mcpx: thin provider layer core — internal/model, model.* settings, intent routing (#1925 slice 1) by dezren39 · Pull Request #2592 · developing-today/code

dezren39 · 2026-10-30T06:08:29Z

Slice 1 of #1925 (Approach C): the thin provider layer core. Queue task 3215.

What lands

pkgs/mcpx/internal/model (new package) — one small interface mcpx owns:

Provider interface: Name() + Generate(ctx, Request) (Response, error); Request carries messages, maxTokens, stop sequences and an optional JSON schema; model ids use the provider/model namespace (SplitModelID).
OpenAI-compatible configured-provider backend (OpenAIProvider): one client, configurable base URL, so OpenAI / a company gateway / LiteLLM / Ollama / LM Studio all work through it. Honours model.baseURL and the key named by model.apiKeyEnv; passes a request schema through as response_format: json_schema.
Sampling backend (SamplingProvider): the pre-existing broker sampling path exposed as a backend over an injected SampleFunc, so sampling survives as one backend among several instead of the only path.
Resolve(Config): none (default) resolves to no provider and no error — configuring nothing changes nothing. auto prefers a live sampling path, then a configured provider (only when the user actually configured something: a set apiKeyEnv variable or an explicit base URL — it never scans the environment for known key names), then fails cleanly. plugin is named in the enum and resolves to unavailable with the reason: that backend is a later slice and depends on build(deps): bump actions/checkout from 3 to 4 #47.

model.* settings through the one registry (internal/settings), no new endpoints — config key, MCPX_* env, CLI flag, /v1/settings, MCP tool, per the prompt.* precedent: model.provider (auto|sampling|plugin|openai|anthropic|google|none, daemon default none, matching prompt.sample: never), model.baseURL, model.apiKeyEnv (opt-in: name the env var; the key itself is never a setting value), model.default, model.budget (hard maxTokens ceiling on the layer; 0 = no layer ceiling), model.timeout. docs/configuration.md inventory regenerated.

Routing: /v1/intent's generation caller (Server.generate) goes through the layer when model.provider is configured: same catalog slice and system prompt, one Generate call, model.budget/model.default applied, and the daemon wires the sampling backend to the elicitation broker (elicit.SampleAsk + Await, bounded by model.timeout). A provider error lands exactly where the scope guard requires: the route returns its deterministic result — the ranked recipes — with the note that no model answered. With model.provider=none (default) the branch is skipped entirely and the existing repair.* provider path is untouched, including script repair.

Docs: docs/consumer.md's "a configured provider … was not built and should not be" passage is rewritten in the same change, per the issue: SEP-2577 names direct provider APIs as the migration from deprecated Sampling, so the sentence now records the reversal, the layer's shape, and the scope guard that survives it.

Deliberately out of this slice (stay on #1925)

model.profiles/model.sets fallback chains, the opencode-plugin backend (#47), sampling.answerLocally, mcpx-as-an-OpenAI-proxy, script-side mcpx.model.generate, native anthropic/google clients (both are served here through their OpenAI-compatible endpoints).

Tests

go test ./... in pkgs/mcpx — new internal/model tests cover the chat-completions request shape (provider prefix stripped, key header, schema passthrough), HTTP error surfacing, the sampling adapter, and the Resolve matrix (none/auto/plugin/sampling-without-path). Settings and daemon suites pass, including the regenerated-docs inventory test and the settings-consumed scan (every model.* setting is read via an accessor in consumerPolicyFrom).

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

coderabbitai · 2026-10-30T06:08:39Z

Warning

Review limit reached

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Next included review available in 53 minutes.

Check out review usage here.

View limit details

Limit details: You’ve used the included review currently available.

Learn how review limits work.

Review configuration:

Run configuration

Source mode: Auto
Repository: Unavailable
Branch: Unavailable
File: No configuration file from this repository applied
Commit: Unavailable
Layers: defaults
Review profile: CHILL
Plan: Advanced
Run ID: b52d9561-309c-4852-a845-b367fa9fff92

Commits

Reviewing files that changed from the base of the PR and between 1ab2f23 and a9ef359.

Files selected for processing: 15

pkgs/mcpx/docs/configuration.md
pkgs/mcpx/docs/consumer.md
pkgs/mcpx/internal/daemon/consumer.go
pkgs/mcpx/internal/daemon/generate_test.go
pkgs/mcpx/internal/daemon/model_layer_test.go
pkgs/mcpx/internal/daemon/routes_consumer.go
pkgs/mcpx/internal/defaults/defaults.go
pkgs/mcpx/internal/defaults/defaults.json
pkgs/mcpx/internal/model/model.go
pkgs/mcpx/internal/model/model_test.go
pkgs/mcpx/internal/model/openai.go
pkgs/mcpx/internal/model/resolve.go
pkgs/mcpx/internal/model/sampling.go
pkgs/mcpx/internal/settings/model.go
pkgs/mcpx/internal/settings/registry.go

Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands.}

chatgpt-codex-connector · 2026-10-30T06:08:39Z

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-30T12:19:18.493443Z	`a9ef359`	New commits

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d33249c30a

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".

