Require secrets for ProviderError redaction and omit unredactable bodies

NewProviderError now takes the secrets to redact as a required []string
instead of an optional variadic. When no non-empty secret is given the
body cannot be redacted, so no excerpt is stored and Body()/Error() never
expose raw vendor output.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

internal/ai/llm/errors.go

@@ -0,0 +1,103 @@
+package llm
+
+import (
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+	"unicode/utf8"
+)
+
+var (
+	// ErrAuth means the vendor rejected the credentials.
+	ErrAuth = errors.New("llm: authentication failed")
+	// ErrRateLimited means the vendor throttled the request. The concrete
+	// error is a *RateLimitedError carrying the requested wait.
+	ErrRateLimited = errors.New("llm: rate limited")
+	// ErrTimeout means the request did not finish within its timeout.
+	ErrTimeout = errors.New("llm: request timed out")
+	// ErrBadOutput means the model replied but the reply is unusable, e.g.
+	// structured output that does not parse.
+	ErrBadOutput = errors.New("llm: unusable model output")
+	// ErrProvider means the vendor returned an unexpected HTTP status. The
+	// concrete error is a *ProviderError carrying the status and a redacted
+	// body excerpt.
+	ErrProvider = errors.New("llm: provider error")
+)
+
+// RateLimitedError is a throttled request. It matches ErrRateLimited.
+type RateLimitedError struct {
+	// RetryAfter is how long the vendor asked callers to wait; 0 means it
+	// did not say.
+	RetryAfter time.Duration
+}
+
+func (e *RateLimitedError) Error() string {
+	if e.RetryAfter == 0 {
+		return ErrRateLimited.Error()
+	}
+	return fmt.Sprintf("%s: retry after %v", ErrRateLimited, e.RetryAfter)
+}
+
+// Is reports whether target is ErrRateLimited.
+func (e *RateLimitedError) Is(target error) bool { return target == ErrRateLimited }
+
+// MaxBodyExcerpt is the longest response-body excerpt, in bytes, a
+// ProviderError keeps.
+const MaxBodyExcerpt = 512
+
+// redacted replaces each secret in a ProviderError body.
+const redacted = "[REDACTED]"
+
+// ProviderError is an unexpected HTTP status from the vendor. It matches
+// ErrProvider. Build it with NewProviderError, which redacts the body.
+type ProviderError struct {
+	status int
+	body   string
+}
+
+// NewProviderError returns the error for an unexpected HTTP status.
+// secrets must hold the API key (and any other credential sent with the
+// request). Every occurrence of each non-empty secret in body is replaced
+// before the body is cut to MaxBodyExcerpt bytes, so no part of a secret
+// survives. If secrets holds no non-empty value the body cannot be
+// redacted, so no excerpt is kept at all.
+func NewProviderError(status int, body []byte, secrets []string) *ProviderError {
+	s := string(body)
+	redactedAny := false
+	for _, secret := range secrets {
+		if secret != "" {
+			s = strings.ReplaceAll(s, secret, redacted)
+			redactedAny = true
+		}
+	}
+	if !redactedAny {
+		return &ProviderError{status: status}
+	}
+	s = strings.TrimSpace(s)
+	if len(s) > MaxBodyExcerpt {
+		cut := MaxBodyExcerpt
+		for cut > 0 && !utf8.RuneStart(s[cut]) {
+			cut--
+		}
+		s = s[:cut] + "…"
+	}
+	return &ProviderError{status: status, body: s}
+}
+
+// Status returns the HTTP status code.
+func (e *ProviderError) Status() int { return e.status }
+
+// Body returns the redacted, truncated response-body excerpt; it is empty
+// when the vendor sent no body or no secret was given to redact it with.
+func (e *ProviderError) Body() string { return e.body }
+
+func (e *ProviderError) Error() string {
+	if e.body == "" {
+		return fmt.Sprintf("%s: HTTP %d", ErrProvider, e.status)
+	}
+	return fmt.Sprintf("%s: HTTP %d: %s", ErrProvider, e.status, e.body)
+}
+
+// Is reports whether target is ErrProvider.
+func (e *ProviderError) Is(target error) bool { return target == ErrProvider }

