Prepare documentation for the v0.8.0 release

README.md

@@ -33,10 +33,13 @@ boundary and constraints that framework work must preserve.
 
 ## Release Guidance
 
-Consumers upgrading from `v0.6.0` to `v0.7.0` should read the
+Consumers upgrading from `v0.7.0` to `v0.8.0` should read the
-[v0.7.0 changelog and migration guide](docs/releases/v0.7.0.md).
+[v0.8.0 changelog and migration guide](docs/releases/v0.8.0.md).
 
 Earlier adopters can consult the
+[v0.7.0 changelog and migration guide](docs/releases/v0.7.0.md).
+
+Consumers upgrading from `v0.5.0` to `v0.6.0` can consult the
 [v0.6.0 changelog and migration guide](docs/releases/v0.6.0.md).
 
 Consumers upgrading from `v0.4.0` to `v0.5.0` can consult the

docs/consumers/pkg-promptkit.md

@@ -173,6 +173,29 @@ semantics. The
 [OpenAI-compatible integration contract](../integrations/openai-compatible-chat.md)
 owns the built-in client's outbound HTTP behavior.
 
+### Repair A Structured Result
+
+Set a small additional-call budget when a structurally invalid result can be
+corrected automatically:
+
+```go
+request.Validation = &promptkit.OutputContract{
+	Format:         promptkit.FormatJSON,
+	ValidationMode: promptkit.ValidationJSONSchema,
+	SchemaPath:     "events.schema.json",
+	RepairAttempts: 1,
+}
+```
+
+Each repair attempt is another model call, so it can increase latency and
+usage; `RunResult.Usage` is cumulative and `Validation.RepairAttempts` reports
+calls actually started. Exhaustion still returns the final failed validation
+result. `basic` validation can also repair an empty candidate, but structural
+validity is not evidence of factual or domain correctness. See the
+[output-contract format reference](../formats.md#output-contract) and
+[`OutputContract` GoDoc](../../types.go) for the exact budget and eligibility
+rules.
+
 ## Inputs, Profiles, And Overrides
 
 Use `File`, `Inline`, or `InlineWithURI` to construct request inputs. A request

docs/formats.md

@@ -126,7 +126,7 @@ outbound integration determines its wire representation.
 | `format` | yes | `text`, `markdown`, or `json`. |
 | `validation_mode` | yes | `none`, `basic`, `json`, or `json_schema`. |
 | `schema_path` | for `json_schema` | Path to a schema in the configured schema source. |
-| `repair_attempts` | no | Integer zero or greater; omitted means zero. |
+| `repair_attempts` | no | Integer from zero through three; omitted means zero. A positive value requires `basic`, `json`, or `json_schema` validation. |
 
 The validation modes behave as follows:
 
@@ -136,9 +136,15 @@ The validation modes behave as follows:
 - `json_schema` requires valid JSON that satisfies the selected schema.
 
 `format` controls output artifact metadata. JSON Schema mode also supplies the
-schema to compatible model clients as structured-output metadata. The public
+schema to compatible model clients as structured-output metadata. Plain `json`
-engine does not install an output repairer, so its validation is single-pass
+validation accepts every valid JSON value and does not request a provider-native
-even when a positive `repair_attempts` value is present.
+JSON-object constraint.
+
+`repair_attempts` counts additional generation calls after a failed validation.
+Zero is single-pass. With a positive eligible budget, Promptkit stops at the
+first valid candidate. If the budget is exhausted, it returns the final
+candidate and its complete failed validation result; generation and operational
+validation failures remain errors. `none` never permits repair.
 
 A request-level `OutputContract` replaces the complete prompt output contract.
 It does not merge individual fields. If its format is empty, Promptkit uses

docs/integrations/openai-compatible-chat.md

@@ -67,7 +67,8 @@ The client conditionally includes:
 - non-empty `service_tier` and effective `reasoning_effort`; an explicitly
   disabled reasoning setting is empty and therefore omitted; and
 - `response_format` for JSON Schema structured output, including its name,
-  strict flag, and schema document.
+  strict flag, and schema document. Plain JSON validation does not add an
+  object-only response constraint.
 
 The engine resolves backend, profile, and request extra-parameter maps by
 whole-map replacement rather than key merging. The resulting effective map is
@@ -99,10 +100,12 @@ drained.
 
 The bounded body must contain exactly one OpenAI-compatible JSON response
 object followed only by JSON whitespace and EOF. The client returns the first
-choice's non-empty message content and maps prompt, completion, total, cached,
+choice's explicitly present string message content, including an empty or
-and cache-write token counts. Invalid or truncated JSON, trailing non-whitespace
+whitespace-only string, and maps prompt, completion, total, cached, and
-data, a second JSON value, absent choices, empty first-choice content, and size
+cache-write token counts. Invalid or truncated JSON, trailing non-whitespace
-overflow are malformed responses and return no partial result.
+data, a second JSON value, absent choices, missing content, `null` content,
+non-string content, and size overflow are malformed responses and return no
+partial result.
 
 For a non-2xx status, Promptkit recognizes one JSON document with a top-level
 object-valued `error` member. Its optional `message` and `type` fields must be

docs/internal/llm.md

@@ -89,6 +89,12 @@ truncation, malformed JSON, trailing data, and a second value are malformed
 responses with no partial result or provider content in the error. Every body
 is closed, and an unbounded oversized stream is not drained.
 
+After framing succeeds, the first choice must contain an explicitly present
+string `message.content`. The string is returned exactly, including empty or
+whitespace-only content. Missing choices, missing or `null` content, and
+non-string content are malformed responses. Output validation and correction
+eligibility remain outside this package.
+
 For a non-success response, `ProviderHTTPError` retains the HTTP status and
 only normalized detail from the bounded recognized envelope. It retains
 `ErrUnexpectedStatus` through unwrapping. The client owns response closure;

docs/internal/overview.md

@@ -11,7 +11,7 @@ contributor workflow and validation.
 
 | Component | Implemented responsibility | References |
 | --- | --- | --- |
-| Root `promptkit` package | Provides the supported engine facade, source, backend-registration, and injection options, public request, result, prompt-inspection, and profile-inspection values, opaque prepared-execution handles, profile construction, extension interfaces, value conversion, redacted formatting, typed capacity and generation error mapping, and engine-local profile-source assembly including application fallbacks. | [Package GoDoc](../../doc.go), [prepared execution](../../prepared_execution.go), [backend API](../../backends.go), [engine assembly](../../engine.go) |
+| Root `promptkit` package | Provides the supported engine facade, source, backend-registration, and injection options, public request, result, prompt-inspection, and profile-inspection values, opaque prepared-execution handles, profile construction, extension interfaces, value conversion, redacted formatting, typed capacity and generation error mapping, engine-local profile-source assembly including application fallbacks, and bounded output-repair assembly. | [Package GoDoc](../../doc.go), [prepared execution](../../prepared_execution.go), [backend API](../../backends.go), [engine assembly](../../engine.go) |
 | `examples/go-library/prepare` | Demonstrates an offline downstream consumer using a prompt file, in-memory profile, inline input, and `Prepare`. It is not a public library package. | [Example program](../../examples/go-library/prepare/main.go) |
 | `examples/go-library/run` | Demonstrates an offline downstream consumer using a prompt file, in-memory profile, inline input, an injected deterministic model client, and `Run`. It is not a public library package. | [Example program](../../examples/go-library/run/main.go) |
 | `internal/backend` | Constructs each engine's immutable registry from the maintained built-in definitions and consumer additions, validates and defensively copies definitions through the shared JSON-value package, and consumes the LLM-owned OpenAI-compatible reserved request-field rule. | [Backend registry](../../internal/backend/registry.go) |
@@ -27,7 +27,7 @@ contributor workflow and validation.
 | `internal/artifact` | Resolves ordinary inline and unrestricted caller-selected file references into copied artifacts with metadata and hashes. | [Internal sources and validation](sources.md) |
 | `internal/validate` | Validates basic, JSON, and JSON Schema output using operating-system filesystem or `fs.FS` schema sources and creates operation-local validation plans with canonical contained schema resources. | [Framework formats](../formats.md#schemas), [internal sources and validation](sources.md) |
 | `internal/llm` | Defines the internal generation boundary and implements outbound OpenAI-compatible chat requests from resolved execution targets, including bounded structured non-success response decoding, successful-response decoding, authentication, deadline handling, and ownership of the OpenAI-compatible reserved request-field policy. | [Internal model client](llm.md) |
-| `internal/usecase` | Resolves prompt definitions and hashes, profiles, backends, and targets for exact inspection and request settings for preparation, and coordinates ordinary execution and one-attempt prepared execution across internal sources, rendering, artifact loading, operation-local validation plans, generation, capacity, and optional repair. | [Internal runner](runner.md), [prepared-execution implementation](../../internal/usecase/prepared_execution.go) |
+| `internal/usecase` | Resolves prompt definitions and hashes, profiles, backends, and targets for exact inspection and request settings for preparation, and coordinates ordinary execution and one-attempt prepared execution across internal sources, rendering, artifact loading, operation-local validation plans, generation, capacity, and bounded repair. | [Internal runner](runner.md), [prepared-execution implementation](../../internal/usecase/prepared_execution.go) |
 
 The root package assembles these internal components without exposing their
 representations. Consumers depend only on the root facade.

docs/internal/runner.md

@@ -23,8 +23,10 @@ built-in backend and validated consumer additions, one engine-local run
 admitter, and a model client wrapped by the same capacity manager. Validation
 plans and provider-facing schema metadata come from the validator's preparation
 interface.
-An output repairer can be injected internally, but the ordinary runner
+The root engine supplies one default output repairer through the explicit
-constructor does not enable one.
+runner constructor, using the same capacity-wrapped client as initial
+generation. The no-repair runner constructor remains available for focused
+internal callers and tests.
 
 Each invocation carries its state in request, prepared-run, and result values.
 The runner has no durable run or session store.
@@ -125,8 +127,8 @@ admitter is an internal unlimited fallback. After successful admission, `Run`
 immediately defers the returned release function, performs the completion
 phase, makes one initial generation call, builds the named output artifact,
 and validates that artifact with the plan compiled during completion. Invalid
-generated content remains a validation result; an inability to perform
+generated content remains a validation result; an inability to generate or
-validation is an operational error.
+perform validation is an operational error.
 