Six findings from the Codex review of d33249c, each fixed in the PR: - P1 routes_consumer: the legacy prov==nil guard no longer gates the model layer; a model.provider-only configuration reaches generate. - P1 consumer/model: auto no longer prefers sampling over a configured provider -- broker presence is not a live sampling path; sampling is auto's fallback and first choice only when named explicitly. - P2 consumer: a model-layer resolution error returns errNoModel instead of falling through to the legacy repair provider. - P2 routes_consumer: handleIntent passes only the explicit request model; repair.model no longer overrides model.default in the layer (the legacy path applies repair.model itself for its label). - P2 consumer: brokerSampler bounds the sampling wait with a context deadline from model.timeout, not only the question TTL that Broker.Get renews. - P2 consumer: the intent session is propagated into the sampling question so Pending session filtering can find it. Tests: model_layer_test.go (resolution-error no-fallback, layer-alone answers, broker session+timeout); TestResolveAuto updated for the new precedence. go test ./internal/model/ ./internal/daemon/ passes on the laptop.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6cb37afc5b

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".

…2592) Round on 6cb37af, all four in the intent/model path: - Explicit provider wins: generate() now takes the request's explicit provider and consults the daemon model layer only when none was supplied -- a request naming provider=claude no longer has its model id sent to the configured model.provider endpoint. - Repair follows generation: handleIntent repairs through the backend that generated the script. A model-layer script is repaired via modelRepairProvider (the layer's backend under the layer's budget); a layer that no longer resolves means no repair, never an unrelated auto-discovered legacy provider. - Budget clamp: modelMaxTokens treats a nonpositive per-call limit as unset, so prompt.maxTokens=-1 can no longer bypass model.budget by being omitted from the request. - Token-limit field: the OpenAI backend sends max_completion_tokens to o-series and GPT-5 models, which reject max_tokens; other models and endpoints keep max_tokens. Tests: TestGenerateExplicitProviderSkipsModelLayer, TestModelMaxTokens, TestModelRepairProviderCarriesLayerBounds, TestTokenLimitParam; the existing chat-completion test now asserts the gpt-5 field. go test ./internal/model/ ./internal/daemon/ pass.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 79aea580d8

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".

…sample lifecycle) P2 openai.go: Generate no longer sends the invented model id "default" -- Resolve threads model.default into the provider as its only fallback, and an unnamed request with no configured default is ErrUnavailable (callers fall back deterministically). model.default settings text stops promising a backend default. P2 consumer.go: brokerSampler publishes sample.opened before Await and sample.answered after it, mirroring the upstream sampling path, so sample.opened hook answerers can answer model.provider=sampling requests. Tests: no-model sends nothing, configured default reaches the endpoint, broker sampler publishes the lifecycle in order.

…tent routing (#1925 slice 1) Queue task 3215. Approach C from #1925: a Provider interface mcpx owns (Generate; provider/model id namespace), an OpenAI-compatible configured-provider backend honouring model.baseURL/model.apiKeyEnv, the broker sampling path exposed as a backend, model.* settings through the one registry with daemon default model.provider=none (no behaviour change when unconfigured), /v1/intent generation routed through the layer with deterministic fallback, and the docs/consumer.md provider sentence reversed in the same change per SEP-2577. Out of slice: profiles/sets, plugin backend (#47), answerLocally, proxy, script-side generate.

Six findings from the Codex review of d33249c, each fixed in the PR: - P1 routes_consumer: the legacy prov==nil guard no longer gates the model layer; a model.provider-only configuration reaches generate. - P1 consumer/model: auto no longer prefers sampling over a configured provider -- broker presence is not a live sampling path; sampling is auto's fallback and first choice only when named explicitly. - P2 consumer: a model-layer resolution error returns errNoModel instead of falling through to the legacy repair provider. - P2 routes_consumer: handleIntent passes only the explicit request model; repair.model no longer overrides model.default in the layer (the legacy path applies repair.model itself for its label). - P2 consumer: brokerSampler bounds the sampling wait with a context deadline from model.timeout, not only the question TTL that Broker.Get renews. - P2 consumer: the intent session is propagated into the sampling question so Pending session filtering can find it. Tests: model_layer_test.go (resolution-error no-fallback, layer-alone answers, broker session+timeout); TestResolveAuto updated for the new precedence. go test ./internal/model/ ./internal/daemon/ passes on the laptop.

…2592) Round on 6cb37af, all four in the intent/model path: - Explicit provider wins: generate() now takes the request's explicit provider and consults the daemon model layer only when none was supplied -- a request naming provider=claude no longer has its model id sent to the configured model.provider endpoint. - Repair follows generation: handleIntent repairs through the backend that generated the script. A model-layer script is repaired via modelRepairProvider (the layer's backend under the layer's budget); a layer that no longer resolves means no repair, never an unrelated auto-discovered legacy provider. - Budget clamp: modelMaxTokens treats a nonpositive per-call limit as unset, so prompt.maxTokens=-1 can no longer bypass model.budget by being omitted from the request. - Token-limit field: the OpenAI backend sends max_completion_tokens to o-series and GPT-5 models, which reject max_tokens; other models and endpoints keep max_tokens. Tests: TestGenerateExplicitProviderSkipsModelLayer, TestModelMaxTokens, TestModelRepairProviderCarriesLayerBounds, TestTokenLimitParam; the existing chat-completion test now asserts the gpt-5 field. go test ./internal/model/ ./internal/daemon/ pass.

