Developing Genkit Go

Middleware

Middleware wraps Generate calls to add cross-cutting behavior (retries, logging, fallback, gating, sandboxed tools) without touching the flow that uses it. Middleware composes, so a single Generate call can stack several behaviors. Built-ins ship in the plugins/middleware package; custom middleware is just a Go struct with two methods.

The mental model

A middleware is a config struct that implements two methods:

type Middleware interface {
    Name() string                            // stable, registered identifier
    New(ctx context.Context) (*Hooks, error) // produces a per-call hook bundle
}

The same struct value the user passes to ai.WithUse is the value the runtime calls New on. There is no separate factory parameter and no embedded base type. Per-call state goes in closures captured by New. Plugin-level state goes on unexported fields of the struct.

New is invoked once per Generate call. The returned *Hooks is reused across every iteration of the tool loop within that call.

type Hooks struct {
    Tools        []Tool                                        // injected for this call
    WrapGenerate func(ctx, *GenerateParams, GenerateNext) ...  // tool-loop iteration
    WrapModel    func(ctx, *ModelParams, ModelNext) ...        // model API call
    WrapTool     func(ctx, *ToolParams, ToolNext) ...          // tool execution
}

A nil hook is a pass-through. Implement only what the middleware needs.

When each hook fires

A Generate call executes a tool loop: model produces output, any tool calls execute, results feed back into a new model call, repeat until the model stops. The hooks fire at three different layers of this loop:

WrapGenerate is the right place for logic that needs to see the whole conversation (rewrites, system-prompt injection, message accumulation). WrapModel is the right place for logic about the model call itself (retry, fallback, caching). WrapTool is the right place for logic about a single tool execution (approval, sandboxing, logging).

Composition order

ai.WithUse(A, B, C) expands to A { B { C { actual } } } at call time. Each layer's next continuation runs the next inner layer:

ai.WithUse(
    &middleware.Retry{MaxRetries: 3},     // outer: retries the whole inner stack
    &middleware.Fallback{Models: ...},    // inner: tries fallback models on failure
)
// effective chain: Retry { Fallback { model } }

Order matters. Retry outside Fallback retries the entire fallback cascade as a unit. Swapped, you'd retry the primary first and only fall back after exhausting retries.

Per-call state

State that should be shared across the hooks of a single Generate call lives in closures captured by New. Each Generate call gets a fresh Hooks bundle, so nothing leaks between calls.

type Counter struct{}

func (Counter) Name() string { return "mine/counter" }

func (Counter) New(ctx context.Context) (*ai.Hooks, error) {
    var modelCalls int
    return &ai.Hooks{
        WrapModel: func(ctx context.Context, p *ai.ModelParams, next ai.ModelNext) (*ai.ModelResponse, error) {
            modelCalls++
            return next(ctx, p)
        },
        WrapGenerate: func(ctx context.Context, p *ai.GenerateParams, next ai.GenerateNext) (*ai.ModelResponse, error) {
            // The same modelCalls variable is visible here because both closures
            // capture it from the enclosing New scope.
            resp, err := next(ctx, p)
            if err == nil {
                log.Printf("iteration %d: %d model calls so far", p.Iteration, modelCalls)
            }
            return resp, err
        },
    }, nil
}

WrapTool may be invoked concurrently for parallel tool calls in the same iteration. Any state it mutates must be guarded:

func (Counter) New(ctx context.Context) (*ai.Hooks, error) {
    var (
        mu        sync.Mutex
        toolCalls int
    )
    return &ai.Hooks{
        WrapTool: func(ctx context.Context, p *ai.ToolParams, next ai.ToolNext) (*ai.MultipartToolResponse, error) {
            mu.Lock()
            toolCalls++
            mu.Unlock()
            return next(ctx, p)
        },
    }, nil
}

WrapGenerate and WrapModel are not called concurrently within a single Generate call.

Plugin-level state

When a middleware needs resources its config can't carry as JSON (an HTTP client, a database handle, a logger), put them on unexported fields of the config struct. The plugin sets them on a prototype, and ai.NewMiddleware captures that prototype in a closure that value-copies it across JSON-dispatched invocations:

type Logger struct {
    Prefix string `json:"prefix,omitempty"`
    out    io.Writer // unexported; preserved across JSON dispatch via value-copy
}

func (Logger) Name() string { return "mine/logger" }

func (l Logger) New(ctx context.Context) (*ai.Hooks, error) {
    return &ai.Hooks{
        WrapModel: func(ctx context.Context, p *ai.ModelParams, next ai.ModelNext) (*ai.ModelResponse, error) {
            start := time.Now()
            resp, err := next(ctx, p)
            fmt.Fprintf(l.out, "%s model call took %s\n", l.Prefix, time.Since(start))
            return resp, err
        },
    }, nil
}

type LoggerPlugin struct{ Out io.Writer }

func (p *LoggerPlugin) Name() string                          { return "logger" }
func (p *LoggerPlugin) Init(ctx context.Context) []api.Action { return nil }

func (p *LoggerPlugin) Middlewares(ctx context.Context) ([]*ai.MiddlewareDesc, error) {
    return []*ai.MiddlewareDesc{
        ai.NewMiddleware("logs model call latency", Logger{out: p.Out}),
    }, nil
}

The Dev UI and other-runtime callers send JSON config; the prototype's value copy preserves out (unexported, not in JSON) while Prefix is overridden by the unmarshaled config.