 Validation preparation and execution honor cancellation at every
 Promptkit-controlled boundary and do not publish a partial plan or result.
@@ -142,17 +144,25 @@ serializing preparation or validation behind the active-generation limit.
 The wrapped model client separately acquires a FIFO active permit only around
 each actual generation call.
 
-When an internal repairer is present, a JSON or JSON Schema content failure can
+After a failed `basic`, JSON, or JSON Schema validation with a positive frozen
-trigger bounded repair attempts. Repair receives the effective execution
+budget, the installed repairer can make a bounded corrective call. Each request
-target, explicit numeric-presence bits, credential, backend identity, session
+starts with a fresh copy of the complete original rendered messages, includes
-ID, validation errors, prior output, and structured-output specification. One
+only the latest nonempty candidate as an assistant message, and appends one
-request constructor supplies those common fields to initial and repair
+corrective user message. Empty candidates omit that assistant message. The
-generation while their rendered prompts remain intentionally distinct. The
+correction carries validation diagnostics as JSON data bounded to 64 KiB; the
-default repairer uses the same wrapped client as initial generation, so each
+full diagnostics remain in the validation result.
-repair reacquires the selected backend's active permit while remaining inside
+
-its original admission lease. Repair never performs a second bounded
+Repair receives the effective execution target, explicit numeric-presence bits,
-admission, and repaired outputs use the operation's existing validation plan.
+credential, backend identity, session ID, and structured-output specification.
-This capability remains internal and is not a public option.
+The same request constructor supplies those common fields to initial and repair
+generation. The default repairer uses the same wrapped client as initial
+generation, so each repair reacquires the selected backend's active permit
+while remaining inside its original admission lease. Repair never performs a
+second bounded admission, and repaired outputs use the operation's existing
+validation plan. The runner stops at the first valid candidate, sums completed
+generation usage, reports calls actually started, and returns the final failed
+validation result on exhaustion. A repair generation failure follows the
+ordinary generation-error category rather than becoming a validation error.
 
 A successful result includes the output artifact and raw output, validation
 state, effective session ID, prompt and rendered-prompt hashes, selected

docs/releases/v0.8.0.md

@@ -0,0 +1,109 @@
+# Promptkit v0.8.0
+
+This supplemental changelog and migration guide summarizes the consumer-facing
+changes from `v0.7.0` to `v0.8.0`. The annotated `v0.8.0` tag is the
+authoritative release record. Exact current contracts belong to the linked
+GoDoc and durable documentation.
+
+## Summary
+
+`v0.8.0` activates Promptkit's bounded output-repair workflow:
+
+- failed nonempty-text, JSON, and JSON Schema validation can make a limited
+  number of corrective model calls;
+- corrective calls preserve the original rendered conversation, effective
+  target, session, structured-output contract, and backend capacity policy;
+- results report cumulative usage and the number of corrective calls actually
+  made; and
+- explicitly empty OpenAI-compatible response content now reaches output
+  validation instead of being classified as a malformed provider envelope.
+
+## Compatibility
+
+This release adds no public declarations or fields and removes none. Existing
+source code remains source-compatible.
+
+The behavior of the existing `OutputContract.RepairAttempts` field and prompt
+YAML `repair_attempts` field has changed. A positive value now authorizes real
+additional model calls after eligible validation failures; earlier releases
+accepted the field but the public engine remained single-pass. Consumers that
+set a positive value should expect additional latency, token usage, and
+provider cost when repair is needed.
+
+Repair budgets must now be between zero and three. A positive budget requires
+`basic`, `json`, or `json_schema` validation. Values above three and a positive
+budget paired with `none` are invalid contracts rather than ignored settings.
+
+An explicitly present empty or whitespace-only string returned by the built-in
+OpenAI-compatible client is now a completed generation candidate. `none`
+validation permits it, while `basic`, `json`, and `json_schema` classify it
+under their ordinary validation rules and may repair it when configured.
+Missing, `null`, or non-string content remains a malformed provider response.
+
+## Upgrade
+
+Update the module dependency with:
+
+```sh
+go get gitea.maximumdirect.net/eric/promptkit@v0.8.0
+go mod tidy
+```
+
+Review every prompt definition and request override that sets a positive repair
+budget. Use zero or omit the field to retain single-pass execution. Ensure each
+positive budget is no greater than three and uses an eligible validation mode,
+then run the consuming project's ordinary and race-enabled tests.
+
+## Bounded Output Repair
+
+`repair_attempts` counts corrective calls in addition to the initial model
+call. Promptkit validates each completed candidate, stops at the first valid
+one, and never exceeds the configured bound. If every candidate remains
+invalid, the run completes successfully with the final candidate and its
+failed validation result rather than returning an operational error.
+
+Each correction starts from the original rendered messages and includes only
+the latest invalid candidate and latest validation diagnostics. JSON Schema
+mode retains the provider-native structured-output request as its first line of
+defense. Promptkit performs only deterministic structural validation; a valid
+response is not necessarily factual or correct for an application's domain.
+
+Usage in the final result is cumulative across the initial response and every
+completed corrective response. `ValidationResult.RepairAttempts` reports the
+number of corrective calls actually made. Corrective generation failures use
+the same public generation-error categories and structured provider details as
+an initial generation failure.
+
+See the [output-contract format reference](../formats.md#output-contract), the
+[consumer repair example](../consumers/pkg-promptkit.md#repair-a-structured-result),
+and the [`OutputContract` and `ValidationResult` GoDoc](../../types.go) for the
+current contracts.
+
+## Explicit Empty Content
+
+The built-in OpenAI-compatible client now distinguishes an explicitly present
+empty string from a missing or malformed `content` field. This aligns built-in
+and injected clients by letting the selected output contract decide whether an
+empty candidate is acceptable, invalid, or eligible for repair.
+
+See the
+[OpenAI-compatible response contract](../integrations/openai-compatible-chat.md#response-handling)
+for the exact envelope behavior.
+
+## Public API Changes
+
+None. This release activates and tightens the documented behavior of existing
+fields.
+
+## Consumer Action
+
+- Remove or set `repair_attempts` to zero where execution must remain
+  single-pass.
+- Keep every positive repair budget at three or fewer and pair it with
+  `basic`, `json`, or `json_schema` validation.
+- Account for additional latency, usage, and provider cost when enabling
+  repair.
+- Continue checking the returned validation status because bounded repair can
+  exhaust without producing a valid candidate.
+- Review workflows that previously treated explicit empty provider content as
+  a generation error.

docs/roadmap/future.md

@@ -38,28 +38,8 @@ consumers.
 
 ## Ideas
 
-### Public bounded output repair
+No ideas are currently awaiting selection. Active feature work belongs in its
-
+focused roadmap rather than this catalog.
-Promptkit should make its bounded output-repair capability available through
-the public engine. A consumer should be able to request a limited number of
-corrective generation attempts when JSON or JSON Schema output fails content
-validation, without having to reproduce Promptkit's generation, validation,
-capacity, and result-accounting orchestration.
-
-- Repair is validation recovery, not a general provider retry, failover, or
-  backoff policy. Transport failures, cancellation, and operational schema or
-  validation errors must retain their ordinary error behavior.
-- Repair must stop after the first valid result or the configured attempt
-  bound. Exhausting the bound should preserve the final invalid result and its
-  validation diagnostics rather than inventing success.
-- Initial generation and every repair attempt must use the same resolved
-  backend, effective execution settings and presence semantics, session,
-  credential boundary, structured-output contract, and backend-capacity
-  policy.
-- Results should report the number of repair attempts and cumulative usage for
-  every model call made by the run.
-- Ordinary and prepared execution should expose coherent behavior, including
-  cancellation, frozen prepared state, error identity, and capacity lifetime.
 
 ## Entry Format
 

engine.go

@@ -430,7 +430,7 @@ func NewEngine(cfg Config, opts ...Option) (*Engine, error) {
 	}
 
 	return &Engine{
-		runner: usecase.NewRunner(
+		runner: usecase.NewRunnerWithRepairer(
 			promptDefs,
 			profiles,
 			backendRegistry,
@@ -438,6 +438,7 @@ func NewEngine(cfg Config, opts ...Option) (*Engine, error) {
 			prompt.NewGoRenderer(),
 			llmClient,
 			validator,
+			usecase.NewDefaultOutputRepairer(llmClient),
 			capacityManager,
 		),
 	}, nil
@@ -639,10 +640,12 @@ func (e *Engine) PrepareExecution(ctx context.Context, req RunRequest) (*Prepare
 // generated output.
 //
 // A content-validation failure is a successful run whose
-// RunResult.Validation has Status ValidationFailed. An inability to perform
+// RunResult.Validation has Status ValidationFailed. When its output contract
-// validation returns an error matching ErrValidation and no partial result.
+// has a positive repair budget, a failed eligible validation can make bounded
-// The public Engine does not perform output repair, so validation is
+// additional model calls and stops at the first valid candidate. Exhaustion
-// single-pass even when OutputContract.RepairAttempts is positive.
+// returns the final failed validation result with cumulative usage and actual
+// repair attempts. An inability to generate or validate returns an error and
+// no partial result.
 //
 // Run can return every error category documented by [Engine.Prepare], plus
 // ErrCapacityExceeded and ErrLLMGenerate. An engine admission rejection is
@@ -686,17 +689,17 @@ func (e *Engine) Run(ctx context.Context, req RunRequest) (*RunResult, error) {
 //
 // The supplied context governs this execution attempt independently of the
 // preparation context. It covers credential revalidation, admission,
-// generation, validation, and any internal repair. Result timing begins after
+// generation, validation, and any bounded output repair. Result timing begins
-// the claim and excludes preparation and consumer-held delay.
+// after the claim and excludes preparation and consumer-held delay.
 //
 // RunPrepared can return ErrInvalidRequest, ErrAPIKeyEnvMissing,
 // ErrCapacityExceeded, ErrLLMGenerate, or ErrValidation as applicable while
 // preserving documented collaborator and context identities. An engine
 // admission rejection is discoverable as [CapacityError] and still matches
 // ErrCapacityExceeded. A built-in OpenAI-compatible non-2xx response is
-// discoverable as [GenerationError]. A completed content-validation rejection
+// discoverable as [GenerationError]. A completed content-validation rejection,
-// is returned in RunResult, not as an operational error. An operational error
+// including repair exhaustion, is returned in RunResult, not as an operational
-// returns no partial RunResult.
+// error. An operational error returns no partial RunResult.
 func (e *Engine) RunPrepared(ctx context.Context, prepared *PreparedExecution) (*RunResult, error) {
 	if e == nil || e.runner == nil {
 		return nil, fmt.Errorf("%w: engine is nil", ErrInvalidConfig)

