Provide a stable pid↔session runtime record for local tooling (e.g. run/<pid>.json)

Summary

Kimi Code rewrites its process title at runtime (observed on 0.34.0: /proc/<pid>/cmdline collapses to kimi/kimi-code plus padding within seconds of start), which destroys the --session <sid> argv evidence. No other stable pid↔session binding exists: there is no session-id env var on the process, the wire-log fd is transient (open around writes only), and neither session_index.jsonl nor state.json records the hosting pid.

Local session-management tooling therefore cannot tell which process holds which session, and any stop/takeover decision becomes a guess.

What exists today

The SessionStart/SessionEnd hooks can serve as a user-side bridge: the payload carries session_id, and the hook process can recover the hosting kimi pid from its ancestry (kimi → sh -c → hook). But this requires every user to hand-configure [[hooks]] rules, and the parentage walk is indirect.

Ask

At session materialization, write a small runtime record keyed by pid — e.g. ~/.kimi-code/run/<pid>.json containing the sessionId (and ideally the process start time) — and remove it on session end. This mirrors what Claude Code provides via ~/.claude/sessions/<pid>.json.

Lighter alternatives that would also unblock local tooling:

• keep the original argv (skip the process-title rewrite), or
• set a session-id env var on the process, or
• record the hosting pid in session_index.jsonl entries.

Context

Verified on Kimi Code 0.34.0 (Linux): title-rewrite forms (kimi on a TTY, kimi-code otherwise), transient wire-log fd, and hook payload/timing (SessionStart fires at a TUI's first prompt, immediately for --prompt; SessionEnd on clean exit). I maintain csctl, a tmux-based local workbench for agent CLI sessions, which would consume this record for safe liveness/takeover decisions.

Start by tracing session materialization and the SessionStart/SessionEnd lifecycle, then inspect how session_index.jsonl, state.json, and the process-title rewrite are handled. Compare the proposed pid-keyed runtime record with the listed alternatives and define cleanup behavior on session end. Done means local tooling can reliably map a live process to its session without user-configured hooks.