Claus Ibsen created CAMEL-26911:
-----------------------------------

             Summary: camel-jbang - process management: camel stop can kill an 
unrelated process through a stale status file, and lost or torn action replies
                 Key: CAMEL-26911
                 URL: https://issues.apache.org/jira/browse/CAMEL-26911
             Project: Camel
          Issue Type: Bug
          Components: camel-jbang
            Reporter: Claus Ibsen
             Fix For: 4.24.0


Found by an analysis of the Camel CLI on the night of 2026-11-12, with the 
unpacked 4.24.0-SNAPSHOT camel-launcher on macOS. *(exp)* means reproduced by 
experiment, *(read)* means traced in the code. Paths are relative to 
{{dsl/camel-jbang/}}, and line numbers are from main at that time. Dependency 
resolution of camel run is covered separately by CAMEL-26901.

h3. 1. camel stop can kill an unrelated process (exp, high)

{{ProcessFinder.scan()}} (camel-jbang-core {{common/ProcessFinder.java}} around 
lines 187-205) treats any live PID that has a {{<pid>-status.json}} in 
{{~/.camel}} as a Camel integration, and never checks that the process is the 
one that wrote the file.

An integration killed with SIGKILL, by the OOM killer, or a crash leaves its 
files behind, and {{camel stop --kill}} itself leaves them (8 files). Once the 
PID is reused by another process:
* {{camel ps}} lists a ghost integration
* {{camel stop <name>}}, or {{camel stop}} for all, sends SIGTERM to whatever 
process owns that PID now, and SIGKILL with {{--kill}}

macOS PIDs wrap at 99999, so this is reachable on a developer machine. 
Reproduced: a {{sleep 600}} with a fake status file was killed by {{camel stop 
ghost}}.

Suggested fix: write the process start instant 
({{ProcessHandle.info().startInstant()}}) into the status file, and only treat 
the PID as the integration when it matches; remove the stale files otherwise. 
{{camel stop --kill}} should also clean up the files of the process it killed.

h3. 2. A slow action blocks stop and status (exp)

One connector thread does the lock-file check, the status writes and the 
actions. During {{camel cmd send --poll --timeout=15000}}, {{camel stop}} took 
16.9s, and {{camel ps}}, {{camel get}} and the TUI showed stale data for that 
time.

h3. 3. camel trace clear and camel receive clear lose the next batch (read + 
parse test)

They write an empty JSON object with no newline ({{CamelTraceAction}} around 
line 229, {{CamelReceiveAction}} around line 284). The connector then appends 
the next JSON object plus a newline, so the line holds two JSON objects and 
fails to parse: the first batch after a clear is lost.

h3. 4. Torn reads of action replies (suspected)

{{writeReply}} writes the reply file in place (only the status file uses a tmp 
file plus atomic move), and {{ActionClient.executeText}} returns any non-blank 
text, so AI and MCP tools can read truncated JSON. The follow loops drop a 
partial last line.

h3. 5. camel log is 12x slower with color (exp)

{{CamelCommand.printColored}} writes through the unbuffered 
{{AnsiConsole.out()}}, and color is on even when the output is not a TTY. 100k 
lines: 4.88s with color vs 0.42s without.

h3. 6. camel log with 2+ integrations moves stack traces away from their 
records (exp)

The comparator that merges the logs gives lines without a timestamp -1, so the 
lines of a stack trace are sorted away from the record they belong to.

h3. 7. A background run writes its log twice with no size limit (exp)

{{<pid>-run.log}} is byte-identical to {{<pid>.log}} (27 MB each after 6 
minutes). {{-run.log}} is left behind after {{--max-seconds}}, {{camel stop 
--no-wait}} or a crash.

h3. 8. camel log --follow never closes the reader of a stopped integration 
(read)

A file descriptor leak for a long follow session. Minor: {{camel ps}} parses 
each status file twice (0.35s flat, fine today).

h3. Came out clean

{{ActionClient}} waits are bounded at 10s, requests are published atomically, 
and a missing {{~/.camel}} is handled.

h3. Suggested order

# 1 (small fix, worst outcome)
# 3 and 4 (lost or torn data for the AI and MCP tools)
# 2, 5, then the rest

_Claude Code on behalf of Claus Ibsen (davsclaus)_



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

Reply via email to