engine_test.go

@@ -3487,6 +3487,7 @@ extra_params:
 
 type fakeLLMClient struct {
 	response  *promptkit.GenerateResponse
+	responses []*promptkit.GenerateResponse
 	err       error
 	requests  []promptkit.GenerateRequest
 }
@@ -3664,5 +3665,12 @@ func (f *fakeLLMClient) Generate(_ context.Context, req promptkit.GenerateReques
 	if f.err != nil {
 		return nil, f.err
 	}
+	if len(f.responses) > 0 {
+		index := len(f.requests) - 1
+		if index >= len(f.responses) {
+			return nil, fmt.Errorf("no response configured for generation %d", index+1)
+		}
+		return f.responses[index], nil
+	}
 	return f.response, nil
 }

generation_error_contract_test.go

@@ -39,6 +39,44 @@ func TestBuiltInGenerationError(t *testing.T) {
 	assertGenerationError(t, err, http.StatusServiceUnavailable, "", "", "")
 }
 
+func TestBuiltInRepairGenerationError(t *testing.T) {
+	const (
+		codeMarker    = "repair-code-marker"
+		typeMarker    = "repair-type-marker"
+		messageMarker = "repair-message-marker"
+	)
+	calls := 0
+	config := contractConfig(frameworkSchemaDir)
+	config.HTTPClient = &http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
+		calls++
+		if calls == 1 {
+			body := `{"choices":[{"message":{"content":"not-json"}}]}`
+			return &http.Response{StatusCode: http.StatusOK, ContentLength: int64(len(body)), Body: io.NopCloser(strings.NewReader(body))}, nil
+		}
+		body := `{"error":{"code":"` + codeMarker + `","type":"` + typeMarker + `","message":"` + messageMarker + `"}}`
+		return &http.Response{StatusCode: http.StatusUnprocessableEntity, ContentLength: int64(len(body)), Body: io.NopCloser(strings.NewReader(body))}, nil
+	})}
+	engine, err := promptkit.NewEngine(config)
+	if err != nil {
+		t.Fatalf("NewEngine: %v", err)
+	}
+	req := generationErrorRunRequest()
+	req.Validation = &promptkit.OutputContract{
+		Format:         promptkit.FormatJSON,
+		ValidationMode: promptkit.ValidationJSON,
+		RepairAttempts: 1,
+	}
+
+	result, err := engine.Run(context.Background(), req)
+	if result != nil {
+		t.Fatalf("Run result = %#v, want nil", result)
+	}
+	if calls != 2 {
+		t.Fatalf("provider calls = %d, want 2", calls)
+	}
+	assertGenerationError(t, err, http.StatusUnprocessableEntity, codeMarker, typeMarker, messageMarker)
+}
+
 func assertGenerationError(t *testing.T, err error, statusCode int, code, providerType, message string) {
 	t.Helper()
 

internal/domain/output_contract.go

@@ -6,6 +6,8 @@ import (
 	"strings"
 )
 
+const maxOutputRepairAttempts = 3
+
 // ValidateOutputContract validates source-neutral output-contract invariants.
 func ValidateOutputContract(contract OutputContract) error {
 	switch contract.Format {
@@ -26,5 +28,11 @@ func ValidateOutputContract(contract OutputContract) error {
 	if contract.RepairAttempts < 0 {
 		return errors.New("repair_attempts must be greater than or equal to 0")
 	}
+	if contract.RepairAttempts > maxOutputRepairAttempts {
+		return fmt.Errorf("repair_attempts must be less than or equal to %d", maxOutputRepairAttempts)
+	}
+	if contract.ValidationMode == ValidationNone && contract.RepairAttempts > 0 {
+		return errors.New("repair_attempts requires basic, json, or json_schema validation")
+	}
 	return nil
 }

internal/domain/output_contract_test.go

@@ -30,9 +30,28 @@ func TestValidateOutputContract(t *testing.T) {
 		}},
 		{name: "empty validation mode", change: func(c *OutputContract) { c.ValidationMode = "" }, wantErr: "validation mode"},
 		{name: "unsupported validation mode", change: func(c *OutputContract) { c.ValidationMode = ValidationMode("unknown") }, wantErr: "validation mode"},
-		{name: "negative repair attempts", change: func(c *OutputContract) { c.RepairAttempts = -1 }, wantErr: "repair_attempts"},
+		{name: "negative repair attempts", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationBasic
+			c.RepairAttempts = -1
+		}, wantErr: "repair_attempts"},
 		{name: "zero repair attempts", change: func(c *OutputContract) { c.RepairAttempts = 0 }},
-		{name: "positive repair attempts", change: func(c *OutputContract) { c.RepairAttempts = 1 }},
+		{name: "one repair attempt", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationBasic
+			c.RepairAttempts = 1
+		}},
+		{name: "maximum repair attempts", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationJSON
+			c.RepairAttempts = 3
+		}},
+		{name: "too many repair attempts", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationJSONSchema
+			c.SchemaPath = "schema.json"
+			c.RepairAttempts = 4
+		}, wantErr: "repair_attempts"},
+		{name: "none validation with repair attempts", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationNone
+			c.RepairAttempts = 1
+		}, wantErr: "repair_attempts"},
 		{name: "json schema empty path", change: func(c *OutputContract) {
 			c.ValidationMode = ValidationJSONSchema
 			c.SchemaPath = ""

internal/llm/openai_compatible_client.go

@@ -179,12 +179,12 @@ func (c *OpenAICompatibleClient) Generate(ctx context.Context, req domain.Genera
 		return nil, fmt.Errorf("%w: no choices returned", ErrMalformedResponse)
 	}
 	content := wireResp.Choices[0].Message.Content
-	if content == "" {
+	if content == nil {
-		return nil, fmt.Errorf("%w: first choice has empty message content", ErrMalformedResponse)
+		return nil, fmt.Errorf("%w: first choice has missing message content", ErrMalformedResponse)
 	}
 
 	return &domain.GenerateResponse{
-		Content: content,
+		Content: *content,
 		Usage: domain.TokenUsage{
 			PromptTokens:     wireResp.Usage.PromptTokens,
 			CompletionTokens: wireResp.Usage.CompletionTokens,
@@ -380,7 +380,7 @@ type openAICacheControl struct {
 
 type openAIChatResponseMessage struct {
 	Role    string  `json:"role"`
-	Content string `json:"content"`
+	Content *string `json:"content"`
 }
 
 type openAIChatResponse struct {

internal/llm/openai_compatible_client_test.go

@@ -765,6 +765,7 @@ func TestOpenAICompatibleClientResponseFraming(t *testing.T) {
 		run  func(*testing.T)
 	}{
 		{name: "usage mapping", run: checkCacheUsageMapping},
+		{name: "content presence", run: checkContentPresence},
 		{name: "common response failures", run: checkCommonResponseFailures},
 		{name: "successful response byte boundary", run: checkSuccessfulResponseByteBoundary},
 		{name: "continuing oversized response", run: checkContinuingOversizedResponse},
@@ -776,6 +777,53 @@ func TestOpenAICompatibleClientResponseFraming(t *testing.T) {
 	}
 }
 
+func checkContentPresence(t *testing.T) {
+	tests := []struct {
+		name    string
+		content string
+	}{
+		{name: "explicit empty string", content: ""},
+		{name: "whitespace string", content: " \n\t "},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			provider := newRecordingProvider(t)
+			provider.respond(http.StatusOK, `{
+  "choices": [{"message": {"content": `+strconv.Quote(tc.content)+`}}],
+  "usage": {
+    "prompt_tokens": 10,
+    "completion_tokens": 20,
+    "total_tokens": 30,
+    "prompt_tokens_details": {"cached_tokens": 4},
+    "cache_write_tokens": 5
+  }
+}`)
+			client := newProviderClient(t, provider, OpenAICompatibleConfig{Model: "model"})
+
+			response, err := client.Generate(context.Background(), ordinaryGenerateRequest())
+			if err != nil {
+				t.Fatalf("generate: %v", err)
+			}
+			if response == nil {
+				t.Fatal("expected response")
+			}
+			if response.Content != tc.content {
+				t.Fatalf("content = %q, want %q", response.Content, tc.content)
+			}
+			if response.Usage != (domain.TokenUsage{
+				PromptTokens:     10,
+				CompletionTokens: 20,
+				TotalTokens:      30,
+				CachedTokens:     4,
+				CacheWriteTokens: 5,
+			}) {
+				t.Fatalf("usage = %+v", response.Usage)
+			}
+		})
+	}
+}
+
 func checkCacheUsageMapping(t *testing.T) {
 	provider := newRecordingProvider(t)
 	provider.respond(http.StatusOK, `{
@@ -1129,6 +1177,9 @@ func checkCommonResponseFailures(t *testing.T) {
 		},
 		{name: "invalid JSON", statusCode: http.StatusOK, body: `{not valid json`, wantErr: ErrMalformedResponse},
 		{name: "missing choices", statusCode: http.StatusOK, body: `{"choices": []}`, wantErr: ErrMalformedResponse},
+		{name: "missing content", statusCode: http.StatusOK, body: `{"choices": [{"message": {}}]}`, wantErr: ErrMalformedResponse},
+		{name: "null content", statusCode: http.StatusOK, body: `{"choices": [{"message": {"content": null}}]}`, wantErr: ErrMalformedResponse},
+		{name: "non-string content", statusCode: http.StatusOK, body: `{"choices": [{"message": {"content": 1}}]}`, wantErr: ErrMalformedResponse},
 	}
 
