fix: persist OpenAI reasoning rejections in chat history

chatd kept rejected encrypted reasoning in process memory, so a restart
or another replica forgot it. The next rejection then marked every
reasoning item in the prompt as rejected, including reasoning produced
after the credential change, and the memory grew until the process
exited.

Reasoning parts now also record the credential that produced them (the
deployment's or the user's personal key) and a rejection epoch, next to
the provider and model. A turn on a different credential leaves out the
other credential's reasoning before sending. When a provider rejects
encrypted reasoning, chatd retries without reasoning as before and
stamps the step's new reasoning with the provider's next epoch, so later
steps on any replica leave out the older reasoning and replay the newer.

A retried step that returns no reasoning records no new epoch, so the
next step is rejected and retried once more.
This commit is contained in:
Michael Suchacz committed 2026-10-12 15:16:09 +00:00
1 parent 678d6f0c1e
commit 5fb3ce1689
18 files changed
+398 -100

No files matched your search

+8
View File
@@ -21311,6 +21311,10 @@ const docTemplate = `{
}
}
},
"provider_credential": {
"description": "ProviderCredential records whether a reasoning part was produced\nwith the deployment's credential or the user's personal API key,\nbecause encrypted reasoning only decrypts for the organization\nbehind that credential. Internal only: stripped before API responses.",
"type": "string"
},
"provider_executed": {
"description": "ProviderExecuted indicates the tool call was executed by\nthe provider (e.g. Anthropic computer use).",
"type": "boolean"
@@ -21333,6 +21337,10 @@ const docTemplate = `{
"reasoning_delta": {
"type": "string"
},
"reasoning_epoch": {
"description": "ReasoningEpoch records how many times the producing provider had\nrejected the chat's encrypted reasoning before this reasoning part\nwas produced. Internal only: stripped before API responses.",
"type": "integer"
},
"result": {
"type": "array",
"items": {
+8
View File
@@ -19241,6 +19241,10 @@
}
}
},
"provider_credential": {
"description": "ProviderCredential records whether a reasoning part was produced\nwith the deployment's credential or the user's personal API key,\nbecause encrypted reasoning only decrypts for the organization\nbehind that credential. Internal only: stripped before API responses.",
"type": "string"
},
"provider_executed": {
"description": "ProviderExecuted indicates the tool call was executed by\nthe provider (e.g. Anthropic computer use).",
"type": "boolean"
@@ -19263,6 +19267,10 @@
"reasoning_delta": {
"type": "string"
},
"reasoning_epoch": {
"description": "ReasoningEpoch records how many times the producing provider had\nrejected the chat's encrypted reasoning before this reasoning part\nwas produced. Internal only: stripped before API responses.",
"type": "integer"
},
"result": {
"type": "array",
"items": {
-5
View File
@@ -211,11 +211,6 @@ type Server struct {
// model that accepted Anthropic's thinking drop_block control.
thinkingDropBlock sync.Map
// rejectedReasoning holds the rejectedReasoningKey of each OpenAI
// reasoning item whose encrypted content a provider rejected. It grows
// only on rejections, which follow a credential change.
rejectedReasoning sync.Map
// Configuration
inFlightChatStaleAfter time.Duration
streamSilenceTimeout time.Duration
+1
View File
@@ -1903,6 +1903,7 @@ var partNulFields = []partNulField{
{name: "ParsedCommands", policy: nulReject, grid: func(p *codersdk.ChatMessagePart) [][]string { return p.ParsedCommands }},
{name: "ProviderIdentity", policy: nulReject, str: func(p *codersdk.ChatMessagePart) *string { return &p.ProviderIdentity }},
{name: "ProviderModel", policy: nulReject, str: func(p *codersdk.ChatMessagePart) *string { return &p.ProviderModel }},
{name: "ProviderCredential", policy: nulReject, str: func(p *codersdk.ChatMessagePart) *string { return &p.ProviderCredential }},
{name: "SourceID", policy: nulReject, str: func(p *codersdk.ChatMessagePart) *string { return &p.SourceID }},
{name: "URL", policy: nulReject, str: func(p *codersdk.ChatMessagePart) *string { return &p.URL }},
{name: "MediaType", policy: nulReject, str: func(p *codersdk.ChatMessagePart) *string { return &p.MediaType }},
+3 -19
View File
@@ -5,7 +5,6 @@ import (
"fmt"
"charm.land/fantasy"
fantasyopenai "charm.land/fantasy/providers/openai"
"cdr.dev/slog/v3"
"github.com/coder/coder/v2/coderd/database"
@@ -136,30 +135,15 @@ func dropOpenAIReasoningParts(
ctx context.Context,
logger slog.Logger,
messages []fantasy.Message,
) []fantasy.Message {
return dropOpenAIReasoningItems(ctx, logger, messages, func(string) bool { return true })
}
// dropOpenAIReasoningItems removes the OpenAI reasoning parts whose item ID
// satisfies drop from a copy of messages. Messages emptied by the drop are
// removed.