internal/ai/llm/errors_test.go

@@ -0,0 +1,121 @@
+package llm
+
+import (
+	"errors"
+	"fmt"
+	"strings"
+	"testing"
+	"time"
+)
+
+func TestSentinelErrorsMatchWhenWrapped(t *testing.T) {
+	for _, sentinel := range []error{ErrAuth, ErrTimeout, ErrBadOutput} {
+		wrapped := fmt.Errorf("openai: generate: %w", sentinel)
+		if !errors.Is(wrapped, sentinel) {
+			t.Errorf("errors.Is(%v, %v) = false", wrapped, sentinel)
+		}
+		for _, other := range []error{ErrAuth, ErrTimeout, ErrBadOutput, ErrRateLimited, ErrProvider} {
+			if other != sentinel && errors.Is(wrapped, other) {
+				t.Errorf("errors.Is(%v, %v) = true", wrapped, other)
+			}
+		}
+	}
+}
+
+func TestRateLimitedError(t *testing.T) {
+	err := fmt.Errorf("anthropic: %w", &RateLimitedError{RetryAfter: 30 * time.Second})
+	if !errors.Is(err, ErrRateLimited) {
+		t.Fatal("errors.Is(err, ErrRateLimited) = false")
+	}
+	if errors.Is(err, ErrProvider) || errors.Is(err, ErrAuth) {
+		t.Fatal("rate-limited error matches an unrelated sentinel")
+	}
+	var rl *RateLimitedError
+	if !errors.As(err, &rl) {
+		t.Fatal("errors.As(err, *RateLimitedError) = false")
+	}
+	if rl.RetryAfter != 30*time.Second {
+		t.Fatalf("RetryAfter = %v, want 30s", rl.RetryAfter)
+	}
+	if got := err.Error(); got != "anthropic: llm: rate limited: retry after 30s" {
+		t.Fatalf("Error() = %q", got)
+	}
+	if got := (&RateLimitedError{}).Error(); got != "llm: rate limited" {
+		t.Fatalf("Error() without RetryAfter = %q", got)
+	}
+}
+
+func TestProviderError(t *testing.T) {
+	err := fmt.Errorf("gemini: %w", NewProviderError(503, []byte(` {"error":"overloaded"} `), []string{"sk-test"}))
+	if !errors.Is(err, ErrProvider) {
+		t.Fatal("errors.Is(err, ErrProvider) = false")
+	}
+	if errors.Is(err, ErrRateLimited) || errors.Is(err, ErrAuth) {
+		t.Fatal("provider error matches an unrelated sentinel")
+	}
+	var pe *ProviderError
+	if !errors.As(err, &pe) {
+		t.Fatal("errors.As(err, *ProviderError) = false")
+	}
+	if pe.Status() != 503 || pe.Body() != `{"error":"overloaded"}` {
+		t.Fatalf("Status, Body = %d, %q", pe.Status(), pe.Body())
+	}
+	if got := err.Error(); got != `gemini: llm: provider error: HTTP 503: {"error":"overloaded"}` {
+		t.Fatalf("Error() = %q", got)
+	}
+	if got := NewProviderError(500, nil, []string{"sk-test"}).Error(); got != "llm: provider error: HTTP 500" {
+		t.Fatalf("Error() without body = %q", got)
+	}
+}
+
+func TestProviderErrorRedactsSecrets(t *testing.T) {
+	const key = "sk-live-0123456789abcdef"
+	body := []byte(`{"error":"invalid key sk-live-0123456789abcdef","echo":"Bearer sk-live-0123456789abcdef"}`)
+	pe := NewProviderError(400, body, []string{key, ""})
+	for _, s := range []string{pe.Error(), pe.Body()} {
+		if strings.Contains(s, key) {
+			t.Fatalf("secret leaked: %q", s)
+		}
+	}
+	want := `{"error":"invalid key [REDACTED]","echo":"Bearer [REDACTED]"}`
+	if pe.Body() != want {
+		t.Fatalf("Body() = %q, want %q", pe.Body(), want)
+	}
+}
+
+func TestProviderErrorRedactsBeforeTruncating(t *testing.T) {
+	const key = "sk-secret-straddling-the-cut"
+	// The key starts just before the excerpt limit, so truncating first
+	// would leave its prefix in the excerpt.
+	body := strings.Repeat("x", MaxBodyExcerpt-5) + key
+	pe := NewProviderError(500, []byte(body), []string{key})
+	if strings.Contains(pe.Body(), "sk-se") {
+		t.Fatalf("secret prefix leaked: %q", pe.Body())
+	}
+}
+
+func TestProviderErrorTruncatesOnRuneBoundary(t *testing.T) {
+	body := strings.Repeat("a", MaxBodyExcerpt-1) + "é" + strings.Repeat("b", 100)
+	pe := NewProviderError(500, []byte(body), []string{"sk-test"})
+	want := strings.Repeat("a", MaxBodyExcerpt-1) + "…"
+	if pe.Body() != want {
+		t.Fatalf("Body() has %d bytes, want %d", len(pe.Body()), len(want))
+	}
+}
+
+func TestProviderErrorOmitsBodyWithoutSecret(t *testing.T) {
+	const key = "sk-live-0123456789abcdef"
+	body := []byte(`{"error":"invalid key sk-live-0123456789abcdef"}`)
+	for _, secrets := range [][]string{nil, {}, {""}} {
+		pe := NewProviderError(401, body, secrets)
+		if pe.Body() != "" {
+			t.Fatalf("secrets %q: Body() = %q, want empty", secrets, pe.Body())
+		}
+		if got := pe.Error(); got != "llm: provider error: HTTP 401" {
+			t.Fatalf("secrets %q: Error() = %q", secrets, got)
+		}
+		if strings.Contains(pe.Error(), key) {
+			t.Fatalf("secrets %q: secret leaked: %q", secrets, pe.Error())
+		}
+	}
+}