 	for _, tc := range tests {

internal/promptdef/repository_test.go

@@ -892,6 +892,38 @@ output:
   format: text
   validation_mode: none
   repair_attempts: -1
+`,
+			wantErr:        true,
+			wantDiagnostic: "repair_attempts",
+		},
+		{
+			name: "repair attempts above maximum",
+			definition: `
+id: normalization-rule
+version: "1"
+messages:
+  - role: user
+    content: test
+output:
+  format: text
+  validation_mode: basic
+  repair_attempts: 4
+`,
+			wantErr:        true,
+			wantDiagnostic: "repair_attempts",
+		},
+		{
+			name: "none validation with repair attempts",
+			definition: `
+id: normalization-rule
+version: "1"
+messages:
+  - role: user
+    content: test
+output:
+  format: text
+  validation_mode: none
+  repair_attempts: 1
 `,
 			wantErr:        true,
 			wantDiagnostic: "repair_attempts",

internal/usecase/output_contract_test.go

@@ -106,6 +106,8 @@ func TestRunnerPreparationRejectsInvalidOutputContractsBeforeCompletion(t *testi
 		{name: "unsupported format", override: domain.OutputContract{Format: "binary", ValidationMode: domain.ValidationNone}},
 		{name: "empty validation mode", override: domain.OutputContract{Format: domain.FormatText}},
 		{name: "negative repair attempts", override: domain.OutputContract{Format: domain.FormatText, ValidationMode: domain.ValidationNone, RepairAttempts: -1}},
+		{name: "repair attempts above maximum", override: domain.OutputContract{Format: domain.FormatText, ValidationMode: domain.ValidationBasic, RepairAttempts: 4}},
+		{name: "none validation with repair attempts", override: domain.OutputContract{Format: domain.FormatText, ValidationMode: domain.ValidationNone, RepairAttempts: 1}},
 		{name: "json schema without path", override: domain.OutputContract{Format: domain.FormatJSON, ValidationMode: domain.ValidationJSONSchema}},
 	}
 

internal/usecase/prepared_execution_test.go

@@ -3,10 +3,12 @@ package usecase
 import (
 	"context"
 	"errors"
+	"fmt"
 	"reflect"
 	"testing"
 
 	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/llm"
 	"gitea.maximumdirect.net/eric/promptkit/internal/validate"
 )
 
@@ -385,19 +387,20 @@ func TestRunnerRunPreparedUsesFrozenValidationForInitialAndRepairOutputs(t *test
 	admitter := &fakeRunAdmitter{}
 	reader := defaultArtifactReader()
 	renderer := defaultRenderer()
+	client := &sequenceLLM{responses: []*domain.GenerateResponse{{
+		Content: `{"broken":true}`,
+		Usage: domain.TokenUsage{
+			PromptTokens: 13, CompletionTokens: 17, TotalTokens: 19,
+			CachedTokens: 23, CacheWriteTokens: 29,
+		},
+	}}}
 	runner := NewRunnerWithRepairer(
 		&fakePromptRepo{def: promptDef(domain.FormatJSON, domain.ValidationJSON, 1)},
 		&fakeExecutionProfileRepo{profiles: map[string]*domain.ExecutionProfile{"exec": defaultExecutionProfile()}},
 		nil,
 		reader,
 		renderer,
-		&fakeLLM{resp: &domain.GenerateResponse{
+		client,
-			Content: `{"broken":true}`,
-			Usage: domain.TokenUsage{
-				PromptTokens: 13, CompletionTokens: 17, TotalTokens: 19,
-				CachedTokens: 23, CacheWriteTokens: 29,
-			},
-		}},
 		validator,
 		repairer,
 		admitter,
@@ -421,6 +424,10 @@ func TestRunnerRunPreparedUsesFrozenValidationForInitialAndRepairOutputs(t *test
 	if validator.directValidateCalls != 0 || repairer.calls != 1 {
 		t.Fatalf("validation/repair calls=(direct=%d repair=%d), want (0, 1)", validator.directValidateCalls, repairer.calls)
 	}
+	if len(client.requests) != 1 || len(repairer.reqs) != 1 ||
+		!reflect.DeepEqual(repairer.reqs[0].OriginalMessages, client.requests[0].Prompt.Messages) {
+		t.Fatalf("initial and repair messages = (%#v, %#v)", client.requests, repairer.reqs)
+	}
 	if result.Validation.Status != domain.ValidationPassed || result.Validation.RepairAttempts != 1 {
 		t.Fatalf("unexpected repaired validation result: %+v", result.Validation)
 	}
@@ -443,6 +450,7 @@ func TestRunnerRunPreparedReleasesAdmissionAcrossExecutionErrors(t *testing.T) {
 	generationFailure := errors.New("generation failed")
 	validationFailure := errors.New("validation failed")
 	repairFailure := errors.New("repair failed")
+	repairInvalidRequest := fmt.Errorf("repair request: %w", llm.ErrInvalidRequest)
 
 	tests := []struct {
 		name          string
@@ -450,6 +458,7 @@ func TestRunnerRunPreparedReleasesAdmissionAcrossExecutionErrors(t *testing.T) {
 		validation    *recordingPreparedValidation
 		repairer      *fakeRepairer
 		wantError     error
+		wantSource    error
 	}{
 		{
 			name:          "generation failure",
@@ -475,7 +484,50 @@ func TestRunnerRunPreparedReleasesAdmissionAcrossExecutionErrors(t *testing.T) {
 				}},
 			},
 			repairer:   &fakeRepairer{err: repairFailure},
-			wantError: ErrValidation,
+			wantError:  ErrLLMGenerate,
+			wantSource: repairFailure,
+		},
+		{
+			name: "repair invalid request",
+			validation: &recordingPreparedValidation{
+				results: []domain.ValidationResult{{
+					Status:  domain.ValidationFailed,
+					Mode:    domain.ValidationJSON,
+					Errors:  []string{"invalid"},
+					IsValid: false,
+				}},
+			},
+			repairer:   &fakeRepairer{err: repairInvalidRequest},
+			wantError:  ErrInvalidRequest,
+			wantSource: repairInvalidRequest,
+		},
+		{
+			name: "repair cancellation",
+			validation: &recordingPreparedValidation{
+				results: []domain.ValidationResult{{
+					Status:  domain.ValidationFailed,
+					Mode:    domain.ValidationJSON,
+					Errors:  []string{"invalid"},
+					IsValid: false,
+				}},
+			},
+			repairer:   &fakeRepairer{err: context.Canceled},
+			wantError:  ErrLLMGenerate,
+			wantSource: context.Canceled,
+		},
+		{
+			name: "repair deadline",
+			validation: &recordingPreparedValidation{
+				results: []domain.ValidationResult{{
+					Status:  domain.ValidationFailed,
+					Mode:    domain.ValidationJSON,
+					Errors:  []string{"invalid"},
+					IsValid: false,
+				}},
+			},
+			repairer:   &fakeRepairer{err: context.DeadlineExceeded},
+			wantError:  ErrLLMGenerate,
+			wantSource: context.DeadlineExceeded,
 		},
 	}
 
@@ -514,6 +566,9 @@ func TestRunnerRunPreparedReleasesAdmissionAcrossExecutionErrors(t *testing.T) {
 			if result != nil || !errors.Is(err, test.wantError) {
 				t.Fatalf("run prepared=(%+v, %v), want %v", result, err, test.wantError)
 			}
+			if test.wantSource != nil && !errors.Is(err, test.wantSource) {
+				t.Fatalf("run prepared error = %v, want source %v", err, test.wantSource)
+			}
 			if len(admitter.backendIDs) != 1 || admitter.releaseCalls != 1 {
 				t.Fatalf(
 					"admission calls=%#v releases=%d, want one each",

internal/usecase/repairer.go

@@ -2,19 +2,29 @@ package usecase
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"strings"
+	"unicode/utf8"
 
 	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
 	"gitea.maximumdirect.net/eric/promptkit/internal/llm"
 )
 
+const (
+	maxRepairDiagnosticBytes = 64 * 1024
+	omittedRepairDiagnostics = "additional validation diagnostics were omitted"
+)
+
+// OutputRepairer generates a corrected candidate after validation fails.
 type OutputRepairer interface {
 	Repair(ctx context.Context, req RepairRequest) (*domain.GenerateResponse, error)
 }
 
+// RepairRequest contains the immutable execution state needed for one correction.
 type RepairRequest struct {
+	OriginalMessages []domain.RenderedMessage
 	PreviousOutput   string
 	ValidationErrors []string
 	SessionID        string
@@ -30,6 +40,7 @@ type defaultOutputRepairer struct {
 	llm llm.Client
 }
 
+// NewDefaultOutputRepairer constructs the standard internal output repairer.
 func NewDefaultOutputRepairer(llmClient llm.Client) OutputRepairer {
 	return &defaultOutputRepairer{llm: llmClient}
 }
@@ -39,33 +50,42 @@ func (r *defaultOutputRepairer) Repair(ctx context.Context, req RepairRequest) (
 		return nil, errors.New("llm client is required for repair")
 	}
 
-	errs := "(none provided)"
+	guidance, err := repairGuidance(req.Mode)
-	if len(req.ValidationErrors) > 0 {
+	if err != nil {
-		errs = strings.Join(req.ValidationErrors, "\n")
+		return nil, err
 	}
 