func dropOpenAIReasoningItems(
ctx context.Context,
logger slog.Logger,
messages []fantasy.Message,
drop func(itemID string) bool,
) []fantasy.Message {
dropped := 0
out := make([]fantasy.Message, 0, len(messages))
for _, msg := range messages {
parts := make([]fantasy.MessagePart, 0, len(msg.Content))
for _, part := range msg.Content {
if reasoning, ok := part.(fantasy.ReasoningPart); ok {
metadata := fantasyopenai.GetReasoningMetadata(reasoning.ProviderOptions)
if metadata != nil && metadata.ItemID != "" && drop(metadata.ItemID) {
dropped++
continue
}
if reasoning, ok := part.(fantasy.ReasoningPart); ok && chatsanitize.HasOpenAIReasoningState(reasoning.ProviderOptions) {
dropped++
continue
}
parts = append(parts, part)
}
+13 -33
View File
@@ -3,7 +3,7 @@ package chatd
import (
"context"
"errors"
"sync"
"sync/atomic"
"charm.land/fantasy"
"github.com/google/uuid"
@@ -18,45 +18,35 @@ func isInvalidEncryptedContentError(err error) bool {
return errors.As(err, &apiErr) && apiErr.Code == "invalid_encrypted_content"
}
type rejectedReasoningKey struct {
chatID uuid.UUID
itemID string
}
// withEncryptedReasoningFallback wraps OpenAI Responses models in an
// encryptedReasoningFallbackModel.
func (p *Server) withEncryptedReasoningFallback(model chatprovider.Model, chatID uuid.UUID) chatprovider.Model {
// encryptedReasoningFallbackModel that sets rejected when it retries.
func (p *Server) withEncryptedReasoningFallback(model chatprovider.Model, chatID uuid.UUID, rejected *atomic.Bool) chatprovider.Model {
if !model.Transport().UsesResponses() {
return model
}
return model.WithLanguageModel(&encryptedReasoningFallbackModel{
LanguageModel: model.LanguageModel(),
logger: p.logger.With(slog.F("chat_id", chatID)),
rejected: &p.rejectedReasoning,
chatID: chatID,
rejected: rejected,
})
}
// encryptedReasoningFallbackModel retries a request without OpenAI reasoning
// when the provider rejects its encrypted content. Encrypted reasoning only
// decrypts for the OpenAI organization that produced it, which can change
// without a provider or model change, for example when a user adds or
// removes a personal API key. The rejected reasoning stays in the chat
// history, so the model remembers its item IDs and leaves them out of later
// requests, while reasoning produced after the change still replays.
// without a provider, model, or credential change, for example when an admin
// replaces the deployment key with one from another organization. The
// rejected reasoning stays in the chat history, so the step stamps its new
// reasoning with the next epoch, and later steps leave out reasoning from
// earlier epochs.
type encryptedReasoningFallbackModel struct {
fantasy.LanguageModel
logger slog.Logger
rejected *sync.Map
chatID uuid.UUID
rejected *atomic.Bool
}
// Stream retries only when the rejection arrives before any output.
func (m *encryptedReasoningFallbackModel) Stream(ctx context.Context, call fantasy.Call) (fantasy.StreamResponse, error) {
call.Prompt = dropOpenAIReasoningItems(ctx, m.logger, call.Prompt, func(itemID string) bool {
_, rejected := m.rejected.Load(rejectedReasoningKey{chatID: m.chatID, itemID: itemID})
return rejected
})
stream, err := m.LanguageModel.Stream(ctx, call)
if err != nil {
return nil, err
@@ -77,19 +67,9 @@ func (m *encryptedReasoningFallbackModel) Stream(ctx context.Context, call fanta
if rejection == nil {
return
}
// The error names a single item, but a credential change
// invalidates all earlier reasoning, so every remaining item
// counts as rejected.
remembered := 0
call.Prompt = dropOpenAIReasoningItems(ctx, m.logger, call.Prompt, func(itemID string) bool {
m.rejected.Store(rejectedReasoningKey{chatID: m.chatID, itemID: itemID}, struct{}{})
remembered++
return true
})
m.logger.Warn(ctx, "retrying model call without rejected encrypted reasoning",
slog.F("rejected_items", remembered),
slogError(rejection),
)
m.rejected.Store(true)
m.logger.Warn(ctx, "retrying model call without rejected encrypted reasoning", slogError(rejection))
call.Prompt = dropOpenAIReasoningParts(ctx, m.logger, call.Prompt)
retry, err := m.LanguageModel.Stream(ctx, call)
if err != nil {
yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeError, Error: err})
+17 -2
View File
@@ -6,6 +6,7 @@ import (
"errors"
"fmt"
"strings"
"sync/atomic"
"time"
"charm.land/fantasy"
@@ -73,6 +74,9 @@ type generationPrepared struct {
// user-facing errors. See chatloop.GenerateAssistantOptions.ErrorProvider.
ResolvedProvider string
ReasoningProvenance reasoningProvenance
// ReasoningRejected reports whether the provider rejected the
// encrypted reasoning replayed into this step's model call.
