MAF-20668: docs(skills): correct the KV cache events guidance in guide-odin

[seongsu-dev]

Summary

Four claims in skills/guide-odin/SKILL.md had gone stale against the runtime-bases they describe. Verified against moreh-dev/moai-inference-preset, which odin's preset-distribution PRD names as the source of truth for presets:

# ISVC_ENABLE_PREFIX_CACHE turns on the engine prefix cache and KV
# cache events. No explicit endpoint: vLLM offsets the KV-events
# port by data_parallel_rank (tcp://*:5557 -> 5557+rank), and the
# heimdall sidecar subscribes at engine_kv_port+rank to match — so a
# constant config is per-rank-safe; a per-rank endpoint double-offsets.

Line	Was	Is
env table	ISVC_USE_KV_EVENTS, default false	ISVC_ENABLE_PREFIX_CACHE, default true (all four vllm runtime-bases: ${ISVC_ENABLE_PREFIX_CACHE:-true})
KV events section	enable with ISVC_USE_KV_EVENTS: "true"	on by default; the variable turns it off
KV events section	"the current preset still resolves this ZMQ endpoint from inferencePoolRefs … the pod exits with an error if either is missing"	no endpoint is set; vLLM offsets 5557 by data-parallel rank and the sidecar subscribes at the matching offset. Nothing else has to be set on the InferenceService
gateway binding aside	"(The older inferencePoolRefs / InferencePool mechanism is deprecated; use the label.)"	removed — the field is being taken out of the Odin API reference

ISVC_USE_KV_EVENTS and inferencePoolRef both return zero matches across the preset repository, which is indexed (a 5557 search returns the two DP runtime-bases).

The claim that the ZMQ topic embeds spec.model.name is dropped rather than reworded: nothing in the runtime-bases sets a topic, and I could not verify the format. An unverified format is worse than none.

Left in place deliberately

InferencePool still appears in three places that should keep it:

• skills/guide-heimdall/evals/evals.json — a negative assertion, that the guide does not use InferencePool. Removing it removes a check.
• website/blog/2025-11-11-…mdx — a dated post. Rewriting its manifests would falsify the record.
• website/versioned_docs/version-v0.4.0/**, version-v0.5.0/** — frozen snapshots; website/AGENTS.md forbids editing them.

Not in this PR

website/docs/reference/odin/api-reference.mdx still carries the inferencePoolRefs row and its sub-table. That content sits inside an AUTO-GENERATED block, so hand-editing it would be overwritten on the next docs sync. The generator is being changed to leave out any field whose description opens a Deprecated: paragraph — odin's says the field "is ignored at runtime" — and the sync will remove the row and sub-table on its next run.

Test

skills/ is not part of the Docusaurus site, so no website build is involved.

[Copilot]

Pull request overview

Updates the Odin guide skill documentation to reflect current vLLM runtime-base behavior for prefix cache + KV cache events, removing stale guidance around deprecated inferencePoolRefs-based configuration.

Changes:

• Replaces ISVC_USE_KV_EVENTS with ISVC_ENABLE_PREFIX_CACHE (default true) in the environment variable table.
• Updates the “KV cache events” section to document that KV events are on by default and can be disabled via ISVC_ENABLE_PREFIX_CACHE=false.
• Removes the deprecated inferencePoolRefs / InferencePool aside from the AIGateway binding section.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

Suppressed comments (1)

skills/guide-odin/SKILL.md:535

• In this section, “value is exactly true” could be misread as referring to a YAML boolean rather than the string value Kubernetes passes to the container. Since env.value is always a string, consider explicitly calling out that only the string "true" enables the feature, and that the default when unset is "true".

The runtime-base enables them only when the value is exactly `true`, so any other value disables them — leaving the variable unset is what keeps them on.