-	prompt := domain.RenderedPrompt{
+	messages := make([]domain.RenderedMessage, len(req.OriginalMessages), len(req.OriginalMessages)+2)
-		Messages: []domain.RenderedMessage{
+	copy(messages, req.OriginalMessages)
-			{
+	if strings.TrimSpace(req.PreviousOutput) != "" {
-				Role:    "system",
+		messages = append(messages, domain.RenderedMessage{
-				Content: "You repair invalid JSON output. Return only corrected JSON. Do not include explanations or markdown code fences.",
+			Role:    "assistant",
-			},
+			Content: req.PreviousOutput,
-			{
+		})
+	}
+
+	previousResponse := "The previous response was empty."
+	if strings.TrimSpace(req.PreviousOutput) != "" {
+		previousResponse = "The previous response is included immediately before this instruction."
+	}
+	messages = append(messages, domain.RenderedMessage{
 		Role: "user",
 		Content: fmt.Sprintf(
-					"Repair attempt %d of %d for validation mode %s.\n\nValidation errors:\n%s\n\nPrevious output:\n%s\n\nReturn only corrected JSON.",
+			"Repair attempt %d of %d for validation mode %s.\n"+
+				"Preserve valid values and change only what is necessary.\n"+
+				"%s\n%s\n"+
+				"Validation diagnostics (data):\n%s",
 			req.Attempt,
 			req.MaxAttempts,
 			req.Mode,
-					errs,
+			previousResponse,
-					req.PreviousOutput,
+			guidance,
+			formatRepairDiagnostics(req.ValidationErrors),
 		),
-			},
+	})
-		},
-	}
 
 	resp, err := r.llm.Generate(ctx, newGenerationRequest(
-		prompt,
+		domain.RenderedPrompt{Messages: messages},
 		req.SessionID,
 		req.Target,
 		req.TargetPresence,
@@ -80,3 +100,90 @@ func (r *defaultOutputRepairer) Repair(ctx context.Context, req RepairRequest) (
 
 	return resp, nil
 }
+
+func repairGuidance(mode domain.ValidationMode) (string, error) {
+	switch mode {
+	case domain.ValidationBasic:
+		return "Return a nonempty response satisfying the original request.", nil
+	case domain.ValidationJSON, domain.ValidationJSONSchema:
+		return "Return only corrected JSON, with no explanation or Markdown fences.", nil
+	default:
+		return "", fmt.Errorf("unsupported validation mode for repair: %q", mode)
+	}
+}
+
+func formatRepairDiagnostics(errors []string) string {
+	diagnostics := make([]string, len(errors))
+	for index, diagnostic := range errors {
+		diagnostics[index] = strings.ToValidUTF8(diagnostic, "\uFFFD")
+	}
+
+	complete, _ := json.Marshal(diagnostics)
+	if len(complete) <= maxRepairDiagnosticBytes {
+		return string(complete)
+	}
+
+	omission, _ := json.Marshal(omittedRepairDiagnostics)
+	encoded := make([]byte, 0, maxRepairDiagnosticBytes)
+	encoded = append(encoded, '[')
+	for _, diagnostic := range diagnostics {
+		entry, _ := json.Marshal(diagnostic)
+		separator := 0
+		if len(encoded) > 1 {
+			separator = 1
+		}
+		available := maxRepairDiagnosticBytes - len(encoded) - separator - 1 - len(omission) - 1
+		if len(entry) <= available {
+			if separator != 0 {
+				encoded = append(encoded, ',')
+			}
+			encoded = append(encoded, entry...)
+			continue
+		}
+
+		if available < len(`""`) {
+			break
+		}
+		if separator != 0 {
+			encoded = append(encoded, ',')
+		}
+		encoded = append(encoded, truncateDiagnosticJSONValue(diagnostic, available)...)
+		break
+	}
+	if len(encoded) > 1 {
+		encoded = append(encoded, ',')
+	}
+	encoded = append(encoded, omission...)
+	encoded = append(encoded, ']')
+	return string(encoded)
+}
+
+func truncateDiagnosticJSONValue(value string, maxBytes int) []byte {
+	if maxBytes < len(`""`) {
+		return nil
+	}
+
+	boundaries := []int{0}
+	for end := 0; end < len(value); {
+		_, size := utf8.DecodeRuneInString(value[end:])
+		end += size
+		if end+len(`""`) > maxBytes {
+			break
+		}
+		boundaries = append(boundaries, end)
+	}
+
+	low, high := 0, len(boundaries)-1
+	best := []byte(`""`)
+	for low <= high {
+		mid := low + (high-low)/2
+		candidate, _ := json.Marshal(value[:boundaries[mid]])
+		if len(candidate) <= maxBytes {
+			best = candidate
+			low = mid + 1
+			continue
+		}
+		high = mid - 1
+	}
+	return best
+}