ReasoningRejected *atomic.Bool
StageModel chatloop.StageModel
@@ -94,6 +98,17 @@ type generationPrepared struct {
Debug *generationDebug
}
// stepReasoningProvenance returns the provenance to stamp on the step's
// reasoning. A rejection starts the provider's next epoch, so later steps
// leave out the reasoning the rejected call replayed.
func (p generationPrepared) stepReasoningProvenance() reasoningProvenance {
provenance := p.ReasoningProvenance
if p.ReasoningRejected != nil && p.ReasoningRejected.Load() {
provenance.Epoch++
}
return provenance
}
// generationCompaction contains compaction inputs prepared for generation.
type generationCompaction struct {
// Override, when non-nil, is the compaction model override resolved at
@@ -967,7 +982,7 @@ func (s *taskStarter) generateAssistant(
logger: s.opts.Logger,
contentVersion: chatprompt.CurrentContentVersion,
hookRewrittenToolCalls: preflight.Overrides,
reasoningProvenance: prepared.ReasoningProvenance,
reasoningProvenance: prepared.stepReasoningProvenance(),
})
if err != nil {
return s.finishGenerationError(ctx, machine, input, err, requireGenerationAttempt(attempt.number))
@@ -1202,7 +1217,7 @@ func (s *taskStarter) executeLocalTools(
toolNameToConfigID: prepared.ToolNameToConfigID,
logger: s.opts.Logger,
contentVersion: chatprompt.CurrentContentVersion,
reasoningProvenance: prepared.ReasoningProvenance,
reasoningProvenance: prepared.stepReasoningProvenance(),
})
if err != nil {
return s.finishGenerationError(ctx, machine, input, err, requireGenerationAttempt(attempt.number))
+13 -6
View File
@@ -175,7 +175,11 @@ func (server *Server) prepareGeneration(
}
model := resolved.model
callConfig := resolved.callConfig
reasoningSource := reasoningProvenance{ProviderIdentity: providerIdentity, Model: model.ModelID()}
reasoningSource := reasoningProvenance{
ProviderIdentity: providerIdentity,
Model: model.ModelID(),
Credential: reasoningCredential(resolved.route.ProviderAuth),
}
modelRoute := resolved.route
currentPlanMode := chat.PlanMode
@@ -329,11 +333,12 @@ func (server *Server) prepareGeneration(
)
// Drop provider-executed tool history produced by a different provider,
// and OpenAI reasoning state produced by a different provider or model,
// before building the prompt. A provider that shares another's wire
// format (e.g. Bedrock and Anthropic) can still reject the other's
// provider-specific blocks, so a mid-chat provider switch must not
// replay them.
// and OpenAI reasoning state produced by a different provider, model,
// or credential, or before the provider's latest rejection of the
// chat's encrypted reasoning, before building the prompt. A provider
// that shares another's wire format (e.g. Bedrock and Anthropic) can
// still reject the other's provider-specific blocks, so a mid-chat
// provider switch must not replay them.
//
// The pending-user segment (trailing user rows the assistant has not
// answered yet) is handled separately from here through prompt
@@ -346,6 +351,7 @@ func (server *Server) prepareGeneration(
// sanitizing everything, since the tail is user-only and the
// sanitizer only rewrites assistant rows.
pendingRowsStart := pendingUserSegmentStart(promptRows)
reasoningSource.Epoch = currentReasoningEpoch(promptRows[:pendingRowsStart], providerIdentity)
sanitizedHead, err := server.sanitizeForeignProviderStateRows(ctx, logger, promptRows[:pendingRowsStart], chat.OwnerID, reasoningSource)
if err != nil {
return generationPrepared{}, err
@@ -907,6 +913,7 @@ func (server *Server) prepareGeneration(
ModelBuildOptions: modelOpts,
ResolvedProvider: resolved.resolvedProvider,
ReasoningProvenance: reasoningSource,
ReasoningRejected: resolved.reasoningRejected,
StageModel: resolved.stageModel(),
ModelConfigID: modelConfig.ID,
CallTemplate: resolved.newCall(),
+118 -16
View File
@@ -376,9 +376,26 @@ func TestOpenAIResponsesRetriesWithoutRejectedEncryptedReasoning(t *testing.T) {
model := insertChatModelConfigWithCallConfig(t, db, user.ID, "openai", "gpt-4o",
codersdk.ChatModelCallConfig{})
factory := chattest.NewMockAIBridgeTransport(t, openAIURL)
server := newOpenAIResponsesTestServer(t, db, ps, func(cfg *chatd.Config) {
cfg.AIBridgeTransportFactory = chatAIGatewayTransportFactoryPointer(factory)
})
newServer := func() *chatd.Server {
return newOpenAIResponsesTestServer(t, db, ps, func(cfg *chatd.Config) {
cfg.AIBridgeTransportFactory = chatAIGatewayTransportFactoryPointer(factory)
})
}
server := newServer()
sendMessage := func(server *chatd.Server, chatID uuid.UUID, text string) {
t.Helper()
_, err := server.SendMessage(ctx, chatd.SendMessageOptions{
ChatID: chatID,
CreatedBy: user.ID,
ModelConfigID: model.ID,
Content: []codersdk.ChatMessagePart{
codersdk.ChatMessageText(text),
},
})
require.NoError(t, err)
waitForChatProcessed(ctx, t, db, chatID, server)
requireResponsesChatWaiting(ctx, t, db, chatID)
}
chat, err := server.CreateChat(ctx, chatd.CreateOptions{
OrganizationID: org.ID,
@@ -392,8 +409,94 @@ func TestOpenAIResponsesRetriesWithoutRejectedEncryptedReasoning(t *testing.T) {
require.NoError(t, err)
waitForChatProcessed(ctx, t, db, chat.ID, server)
requireResponsesChatWaiting(ctx, t, db, chat.ID)
sendMessage(server, chat.ID, "list the templates")
for _, text := range []string{"list the templates", "thanks"} {
// The next turn runs on a server that never saw the rejection, like
// a restarted or different replica.
