Skip to content

docs: resync skills with smg HEAD (v1.6.0) - #20

Merged
slin1237 merged 1 commit into
mainfrom
docs/resync-skills-v1.6.0
Jul 15, 2026
Merged

docs: resync skills with smg HEAD (v1.6.0)#20
slin1237 merged 1 commit into
mainfrom
docs/resync-skills-v1.6.0

Conversation

@slin1237

Copy link
Copy Markdown
Member

Summary

Skills were last verified against smg on 2026-06-17; ~103 commits have landed since (HEAD 72ec2bfa, 2026-07-15). Audited every recipe against the current code (not commit messages) via parallel per-subsystem audits and fixed the drift. 15 files changed, 23 confirmed drifts plus one architecture-coverage addition.

Highlights

Routing policies (9 → 10, passthrough added) — rippled through map/SKILL.md (×2), routing-policy.md, bindings-update.md (PolicyType/PolicyConfig).

gRPC router refactor (#1923) — documented the unified Mode-parameterized GrpcRouter + RequestPipeline/PipelineStage architecture in grpc-backend.md (pd_router.rs removed, epd_encodecommon/stages/encode.rs), and added the TokenSpeedEncoderClient EPD-encode exception. This was a coverage gap, not a wrong claim — the recipes never described the router internals.

Factual fixes

  • multimodal: drop non-existent ::new_dynamic (::new is generic over dimensionality); add AudioUrl/AudioData to MediaContentPart; scope the registry-ordering rule to the spec Vec (processor registry is a HashMap); fetch_videoArc<VideoClip>.
  • mcp: rmcp 0.8.x1.7; SSE client transport dropped — model new transports on the Streamable arm; drop SseClientTransport from the anti-pattern.
  • rate-limit: rate_limit_tokens_per_second = 0 does not yield semaphore mode — coerced to max_concurrent_requests via .filter(|&t| t > 0).
  • discovery: add the Encode pod type / --encode-selector; Router is an is_router overlay, not a PodType.
  • provider-api: ProviderType gains Custom(String).
  • config-plumbing: to_server_config does take a RouterConfig and uses .or_else() for discovery-derived fields.
  • implement: realtime router path → routers/common/realtime/.
  • wasm: OnError attach point is defined but not implemented.
  • auth: JTI replay check is opt-in / off by default.
  • storage-backend: data-connector v2.3.0 → v2.3.2.
  • review-pr: correct the e2e pytest marker set (model/workers/gateway/e2e/slow).
  • map: add the mm_rdma (smg-mm-rdma) crate row; fix the bindings param anchor.

Verified in sync (no changes)

scheduler, tenancy, tool-parser, reasoning-parser, observability, kv-index, mesh, contribute gate, review anti-patterns. Counts re-confirmed: 14 tool parsers, 8 reasoning families, Python Router now 117 params.

Known non-changes (flagged, not drift)

  1. contribute/SKILL.md uses clippy --workspace; review-pr/CONTRIBUTING.md don't — both valid per their respective sources, just internally inconsistent.
  2. contribute/SKILL.md documents no branch-name regex — a coverage gap; the enforced patterns live in the pre-commit branch-name-check hook and pr-naming-check.yml.

Skills were last verified against smg on 2026-06-17; ~103 commits have
landed since (HEAD 72ec2bfa). Audited every recipe against the current
code (not commit messages) via parallel per-subsystem audits and fixed
the drift.

Routing policies (9 -> 10, `passthrough` added):
- map: "9 routing policies" / "9 load-balancing policies" -> 10.
- routing-policy: add `passthrough` to the impl list.
- bindings-update: PolicyType/PolicyConfig list -> 10 (add Passthrough).

gRPC router refactor (#1923, unified regular/PD/EPD):
- grpc-backend: document the Mode-parameterized GrpcRouter +
  RequestPipeline/PipelineStage architecture (pd_router.rs removed,
  epd_encode -> common/stages/encode.rs); add the TokenSpeedEncoderClient
  EPD-encode exception (not a GrpcClient enum variant).

Factual fixes:
- multimodal: drop non-existent `::new_dynamic` (`::new` is generic over
  dimensionality); add AudioUrl/AudioData to MediaContentPart; scope the
  registry-ordering rule to the spec Vec (processor registry is a HashMap);
  fetch_video returns Arc<VideoClip>.
- mcp: rmcp 0.8.x -> 1.7; SSE client transport dropped, model new
  transports on the Streamable arm; drop SseClientTransport from the
  anti-pattern list.
- rate-limit: `rate_limit_tokens_per_second = 0` does NOT yield semaphore
  mode; it is coerced to max_concurrent_requests via .filter(|&t| t > 0).
- discovery: add the Encode pod type / --encode-selector; note Router is
  an is_router overlay, not a PodType.
- provider-api: ProviderType gains Custom(String).
- config-plumbing: to_server_config does take a RouterConfig and uses
  .or_else() for discovery-derived fields.
- implement: realtime router path -> routers/common/realtime/.
- wasm: OnError attach point is defined but not implemented.
- auth: JTI replay check is opt-in / off by default.
- storage-backend: data-connector v2.3.0 -> v2.3.2.
- review-pr: correct the e2e pytest marker set (model/workers/gateway/
  e2e/slow).
- map: add the mm_rdma (smg-mm-rdma) crate row; bindings param anchor.

Verified in sync (no changes): scheduler, tenancy, tool-parser,
reasoning-parser, observability, kv-index, mesh, contribute gate,
review anti-patterns.

Signed-off-by: Simo Lin <25425177+slin1237@users.noreply.github.com>
@slin1237
slin1237 merged commit e119304 into main Jul 15, 2026
2 checks passed
@slin1237
slin1237 deleted the docs/resync-skills-v1.6.0 branch July 15, 2026 20:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant