Params Pipeline — Reviewer Cluster

Scope

Intent

v1's PluginBuilder had a giant if/else tree that selected plugins by sniffing provider id, model name, and assistant settings. v2 expresses the same matrix as a list of RequestFeatures, each with its own applies gate and contribution methods. Same decision tree — table- driven instead of branch-assembled.

Architecture: docs/references/ai/params-pipeline.md.

Key changes

RequestFeature interface

interface RequestFeature {
  readonly name: string
  applies?(scope: RequestScope): boolean
  contributeModelAdapters?(scope: RequestScope): AiPlugin<any, any>[]
  contributeHooks?(scope: RequestScope): Partial<AgentLoopHooks>
}

Each feature contributes plugins (AI SDK middlewares) and/or hooks (Agent loop callbacks). applies errors are caught and treated as false.

INTERNAL_FEATURES list (order matters)

Listed in features/internalFeatures.ts. Mirrors the prior PluginBuilder.buildPlugins ordering — important pairs:

• pdfCompatibilityFeature before anthropicCacheFeature (cache marker must not see file parts that PDF conversion will rewrite).
• reasoningExtractionFeature before simulateStreamingFeature.

Caller-supplied extraFeatures (e.g. AiService analytics) run after the internal set.

Per-feature complexity

Each file averages ~50 LOC; the gate and the contribution are short.

RequestScope

interface RequestScope extends ToolApplyScope {
  request, signal, registry, assistant, model, provider,
  capabilities, sdkConfig, endpointType, aiSdkProviderId,
  requestContext, mcpToolIds
}

Pre-computed once at the top of buildAgentParams and passed read-only to every feature. Adding a new field is a contract change; review any PR that adds one.

assembleSystemPrompt

Composes:

1. assistant.prompt (variable-replaced via replacePromptVariables).
2. The <DEFERRED_TOOLS> system-prompt section enumerating namespaces that tool_search covers — added only when tool_search is in the final tool set.

buildAgentOptions

Assembles AgentOptions (per-request AI SDK settings). Notable bits:

• customParameters are split via extractAiSdkStandardParams into:
  • AI-SDK standard params (top-level on AgentOptions)
  • Provider-scoped params (merged into providerOptions[aiSdkProviderId] via mergeCustomProviderParameters)
• stopWhen is stepCountIs(N) where N reads assistant.settings.maxToolCalls clamped to [MIN_TOOL_CALLS, MAX_TOOL_CALLS].
• repairToolCall is built from createAiRepair(...) — see tool-cluster.md.

Invariants

• Features never mutate RequestScope.
• Feature order in INTERNAL_FEATURES is significant — don't reorder without testing the pairs called out above.
• applies predicates are synchronous; async gating happens elsewhere (e.g. registry sync).

Validation

• params/__tests__/composeHooks.test.ts (167)
• params/__tests__/assembleSystemPrompt.test.ts (107)
• params/__tests__/collectFromFeatures.test.ts (123)
• params/features/__tests__/internalFeatures.test.ts (201) — gates fire on the right shapes
• params/features/__tests__/deepseekDsmlParserPlugin.test.ts (462) — reasoning-extraction parser

Follow-ups (out of scope)

• See also Present-tense consumers only memory: don't add feature-level config fields without a real call site.