require.NoError(t, server.Close())
sendMessage(newServer(), chat.ID, "thanks")
// Only the first request after the change is rejected.
requests := recorder.all()
require.Len(t, requests, 5)
requireInlineReasoningItem(t, requests[1].Prompt, staleID, staleBlob, "")
retry := requests[2]
require.NotEmpty(t, retry.Prompt)
require.NotContains(t, promptItemTypes(retry.Prompt), "reasoning")
// requests[3] is the step after the tool call; requests[4] is the next
// turn on the fresh server.
for _, request := range requests[3:] {
requireInlineReasoningItem(t, request.Prompt, freshID, freshBlob, "")
requireNoResponsesProviderItemReplay(t, request.Prompt, staleID)
}
}
func TestOpenAIResponsesReasoningNotReplayedAfterCredentialChange(t *testing.T) {
t.Parallel()
db, ps := dbtestutil.NewDB(t)
ctx := testutil.Context(t, testutil.WaitLong)
const (
userAPIKey = "sk-user-reasoning"
deploymentID = "rs_deployment_key"
deploymentBlob = "encrypted-by-deployment-organization"
personalKeyID = "rs_personal_key"
personalKeyBlob = "encrypted-by-personal-organization"
)
var (
recorder responsesRequestRecorder
authMu sync.Mutex
authByCall = map[int]string{}
)
openAIURL := chattest.NewOpenAI(t, func(req *chattest.OpenAIRequest) chattest.OpenAIResponse {
if !req.Stream {
return chattest.OpenAINonStreamingResponse("title")
}
n := recorder.record(req)
authMu.Lock()
authByCall[n] = req.Header.Get("Authorization")
authMu.Unlock()
resp := chattest.OpenAIStreamingResponse(chattest.OpenAITextChunks("answer")...)
switch n {
case 1:
resp.Reasoning = &chattest.OpenAIReasoningItem{ID: deploymentID, EncryptedContent: deploymentBlob}
case 2:
resp.Reasoning = &chattest.OpenAIReasoningItem{ID: personalKeyID, EncryptedContent: personalKeyBlob}
}
return resp
})
user, org, _ := seedChatDependenciesWithProvider(t, db, "openai", openAIURL)
model := insertChatModelConfigWithCallConfig(t, db, user.ID, "openai", "gpt-4o",
codersdk.ChatModelCallConfig{})
factory := chattest.NewMockAIBridgeTransport(t, openAIURL)
server := newOpenAIResponsesTestServer(t, db, ps, func(cfg *chatd.Config) {
cfg.AIBridgeTransportFactory = chatAIGatewayTransportFactoryPointer(factory)
})
chat, err := server.CreateChat(ctx, chatd.CreateOptions{
OrganizationID: org.ID,
OwnerID: user.ID,
Title: uniqueResponsesTitle(t, "credential-change"),
ModelConfigID: model.ID,
InitialUserContent: []codersdk.ChatMessagePart{
codersdk.ChatMessageText("hello"),
},
})
require.NoError(t, err)
waitForChatProcessed(ctx, t, db, chat.ID, server)
requireResponsesChatWaiting(ctx, t, db, chat.ID)
_, err = db.UpsertUserAIProviderKey(ctx, database.UpsertUserAIProviderKeyParams{
ID: uuid.New(),
UserID: user.ID,
AIProviderID: model.AIProviderID.UUID,
APIKey: userAPIKey,
})
require.NoError(t, err)
for _, text := range []string{"switched keys", "thanks"} {
_, err = server.SendMessage(ctx, chatd.SendMessageOptions{
ChatID: chat.ID,
CreatedBy: user.ID,
@@ -407,19 +510,18 @@ func TestOpenAIResponsesRetriesWithoutRejectedEncryptedReasoning(t *testing.T) {
requireResponsesChatWaiting(ctx, t, db, chat.ID)
}
// Only the first request after the change is rejected.
requests := recorder.all()
require.Len(t, requests, 5)
requireInlineReasoningItem(t, requests[1].Prompt, staleID, staleBlob, "")
retry := requests[2]
require.NotEmpty(t, retry.Prompt)
require.NotContains(t, promptItemTypes(retry.Prompt), "reasoning")
// requests[3] continues the tool step from in-memory history;
// requests[4] rebuilds the prompt from persisted messages.
for _, request := range requests[3:] {
requireInlineReasoningItem(t, request.Prompt, freshID, freshBlob, "")
requireNoResponsesProviderItemReplay(t, request.Prompt, staleID)
}
require.Len(t, requests, 3)
authMu.Lock()
require.NotEqual(t, "Bearer "+userAPIKey, authByCall[1])
require.Equal(t, "Bearer "+userAPIKey, authByCall[2])
authMu.Unlock()
// Reasoning from the deployment's key is left out before the
// personal key's organization can reject it.
