feat(hub): map kimi 0.31 wire events — cancel, compaction, plan/swarm/permission modes

[agentfleets]

Summary

Updates the kimi-code M4 wire-tail mapper (hub/internal/drivers/local_log_tail/kimi_code) for the wire event vocabulary added in kimi-code 0.31.0. Verified against live 0.31.0 wire.jsonl captures on a dev machine (~/.kimi-code/sessions/wd_*/session_*/agents/main/wire.jsonl, 21991-line busy session).

1.4 → 1.5 is vocabulary growth, not a flag day. 0.31.0 still writes protocol_version: "1.5" and every 0.28.x event shape we already parse is byte-identical (content.part, tool.call args, tool.result result.output, step.end finishReason, usage.record, todo store). The SupportedProtocolVersion major-version gate (any 1.x) is unchanged, as is the launch gate in launch_m4_kimi.go. What's new is 10 event types that today fall into the mapper's default: arm and each become a noisy system{subtype:unknown_type} event (hundreds per busy session — 133 turn.steer alone in the sampled wire) while missing real signal — most importantly the cancel terminal.

Mapping table implemented

wire type	maps to	rationale
turn.cancel	turn.result {status:"cancelled", stop_reason:"cancelled"} (main agent only)	The correctness fix. kimi writes no step.end end_turn for an interrupted turn (verified: cancel → next line is the next turn's prompt), so a cancelled turn left the session stuck busy forever. The payload is exactly the ACP driver's eager-cancel vocabulary (driver_acp.go Input("cancel")): mobile's busy-walker terminates on the kind, the hub digest reads stop_reason (cancelled ∈ normalTurnStopReasons — an intentional end, not an integrity finding), and status matches what every ACP engine posts on cancel, so digest + mobile failure accounting treat both engines identically. Subagent guard mirrors mapStepEnd (#374).
turn.steer (any origin)	explicit drop	background_task (129/133 in capture) is the engine's own task-notification envelope — CLI-internal noise. user steers: same duplication rationale as turn.prompt — hub-originated input arrives via tmux send-keys and was already posted as input.text by the hub, so mapping would double the bubble; pane-typed steers are the accepted loss, same as pane-typed prompts. Unknown future origins drop quietly. Value over default: = no unknown_type noise + a stated decision. No plan-chain side effects (steers redirect the in-flight turn; verified turnId unchanged).
turn.prompt origin.kind="task"	explicit drop (still arms plan chain)	Goal-mode/cron engine tasks + background-task notices. They open a real engine turn, so the per-turn plan chain still re-arms; the body is an engine-generated notification envelope, not user prose.
context.apply_compaction	ONE system {subtype:"compaction", summary(≤500 chars), tokens_before, tokens_after, compacted_count, text}	Clients render it as a system card today (the text line lands on the default renderer); the fields keep structure for a future fold marker. Mirrors claude mapper's compact_boundary arm (single producer=system notice, off the busy-inference path). Summary is rune-bounded so one event can't dominate the transcript/event store.
full_compaction.begin / full_compaction.complete	drop	They bracket the apply event that carries the signal — one card per compaction, not three (15 trios in the sampled wire).
plan_mode.enter/cancel/exit	system {subtype:"plan_mode", phase, id?, text}	No dedicated mode event kind exists hub-side — the ACP driver forwards current_mode_update as a verbatim producer=system frame — so a system subtype is the reuse-consistent shape. exit handled though not yet observed (enter/cancel verified live).
permission.set_mode	system {subtype:"permission_mode", mode, text}	Mid-session complement to the launch-time permission_mode on session.init. Modes observed live: yolo / auto / manual.
swarm_mode.enter/exit	system {subtype:"swarm_mode", phase, trigger?, text}	Enter carries trigger ("tool" observed).
mcp.tools_discovered	drop	Tool catalog noise, same class as tools.set_active_tools / llm.tools_snapshot.
anything else	system {subtype:unknown_type}	Drift policy unchanged for genuinely unknown future types.

Evidence (local 0.31.0 wire samples)

• Type histogram of the 21991-line main wire: turn.steer ×133 (129 background_task + 4 user), turn.prompt ×111 (109 user + 2 task), full_compaction.* ×15 trios, permission.set_mode ×5, swarm_mode.* ×4 pairs, plan_mode.* ×1 pair, turn.cancel ×2, mcp.tools_discovered ×1, metadata.protocol_version="1.5".
• Cancel context: llm.request (turn 19.11) → turn.cancel → next line is turn.prompt (turn 20) — no end_turn for the interrupted turn, confirming the stuck-busy mechanism.
• New testdata/wire_main_0_31.jsonl holds redacted real lines for every new type (user text replaced with placeholders); wire_main.jsonl stays as the 0.28.1 pin.

Tests

• mapper_test.go extended per type: exact payload shapes (esp. the cancel → turn.result vocabulary), explicit drops emit NOTHING (incl. no unknown_type), plan chain (turnSeq/planMsgID) provably unaffected by every new type, subagent guard on turn.cancel, task-origin prompt still re-arms, 500-rune summary truncation, and a 0.31 fixture replay pinning the kind histogram.
• go test ./internal/drivers/local_log_tail/... ./internal/hostrunner/... — green except 2 hostrunner failures that reproduce identically on pristine origin/main on this machine (TestWriteKimiTSMCPConfig_FreshWorkdir — local TermiPod.app browser bridge injection; TestDatasetExportRRD_EndToEndWithAStubExporter — macOS /var→/private/var symlink). Pre-existing, environment-induced, unrelated to this diff.
• go vet clean on the touched packages.
• Full go test ./... result noted in the PR comments if it differs (the 8 TestLintGovernedActions_* failures on macOS bash 3.2 are pre-existing and green in CI).