Composition with WithUse

response, _ := genkit.Generate(ctx, g,
    ai.WithModelName("googleai/gemini-flash-latest"),
    ai.WithPrompt("Explain quantum computing."),
    ai.WithUse(
        Logger{Prefix: "[trace]"},
        &middleware.Retry{MaxRetries: 3},
    ),
)

No registration is required for pure-Go use. WithUse calls each value's New directly on a fast path; the registry is only consulted for JSON-dispatched calls (Dev UI or cross-runtime). Registration is what makes a middleware visible to the Dev UI and addressable by name.

Inline middleware

For ad-hoc middleware that does not need Dev UI visibility or a named type, use ai.MiddlewareFunc:

ai.WithUse(ai.MiddlewareFunc(func(ctx context.Context) (*ai.Hooks, error) {
    return &ai.Hooks{
        WrapModel: func(ctx context.Context, p *ai.ModelParams, next ai.ModelNext) (*ai.ModelResponse, error) {
            log.Printf("model call: %s", p.Request.Messages[len(p.Request.Messages)-1].Text())
            return next(ctx, p)
        },
    }, nil
}))

The adapter satisfies Middleware with a placeholder Name() of "inline". Inline middleware is resolved on the fast path and never touches the registry, so the placeholder name is fine.

Application-owned middleware

Use genkit.DefineMiddleware to register a middleware your application owns directly. Registration surfaces it in the Dev UI and lets cross-runtime callers reference it by name:

genkit.DefineMiddleware(g, "logs model call latency", Logger{out: os.Stderr})

// Lookup by name (mostly for inspection / cross-runtime dispatch).
desc := genkit.LookupMiddleware(g, "mine/logger")

For application code, DefineMiddleware is the typical entry point. For plugin authors, ai.NewMiddleware (no registration) plus MiddlewarePlugin.Middlewares() is the typical entry point. genkit.Init registers the returned descriptors automatically.

Built-in middleware

The plugins/middleware package bundles six production-ready implementations (Retry, Fallback, ToolApproval, Filesystem, Skills, SoftToolErrors); the experimental plugins/middleware/exp package (imported as middlewarex) adds ContextCompression, Budget, Agents, and Artifacts. Register the plugin once during Init to make them visible to the Dev UI:

import "github.com/firebase/genkit/go/plugins/middleware"

g := genkit.Init(ctx, genkit.WithPlugins(
    &googlegenai.GoogleAI{},
    &middleware.Middleware{},
))

Retry

Retries failed model API calls with exponential backoff and jitter. Hooks WrapModel.

ai.WithUse(&middleware.Retry{
    MaxRetries:     3,     // default 3
    InitialDelayMs: 1000,  // default 1000
    MaxDelayMs:     60000, // default 60000
    BackoffFactor:  2,     // default 2
    NoJitter:       false, // default false
    // Statuses (default: UNAVAILABLE, DEADLINE_EXCEEDED, RESOURCE_EXHAUSTED, ABORTED, INTERNAL)
    // Statuses: []core.StatusName{core.UNAVAILABLE, core.RESOURCE_EXHAUSTED},
})

Non-GenkitError errors (network, parse, etc.) are always retried. GenkitError errors are retried only if their status is in Statuses. The backoff respects ctx.Done(): a canceled context aborts the retry loop with the last error.

Fallback

Tries alternative models when the primary fails with a fallback-eligible status. Hooks WrapModel.

ai.WithUse(&middleware.Fallback{
    Models: []ai.ModelRef{
        googlegenai.ModelRef("googleai/gemini-flash-latest", nil),
        googlegenai.ModelRef("vertexai/gemini-flash-latest", nil),
    },
    // default Statuses: UNAVAILABLE, DEADLINE_EXCEEDED, RESOURCE_EXHAUSTED,
    // ABORTED, INTERNAL, NOT_FOUND, UNIMPLEMENTED
})

Each fallback model uses its own ModelRef.Config() verbatim; the original request's config is not inherited. Compose with Retry outside to retry the whole cascade, or inside to retry just the primary before falling back.