require.NotContains(t, promptItemTypes(requests[1].Prompt), "reasoning")
requireNoResponsesProviderItemReplay(t, requests[1].Prompt, deploymentID)
requireInlineReasoningItem(t, requests[2].Prompt, personalKeyID, personalKeyBlob, "")
requireNoResponsesProviderItemReplay(t, requests[2].Prompt, deploymentID)
}
func TestOpenAIResponsesAdvisorPromptOmitsReasoning(t *testing.T) {
+2
View File
@@ -55,6 +55,8 @@ func buildCommitStepMessages(input buildCommitStepMessagesInput) (stepMessagesFo
if assistantParts[i].Type == codersdk.ChatMessagePartTypeReasoning {
assistantParts[i].ProviderIdentity = input.reasoningProvenance.ProviderIdentity
assistantParts[i].ProviderModel = input.reasoningProvenance.Model
assistantParts[i].ProviderCredential = input.reasoningProvenance.Credential
assistantParts[i].ReasoningEpoch = input.reasoningProvenance.Epoch
}
}
+13 -4
View File
@@ -30,10 +30,15 @@ func TestBuildCommitStepMessages_AssistantTextAndReasoning(t *testing.T) {
startedAt := time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC)
completedAt := startedAt.Add(2 * time.Second)
got, err := buildCommitStepMessages(buildCommitStepMessagesInput{
modelConfigID: modelConfigID,
contentVersion: chatprompt.CurrentContentVersion,
logger: slog.Make(),
reasoningProvenance: reasoningProvenance{ProviderIdentity: "provider-a", Model: "model-a"},
modelConfigID: modelConfigID,
contentVersion: chatprompt.CurrentContentVersion,
logger: slog.Make(),
reasoningProvenance: reasoningProvenance{
ProviderIdentity: "provider-a",
Model: "model-a",
Credential: reasoningCredentialUser,
Epoch: 2,
},
step: stepData{
Content: []fantasy.Content{
fantasy.ReasoningContent{Text: "thinking"},
@@ -59,10 +64,14 @@ func TestBuildCommitStepMessages_AssistantTextAndReasoning(t *testing.T) {
require.Equal(t, completedAt, requireNotNilTime(t, parts[0].CompletedAt))
require.Equal(t, "provider-a", parts[0].ProviderIdentity)
require.Equal(t, "model-a", parts[0].ProviderModel)
require.Equal(t, reasoningCredentialUser, parts[0].ProviderCredential)
require.Equal(t, 2, parts[0].ReasoningEpoch)
require.Equal(t, codersdk.ChatMessagePartTypeText, parts[1].Type)
require.Equal(t, "hello", parts[1].Text)
require.Empty(t, parts[1].ProviderIdentity)
require.Empty(t, parts[1].ProviderModel)
require.Empty(t, parts[1].ProviderCredential)
require.Zero(t, parts[1].ReasoningEpoch)
}
func TestBuildCommitStepMessages_LocalToolResultsBecomeToolMessages(t *testing.T) {
+6 -1
View File
@@ -2,6 +2,7 @@ package chatd
import (
"context"
"sync/atomic"
"charm.land/fantasy"
"golang.org/x/xerrors"
@@ -72,6 +73,9 @@ type resolvedModelCall struct {
resolvedEffort string
route aiGatewayModelRoute
debugEnabled bool
// reasoningRejected reports whether the provider rejected encrypted
// reasoning replayed into a call to model.
reasoningRejected *atomic.Bool
}
// stageModel uses the configured provider type rather than
@@ -182,7 +186,8 @@ func (p *Server) resolveModelCall(ctx context.Context, spec modelCallSpec) (reso
return resolvedModelCall{}, xerrors.Errorf("create model: %w", err)
}
model = p.withThinkingDropBlock(model, out.route.Provider.ID, clientCallConfig, spec.chat.ID)
model = p.withEncryptedReasoningFallback(model, spec.chat.ID)
out.reasoningRejected = &atomic.Bool{}
model = p.withEncryptedReasoningFallback(model, spec.chat.ID, out.reasoningRejected)
if out.debugEnabled {
model = model.WithLanguageModel(chatdebug.WrapModel(model.LanguageModel(), debugSvc, chatdebug.RecorderOptions{
+61 -7
View File
@@ -38,13 +38,62 @@ func modelConfigProviderIdentity(modelConfig database.ChatModelConfig, normalize
return normalizedProvider
}
// reasoningProvenance identifies the provider instance and model that
// produced OpenAI reasoning state. Reasoning item IDs only resolve on the
// issuing provider instance, and encrypted content can be rejected by a
// different model.
const (
reasoningCredentialDeployment = "deployment"
reasoningCredentialUser = "user"
)
// reasoningProvenance identifies where OpenAI reasoning state was produced.