internal/usecase/repairer_test.go

@@ -0,0 +1,349 @@
+package usecase
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"reflect"
+	"strings"
+	"testing"
+	"unicode/utf8"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+type recordingRepairClient struct {
+	requests []domain.GenerateRequest
+	response *domain.GenerateResponse
+	err      error
+}
+
+func (c *recordingRepairClient) Generate(_ context.Context, req domain.GenerateRequest) (*domain.GenerateResponse, error) {
+	c.requests = append(c.requests, req)
+	return c.response, c.err
+}
+
+func TestDefaultOutputRepairerBuildsFullContextRequest(t *testing.T) {
+	client := &recordingRepairClient{response: &domain.GenerateResponse{Content: "corrected"}}
+	repairer := NewDefaultOutputRepairer(client)
+	original := []domain.RenderedMessage{
+		{Role: "system", Content: "Follow the task.", CacheControl: &domain.CacheControl{Type: domain.CacheControlEphemeral, TTL: "1h"}},
+		{Role: "user", Content: "Summarize the report."},
+	}
+	before := append([]domain.RenderedMessage(nil), original...)
+	previous := strings.Repeat("candidate ", 12_000)
+	target := domain.ExecutionTarget{BackendID: "backend", Endpoint: "https://provider.example/v1", Model: "model"}
+	presence := domain.ExecutionTargetPresence{Temperature: true, TopP: true}
+	structured := &domain.StructuredOutputSpec{}
+
+	response, err := repairer.Repair(context.Background(), RepairRequest{
+		OriginalMessages: original,
+		PreviousOutput:   previous,
+		ValidationErrors: []string{"invalid JSON"},
+		SessionID:        "session",
+		Target:           target,
+		TargetPresence:   presence,
+		StructuredOutput: structured,
+		Attempt:          1,
+		MaxAttempts:      3,
+		Mode:             domain.ValidationJSONSchema,
+	})
+	if err != nil {
+		t.Fatalf("repair: %v", err)
+	}
+	if response == nil || response.Content != "corrected" {
+		t.Fatalf("response = %+v", response)
+	}
+	if !reflect.DeepEqual(original, before) {
+		t.Fatalf("original messages changed: got %#v, want %#v", original, before)
+	}
+	if len(client.requests) != 1 {
+		t.Fatalf("generation requests = %d, want 1", len(client.requests))
+	}
+
+	request := client.requests[0]
+	if request.Prompt.SessionID != "session" || !reflect.DeepEqual(request.Target, target) || request.TargetPresence != presence || request.StructuredOutput != structured {
+		t.Fatalf("generation request fields = %+v", request)
+	}
+	if len(request.Prompt.Messages) != len(original)+2 {
+		t.Fatalf("message count = %d, want %d", len(request.Prompt.Messages), len(original)+2)
+	}
+	if !reflect.DeepEqual(request.Prompt.Messages[:len(original)], original) {
+		t.Fatalf("original messages = %#v, want %#v", request.Prompt.Messages[:len(original)], original)
+	}
+	assistant := request.Prompt.Messages[len(original)]
+	if assistant.Role != "assistant" || assistant.Content != previous {
+		t.Fatalf("assistant candidate = %+v", assistant)
+	}
+	correction := request.Prompt.Messages[len(original)+1]
+	if correction.Role != "user" || !strings.Contains(correction.Content, "Repair attempt 1 of 3") ||
+		!strings.Contains(correction.Content, "Preserve valid values") ||
+		!strings.Contains(correction.Content, "Return only corrected JSON") {
+		t.Fatalf("correction message = %q", correction.Content)
+	}
+	if diagnostics := repairDiagnosticsFromMessage(t, correction.Content); !reflect.DeepEqual(diagnostics, []string{"invalid JSON"}) {
+		t.Fatalf("diagnostics = %#v", diagnostics)
+	}
+}
+
+func TestDefaultOutputRepairerDoesNotAccumulateCandidates(t *testing.T) {
+	client := &recordingRepairClient{response: &domain.GenerateResponse{Content: "corrected"}}
+	repairer := NewDefaultOutputRepairer(client)
+	backing := make([]domain.RenderedMessage, 1, 4)
+	backing[0] = domain.RenderedMessage{Role: "user", Content: "Original task"}
+	before := append([]domain.RenderedMessage(nil), backing...)
+
+	for _, candidate := range []string{"first invalid", "second invalid"} {
+		_, err := repairer.Repair(context.Background(), RepairRequest{
+			OriginalMessages: backing,
+			PreviousOutput:   candidate,
+			Attempt:          1,
+			MaxAttempts:      3,
+			Mode:             domain.ValidationJSON,
+		})
+		if err != nil {
+			t.Fatalf("repair %q: %v", candidate, err)
+		}
+	}
+	if !reflect.DeepEqual(backing, before) {
+		t.Fatalf("caller messages changed: got %#v, want %#v", backing, before)
+	}
+	if len(client.requests) != 2 {
+		t.Fatalf("generation requests = %d, want 2", len(client.requests))
+	}
+	for index, request := range client.requests {
+		messages := request.Prompt.Messages
+		if len(messages) != 3 || messages[0] != backing[0] || messages[1].Role != "assistant" || messages[1].Content != []string{"first invalid", "second invalid"}[index] {
+			t.Fatalf("request %d messages = %#v", index, messages)
+		}
+	}
+}
+
+func TestDefaultOutputRepairerOmitsEmptyCandidateMessage(t *testing.T) {
+	for _, tc := range []struct {
+		name      string
+		candidate string
+	}{
+		{name: "empty", candidate: ""},
+		{name: "whitespace", candidate: " \n\t "},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			client := &recordingRepairClient{response: &domain.GenerateResponse{Content: "corrected"}}
+			repairer := NewDefaultOutputRepairer(client)
+			_, err := repairer.Repair(context.Background(), RepairRequest{
+				OriginalMessages: []domain.RenderedMessage{{Role: "user", Content: "Original task"}},
+				PreviousOutput:   tc.candidate,
+				Attempt:          1,
+				MaxAttempts:      1,
+				Mode:             domain.ValidationBasic,
+			})
+			if err != nil {
+				t.Fatalf("repair: %v", err)
+			}
+			messages := client.requests[0].Prompt.Messages
+			if len(messages) != 2 || messages[1].Role != "user" || !strings.Contains(messages[1].Content, "previous response was empty") {
+				t.Fatalf("messages = %#v", messages)
+			}
+			if strings.Contains(messages[1].Content, tc.candidate) && tc.candidate != "" {
+				t.Fatalf("correction message repeated whitespace candidate: %q", messages[1].Content)
+			}
+		})
+	}
+}
+
+func TestDefaultOutputRepairerUsesModeSpecificGuidance(t *testing.T) {
+	tests := []struct {
+		mode domain.ValidationMode
+		want string
+	}{
+		{mode: domain.ValidationBasic, want: "Return a nonempty response"},
+		{mode: domain.ValidationJSON, want: "Return only corrected JSON"},
+		{mode: domain.ValidationJSONSchema, want: "Return only corrected JSON"},
+	}
+
+	for _, tc := range tests {
+		t.Run(string(tc.mode), func(t *testing.T) {
+			client := &recordingRepairClient{response: &domain.GenerateResponse{Content: "corrected"}}
+			_, err := NewDefaultOutputRepairer(client).Repair(context.Background(), RepairRequest{Mode: tc.mode, Attempt: 1, MaxAttempts: 1})
+			if err != nil {
+				t.Fatalf("repair: %v", err)
+			}
+			if !strings.Contains(client.requests[0].Prompt.Messages[0].Content, tc.want) {
+				t.Fatalf("correction message = %q", client.requests[0].Prompt.Messages[0].Content)
+			}
+		})
+	}
+
+	client := &recordingRepairClient{response: &domain.GenerateResponse{Content: "unexpected"}}
+	_, err := NewDefaultOutputRepairer(client).Repair(context.Background(), RepairRequest{Mode: domain.ValidationNone})
+	if err == nil || !strings.Contains(err.Error(), "unsupported validation mode") {
+		t.Fatalf("unsupported mode error = %v", err)
+	}
+	if len(client.requests) != 0 {
+		t.Fatalf("generation requests = %d, want 0", len(client.requests))
+	}
+}
+
+func TestFormatRepairDiagnosticsBoundsAndPreservesData(t *testing.T) {
+	tests := []struct {
+		name         string
+		errors       []string
+		wantOmission bool
+		check        func(*testing.T, []string)
+	}{
+		{
+			name:   "below limit",
+			errors: []string{"first", "second"},
+			check: func(t *testing.T, got []string) {
+				t.Helper()
+				if !reflect.DeepEqual(got, []string{"first", "second"}) {
+					t.Fatalf("diagnostics = %#v", got)
+				}
+			},
+		},
+		{
+			name:   "at limit",
+			errors: []string{strings.Repeat("x", maxRepairDiagnosticBytes-4)},
+			check: func(t *testing.T, got []string) {
+				t.Helper()
+				if len(got) != 1 || len(got[0]) != maxRepairDiagnosticBytes-4 {
+					t.Fatalf("diagnostics lengths = %#v", got)
+				}
+			},
+		},
+		{
+			name:         "multibyte truncation preserves prior entries",
+			errors:       []string{"first", strings.Repeat("界", maxRepairDiagnosticBytes)},
+			wantOmission: true,
+			check: func(t *testing.T, got []string) {
+				t.Helper()
+				if len(got) != 3 || got[0] != "first" || !strings.HasPrefix(strings.Repeat("界", maxRepairDiagnosticBytes), got[1]) {
+					t.Fatalf("diagnostics = %#v", got)
+				}
+			},
+		},
+		{
+			name:         "many diagnostics",
+			errors:       manyRepairDiagnostics(),
+			wantOmission: true,
+			check: func(t *testing.T, got []string) {
+				t.Helper()
+				if len(got) < 2 || got[0] != manyRepairDiagnostics()[0] {
+					t.Fatalf("diagnostics = %#v", got)
+				}
+			},
+		},
+		{
+			name: "no room for partial diagnostic",
+			errors: func() []string {
+				omission, _ := json.Marshal(omittedRepairDiagnostics)
+				return []string{
+					strings.Repeat("x", maxRepairDiagnosticBytes-len(omission)-5),
+					strings.Repeat("y", 128),
+				}
+			}(),
+			wantOmission: true,
+			check: func(t *testing.T, got []string) {
+				t.Helper()
+				if len(got) != 2 || got[1] != omittedRepairDiagnostics {
+					t.Fatalf("diagnostics = %#v", got)
+				}
+			},
+		},
+		{
+			name:   "invalid UTF-8",
+			errors: []string{"broken\xffinput"},
+			check: func(t *testing.T, got []string) {
+				t.Helper()
+				if !reflect.DeepEqual(got, []string{"broken\uFFFDinput"}) {
+					t.Fatalf("diagnostics = %#v", got)
+				}
+			},
+		},
+		{
+			name:         "one huge diagnostic",
+			errors:       []string{strings.Repeat("x", maxRepairDiagnosticBytes*2)},
+			wantOmission: true,
+			check: func(t *testing.T, got []string) {
+				t.Helper()
+				if len(got) != 2 || !strings.HasPrefix(strings.Repeat("x", maxRepairDiagnosticBytes*2), got[0]) {
+					t.Fatalf("diagnostics = %#v", got)
+				}
+			},
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			before := append([]string(nil), tc.errors...)
+			encoded := formatRepairDiagnostics(tc.errors)
+			if len(encoded) > maxRepairDiagnosticBytes || !utf8.ValidString(encoded) {
+				t.Fatalf("encoded diagnostic length/UTF-8 = (%d, %v)", len(encoded), utf8.ValidString(encoded))
+			}
+			var got []string
+			if err := json.Unmarshal([]byte(encoded), &got); err != nil {
+				t.Fatalf("decode diagnostics: %v; encoded=%q", err, encoded)
+			}
+			if !reflect.DeepEqual(tc.errors, before) {
+				t.Fatalf("input diagnostics changed: got %#v, want %#v", tc.errors, before)
+			}
+			if hasOmission := len(got) > 0 && got[len(got)-1] == omittedRepairDiagnostics; hasOmission != tc.wantOmission {
+				t.Fatalf("omission = %v, want %v; diagnostics=%#v", hasOmission, tc.wantOmission, got)
+			}
+			tc.check(t, got)
+		})
+	}
+}
+
+func manyRepairDiagnostics() []string {
+	diagnostics := make([]string, 1_000)
+	for index := range diagnostics {
+		diagnostics[index] = fmt.Sprintf("diagnostic %04d %s", index, strings.Repeat("x", 128))
+	}
+	return diagnostics
+}
+
+func TestDefaultOutputRepairerPropagatesGenerationFailures(t *testing.T) {
+	expected := errors.New("generation failed")
+	for _, tc := range []struct {
+		name   string
+		client *recordingRepairClient
+		want   error
+	}{
+		{name: "nil client", want: nil},
+		{name: "generation error", client: &recordingRepairClient{err: expected}, want: expected},
+		{name: "nil response", client: &recordingRepairClient{}, want: nil},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			var repairer OutputRepairer
+			if tc.client == nil {
+				repairer = NewDefaultOutputRepairer(nil)
+			} else {
+				repairer = NewDefaultOutputRepairer(tc.client)
+			}
+			response, err := repairer.Repair(context.Background(), RepairRequest{Mode: domain.ValidationJSON})
+			if response != nil || err == nil {
+				t.Fatalf("response/error = (%+v, %v)", response, err)
+			}
+			if tc.want != nil && !errors.Is(err, tc.want) {
+				t.Fatalf("error = %v, want %v", err, tc.want)
+			}
+		})
+	}
+}
+
+func repairDiagnosticsFromMessage(t *testing.T, message string) []string {
+	t.Helper()
+	const marker = "Validation diagnostics (data):\n"
+	index := strings.Index(message, marker)
+	if index < 0 {
+		t.Fatalf("missing diagnostics marker in %q", message)
+	}
+	encoded := message[index+len(marker):]
+	var diagnostics []string
+	if err := json.Unmarshal([]byte(encoded), &diagnostics); err != nil {
+		t.Fatalf("decode diagnostics: %v", err)
+	}
+	return diagnostics
+}

internal/usecase/runner.go

@@ -184,10 +184,10 @@ func (r *Runner) executePreparedRun(
 		prepared.StructuredOutput,
 	))
 	if err != nil {
-		if errors.Is(err, llm.ErrInvalidRequest) {
+		return nil, wrapGenerationError(err)
-			return nil, fmt.Errorf("%w: %w", ErrInvalidRequest, err)
 	}
-		return nil, fmt.Errorf("%w: %w", ErrLLMGenerate, err)
+	if genResp == nil {
+		return nil, fmt.Errorf("%w: model returned nil response", ErrLLMGenerate)
 	}
 	usage := genResp.Usage
 
@@ -203,6 +203,7 @@ func (r *Runner) executePreparedRun(
 			attemptsUsed++
 
 			repairResp, repairErr := r.repairer.Repair(ctx, RepairRequest{
+				OriginalMessages: prepared.Messages,
 				PreviousOutput:   genResp.Content,
 				ValidationErrors: validationResult.Errors,
 				SessionID:        prepared.SessionID,
@@ -214,10 +215,10 @@ func (r *Runner) executePreparedRun(
 				Mode:             prepared.OutputContract.ValidationMode,
 			})
 			if repairErr != nil {
-				return nil, fmt.Errorf("%w: %w", ErrValidation, repairErr)
+				return nil, wrapGenerationError(repairErr)
 			}
 			if repairResp == nil {
-				return nil, fmt.Errorf("%w: repairer returned nil response", ErrValidation)
+				return nil, fmt.Errorf("%w: repairer returned nil response", ErrLLMGenerate)
 			}
 
 			genResp = repairResp
@@ -257,6 +258,13 @@ func (r *Runner) executePreparedRun(
 	}, nil
 }
 