ToolApproval

Holds any tool call not in the allow list, exposing approval as a human-in-the-loop step. Hooks WrapTool.

approval := ai.WithUse(&middleware.ToolApproval{
    AllowedTools: []string{"lookup", "search"}, // anything else is held
})

Claim holds with middleware.ToolApprovalInterrupted (not the tool's own Interrupted) and answer with a typed middleware.ToolCallDecision. Keep the same middleware on the resume call:

var parts []*ai.Part
for _, part := range resp.Interrupts() {
    if call, ok := middleware.ToolApprovalInterrupted(part); ok {
        decision := middleware.ToolCallDecision{Approved: confirm(part.ToolRequest.Name, call.Input)}
        if !decision.Approved {
            decision.Reason = "the user declined" // the model sees this as the tool's error
        }
        parts = append(parts, call.Restart(decision))
    }
}
resp, err = genkit.Generate(ctx, g,
    ai.WithMessages(resp.History()...),
    ai.WithTools(writeFile),
    approval,
    ai.WithResume(parts...),
)

Silence never approves: any restart without Approved: true refuses the call, and the model receives the refusal as a tool error and continues. Answer a hold with Restart, not Respond.

Judge model. Let a model decide calls the allowlist does not cover (auto-approval). It answers allow, deny (the model sees the denial), or ask (held for you), and fails closed:

ai.WithUse(&middleware.ToolApproval{
    AllowedTools: []string{"list_files"},
    Judge:        googlegenai.ModelRef("googleai/gemini-flash-lite-latest", nil),
    JudgePolicy:  "Never allow deleting anything outside build/.",
})

SoftToolErrors

Returns tool errors to the model as the tool's output ({"error": "..."}) instead of stopping the loop, so the model can fix its input. Useful for tools you don't own, like MCP tools. Interrupts and cancellation still stop the loop.

ai.WithUse(&middleware.SoftToolErrors{})                         // every tool
ai.WithUse(&middleware.SoftToolErrors{Tools: []string{"search"}}) // only these

For a tool you own, return tool.Fail(ctx, err) for just the errors the model can fix. See tools.md.

ContextCompression (experimental, middlewarex)

Keeps long chats and tool loops inside a context budget. The model receives a compressed view; resp.History() keeps every original message, with compression state in metadata.

import middlewarex "github.com/firebase/genkit/go/plugins/middleware/exp"

ai.WithUse(
    &middlewarex.ContextCompression{
        MaxInputTokens:        80_000,
        DedupeToolResponses:   &middlewarex.CompressionDedupe{},
        TruncateToolResponses: &middlewarex.CompressionToolTruncation{MaxChars: 2000},
        Summarize: &middlewarex.CompressionSummarizer{
            Model: googlegenai.ModelRef("googleai/gemini-flash-lite-latest", nil),
        },
        // also: MaxMessages, PreserveRecent (4), MaxToolResponseChars (400000),
        // SkipSummarizationThreshold, ReplaceHistory
    },
    &middleware.Retry{}, // inside compression, so a retry resends the compressed view
)

List it after context-adding middleware (Skills, Filesystem, Agents, Artifacts) and before Retry/Fallback. middlewarex.ResolveCompressedHistory(msgs) returns what the model received. Set MaxInputTokens well above a typical first turn.

Budget (experimental, middlewarex)

Caps what a run (or an agent session) may spend. Each non-zero field of Limit (an ai.GenerationUsage) is a cap; Custom is capped key by key.

resp, err := genkit.Generate(ctx, g,
    ai.WithPrompt("Research the topic and summarize it."),
    ai.WithTools(search),
    ai.WithUse(&middlewarex.Budget{Limit: ai.GenerationUsage{
        TotalTokens: 200_000,
        Custom:      map[string]float64{"cost": 0.50},
    }}),
)
if errors.Is(err, ai.ErrBudgetExceeded) {
    log.Println(resp.FinishReason) // aborted; resp.History() keeps the rounds that ran
}

Checked before each model call, so a run can go over by one turn. Subagents count toward the run. Scope: middlewarex.BudgetScopeSession caps the whole agent session (needs a session store). Interrupt: true pauses instead of stopping; claim with middlewarex.BudgetInterrupted(part) and restart with middlewarex.BudgetDecision{Approved: true}.

Every run reports usage: resp.Usage is the last model call, resp.TotalUsage sums every call in the run (nested ones included). Agents report TurnEnd.Usage, out.Usage, and out.State.Usage (session total).