internal/ai/llm/fake.go

@@ -0,0 +1,127 @@
+package llm
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"slices"
+	"sync"
+)
+
+// Fake is a scripted Provider for tests. Each method returns its queued
+// results in order, one per call, and records every Request Generate
+// receives. A call with nothing left in its queue fails, so an unexpected
+// call is never silently answered. It is safe for concurrent use.
+type Fake struct {
+	mu       sync.Mutex
+	generate []GenerateResult
+	models   []ListModelsResult
+	verify   []error
+	requests []Request
+}
+
+var _ Provider = (*Fake)(nil)
+
+// GenerateResult is one scripted Generate outcome.
+type GenerateResult struct {
+	Response Response
+	Err      error
+}
+
+// ListModelsResult is one scripted ListModels outcome.
+type ListModelsResult struct {
+	Models []Model
+	Err    error
+}
+
+// QueueGenerate appends results to the Generate script.
+func (f *Fake) QueueGenerate(results ...GenerateResult) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.generate = append(f.generate, results...)
+}
+
+// QueueListModels appends results to the ListModels script.
+func (f *Fake) QueueListModels(results ...ListModelsResult) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.models = append(f.models, results...)
+}
+
+// QueueVerify appends results to the Verify script; a nil entry is success.
+func (f *Fake) QueueVerify(results ...error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.verify = append(f.verify, results...)
+}
+
+// Generate records req and returns the next scripted result.
+func (f *Fake) Generate(_ context.Context, req Request) (Response, error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.requests = append(f.requests, cloneRequest(req))
+	if len(f.generate) == 0 {
+		return Response{}, fmt.Errorf("llm: fake: unscripted Generate call %d", len(f.requests))
+	}
+	r := f.generate[0]
+	f.generate = f.generate[1:]
+	return r.Response, r.Err
+}
+
+// ListModels returns the next scripted result.
+func (f *Fake) ListModels(context.Context) ([]Model, error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if len(f.models) == 0 {
+		return nil, fmt.Errorf("llm: fake: unscripted ListModels call")
+	}
+	r := f.models[0]
+	f.models = f.models[1:]
+	return slices.Clone(r.Models), r.Err
+}
+
+// Verify returns the next scripted result.
+func (f *Fake) Verify(context.Context) error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if len(f.verify) == 0 {
+		return fmt.Errorf("llm: fake: unscripted Verify call")
+	}
+	err := f.verify[0]
+	f.verify = f.verify[1:]
+	return err
+}
+
+// Requests returns a copy of every Request Generate received, in order.
+func (f *Fake) Requests() []Request {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	out := make([]Request, len(f.requests))
+	for i, r := range f.requests {
+		out[i] = cloneRequest(r)
+	}
+	return out
+}
+
+// Remaining reports how many scripted results of each kind are unused, so a
+// test can assert its script was consumed.
+func (f *Fake) Remaining() (generate, listModels, verify int) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return len(f.generate), len(f.models), len(f.verify)
+}
+
+// cloneRequest deep-copies r so a recorded Request is unaffected by later
+// changes the caller makes to its slices and pointers.
+func cloneRequest(r Request) Request {
+	r.Messages = slices.Clone(r.Messages)
+	if r.Temperature != nil {
+		t := *r.Temperature
+		r.Temperature = &t
+	}
+	if r.Schema != nil {
+		s := JSONSchema{Name: r.Schema.Name, Schema: json.RawMessage(slices.Clone([]byte(r.Schema.Schema)))}
+		r.Schema = &s
+	}
+	return r
+}

internal/ai/llm/fake_test.go

@@ -0,0 +1,138 @@
+package llm
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"sync"
+	"testing"
+	"time"
+)
+
+func TestFakeGenerateReturnsScriptInOrder(t *testing.T) {
+	ctx := context.Background()
+	var f Fake
+	f.QueueGenerate(
+		GenerateResult{Response: Response{Text: "first", Model: "m1"}},
+		GenerateResult{Err: &RateLimitedError{RetryAfter: time.Second}},
+	)
+	f.QueueGenerate(GenerateResult{Response: Response{JSON: json.RawMessage(`{"ok":true}`), FinishReason: "stop"}})
+
+	req := Request{Messages: []Message{{Role: RoleUser, Content: "one"}}}
+	resp, err := f.Generate(ctx, req)
+	if err != nil || resp.Text != "first" || resp.Model != "m1" {
+		t.Fatalf("call 1 = %+v, %v", resp, err)
+	}
+	if _, err := f.Generate(ctx, req); !errors.Is(err, ErrRateLimited) {
+		t.Fatalf("call 2 err = %v, want ErrRateLimited", err)
+	}
+	resp, err = f.Generate(ctx, req)
+	if err != nil || string(resp.JSON) != `{"ok":true}` || resp.FinishReason != "stop" {
+		t.Fatalf("call 3 = %+v, %v", resp, err)
+	}
+	if _, err := f.Generate(ctx, req); err == nil {
+		t.Fatal("unscripted call 4 succeeded")
+	}
+	if g, l, v := f.Remaining(); g != 0 || l != 0 || v != 0 {
+		t.Fatalf("Remaining() = %d, %d, %d", g, l, v)
+	}
+	if n := len(f.Requests()); n != 4 {
+		t.Fatalf("recorded %d requests, want 4 (unscripted calls are recorded too)", n)
+	}
+}
+
+func TestFakeRecordsRequests(t *testing.T) {
+	var f Fake
+	f.QueueGenerate(GenerateResult{}, GenerateResult{})
+	temp := 0.2
+	first := Request{
+		System:      "be brief",
+		Messages:    []Message{{Role: RoleUser, Content: "plan dinners"}},
+		Temperature: &temp,
+		MaxTokens:   100,
+		Schema:      &JSONSchema{Name: "plan", Schema: json.RawMessage(`{"type":"object"}`)},
+		Timeout:     5 * time.Second,
+	}
+	if _, err := f.Generate(context.Background(), first); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := f.Generate(context.Background(), Request{System: "second"}); err != nil {
+		t.Fatal(err)
+	}
+
+	// Later changes by the caller must not alter what was recorded.
+	first.Messages[0].Content = "changed"
+	temp = 1
+	first.Schema.Schema[2] = 'X'
+
+	got := f.Requests()
+	if len(got) != 2 || got[1].System != "second" {
+		t.Fatalf("Requests() = %+v", got)
+	}
+	r := got[0]
+	if r.System != "be brief" || r.Messages[0].Content != "plan dinners" || *r.Temperature != 0.2 ||
+		r.MaxTokens != 100 || r.Schema.Name != "plan" || string(r.Schema.Schema) != `{"type":"object"}` ||
+		r.Timeout != 5*time.Second {
+		t.Fatalf("recorded request = %+v", r)
+	}
+
+	// Changing a returned copy must not alter the record either.
+	got[0].Messages[0].Content = "mutated"
+	if f.Requests()[0].Messages[0].Content != "plan dinners" {
+		t.Fatal("Requests() returned shared state")
+	}
+}
+
+func TestFakeListModelsAndVerify(t *testing.T) {
+	ctx := context.Background()
+	var f Fake
+	f.QueueListModels(
+		ListModelsResult{Models: []Model{{ID: "a", DisplayName: "A"}, {ID: "b", DisplayName: "b"}}},
+		ListModelsResult{Err: ErrAuth},
+	)
+	f.QueueVerify(nil, fmt.Errorf("wrapped: %w", ErrAuth))
+
+	models, err := f.ListModels(ctx)
+	if err != nil || len(models) != 2 || models[0].ID != "a" {
+		t.Fatalf("ListModels 1 = %+v, %v", models, err)
+	}
+	if _, err := f.ListModels(ctx); !errors.Is(err, ErrAuth) {
+		t.Fatalf("ListModels 2 err = %v, want ErrAuth", err)
+	}
+	if _, err := f.ListModels(ctx); err == nil {
+		t.Fatal("unscripted ListModels succeeded")
+	}
+	if err := f.Verify(ctx); err != nil {
+		t.Fatalf("Verify 1 = %v", err)
+	}
+	if err := f.Verify(ctx); !errors.Is(err, ErrAuth) {
+		t.Fatalf("Verify 2 = %v, want ErrAuth", err)
+	}
+	if err := f.Verify(ctx); err == nil {
+		t.Fatal("unscripted Verify succeeded")
+	}
+}
+
+func TestFakeConcurrentUse(t *testing.T) {
+	const n = 50
+	var f Fake
+	for i := range n {
+		f.QueueGenerate(GenerateResult{Response: Response{Text: fmt.Sprint(i)}})
+	}
+	var wg sync.WaitGroup
+	for range n {
+		wg.Go(func() {
+			if _, err := f.Generate(context.Background(), Request{}); err != nil {
+				t.Error(err)
+			}
+		})
+	}
+	wg.Wait()
+	if len(f.Requests()) != n {
+		t.Fatalf("recorded %d requests, want %d", len(f.Requests()), n)
+	}
+	if g, _, _ := f.Remaining(); g != 0 {
+		t.Fatalf("%d Generate results unused", g)
+	}
+}