// Reasoning item IDs only resolve on the issuing provider instance, and
// encrypted content can be rejected by a different model or by the
// organization behind a different credential. Epoch counts the provider's
// rejections of the chat's encrypted reasoning, so reasoning replayed into
// a rejected call is never replayed again.
type reasoningProvenance struct {
ProviderIdentity string
Model string
Credential string
Epoch int
}
// reasoningCredential names the credential an AI Gateway route sends.
func reasoningCredential(auth aiGatewayProviderAuth) string {
if len(auth.Headers) > 0 {
return reasoningCredentialUser
}
return reasoningCredentialDeployment
}
// unstampedReasoningProvenance is the provenance assumed for reasoning saved
// without a stamp: the row's model config, the target's credential, and the
// first epoch.
func unstampedReasoningProvenance(rowOrigin, target reasoningProvenance) reasoningProvenance {
return reasoningProvenance{
ProviderIdentity: rowOrigin.ProviderIdentity,
Model: rowOrigin.Model,
Credential: target.Credential,
}
}
// currentReasoningEpoch returns the highest epoch stamped on providerIdentity's
// reasoning in rows.
func currentReasoningEpoch(rows []database.ChatMessage, providerIdentity string) int {
epoch := 0
for _, row := range rows {
if row.Role != database.ChatMessageRoleAssistant || !bytes.Contains(row.Content.RawMessage, []byte(`"reasoning_epoch"`)) {
continue
}
parts, err := chatprompt.ParseContent(row)
if err != nil {
continue
}
for _, part := range parts {
if part.Type == codersdk.ChatMessagePartTypeReasoning && part.ProviderIdentity == providerIdentity {
epoch = max(epoch, part.ReasoningEpoch)
}
}
}
return epoch
}
// stripForeignProviderStateRows drops provider-executed tool blocks (calls and
@@ -76,7 +125,7 @@ func stripForeignProviderStateRows(
origin := originOf(row.ModelConfigID)
nativeRow := origin.ProviderIdentity == target.ProviderIdentity
// Only stamped reasoning parts can be foreign when the row origin matches.
if origin == target && !bytes.Contains(row.Content.RawMessage, []byte(`"provider_identity"`)) {
if unstampedReasoningProvenance(origin, target) == target && !bytes.Contains(row.Content.RawMessage, []byte(`"provider_identity"`)) {
out = append(out, row)
continue
}
@@ -126,9 +175,14 @@ func stripForeignProviderStateRows(
}
func isForeignReasoning(part codersdk.ChatMessagePart, rowOrigin, target reasoningProvenance) bool {
source := rowOrigin
source := unstampedReasoningProvenance(rowOrigin, target)
if part.ProviderIdentity != "" {
source = reasoningProvenance{ProviderIdentity: part.ProviderIdentity, Model: part.ProviderModel}
source = reasoningProvenance{
ProviderIdentity: part.ProviderIdentity,
Model: part.ProviderModel,
Credential: part.ProviderCredential,
Epoch: part.ReasoningEpoch,
}
}
return source != target && hasOpenAIReasoningState(part)
}
@@ -332,6 +332,65 @@ func TestStripForeignProviderStateRows(t *testing.T) {
require.Equal(t, rows, got)
require.Zero(t, stats)
})
stampedFull := func(source reasoningProvenance) codersdk.ChatMessagePart {
p := stampedReasoning(source)
p.ProviderCredential = source.Credential
p.ReasoningEpoch = source.Epoch
return p
}
t.Run("stamped reasoning from another credential dropped", func(t *testing.T) {
t.Parallel()
deployment := openAIInstance
deployment.Credential = reasoningCredentialDeployment
user := openAIInstance
user.Credential = reasoningCredentialUser
rows := []database.ChatMessage{
assistantRow(t, openAIInstanceCfg, stampedFull(deployment), text("deployment")),
assistantRow(t, openAIInstanceCfg, stampedFull(user), text("user")),
}
got, stats := stripForeignProviderStateRows(rows, user, resolver)
require.Len(t, got, 2)
require.Equal(t, []codersdk.ChatMessagePart{text("deployment")}, partsOf(t, got[0]))
require.Equal(t, rows[1], got[1])
require.Equal(t, providerSwitchStripStats{RemovedReasoning: 1}, stats)
})
t.Run("reasoning from an earlier epoch dropped", func(t *testing.T) {
t.Parallel()
current := openAIInstance
current.Epoch = 1
rows := []database.ChatMessage{
assistantRow(t, openAIInstanceCfg, openAIReasoning, text("unstamped")),
assistantRow(t, openAIInstanceCfg, stampedFull(openAIInstance), text("rejected")),
assistantRow(t, openAIInstanceCfg, stampedFull(current), text("current")),
}
got, stats := stripForeignProviderStateRows(rows, current, resolver)
require.Len(t, got, 3)
require.Equal(t, []codersdk.ChatMessagePart{text("unstamped")}, partsOf(t, got[0]))
require.Equal(t, []codersdk.ChatMessagePart{text("rejected")}, partsOf(t, got[1]))
require.Equal(t, rows[2], got[2])
require.Equal(t, providerSwitchStripStats{RemovedReasoning: 2}, stats)
})
t.Run("current epoch is per provider", func(t *testing.T) {
t.Parallel()
epoch := func(source reasoningProvenance, n int) reasoningProvenance {
source.Epoch = n
return source
}
otherModelAhead := epoch(openAIInstanceOtherModel, 3)
rows := []database.ChatMessage{
userRow(t, "hi"),
assistantRow(t, openAIInstanceCfg, stampedFull(epoch(openAIInstance, 2)), openAIReasoning),
assistantRow(t, openAIInstanceOtherModelCfg, stampedFull(otherModelAhead), text("done")),
assistantRow(t, openAIInstanceCfg, stampedFull(epoch(otherOpenAIInstance, 5))),
}
require.Equal(t, 3, currentReasoningEpoch(rows, openAIInstance.ProviderIdentity))
require.Equal(t, 5, currentReasoningEpoch(rows, otherOpenAIInstance.ProviderIdentity))
require.Zero(t, currentReasoningEpoch(rows, vllm.ProviderIdentity))
})
}
func TestModelConfigProviderIdentity(t *testing.T) {
+11
View File
@@ -514,6 +514,15 @@ type ChatMessagePart struct {
// because a model config's model can change after the message was
// written. Internal only: stripped before API responses.
