge via auto-merge.sh.

Three fixes after PR #20 daemon crashed in first production deploy on agent host. Symptoms (from agent journalctl): - Service crashed exit 1 ~100s after each msg arrived - systemd restarted, daemon picked same msg again (offset never persisted) - Logs showed only "msg from..." and two HTTP 200s (ack + maybe response), no "reply sent" Root causes: 1. `set -e` killed the daemon on any non-zero exit anywhere in the per-message path. Removed (kept -u, pipefail). 2. `save_offset` ran AFTER `process_message`; if process crashed first, offset lost, infinite loop on the same msg. 3. The main loop iterated updates via `... | jq | while read`, creating a subshell that lost state (offset writes from subshell visible but parent died first). Fix: - Drop `-e`, keep `-u -o pipefail` - Move `save_offset` BEFORE `process_message` - Iterate via tempfile + redirected while-read (no subshell) - Add `|| fallback` on jq calls so malformed Telegram responses don't kill the loop - Per-message errors get logged and skipped, daemon stays alive Verified locally: `claude -p reply ok` returns in ~3s on agent (Xeon E5-2630 v2 with AVX). The failing component is the daemon shell, not claude or systemd. Tests: shell only. shell-lint validates the daemon. Auto-merge via auto-merge.sh.

navigator added the

area:infra

phase:1

labels

2026-05-24 13:30:02 -03:00

navigator added 1 commit

2026-05-24 13:30:02 -03:00

fix(ops): make telegram-claude-bridge daemon resilient

build / scalafmt-check (push) Successful in 4s

Details

build / sbt-compile (push) Successful in 4s

Details

build / shell-lint (push) Successful in 30s

Details

build / scalafmt-check (pull_request) Successful in 10s

Details

build / sbt-compile (pull_request) Successful in 13s

Details

build / shell-lint (pull_request) Successful in 24s

Details

975dbee2e5

Three fixes after first deploy crashed in production:

1. Remove 'set -e' (keep -u -o pipefail). The main loop could die when
   a per-message command returned non-zero, killing the whole daemon
   instead of just logging and continuing.

2. Save offset BEFORE process_message, not after. Previously a crash
   in process_message would lose the offset write, causing infinite
   re-processing of the same message on systemd restart.

3. Iterate updates via tempfile instead of pipe-to-while. The pipe
   creates a subshell where save_offset's file write was lost when
   the parent terminated.

Also added '|| fallback' suffixes on jq parses so malformed
Telegram responses don't kill the loop.

Verified on agent: 'claude -p reply ok' returns in ~3s. Bridge itself
was the failing component; fixes target only the daemon shell, not
claude or systemd unit.