internal/ai/llm/llm.go

@@ -0,0 +1,132 @@
+// Package llm is the single interface MealQuest uses to talk to a language
+// model. Every vendor adapter implements Provider over plain net/http (no
+// vendor SDKs) and reports failures with this package's typed errors, so
+// callers decide what to do from errors.Is and errors.As alone.
+//
+// Errors from this package and its adapters never contain an API key.
+package llm
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"time"
+)
+
+// Provider is a language-model vendor.
+type Provider interface {
+	// Generate sends one request and returns the model's reply.
+	Generate(ctx context.Context, req Request) (Response, error)
+	// ListModels returns the models the configured credentials can use.
+	ListModels(ctx context.Context) ([]Model, error)
+	// Verify checks that the configured credentials are accepted.
+	Verify(ctx context.Context) error
+}
+
+// Role is the author of a Message.
+type Role string
+
+// The roles a Message may have. The system prompt is Request.System, not a
+// message.
+const (
+	RoleUser      Role = "user"
+	RoleAssistant Role = "assistant"
+)
+
+// Message is one turn of a conversation.
+type Message struct {
+	Role    Role
+	Content string
+}
+
+// JSONSchema is a named structured-output contract: when a Request carries
+// one, the model must reply with JSON matching Schema.
+type JSONSchema struct {
+	// Name identifies the contract to the vendor, e.g. "side_quests".
+	Name string
+	// Schema is a JSON Schema object.
+	Schema json.RawMessage
+}
+
+// Request is one generation request.
+type Request struct {
+	// System is the system prompt; empty means none.
+	System string
+	// Messages is the conversation, oldest first. It must not be empty.
+	Messages []Message
+	// Temperature is the sampling temperature; nil means the vendor default.
+	Temperature *float64
+	// MaxTokens caps the reply length; 0 means the vendor default.
+	MaxTokens int
+	// Schema, when set, requests structured output; the reply is then in
+	// Response.JSON.
+	Schema *JSONSchema
+	// Timeout bounds the request; 0 means only ctx bounds it.
+	Timeout time.Duration
+}
+
+// Validate reports whether r is well formed. Adapters call it before
+// contacting the vendor.
+func (r Request) Validate() error {
+	if len(r.Messages) == 0 {
+		return errors.New("llm: request has no messages")
+	}
+	for i, m := range r.Messages {
+		if m.Role != RoleUser && m.Role != RoleAssistant {
+			return fmt.Errorf("llm: message %d has unknown role %q", i, m.Role)
+		}
+	}
+	if r.Temperature != nil && *r.Temperature < 0 {
+		return fmt.Errorf("llm: temperature %v is negative", *r.Temperature)
+	}
+	if r.MaxTokens < 0 {
+		return fmt.Errorf("llm: max tokens %d is negative", r.MaxTokens)
+	}
+	if r.Timeout < 0 {
+		return fmt.Errorf("llm: timeout %v is negative", r.Timeout)
+	}
+	if r.Schema != nil {
+		if r.Schema.Name == "" {
+			return errors.New("llm: schema has no name")
+		}
+		var obj map[string]json.RawMessage
+		if err := json.Unmarshal(r.Schema.Schema, &obj); err != nil || obj == nil {
+			return fmt.Errorf("llm: schema %q is not a JSON object", r.Schema.Name)
+		}
+	}
+	return nil
+}
+
+// Usage is the token count of one Generate call.
+type Usage struct {
+	InputTokens  int
+	OutputTokens int
+}
+
+// Total returns InputTokens + OutputTokens.
+func (u Usage) Total() int { return u.InputTokens + u.OutputTokens }
+
+// Response is the model's reply to a Request.
+type Response struct {
+	// Text is the reply text.
+	Text string
+	// JSON is the structured reply when the Request carried a Schema, and
+	// nil otherwise.
+	JSON json.RawMessage
+	// Usage is the tokens the call consumed.
+	Usage Usage
+	// Model is the model that answered, as the vendor reports it.
+	Model string
+	// FinishReason is why generation stopped, as the vendor reports it.
+	FinishReason string
+}
+
+// Model is a model a Provider offers.
+type Model struct {
+	// ID is the vendor's model identifier, used in requests.
+	ID string
+	// DisplayName is a human-readable name; it equals ID when the vendor
+	// gives none.
+	DisplayName string
+}