ProviderModel string `json:"provider_model,omitempty" typescript:"-"`
// ProviderCredential records whether a reasoning part was produced
// with the deployment's credential or the user's personal API key,
// because encrypted reasoning only decrypts for the organization
// behind that credential. Internal only: stripped before API responses.
ProviderCredential string `json:"provider_credential,omitempty" typescript:"-"`
// ReasoningEpoch records how many times the producing provider had
// rejected the chat's encrypted reasoning before this reasoning part
// was produced. Internal only: stripped before API responses.
ReasoningEpoch int `json:"reasoning_epoch,omitempty" typescript:"-"`
// ProviderExecuted indicates the tool call was executed by
// the provider (e.g. Anthropic computer use).
ProviderExecuted bool `json:"provider_executed,omitempty" variants:"tool-call?,tool-result?"`
@@ -602,6 +611,8 @@ func (p *ChatMessagePart) StripInternal() {
p.ProviderMetadata = nil
p.ProviderIdentity = ""
p.ProviderModel = ""
p.ProviderCredential = ""
p.ReasoningEpoch = 0
if p.FileID.Valid {
p.Data = nil
}
+13 -7
View File
@@ -199,18 +199,22 @@ func TestChatMessagePart_StripInternal(t *testing.T) {
t.Run("StripsProviderMetadata", func(t *testing.T) {
t.Parallel()
part := codersdk.ChatMessagePart{
Type: codersdk.ChatMessagePartTypeToolCall,
ToolCallID: "call-1",
ToolName: "some_tool",
Args: json.RawMessage(`{"key":"value"}`),
ProviderMetadata: json.RawMessage(`{"type":"ephemeral"}`),
ProviderIdentity: "provider-1",
ProviderModel: "model-1",
Type: codersdk.ChatMessagePartTypeToolCall,
ToolCallID: "call-1",
ToolName: "some_tool",
Args: json.RawMessage(`{"key":"value"}`),
ProviderMetadata: json.RawMessage(`{"type":"ephemeral"}`),
ProviderIdentity: "provider-1",
ProviderModel: "model-1",
ProviderCredential: "user",
ReasoningEpoch: 2,
}
part.StripInternal()
assert.Nil(t, part.ProviderMetadata)
assert.Empty(t, part.ProviderIdentity)
assert.Empty(t, part.ProviderModel)
assert.Empty(t, part.ProviderCredential)
assert.Zero(t, part.ReasoningEpoch)
// Public fields preserved.
assert.Equal(t, codersdk.ChatMessagePartTypeToolCall, part.Type)
assert.Equal(t, "call-1", part.ToolCallID)
@@ -309,6 +313,8 @@ func TestChatMessagePartVariantTags(t *testing.T) {
"provider_metadata": "internal only, stripped by db2sdk before API responses",
"provider_identity": "internal only, stripped by db2sdk before API responses",
"provider_model": "internal only, stripped by db2sdk before API responses",
"provider_credential": "internal only, stripped by db2sdk before API responses",
"reasoning_epoch": "internal only, stripped by db2sdk before API responses",
"context_file_content": "internal only, stripped before API responses (typescript:\"-\")",
"context_file_os": "internal only, used during prompt expansion (typescript:\"-\")",
"context_file_directory": "internal only, used during prompt expansion (typescript:\"-\")",
+22
View File
@@ -2903,6 +2903,7 @@ curl -X GET http://coder-server:8080/api/v2/chats/{chat}/messages \
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -2910,6 +2911,7 @@ curl -X GET http://coder-server:8080/api/v2/chats/{chat}/messages \
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -2986,6 +2988,7 @@ curl -X GET http://coder-server:8080/api/v2/chats/{chat}/messages \
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -2993,6 +2996,7 @@ curl -X GET http://coder-server:8080/api/v2/chats/{chat}/messages \
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -3143,6 +3147,7 @@ curl -X POST http://coder-server:8080/api/v2/chats/{chat}/messages \
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -3150,6 +3155,7 @@ curl -X POST http://coder-server:8080/api/v2/chats/{chat}/messages \
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -3225,6 +3231,7 @@ curl -X POST http://coder-server:8080/api/v2/chats/{chat}/messages \
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -3232,6 +3239,7 @@ curl -X POST http://coder-server:8080/api/v2/chats/{chat}/messages \
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -3308,6 +3316,7 @@ curl -X POST http://coder-server:8080/api/v2/chats/{chat}/messages \
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -3315,6 +3324,7 @@ curl -X POST http://coder-server:8080/api/v2/chats/{chat}/messages \
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -3450,6 +3460,7 @@ curl -X PATCH http://coder-server:8080/api/v2/chats/{chat}/messages/{message} \
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -3457,6 +3468,7 @@ curl -X PATCH http://coder-server:8080/api/v2/chats/{chat}/messages/{message} \
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -3532,6 +3544,7 @@ curl -X PATCH http://coder-server:8080/api/v2/chats/{chat}/messages/{message} \
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -3539,6 +3552,7 @@ curl -X PATCH http://coder-server:8080/api/v2/chats/{chat}/messages/{message} \
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -4125,6 +4139,7 @@ curl -X GET http://coder-server:8080/api/v2/chats/{chat}/stream \
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -4132,6 +4147,7 @@ curl -X GET http://coder-server:8080/api/v2/chats/{chat}/stream \
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -4206,6 +4222,7 @@ curl -X GET http://coder-server:8080/api/v2/chats/{chat}/stream \
