#572 ca.sh codex launcher: run codex on llmmsg-srv (ca= identity, sandbox-off, pull-only inbound, push-to-ccs) - fleet

#572 · llmmsg-srv

ca.sh codex launcher: run codex on llmmsg-srv (ca= identity, sandbox-off, pull-only inbound, push-to-ccs) - fleet

Ref: #572 (#572)
Project: llmmsg-srv
Status: inProgress
Priority: high
Type: epic
Assigned: pm-llmmsgsrv-cc
Created by: —
Created: 2026-05-29T02:58:17.582Z
Updated: 2026-06-05T06:51:02.451Z

Sub-items (0/5 done · 0%)

ref	title	status	priority	assignee
#573	Design + LIVE-TEST ca.sh: reads ca= from .agent-name, codex --dangerously-bypass-approvals-and-sandbox, injects -c mcp_servers.llmmsg-srv.env.LLMMSG_AGENT_NAME + LLMMSG_HUB_URL; PROVE shim auto-registers under ca name (roster check) before calling done	backlog	high	coder-llmmsgsrv-cc
#574	Land reviewed ca.sh in sh.git (after PM draft review); commit+push	backlog	high	bin-whey-cc
#575	Deploy/verify ca.sh on venus (sh.git pull); confirm a codex test session registers + hub URL resolves remote	backlog	normal	nw-venus-cc
#577	Fix static ~/.codex/config.toml footgun: wrong key LLMMSG_AGENT->LLMMSG_AGENT_NAME + LLMMSG_SRV_HOST 127.0.0.1->public; bare codex (not via ca.sh) currently falls back to cc= in /opt/llmmsg-srv = registers as pm-llmmsgsrv-cc on loopback. Non-blocking; ca.sh -c overrides already bypass it	backlog	normal	coder-llmmsgsrv-cc
#576	lezama ca.sh deploy: GCABA box - get Elazar explicit OK before nw-lezama-cc syncs; codex AGENTS.md convention (aro_join on start, cm/cms -> read_unread)	deferred	normal	pm-llmmsgsrv-cc

+ Add sub-item

Questions

No questions.

Event log

2026-05-29T02:58:17.584Z · created

wi cli
2026-06-05T06:51:02.451Z · statusChanged · wi-cli-whey

status=inProgress
2026-06-06T07:42:30.970Z · commented · wi-cli-whey

venus ca unblock LANDED+verified: appended [mcp_servers.llmmsg-srv] block to venus ~/.codex/config.toml (command=absolute pnpm node, args=/opt/llmmsg-srv/hub/llmmsg-srv-mcp.mjs, env LLMMSG_HUB_URL=http://172.27.178.27:9703 + LLMMSG_AGENT_NAME=audit-mars-ca). nw-venus-cc applied verbatim (whey has no direct ssh to venus), TOML validated, shim path resolves, backup config.toml.bak.20260606-0428. Elazar relaunched ca.sh; audit-mars-ca re-registered + actively polling (newest last_seen in roster) = MCP mount confirmed, no more slow stdio fallback. REMAINING for #572: fold wiring into ca.sh as per-launch -c override keyed to resolve_agent_label (kills global-hardcode footgun; supersedes the static config.toml block since -c beats config file). Also noted: ca.sh exports LLMMSG_AGENT not LLMMSG_AGENT_NAME (wrong var) - the -c override fix moots it.
2026-06-06T07:59:38.762Z · commented · wi-cli-whey

ca.sh resume-trap (root cause of stale-session/no-mount): ca.sh:45-46 does blind 'exec codex resume --last' whenever $cwd/.codex state dir exists, else fresh. So any config.toml/MCP change is silently swallowed - the agent keeps resuming the pre-change session and never mounts the new server. Verified on venus: codex-cli 0.137.0, 'codex mcp list' in audit-mars cwd shows llmmsg-srv enabled (mount IS valid), but the running audit-mars-ca was a stale pre-04:28 session improvising a hand-rolled node MCP client because ca.sh resumed it. AGENTS.md/commons all use native tools (no hand-roll instruction) - confirmed not an AGENTS.md issue. Immediate unblock given to Elazar: mv .codex aside + rerun ca.sh = fresh session picks up native mount. STRUCTURAL FIX for #572: add a --fresh/--new flag to ca.sh (and/or detect config-newer-than-session) so a launcher restart actually re-inits MCP instead of resuming across a config change. Pairs with the per-launch -c override item already on #572.
2026-06-06T08:14:07.089Z · commented · wi-cli-whey

CORRECTION to prior 'ca.sh resume-trap' note: empirically FALSE. nw-venus-cc inspected the live codex process on venus (PID 998252, started via resume --last in Elazar's pts/32) and found a CHILD process = node /opt/llmmsg-srv/hub/llmmsg-srv-mcp.mjs with env LLMMSG_AGENT_NAME=audit-mars-ca + LLMMSG_HUB_URL=172.27.178.27:9703 = the config.toml [mcp_servers.llmmsg-srv] native mount. So codex 'resume --last' DOES re-read config.toml mcp_servers and spawn the MCP server at process start - a config/MCP change IS picked up on the next ca.sh launch even via the resume branch. The earlier hand-rolled 'node <<NODE' the agent ran was conversational improvisation ON TOP of a working mount (and/or pre-mount scrollback), NOT a missing mount. Therefore: DROP the '--fresh flag needed so restart re-inits MCP' rationale - unneeded; resume already re-inits. Remaining valid #572 work: (1) per-launch -c override to set LLMMSG_AGENT_NAME per agent-CWD so the global config.toml hardcode (currently audit-mars-ca) doesn't mis-identify other ca agents; (2) ca.sh exports LLMMSG_AGENT not LLMMSG_AGENT_NAME - the -c override or a config.toml env moots it. venus audit-mars-ca is now live on the native mount; mars join uses wi=725 (MARS-38), directed by pm-mars-cc.