internal/ai/llm/llm_test.go

@@ -0,0 +1,44 @@
+package llm
+
+import (
+	"encoding/json"
+	"testing"
+)
+
+func TestRequestValidate(t *testing.T) {
+	neg := -0.1
+	valid := func() Request {
+		return Request{
+			Messages: []Message{{Role: RoleUser, Content: "hi"}, {Role: RoleAssistant, Content: "hello"}},
+			Schema:   &JSONSchema{Name: "plan", Schema: json.RawMessage(`{"type":"object"}`)},
+		}
+	}
+	if err := valid().Validate(); err != nil {
+		t.Fatalf("valid request: %v", err)
+	}
+	cases := map[string]func(*Request){
+		"no messages":    func(r *Request) { r.Messages = nil },
+		"unknown role":   func(r *Request) { r.Messages[0].Role = "system" },
+		"negative temp":  func(r *Request) { r.Temperature = &neg },
+		"negative max":   func(r *Request) { r.MaxTokens = -1 },
+		"negative time":  func(r *Request) { r.Timeout = -1 },
+		"unnamed schema": func(r *Request) { r.Schema.Name = "" },
+		"schema array":   func(r *Request) { r.Schema.Schema = json.RawMessage(`[]`) },
+		"schema null":    func(r *Request) { r.Schema.Schema = json.RawMessage(`null`) },
+		"schema invalid": func(r *Request) { r.Schema.Schema = json.RawMessage(`{`) },
+		"schema empty":   func(r *Request) { r.Schema.Schema = nil },
+	}
+	for name, mutate := range cases {
+		r := valid()
+		mutate(&r)
+		if err := r.Validate(); err == nil {
+			t.Errorf("%s: Validate() = nil", name)
+		}
+	}
+}
+
+func TestUsageTotal(t *testing.T) {
+	if got := (Usage{InputTokens: 12, OutputTokens: 30}).Total(); got != 42 {
+		t.Fatalf("Total() = %d, want 42", got)
+	}
+}

internal/jobs/jobs.go

@@ -68,6 +68,14 @@ var codePattern = regexp.MustCompile(`^[a-z][a-z0-9_]{0,63}$`)
 // ErrNotFound is returned by Get for an unknown job id.
 var ErrNotFound = errors.New("jobs: job not found")
 
+// ErrUnknownKind is returned by Runner.Enqueue for a kind with no
+// registered handler.
+var ErrUnknownKind = errors.New("jobs: unknown job kind")
+
+// ErrInvalidPayload is returned by Runner.Enqueue for a payload that does
+// not decode into the kind's payload type.
+var ErrInvalidPayload = errors.New("jobs: payload does not decode")
+
 // Failure is a handler error reporting a failure code the household may
 // see. Code must be one the job's kind declared at registration; the runner
 // stores it as the job's error. Any other handler error, including a Failure
@@ -106,6 +114,13 @@ type Job struct {
 // never quotes an error.
 type runFunc func(ctx context.Context, payload string, publish func(json.RawMessage) error) (result json.RawMessage, code, reason string)
 
+// handler is a registered kind: run runs a job, and check reports whether
+// a payload decodes into the kind's payload type.
+type handler struct {
+	run   runFunc
+	check func(payload []byte) bool
+}
+
 // Runner owns the job handlers and the worker pool.
 type Runner struct {
 	st      *store.Store
@@ -118,7 +133,7 @@ type Runner struct {
 	claim func(ctx context.Context, now sql.NullString) (db.Job, error)
 
 	mu       sync.Mutex
-	handlers map[string]runFunc
+	handlers map[string]handler
 	started  bool
 
 	wg sync.WaitGroup
@@ -141,7 +156,7 @@ func NewRunner(st *store.Store, logger *slog.Logger, workers int) (*Runner, erro
 		workers:  workers,
 		wake:     make(chan struct{}, workers),
 		claim:    db.New(st.Writer).ClaimNextJob,
-		handlers: make(map[string]runFunc),
+		handlers: make(map[string]handler),
 	}, nil
 }
 
