feat(provider): MTP posture and acceptance counters on the local /metrics endpoint

[Gajesh2007]

Problem

Bug 2 from Layr-Labs/mlx-swift-lm#106's Discovered bugs: MTP acceptance is unobservable outside Datadog Logs. The counters exist in-engine (CBv2MTPMetrics, polled via EngineV2Bridge.mtpStatusSnapshot()), but the only surface was the engine_v2_slot_posture telemetry event → Datadog Logs. The provider's local HTTP server exposes /metrics with no MTP counters at all, and the os_log store is inaccessible in sandboxes — so a headless operator could not observe whether MTP is producing drafts, what the acceptance ratio is, or that MTP silently never activated (inline_artifact_invalid, the sibling bug fixed in #618; inert_kv_unsupported; …).

Change

/metrics now appends provider-owned per-model lines behind the upstream MLXLMServer body, in the same Prometheus text style (# TYPE + bare lines):

mtp_enabled{model="…"} 0|1
mtp_active{model="…"} 0|1
mtp_rounds_total{model="…"} N
mtp_tokens_proposed_total{model="…"} N
mtp_tokens_accepted_total{model="…"} N
mtp_inactive_reason{model="…",reason="inline_artifact_invalid"} 1   # omitted when productively running

Counters are cumulative per slot lifetime (standard counter contract, matching the telemetry event); acceptance rate = accepted_total / proposed_total. mtp_inactive_reason carries the existing MTPFallbackReason enum — the same status inline_artifact_invalid rides in — so a silently-disabled drafter is operationally visible.

Mechanics: the upstream /metrics route lives in MLXLMServer (ServerMetrics.prometheusText()), whose router is already built when the provider wraps it — so a new LocalMetricsResponder serves GET /metrics itself (same public MLXOpenAIService.prometheusMetrics() entry point + the MTP lines), exactly the interception pattern LocalChatUploadResponder established. Wired through both local endpoints via a mtpSlots closure: the StandaloneServer slot cache (--local) and the unified ProviderLoop.modelSlots (--local-endpoint) — the same bridge snapshot the capacity tick already feeds into DaemonSlotPostureBuilder, so /metrics, daemon-state, and telemetry cannot describe a slot differently.

Deliberately not touched: the coordinator telemetry wire types (Go/Swift/TS three-language sync — out of scope); the daemon-state SlotPosture already carries mtpEnabled/mtpActive/mtpInactiveReason, so darkbloom status needed no change.

Behavior

flowchart LR
  subgraph Before
    A1["CBv2MTPMetrics (in-engine)"] --> B1["engine_v2_slot_posture event"] --> C1["Datadog Logs ONLY<br/>headless/sandboxed operator: blind<br/>silently-disabled MTP invisible"]
    D1["GET /metrics"] --> E1["upstream mlx_server_* counters only"]
  end
  subgraph After
    A2["CBv2MTPMetrics (in-engine)"] --> B2["engine_v2_slot_posture event -> Datadog (unchanged)"]
    A2 --> D2["GET /metrics"] --> E2["mlx_server_* body +<br/>mtp_enabled / mtp_active / mtp_rounds_total /<br/>mtp_tokens_proposed_total / mtp_tokens_accepted_total /<br/>mtp_inactive_reason{reason=…}"]
  end

Loading

Code

flowchart LR
  subgraph Before
    R1["auth -> CORS -> chat-upload -> upstream router"] --> M1["/metrics = ServerMetrics.prometheusText() verbatim"]
  end
  subgraph After
    R2["auth -> CORS -> LocalMetricsResponder -> chat-upload -> upstream router"] --> M2["/metrics = prometheusMetrics() + MTPPrometheusRenderer.render(mtpSlots())"]
    S2[StandaloneServer.mtpSlotMetricsSamples] --> M2
    P2[ProviderLoop.mtpSlotMetricsSamplesForLocal] --> M2
  end

Loading

Pin independence (#616)

Pure provider-swift change, based on master (which already carries the #616 pin bump to mlx-swift-lm d06da67). It uses no field introduced by mlx-swift-lm#106 — CBv2MTPMetrics.rounds/draftedTokens/acceptedTokens/active and MLXOpenAIService.prometheusMetrics() all predate it — so it also compiles against the pre-#106 pin (c04afb4); no availability gating needed.

Tests

Targeted filters green (swift test, debug, with source-matched metallib per make provider-test):

• mtpPrometheusRendererEmitsPostureAndCounters — stubbed active snapshot (7 rounds, 21 proposed, 14 accepted) and a stubbed inline_artifact_invalid slot: all six families render, healthy slots omit mtp_inactive_reason, label values are Prometheus-escaped, empty slot list renders nothing (upstream body stays byte-identical).
• localMetricsEndpointAppendsMTPLinesToUpstreamBody — full responder stack via makeLocalInferenceApplication with a stubbed mtpSlots closure: GET /metrics returns 200 text/plain, the upstream mlx_server_* body intact, and the stubbed counters (3/9/6) appear.
• standaloneServerMetricsWithoutSlotsServesUpstreamBodyOnly — cold standalone server: upstream body only, no mtp_ lines.
• Plus the 16-test standalone-server/CORS/auth/upload filter unchanged.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
d-inference	Ready	Preview	Aug 13, 2026 11:41pm
d-inference-console-ui-dev	Ready	Preview	Aug 13, 2026 11:41pm
d-inference-landing	Ready	Preview	Aug 13, 2026 11:41pm

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8678aa4af8

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".