+func wrapGenerationError(err error) error {
+	if errors.Is(err, llm.ErrInvalidRequest) {
+		return fmt.Errorf("%w: %w", ErrInvalidRequest, err)
+	}
+	return fmt.Errorf("%w: %w", ErrLLMGenerate, err)
+}
+
 func addTokenUsage(total, next domain.TokenUsage) domain.TokenUsage {
 	return domain.TokenUsage{
 		PromptTokens:     total.PromptTokens + next.PromptTokens,
@@ -524,7 +532,12 @@ func (r *Runner) shouldAttemptRepair(contract domain.OutputContract, validationR
 	if validationResult.Status != domain.ValidationFailed {
 		return false
 	}
-	return contract.ValidationMode == domain.ValidationJSON || contract.ValidationMode == domain.ValidationJSONSchema
+	switch contract.ValidationMode {
+	case domain.ValidationBasic, domain.ValidationJSON, domain.ValidationJSONSchema:
+		return true
+	default:
+		return false
+	}
 }
 
 func mergeExecutionTarget(base domain.ExecutionTarget, override domain.ExecutionTarget) domain.ExecutionTarget {

internal/usecase/runner_test.go

@@ -295,8 +295,10 @@ func (c *controlledRepairLLM) Generate(
 	ctx context.Context,
 	req domain.GenerateRequest,
 ) (*domain.GenerateResponse, error) {
-	isRepair := len(req.Prompt.Messages) > 0 &&
+	messages := req.Prompt.Messages
-		strings.HasPrefix(req.Prompt.Messages[0].Content, "You repair invalid JSON")
+	isRepair := len(messages) >= 2 &&
+		messages[len(messages)-2].Role == "assistant" &&
+		messages[len(messages)-1].Role == "user"
 	c.mu.Lock()
 	c.calls++
 	c.active++
@@ -2143,6 +2145,8 @@ func TestRunnerRepairStateMachine(t *testing.T) {
 		TopP:           true,
 		TimeoutSeconds: true,
 	}
+	emptyThenValid := responses(2)
+	emptyThenValid[0].Content = " \t "
 
 	tests := []struct {
 		name              string
@@ -2165,12 +2169,13 @@ func TestRunnerRepairStateMachine(t *testing.T) {
 			wantStatus:        domain.ValidationPassed,
 		},
 		{
-			name:              "basic failure is ineligible despite budget",
+			name:              "empty basic output repairs successfully",
 			mode:              domain.ValidationBasic,
 			budget:            3,
-			validationResults: []domain.ValidationResult{failed(domain.ValidationBasic, "empty output")},
+			validationResults: []domain.ValidationResult{failed(domain.ValidationBasic, "empty output"), passed(domain.ValidationBasic)},
-			responses:         responses(1),
+			responses:         emptyThenValid,
-			wantStatus:        domain.ValidationFailed,
+			wantRepairs:       1,
+			wantStatus:        domain.ValidationPassed,
 		},
 		{
 			name:   "inherited numeric values remain absent",
@@ -2202,7 +2207,7 @@ func TestRunnerRepairStateMachine(t *testing.T) {
 		{
 			name:   "successful repair stops below larger budget",
 			mode:   domain.ValidationJSON,
-			budget: 4,
+			budget: 3,
 			validationResults: []domain.ValidationResult{
 				failed(domain.ValidationJSON, "candidate zero"),
 				failed(domain.ValidationJSON, "candidate one"),
@@ -2305,6 +2310,9 @@ func TestRunnerRepairStateMachine(t *testing.T) {
 					!reflect.DeepEqual(req.ValidationErrors, tc.validationResults[index].Errors) {
 					t.Fatalf("repair request %d prior state = %+v", index, req)
 				}
+				if !reflect.DeepEqual(req.OriginalMessages, initialRequest.Prompt.Messages) {
+					t.Fatalf("repair request %d original messages drifted: %#v", index, req.OriginalMessages)
+				}
 				if req.TargetPresence != tc.wantPresence || !reflect.DeepEqual(req.Target, initialRequest.Target) ||
 					req.SessionID != initialRequest.Prompt.SessionID ||
 					!reflect.DeepEqual(req.StructuredOutput, initialRequest.StructuredOutput) {
@@ -2318,8 +2326,19 @@ func TestRunnerRepairStateMachine(t *testing.T) {
 					!reflect.DeepEqual(generated.StructuredOutput, initialRequest.StructuredOutput) {
 					t.Fatalf("repair generation request %d common fields drifted: %+v", index, generated)
 				}
-				if reflect.DeepEqual(generated.Prompt.Messages, initialRequest.Prompt.Messages) {
+				expectedMessages := len(initialRequest.Prompt.Messages) + 1
-					t.Fatalf("repair generation request %d reused the initial prompt", index)
+				if strings.TrimSpace(tc.responses[index].Content) != "" {
+					expectedMessages++
+				}
+				if len(generated.Prompt.Messages) != expectedMessages ||
+					generated.Prompt.Messages[len(generated.Prompt.Messages)-1].Role != "user" {
+					t.Fatalf("repair generation request %d messages = %#v", index, generated.Prompt.Messages)
+				}
+				if strings.TrimSpace(tc.responses[index].Content) != "" {
+					assistant := generated.Prompt.Messages[len(generated.Prompt.Messages)-2]
+					if assistant.Role != "assistant" || assistant.Content != tc.responses[index].Content {
+						t.Fatalf("repair generation request %d candidate = %+v", index, assistant)
+					}
 				}
 			}
 

internal/validate/standard_validator.go

@@ -176,7 +176,6 @@ func validateArtifact(ctx context.Context, artifact *domain.Artifact, contract d
 	res := domain.ValidationResult{
 		Mode:       contract.ValidationMode,
 		SchemaPath: contract.SchemaPath,
-		RepairAttempts: contract.RepairAttempts,
 	}
 
 	if artifact == nil {

internal/validate/standard_validator_test.go

@@ -64,6 +64,43 @@ func TestStandardValidatorBasicFailureEmpty(t *testing.T) {
 	}
 }
 
+func TestStandardValidatorReportsZeroRepairAttempts(t *testing.T) {
+	v := NewStandardValidator("")
+
+	for _, tc := range []struct {
+		name     string
+		body     string
+		contract domain.OutputContract
+	}{
+		{
+			name: "passed basic validation",
+			body: "answer",
+			contract: domain.OutputContract{
+				ValidationMode: domain.ValidationBasic,
+				RepairAttempts: 3,
+			},
+		},
+		{
+			name: "failed JSON validation",
+			body: `{"answer":`,
+			contract: domain.OutputContract{
+				ValidationMode: domain.ValidationJSON,
+				RepairAttempts: 3,
+			},
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			res, err := v.Validate(context.Background(), &domain.Artifact{Body: []byte(tc.body)}, tc.contract)
+			if err != nil {
+				t.Fatalf("validate: %v", err)
+			}
+			if res.RepairAttempts != 0 {
+				t.Fatalf("repair attempts = %d, want 0", res.RepairAttempts)
+			}
+		})
+	}
+}
+
 func TestStandardValidatorJSONSuccess(t *testing.T) {
 	v := NewStandardValidator("")
 

output_contract_contract_test.go

@@ -22,13 +22,36 @@ func TestPreparationRejectsInvalidOutputContractWithPublicError(t *testing.T) {
 		t.Fatalf("construct engine: %v", err)
 	}
 
-	req := promptkit.RunRequest{
+	for _, tc := range []struct {
-		PromptID: "prompt",
+		name     string
-		Validation: &promptkit.OutputContract{
+		contract promptkit.OutputContract
+	}{
+		{
+			name: "unsupported format",
+			contract: promptkit.OutputContract{
 				Format:         promptkit.OutputFormat("binary"),
 				ValidationMode: promptkit.ValidationNone,
 			},
-	}
+		},
+		{
+			name: "repair attempts above maximum",
+			contract: promptkit.OutputContract{
+				Format:         promptkit.FormatText,
+				ValidationMode: promptkit.ValidationBasic,
+				RepairAttempts: 4,
+			},
+		},
+		{
+			name: "none validation with repair attempts",
+			contract: promptkit.OutputContract{
+				Format:         promptkit.FormatText,
+				ValidationMode: promptkit.ValidationNone,
+				RepairAttempts: 1,
+			},
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			req := promptkit.RunRequest{PromptID: "prompt", Validation: &tc.contract}
 
 			prepared, err := engine.Prepare(context.Background(), req)
 			if prepared != nil {
@@ -45,4 +68,6 @@ func TestPreparationRejectsInvalidOutputContractWithPublicError(t *testing.T) {
 			if !errors.Is(err, promptkit.ErrInvalidRequest) {
 				t.Fatalf("prepare execution error = %v, want ErrInvalidRequest", err)
 			}
+		})
+	}
 }

prepared_execution_contract_test.go

@@ -256,6 +256,50 @@ func TestPreparedExecutionLifecycleAndEngineBinding(t *testing.T) {
 	}
 }
 