@@ -179,11 +194,9 @@ func Register[P any, R Public](r *Runner, kind string, h Handler[P, R], failureC
 	if _, ok := r.handlers[kind]; ok {
 		return nil, fmt.Errorf("jobs: register %q: kind already registered", kind)
 	}
-	r.handlers[kind] = func(ctx context.Context, payload string, publish func(json.RawMessage) error) (json.RawMessage, string, string) {
-		var p P
-		dec := json.NewDecoder(bytes.NewReader([]byte(payload)))
-		dec.DisallowUnknownFields()
-		if err := dec.Decode(&p); err != nil {
+	run := func(ctx context.Context, payload string, publish func(json.RawMessage) error) (json.RawMessage, string, string) {
+		p, err := decodePayload[P]([]byte(payload))
+		if err != nil {
 			return nil, CodeInternal, "payload does not decode"
 		}
 		result, err := h(ctx, p, func(partial R) error {
@@ -209,9 +222,23 @@ func Register[P any, R Public](r *Runner, kind string, h Handler[P, R], failureC
 		}
 		return data, "", ""
 	}
+	check := func(payload []byte) bool {
+		_, err := decodePayload[P](payload)
+		return err == nil
+	}
+	r.handlers[kind] = handler{run: run, check: check}
 	return &Kind[P]{r: r, name: kind}, nil
 }
 
+// decodePayload decodes payload into P, refusing unknown fields.
+func decodePayload[P any](payload []byte) (P, error) {
+	var p P
+	dec := json.NewDecoder(bytes.NewReader(payload))
+	dec.DisallowUnknownFields()
+	err := dec.Decode(&p)
+	return p, err
+}
+
 // publicJSON returns v's PublicJSON, refusing anything that is not one JSON
 // value. Its error never quotes PublicJSON's error.
 func publicJSON(v Public) (json.RawMessage, error) {
@@ -229,19 +256,43 @@ func (k *Kind[P]) Enqueue(ctx context.Context, payload P) (string, error) {
 	if err != nil {
 		return "", fmt.Errorf("jobs: enqueue %s: encode payload: %w", k.name, err)
 	}
+	return k.r.enqueue(ctx, k.name, data)
+}
+
+// Enqueue stores a queued job of kind with its JSON payload and returns its
+// id, as Kind.Enqueue does for a kind known at compile time. It returns
+// ErrUnknownKind for a kind with no registered handler and
+// ErrInvalidPayload for a payload that does not decode into the kind's
+// payload type; neither stores a job.
+func (r *Runner) Enqueue(ctx context.Context, kind string, payload json.RawMessage) (string, error) {
+	r.mu.Lock()
+	h, ok := r.handlers[kind]
+	r.mu.Unlock()
+	if !ok {
+		return "", ErrUnknownKind
+	}
+	if !h.check(payload) {
+		return "", ErrInvalidPayload
+	}
+	return r.enqueue(ctx, kind, payload)
+}
+
+// enqueue stores a queued job of kind with its encoded payload and wakes a
+// worker.
+func (r *Runner) enqueue(ctx context.Context, kind string, payload []byte) (string, error) {
 	id := store.NewID()
 	now := store.Now().String()
-	if err := db.New(k.r.st.Writer).CreateJob(ctx, db.CreateJobParams{
+	if err := db.New(r.st.Writer).CreateJob(ctx, db.CreateJobParams{
 		ID:        id,
-		Kind:      k.name,
-		Payload:   string(data),
+		Kind:      kind,
+		Payload:   string(payload),
 		CreatedAt: now,
 		UpdatedAt: now,
 	}); err != nil {
-		return "", fmt.Errorf("jobs: enqueue %s: %w", k.name, err)
+		return "", fmt.Errorf("jobs: enqueue %s: %w", kind, err)
 	}
 	select {
-	case k.r.wake <- struct{}{}:
+	case r.wake <- struct{}{}:
 	default: // every worker already has a token and will look again
 	}
 	return id, nil
@@ -365,7 +416,7 @@ func (r *Runner) run(ctx context.Context, job db.Job) {
 		return nil
 	}
 
-	result, code, reason := h(ctx, job.Payload, publish)
+	result, code, reason := h.run(ctx, job.Payload, publish)
 	if code != "" {
 		if ctx.Err() != nil {
 			logger.WarnContext(write, "job interrupted by shutdown; it fails at the next start")

internal/jobs/jobs_test.go

@@ -568,3 +568,39 @@ func TestClaimErrorIsRetried(t *testing.T) {
 		t.Fatalf("claim error text logged: %s", out)
 	}
 }
+
+func TestRunnerEnqueueByKindName(t *testing.T) {
+	t.Parallel()
+	st := migratedStore(t)
+	r := newRunner(t, st, 1)
+	register(t, r, "count", func(_ context.Context, p countPayload, _ func(countResult) error) (countResult, error) {
+		return countResult{Done: p.Steps}, nil
+	})
+	ctx := context.Background()
+
+	if _, err := r.Enqueue(ctx, "no_such_kind", json.RawMessage(`{}`)); !errors.Is(err, ErrUnknownKind) {
+		t.Errorf("unknown kind error = %v, want ErrUnknownKind", err)
+	}
+	for _, payload := range []string{`{"steps":"three"}`, `{"other":1}`, `[]`, ``} {
+		if _, err := r.Enqueue(ctx, "count", json.RawMessage(payload)); !errors.Is(err, ErrInvalidPayload) {
+			t.Errorf("payload %q error = %v, want ErrInvalidPayload", payload, err)
+		}
+	}
+	var n int
+	if err := st.Reader.QueryRow("SELECT COUNT(*) FROM jobs").Scan(&n); err != nil {
+		t.Fatal(err)
+	}
+	if n != 0 {
+		t.Errorf("refused enqueues stored %d jobs, want 0", n)
+	}
+
+	id, err := r.Enqueue(ctx, "count", json.RawMessage(`{"steps":3}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	start(t, r)
+	job := waitFor(t, r, id, finished)
+	if job.Kind != "count" || job.Status != StatusSucceeded || string(job.Result) != `{"done":3}` {
+		t.Errorf("job = %+v (result %s), want succeeded with {\"done\":3}", job, job.Result)
+	}
+}

internal/server/jobs.go

@@ -1,7 +1,11 @@
 package server
 
 import (
+	"bytes"
+	"encoding/json"
 	"errors"
+	"fmt"
+	"io"
 	"log/slog"
 	"net/http"
 