Filesystem

Grants the model scoped file access under a single root directory via list_files, read_file, plus write_file and search_and_replace when writes are enabled. Hooks WrapGenerate and WrapTool and contributes Tools.

ai.WithUse(&middleware.Filesystem{
    RootDir:          "./workspace",
    AllowWriteAccess: true,
    ToolNamePrefix:   "",  // set distinct prefixes if attaching multiple Filesystem middlewares
})

Path safety is enforced by os.Root (Go 1.25+), which rejects any path resolving outside the root, including via .., absolute paths, or symlinks. read_file returns its content as a queued user message on the next turn (not as the tool's direct output) so binary types like images can be inlined as media parts.

Skills

Exposes a local library of SKILL.md files as loadable system instructions. Hooks WrapGenerate and contributes a use_skill tool.

ai.WithUse(&middleware.Skills{SkillPaths: []string{"skills"}}) // default: ["skills"]

A skill is a directory containing SKILL.md, optionally with YAML frontmatter (name, description). The middleware injects a system prompt listing available skills, and the model calls use_skill("name") to pull the skill body into the conversation on demand. Heavier persona instructions stay off the hot path until actually loaded.

Practical patterns

Streaming-aware middleware

If your WrapGenerate or WrapModel hook emits its own messages (injected user content, system updates), use the streaming callback and the MessageIndex cursor in GenerateParams:

WrapGenerate: func(ctx context.Context, p *ai.GenerateParams, next ai.GenerateNext) (*ai.ModelResponse, error) {
    if p.Callback != nil {
        _ = p.Callback(ctx, &ai.ModelResponseChunk{
            Role:    ai.RoleUser,
            Index:   p.MessageIndex,
            Content: []*ai.Part{ai.NewTextPart("[injected context]")},
        })
        p.MessageIndex++ // advance so downstream middleware and the model see the shifted index
    }
    p.Request.Messages = append(p.Request.Messages, ai.NewUserMessage(ai.NewTextPart("[injected context]")))
    return next(ctx, p)
},

Filesystem does this to deliver read_file content to the model while preserving streamed-chunk ordering.

Adding tools from middleware

Hooks.Tools registers extra tools for the duration of the call without the user wiring them through ai.WithTools. Useful when the middleware's hooks and tools work as a pair (e.g., Filesystem's read/write tools, Skills's use_skill tool):

return &ai.Hooks{
    Tools:    []ai.Tool{myTool},
    WrapTool: myInterceptor, // intercepts both myTool and any user-supplied tools
}, nil

Duplicate tool names across user-supplied tools and middleware-contributed tools error out at call setup; the call won't run.

Interrupting from WrapTool

A WrapTool hook pauses a call with tool.Interrupt(ctx, data) (from github.com/firebase/genkit/go/ai/tool), reads its answer on the restart with tool.ResumeData[T](ctx), and skips calls it already let through with tool.Released(ctx). (ai.NewToolInterruptError was removed in v1.14.0.)

WrapTool: func(ctx context.Context, p *ai.ToolParams, next ai.ToolNext) (*ai.MultipartToolResponse, error) {
    if shouldGate(p.Tool.Name()) && !tool.Released(ctx) {
        if d, ok := tool.ResumeData[MyDecision](ctx); ok {
            if !d.Approved {
                return nil, errors.New("declined")
            }
            return next(ctx, p)
        }
        return nil, tool.Interrupt(ctx, map[string]any{"message": "needs approval"})
    }
    return next(ctx, p)
},

Export a typed claim for your hold by wrapping ai.MiddlewareInterrupted[Res](name, part), as ToolApprovalInterrupted does. ToolApproval uses this pattern.

Modifying the request safely

p.Request is the live request for the iteration. Mutating it in place affects later layers. If the change should be visible only to the inner layer, copy first:

WrapModel: func(ctx context.Context, p *ai.ModelParams, next ai.ModelNext) (*ai.ModelResponse, error) {
    req := *p.Request
    req.Messages = append([]*ai.Message(nil), p.Request.Messages...)
    req.Messages = append(req.Messages, extraSystemMessage)
    p.Request = &req
    return next(ctx, p)
},

Skills.injectSkillsPrompt shows the same pattern for ModelRequest cloning.

Idempotent re-injection across iterations

WrapGenerate runs once per tool-loop iteration. If you inject content into the request, you'll inject it on every iteration unless you mark and detect what you've already added. Skills tags its system prompt part with metadata (skills-instructions: true) and refreshes that one part in place rather than appending a new one each turn.

Migration note

The legacy ai.ModelMiddleware / ai.WithMiddleware API is preserved and marked deprecated. Prefer ai.Middleware / ai.WithUse, which adds WrapGenerate and WrapTool hooks plus Hooks.Tools for dynamically injected tools.

See also

• tools.md for tool definition, interrupt/restart machinery used by ToolApproval.

• Sample sources under go/samples/basic-middleware/: retry-fallback, filesystem, skills.

• The plugins/middleware package source for reference implementations.