+func TestPreparedExecutionRepairsEmptyBasicOutput(t *testing.T) {
+	client := &preparedRecordingClient{responses: []*promptkit.GenerateResponse{
+		{
+			Content: "",
+			Usage:   promptkit.TokenUsage{PromptTokens: 3, CompletionTokens: 5, TotalTokens: 8},
+		},
+		{
+			Content: "Corrected summary.",
+			Usage:   promptkit.TokenUsage{PromptTokens: 7, CompletionTokens: 11, TotalTokens: 18},
+		},
+	}}
+	engine := newPreparedContractEngine(t, client, "Summarize the source.")
+	prepared, err := engine.PrepareExecution(context.Background(), promptkit.RunRequest{
+		PromptID: "prepared",
+		Validation: &promptkit.OutputContract{
+			Format:         promptkit.FormatMarkdown,
+			ValidationMode: promptkit.ValidationBasic,
+			RepairAttempts: 1,
+		},
+	})
+	if err != nil {
+		t.Fatalf("prepare execution: %v", err)
+	}
+	details := prepared.Details()
+
+	result, err := engine.RunPrepared(context.Background(), prepared)
+	if err != nil {
+		t.Fatalf("run prepared: %v", err)
+	}
+	if result.RawOutput != "Corrected summary." || result.Validation.Status != promptkit.ValidationPassed ||
+		result.Validation.RepairAttempts != 1 || result.Usage != (promptkit.TokenUsage{PromptTokens: 10, CompletionTokens: 16, TotalTokens: 26}) {
+		t.Fatalf("repaired result = %+v", result)
+	}
+	requests := client.snapshot()
+	if len(requests) != 2 || len(requests[1].Prompt.Messages) != len(details.Messages)+1 ||
+		!reflect.DeepEqual(requests[1].Prompt.Messages[:len(details.Messages)], details.Messages) ||
+		requests[1].Prompt.Messages[len(requests[1].Prompt.Messages)-1].Role != "user" {
+		t.Fatalf("prepared repair requests = %#v", requests)
+	}
+	if _, err := engine.RunPrepared(context.Background(), prepared); !errors.Is(err, promptkit.ErrInvalidRequest) {
+		t.Fatalf("second RunPrepared error = %v, want ErrInvalidRequest", err)
+	}
+}
+
 func TestPreparedExecutionConcurrentClaimAllowsOneGeneration(t *testing.T) {
 	release := make(chan struct{})
 	client := &preparedRecordingClient{
@@ -673,6 +717,7 @@ func (r *mutablePreparedArtifactReader) callCount() int {
 type preparedRecordingClient struct {
 	mu        sync.Mutex
 	response  *promptkit.GenerateResponse
+	responses []*promptkit.GenerateResponse
 	err       error
 	requests  []promptkit.GenerateRequest
 	started   chan struct{}
@@ -700,6 +745,12 @@ func (c *preparedRecordingClient) Generate(
 	if c.err != nil {
 		return nil, c.err
 	}
+	if len(c.responses) > 0 {
+		if index := len(c.requests) - 1; index < len(c.responses) {
+			return c.responses[index], nil
+		}
+		return nil, fmt.Errorf("no response configured for generation %d", len(c.requests))
+	}
 	return c.response, nil
 }
 

public_contract_test.go

@@ -844,7 +844,7 @@ func TestBackendExtraParamsAreDeeplyCopiedAtConstructionAndLookup(t *testing.T)
 	}
 }
 
-func TestEngineValidationIsSinglePass(t *testing.T) {
+func TestEngineValidationWithZeroRepairBudgetIsSinglePass(t *testing.T) {
 	client := &fakeLLMClient{
 		response: &promptkit.GenerateResponse{Content: "not-json"},
 	}
@@ -859,7 +859,7 @@ func TestEngineValidationIsSinglePass(t *testing.T) {
 		Validation: &promptkit.OutputContract{
 			Format:         promptkit.FormatJSON,
 			ValidationMode: promptkit.ValidationJSON,
-			RepairAttempts: 3,
+			RepairAttempts: 0,
 		},
 	})
 	if err != nil {
@@ -874,6 +874,87 @@ func TestEngineValidationIsSinglePass(t *testing.T) {
 	}
 }
 
+func TestEngineRunRepairsJSONSchemaOutput(t *testing.T) {
+	client := &fakeLLMClient{responses: []*promptkit.GenerateResponse{
+		{
+			Content: "{}",
+			Usage:   promptkit.TokenUsage{PromptTokens: 2, CompletionTokens: 3, TotalTokens: 5, CachedTokens: 7, CacheWriteTokens: 11},
+		},
+		{
+			Content: `{"events":[{"title":"Repaired event"}]}`,
+			Usage:   promptkit.TokenUsage{PromptTokens: 13, CompletionTokens: 17, TotalTokens: 19, CachedTokens: 23, CacheWriteTokens: 29},
+		},
+	}}
+	engine := newContractEngineWithOptions(t, frameworkSchemaDir, promptkit.WithLLMClient(client))
+
+	result, err := engine.Run(context.Background(), promptkit.RunRequest{
+		PromptID:  frameworkStructuredEventsPromptID,
+		SessionID: " repair-session ",
+		Inputs: map[string]promptkit.ArtifactRef{
+			"transcript": promptkit.Inline("Rin opens the gate."),
+			"glossary":   promptkit.Inline("gate: A guarded passage."),
+		},
+		Validation: &promptkit.OutputContract{
+			Format:         promptkit.FormatJSON,
+			ValidationMode: promptkit.ValidationJSONSchema,
+			SchemaPath:     "structured_events.schema.json",
+			RepairAttempts: 1,
+		},
+	})
+	if err != nil {
+		t.Fatalf("run: %v", err)
+	}
+	if result.RawOutput != client.responses[1].Content || result.Validation.Status != promptkit.ValidationPassed ||
+		result.Validation.RepairAttempts != 1 {
+		t.Fatalf("repaired result = %+v", result)
+	}
+	wantUsage := promptkit.TokenUsage{PromptTokens: 15, CompletionTokens: 20, TotalTokens: 24, CachedTokens: 30, CacheWriteTokens: 40}
+	if result.Usage != wantUsage {
+		t.Fatalf("usage = %+v, want %+v", result.Usage, wantUsage)
+	}
+	if len(client.requests) != 2 {
+		t.Fatalf("generation calls = %d, want 2", len(client.requests))
+	}
+	initial, repaired := client.requests[0], client.requests[1]
+	if initial.Prompt.SessionID != "repair-session" || repaired.Prompt.SessionID != initial.Prompt.SessionID ||
+		!reflect.DeepEqual(repaired.Target, initial.Target) || repaired.TargetPresence != initial.TargetPresence ||
+		!reflect.DeepEqual(repaired.StructuredOutput, initial.StructuredOutput) {
+		t.Fatalf("generation request state drifted: initial=%+v repaired=%+v", initial, repaired)
+	}
+	if initial.StructuredOutput == nil || initial.StructuredOutput.JSONSchema == nil {
+		t.Fatalf("expected structured output on initial request: %+v", initial)
+	}
+}
+
+func TestEngineRunReturnsFinalResultAfterRepairExhaustion(t *testing.T) {
+	client := &fakeLLMClient{responses: []*promptkit.GenerateResponse{
+		{Content: "not-json", Usage: promptkit.TokenUsage{PromptTokens: 2, CompletionTokens: 3, TotalTokens: 5}},
+		{Content: "still-not-json", Usage: promptkit.TokenUsage{PromptTokens: 7, CompletionTokens: 11, TotalTokens: 18}},
+	}}
+	engine := newContractEngineWithOptions(t, frameworkSchemaDir, promptkit.WithLLMClient(client))
+
+	result, err := engine.Run(context.Background(), promptkit.RunRequest{
+		PromptID: frameworkMarkdownSummaryPromptID,
+		Inputs: map[string]promptkit.ArtifactRef{
+			"transcript": promptkit.Inline("Rin opens the gate."),
+			"glossary":   promptkit.Inline("gate: A guarded passage."),
+		},
+		Validation: &promptkit.OutputContract{
+			Format:         promptkit.FormatJSON,
+			ValidationMode: promptkit.ValidationJSON,
+			RepairAttempts: 1,
+		},
+	})
+	if err != nil || result == nil {
+		t.Fatalf("run = (%+v, %v), want exhausted result", result, err)
+	}
+	if result.RawOutput != "still-not-json" || result.Validation.Status != promptkit.ValidationFailed ||
+		result.Validation.RepairAttempts != 1 || len(result.Validation.Errors) == 0 ||
+		result.Usage != (promptkit.TokenUsage{PromptTokens: 9, CompletionTokens: 14, TotalTokens: 23}) {
+		t.Fatalf("exhausted result = %+v", result)
+	}
+}
+
 func TestRepeatedOptionsUseLastValueInEachCategory(t *testing.T) {
 	profile := promptkit.Profile{ID: "profile", Endpoint: "http://example.test/v1", Model: "model"}
 

types.go

@@ -565,8 +565,8 @@ type ExecutionTargetPresence struct {
 // JSON representation.
 //
 // A non-nil RunRequest.Validation replaces the complete prompt contract. It
-// does not merge fields. The public Engine validates generated output once and
+// does not merge fields. The public Engine performs bounded correction after a
-// does not install an output repairer.
+// failed eligible validation when RepairAttempts is positive.
 type OutputContract struct {
 	// Format selects generated artifact metadata. An empty value in a non-nil
 	// request replacement defaults to FormatText.
@@ -577,9 +577,9 @@ type OutputContract struct {
 	// SchemaPath is required when ValidationMode is ValidationJSONSchema and is
 	// ignored by other modes.
 	SchemaPath string `json:"schema_path"`
-	// RepairAttempts is a non-negative requested repair limit. Zero requests no
+	// RepairAttempts is an additional generation-call budget from zero through
-	// repairs. The public Engine performs no repairs even when this value is
+	// three. Zero is single-pass. A positive value is valid only with basic,
-	// positive, so its runs report zero attempts used.
+	// json, or json_schema validation.
 	RepairAttempts int `json:"repair_attempts"`
 }
 
@@ -595,8 +595,8 @@ type ValidationResult struct {
 	Errors []string `json:"errors,omitempty"`
 	// SchemaPath is the effective schema path for JSON Schema validation.
 	SchemaPath string `json:"schema_path,omitempty"`
-	// RepairAttempts is the number of repairs actually attempted. It is always
+	// RepairAttempts is the number of corrective generation calls actually
-	// zero for the public Engine.
+	// started for this result.
 	RepairAttempts int `json:"repair_attempts"`
 	// IsValid is true for ValidationPassed and ValidationSkipped and false for
 	// ValidationFailed.
@@ -714,9 +714,8 @@ type GenerateRequest struct {
 // GenerateResponse is returned by an injected LLM client and has a stable JSON
 // representation.
 type GenerateResponse struct {
-	// Content is the generated output. It must be non-empty when using the
+	// Content is the generated output. It may be explicitly empty; Promptkit
-	// built-in client; injected clients may return empty content for Promptkit
+	// applies the effective output contract to classify it.
-	// validation to classify.
 	Content string `json:"content"`
 	// Usage is the client's token accounting.
 	Usage TokenUsage `json:"usage"`