…sample lifecycle) P2 openai.go: Generate no longer sends the invented model id "default" -- Resolve threads model.default into the provider as its only fallback, and an unnamed request with no configured default is ErrUnavailable (callers fall back deterministically). model.default settings text stops promising a backend default. P2 consumer.go: brokerSampler publishes sample.opened before Await and sample.answered after it, mirroring the upstream sampling path, so sample.opened hook answerers can answer model.provider=sampling requests. Tests: no-model sends nothing, configured default reaches the endpoint, broker sampler publishes the lifecycle in order.

dezren39 · 2026-10-30T12:18:44Z

🤖 Agent note · parked

Agent: queue worker (repo-worker-code), Lane E shift
Session / ID: queue task 3215 (resume)
Found via: queue claim
Task: Advanced PR mcpx: thin provider layer core — internal/model, model.* settings, intent routing (#1925 slice 1) #2592 per the recorded next step, then parked for CI:
- Rebased onto current main — clean (4 commits replayed, no conflicts) on the laptop worktree; force-pushed with lease: 56e0dcb3 → a9ef359d.
- Base-drift diagnosis confirmed for one package: go test ./internal/mcpclient/ on the rebased worktree passes (8.0s, laptop) — its CI failure was drift, not this slice. The e2e package (305s in CI) was not re-run locally; the fresh CI run decides it.
- CI triggered on a9ef359d: mcpx workflow run is queued (fleet runner backlog — runs have queued ~30+ min today). Parked 20 min per the CI rule.
- Individual failing test names from the old run remain unavailable (not in job logs; no artifacts retained) — if the fresh e2e run still fails, its gotest step summary is the naming source.
Expectation: on resume — read PR mcpx: thin provider layer core — internal/model, model.* settings, intent routing (#1925 slice 1) #2592 checks fresh; if green (and any new Codex round on the rebased head is dispositioned), merge per standing criteria; if e2e alone is red, name + fix its tests in the worktree.
Links: PR mcpx: thin provider layer core — internal/model, model.* settings, intent routing (#1925 slice 1) #2592 · branch queue/3215-1925-model-provider-core · worktree ~/git/fleet/3215-1925-model-provider-core (laptop)

dezren39 requested a review from a team as a code owner October 30, 2026 06:08

dezren39 mentioned this pull request Oct 30, 2026

mcpx: where the model comes from — a provider layer mcpx owns, model profiles, and life after MCP Sampling's deprecation #1925

Open

chatgpt-codex-connector Bot reviewed Oct 30, 2026

View reviewed changes

Comment thread pkgs/mcpx/internal/daemon/consumer.go Resolved

Comment thread pkgs/mcpx/internal/model/openai.go Resolved

Comment thread pkgs/mcpx/internal/daemon/consumer.go Resolved

Comment thread pkgs/mcpx/internal/daemon/consumer.go Resolved

dezren39 mentioned this pull request Oct 30, 2026

mcpx: accept a flake ref as a program and as an execution environment #1921

Open

chatgpt-codex-connector Bot reviewed Oct 30, 2026

View reviewed changes

Comment thread pkgs/mcpx/internal/model/openai.go Resolved

Comment thread pkgs/mcpx/internal/daemon/consumer.go Resolved

dezren39 mentioned this pull request Oct 30, 2026

ci: Actions check runs stuck in queued across PRs — no runner pickup since ~05:00Z 2026-10-28 #1599

Closed

github-actions Bot assigned dezren39 Oct 30, 2026

dezren39 added 4 commits October 30, 2026 07:16

dezren39 force-pushed the queue/3215-1925-model-provider-core branch from 56e0dcb to a9ef359 Compare October 30, 2026 12:16

This was referenced Oct 30, 2026

mcpx: Phase 1 of #2077 — downgrade rich enum schemas for pre-2025-11-25 hosts (multi-select withheld, not flattened) #2078

Closed

mcpx: downgrade titled single-select enums for pre-2025-11-25 hosts, withhold multi-select (#2078) #2719

Merged

dezren39 merged commit 5d173d8 into main Oct 30, 2026
4 of 5 checks passed

dezren39 deleted the queue/3215-1925-model-provider-core branch October 30, 2026 19:38

Conversation

dezren39 commented Oct 30, 2026 • edited by blacksmith-sh Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What lands

Deliberately out of this slice (stay on #1925)

Tests

Uh oh!

coderabbitai Bot commented Oct 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review limit reached

Uh oh!

chatgpt-codex-connector Bot commented Oct 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codex Review Summary

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

dezren39 commented Oct 30, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

dezren39 commented Oct 30, 2026 •

edited by blacksmith-sh Bot

Loading

coderabbitai Bot commented Oct 30, 2026 •

edited

Loading

chatgpt-codex-connector Bot commented Oct 30, 2026 •

edited

Loading