@@ -10,6 +14,80 @@ import (
 	"code.cubecraftlabs.com/Maniacs_Mansion/Meal-Quest/internal/jobs"
 )
 
+// maxJobBody bounds a POST /jobs request body.
+const maxJobBody = 64 << 10
+
+// jobRequest is the POST /jobs body. Payload is required; a job kind that
+// takes no input is sent the empty object {}.
+type jobRequest struct {
+	Kind    string          `json:"kind"`
+	Payload json.RawMessage `json:"payload"`
+}
+
+// jobCreated is the POST /jobs response.
+type jobCreated struct {
+	ID string `json:"id"`
+}
+
+// postJob serves POST /api/v1/jobs: it enqueues a job of a registered kind
+// and returns 202 with its id, which GET /api/v1/jobs/{id} reports on.
+func postJob(logger *slog.Logger, runner *jobs.Runner) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		req, err := decodeJobRequest(http.MaxBytesReader(w, r.Body, maxJobBody))
+		if err != nil {
+			writeError(w, r, logger, http.StatusBadRequest, "invalid_request", err.Error())
+			return
+		}
+		id, err := runner.Enqueue(r.Context(), req.Kind, req.Payload)
+		switch {
+		case errors.Is(err, jobs.ErrUnknownKind):
+			writeError(w, r, logger, http.StatusBadRequest, "unknown_kind", "unknown job kind")
+			return
+		case errors.Is(err, jobs.ErrInvalidPayload):
+			writeError(w, r, logger, http.StatusBadRequest, "invalid_payload", "payload does not match the job kind")
+			return
+		case err != nil:
+			logger.ErrorContext(r.Context(), "enqueue job", "kind", req.Kind, "error", err)
+			writeError(w, r, logger, http.StatusInternalServerError, "internal", "internal error")
+			return
+		}
+		writeJSON(w, r, logger, http.StatusAccepted, jobCreated{ID: id})
+	}
+}
+
+// decodeJobRequest reads a single JSON object with a non-empty kind and a
+// payload, refusing unknown fields, an omitted or null payload and trailing
+// data.
+func decodeJobRequest(body io.Reader) (jobRequest, error) {
+	data, err := io.ReadAll(body)
+	if err != nil {
+		var tooLarge *http.MaxBytesError
+		if errors.As(err, &tooLarge) {
+			return jobRequest{}, fmt.Errorf("request body exceeds %d bytes", tooLarge.Limit)
+		}
+		return jobRequest{}, errors.New("cannot read request body")
+	}
+	var req jobRequest
+	dec := json.NewDecoder(bytes.NewReader(data))
+	dec.DisallowUnknownFields()
+	if err := dec.Decode(&req); err != nil {
+		return jobRequest{}, errors.New("request body must be a JSON object with kind and payload")
+	}
+	if _, err := dec.Token(); err != io.EOF {
+		return jobRequest{}, errors.New("request body must be a single JSON object")
+	}
+	if req.Kind == "" {
+		return jobRequest{}, errors.New("kind: required")
+	}
+	if req.Payload == nil {
+		return jobRequest{}, errors.New("payload: required")
+	}
+	if bytes.Equal(req.Payload, []byte("null")) {
+		return jobRequest{}, errors.New("payload: must not be null")
+	}
+	return req, nil
+}
+
 // getJob serves GET /api/v1/jobs/{id}: the job's id, kind, status, latest
 // result and error.
 func getJob(logger *slog.Logger, runner *jobs.Runner) http.HandlerFunc {

internal/server/jobs_test.go

@@ -7,6 +7,7 @@ import (
 	"net/http"
 	"net/http/httptest"
 	"slices"
+	"strings"
 	"testing"
 	"time"
 
