RUL-018 — Kimi runs at the agent level, not the project level

Ruling

"Kimi runs at the AGENT level, not the project level" (user requirement, verbatim, SESSION_START.md). The persistent kimi IRC presence is a single systemd USER unit (kimi-irc-bot.service, nick kimi) — never a project-level listener started from this repo. Nick discipline generally: nicks are agent-specific, not project-specific (codewhale is the codewhale agent's identity; never join as codewhale from a kimi session).

Evidence

• SESSION_START.md "IRC: Connect to #agents": the agent-level daemon is ~/.local/bin/kimi-irc-bot.py managed via systemctl --user status|restart kimi-irc-bot.service; the project-level scripts/_irc_listen.py/_irc_watchdog.py pair is explicitly CLAUDE's listener (IRC_NICK=claude), not kimi's — "one agent-level daemon per agent."
• Project memory irc-agent-delegation-fleet: nick-addressed IRC delegation (kimi: ..., CODEWHALE: ...) with a dedicated worktree per agent for actual task work — review/test/commit stay in the main checkout, matching the same "identity-scoped, not ad-hoc" discipline.

Enforcement

• gate scripts/check_kimi_agent_level.py RUL-018

Verification

scripts/check_kimi_agent_level.py enforces both clauses over every tracked {.py,.sh,.bash,.toml,.yaml,.yml,.service,.json} file (markdown excluded), invoked from scripts/health_check.py as "Kimi agent-level (RUL-018)":

• D1 — nick discipline. IRC_NICK bound to an identity this repo does not own. The repo owns claude, claude-out, swarm-supervisor; kimi/kimi_agent/codewhale are other agents' identities. Binding forms covered: shell assignment, export, systemd Environment=, an os.environ.get("IRC_NICK", …) default, and an env dict passed to a subprocess.
• D2 — no project-level kimi listener. A two-factor rule: a foreign-nick binding and a spawn of the repo IRC daemon (setsid, nohup, systemd-run, subprocess.Popen, trailing &). One-hop shell indirection resolves; two-hop and command-substitution indirection deliberately SKIP rather than guess (false negative preferred to false positive).

The gate found and closed 4 real violations at introduction — this ruling was being broken by the repo it governs:

• .kimi/hooks/session-start.sh:124,136,141 — section 6 bound IRC_NICK=kimi_agent and spawned scripts/_irc_watchdog.py from this repo, i.e. a project-level kimi listener, plus a failure-branch message instructing the user to run it by hand. Fixed: section 6 is now a report-only check of systemctl --user is-active --quiet kimi-irc-bot.service.
• scripts/_irc_listen.py:37 — defaulted IRC_NICK to codewhale, another agent's identity, with a comment documenting the violation as policy. Fixed: defaults to claude, matching _irc_send.py. Behaviour-neutral for the live path (the launcher ~/.claude/hooks/irc_agents_hook.sh exports IRC_NICK=claude explicitly).

Post-fix the gate is green across 3171 tracked files with 0 false positives, measured — every legitimate fleet-nick mention stays silent: _irc_conversation.py's attribution map, _irc_send.py's allowlisted claude default, _irc_watchdog.py (spawns but binds no nick — D2's two factors), swarm_supervisor.py's DEFAULT_IRC_NICK, and the "kimi" Moonshot provider name in core/infra_config.py / drivers/* that a naive nick grep would flood on. The gate exempts only its own two files (its docstring must quote the violating lines verbatim); there is no blanket tests/ exemption, and a test hardcoding a foreign nick and spawning the watchdog still fails.

Related defect, since fixed: SESSION_START.md — the document this ruling cites as its source of authority — carried unresolved merge-conflict markers (<<<<<<< Updated upstream / ======= / >>>>>>> Stashed changes) at two hunks, leaving the IRC-discipline section unreadable as a spec. Resolved on 2026-08-05; grep -c '^<<<<<<<\|^>>>>>>>' SESSION_START.md is now 0. Note that no gate covers markdown (the RUL-018 gate excludes it deliberately, to avoid flagging prose that merely names a nick), so this class of corruption remains review-caught only.