@@ -151,3 +152,95 @@ func TestGetJobRequiresSession(t *testing.T) {
 		t.Errorf("status = %d, want 401", rec.Code)
 	}
 }
+
+// TestPostJobThenGetProgressAndResult drives the whole flow over the
+// router: POST enqueues a progressive "draft" job, GET reads its partial
+// result while the handler is held mid-run, then its final result.
+func TestPostJobThenGetProgressAndResult(t *testing.T) {
+	t.Parallel()
+	release := make(chan struct{})
+	runner, _ := startedRunner(t, release, nil)
+	router := newTestRouterWith(t, "dev", newTestSettings(t), runner)
+
+	rec, body := serve(t, router, signedIn(http.MethodPost, "/api/v1/jobs", strings.NewReader(`{"kind":"draft","payload":{}}`)))
+	if rec.Code != http.StatusAccepted {
+		t.Fatalf("status = %d, want 202: %s", rec.Code, rec.Body)
+	}
+	id, _ := body["id"].(string)
+	if keys := slices.Sorted(maps.Keys(body)); !slices.Equal(keys, []string{"id"}) || id == "" {
+		t.Fatalf("body = %v, want only a job id", body)
+	}
+
+	mid := pollJob(t, router, id, func(b map[string]any) bool { return resultDays(b) == 1 })
+	if mid["id"] != id || mid["kind"] != "draft" || mid["status"] != "running" || mid["error"] != nil {
+		t.Errorf("mid-run body = %v, want running draft job %s with no error", mid, id)
+	}
+
+	close(release)
+	done := pollJob(t, router, id, func(b map[string]any) bool { return b["status"] == "succeeded" })
+	if done["id"] != id || resultDays(done) != 7 || done["error"] != nil {
+		t.Errorf("finished body = %v, want result {\"days\":7} and no error", done)
+	}
+}
+
+func TestPostJobRefusesBadRequests(t *testing.T) {
+	t.Parallel()
+	runner, _ := startedRunner(t, nil, nil)
+	router := newTestRouterWith(t, "dev", newTestSettings(t), runner)
+
+	for _, tc := range []struct {
+		name, body, code, message string
+	}{
+		{"unknown kind", `{"kind":"no_such_kind","payload":{}}`, "unknown_kind", "unknown job kind"},
+		{"payload for another type", `{"kind":"draft","payload":{"days":3}}`, "invalid_payload", "payload does not match the job kind"},
+		{"payload not an object", `{"kind":"draft","payload":[1]}`, "invalid_payload", "payload does not match the job kind"},
+		{"missing payload", `{"kind":"draft"}`, "invalid_request", "payload: required"},
+		{"null payload", `{"kind":"draft","payload":null}`, "invalid_request", "payload: must not be null"},
+		{"missing kind", `{"payload":{}}`, "invalid_request", "kind: required"},
+		{"empty body", ``, "invalid_request", "request body must be a JSON object with kind and payload"},
+		{"not an object", `["draft"]`, "invalid_request", "request body must be a JSON object with kind and payload"},
+		{"unknown field", `{"kind":"draft","payload":{},"extra":1}`, "invalid_request", "request body must be a JSON object with kind and payload"},
+		{"trailing data", `{"kind":"draft","payload":{}} {}`, "invalid_request", "request body must be a single JSON object"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+			rec := httptest.NewRecorder()
+			router.ServeHTTP(rec, signedIn(http.MethodPost, "/api/v1/jobs", strings.NewReader(tc.body)))
+			if rec.Code != http.StatusBadRequest {
+				t.Fatalf("status = %d, want 400: %s", rec.Code, rec.Body)
+			}
+			var body errorBody
+			if err := json.Unmarshal(rec.Body.Bytes(), &body); err != nil {
+				t.Fatal(err)
+			}
+			if want := (errorBody{Error: apiError{Code: tc.code, Message: tc.message}}); body != want {
+				t.Errorf("body = %+v, want %+v", body, want)
+			}
+		})
+	}
+}
+
+func TestPostJobNeedsSessionAndCSRF(t *testing.T) {
+	t.Parallel()
+	router := newTestRouter(t, "dev")
+	const body = `{"kind":"noop","payload":{}}`
+
+	rec := httptest.NewRecorder()
+	router.ServeHTTP(rec, httptest.NewRequest(http.MethodPost, "/api/v1/jobs", strings.NewReader(body)))
+	if rec.Code != http.StatusUnauthorized || rec.Body.String() != unauthenticatedJSON {
+		t.Errorf("no session: %d %s, want 401 %s", rec.Code, rec.Body, unauthenticatedJSON)
+	}
+
+	for _, header := range []string{"", "wrong-token"} {
+		req := httptest.NewRequest(http.MethodPost, "/api/v1/jobs", strings.NewReader(body))
+		req.AddCookie(validSessionCookie)
+		if header != "" {
+			req.Header.Set(CSRFHeader, header)
+		}
+		rec := httptest.NewRecorder()
+		router.ServeHTTP(rec, req)
+		if rec.Code != http.StatusForbidden || rec.Body.String() != csrfJSON {
+			t.Errorf("CSRF header %q: %d %s, want 403 %s", header, rec.Code, rec.Body, csrfJSON)
+		}
+	}
+}

internal/server/server.go

@@ -30,7 +30,7 @@ type RouteMounter interface {
 // only from a peer in trusted (see resolveOrigin). version is reported by
 // GET /health; auth mounts the /auth routes; sessions guards /api/v1 and
 // /media; settings serves /api/v1/settings/household; runner serves
-// /api/v1/jobs/{id}; assets is the web build
+// POST /api/v1/jobs and GET /api/v1/jobs/{id}; assets is the web build
 // served for every other GET. /health, /auth and the web build are public.
 // Requests to /api/v1 other than GET also need the session's CSRF token in CSRFHeader.
 func NewRouter(logger *slog.Logger, version string, trusted forwarded.Trusted, auth RouteMounter, sessions SessionLookup, settings *household.SettingsService, runner *jobs.Runner, assets fs.FS) (http.Handler, error) {
@@ -68,6 +68,7 @@ func NewRouter(logger *slog.Logger, version string, trusted forwarded.Trusted, a
 	apiV1.Get("/me", me(logger))
 	apiV1.Get("/settings/household", getHouseholdSettings(logger, settings))
 	apiV1.Patch("/settings/household", patchHouseholdSettings(logger, settings))
+	apiV1.Post("/jobs", postJob(logger, runner))
 	apiV1.Get("/jobs/{id}", getJob(logger, runner))
 	r.Route("/api", func(r chi.Router) {
 		r.MethodNotAllowed(methodNotAllowed)