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -4213,6 +4230,7 @@ curl -X GET http://coder-server:8080/api/v2/chats/{chat}/stream \
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -4274,6 +4292,7 @@ curl -X GET http://coder-server:8080/api/v2/chats/{chat}/stream \
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -4281,6 +4300,7 @@ curl -X GET http://coder-server:8080/api/v2/chats/{chat}/stream \
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -4378,11 +4398,13 @@ Status Code **200**
| `»»» media_type` | string | false | | |
| `»»» name` | string | false | | |
| `»»» parsed_commands` | array | false | | Parsed commands holds parsed programs from an execute tool call's shell command, one entry per simple command in source order. Each entry is [program] or [program, arg] where arg is the first non-flag positional argument. Program names are normalized to their base name (e.g. /usr/bin/go becomes go). Only populated when ToolName is "execute" and the command parses successfully; nil otherwise. |
| `»»» provider_credential` | string | false | | Provider credential records whether a reasoning part was produced with the deployment's credential or the user's personal API key, because encrypted reasoning only decrypts for the organization behind that credential. Internal only: stripped before API responses. |
| `»»» provider_executed` | boolean | false | | Provider executed indicates the tool call was executed by the provider (e.g. Anthropic computer use). |
| `»»» provider_identity` | string | false | | Provider identity records the provider that produced a reasoning part, because a model config's provider can change after the message was written. Internal only: stripped before API responses. |
| `»»» provider_metadata` | array | false | | Provider metadata holds provider-specific response metadata (e.g. Anthropic cache control hints) as raw JSON. Internal only: stripped by db2sdk before API responses. |
| `»»» provider_model` | string | false | | Provider model records the model that produced a reasoning part, because a model config's model can change after the message was written. Internal only: stripped before API responses. |
| `»»» reasoning_delta` | string | false | | |
| `»»» reasoning_epoch` | integer | false | | Reasoning epoch records how many times the producing provider had rejected the chat's encrypted reasoning before this reasoning part was produced. Internal only: stripped before API responses. |
| `»»» result` | array | false | | |
| `»»» result_delta` | string | false | | |
| `»»» result_reset` | boolean | false | | |
+30
View File
@@ -3238,6 +3238,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -3245,6 +3246,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -3335,6 +3337,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -3342,6 +3345,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -3393,11 +3397,13 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
| `media_type` | string | false | | |
| `name` | string | false | | |
| `parsed_commands` | array of array | false | | Parsed commands holds parsed programs from an execute tool call's shell command, one entry per simple command in source order. Each entry is [program] or [program, arg] where arg is the first non-flag positional argument. Program names are normalized to their base name (e.g. /usr/bin/go becomes go). Only populated when ToolName is "execute" and the command parses successfully; nil otherwise. |
| `provider_credential` | string | false | | Provider credential records whether a reasoning part was produced with the deployment's credential or the user's personal API key, because encrypted reasoning only decrypts for the organization behind that credential. Internal only: stripped before API responses. |
| `provider_executed` | boolean | false | | Provider executed indicates the tool call was executed by the provider (e.g. Anthropic computer use). |
| `provider_identity` | string | false | | Provider identity records the provider that produced a reasoning part, because a model config's provider can change after the message was written. Internal only: stripped before API responses. |
| `provider_metadata` | array of integer | false | | Provider metadata holds provider-specific response metadata (e.g. Anthropic cache control hints) as raw JSON. Internal only: stripped by db2sdk before API responses. |
| `provider_model` | string | false | | Provider model records the model that produced a reasoning part, because a model config's model can change after the message was written. Internal only: stripped before API responses. |
| `reasoning_delta` | string | false | | |
| `reasoning_epoch` | integer | false | | Reasoning epoch records how many times the producing provider had rejected the chat's encrypted reasoning before this reasoning part was produced. Internal only: stripped before API responses. |
| `result` | array of integer | false | | |
| `result_delta` | string | false | | |
| `result_reset` | boolean | false | | |
@@ -3513,6 +3519,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -3520,6 +3527,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -3596,6 +3604,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -3603,6 +3612,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -4979,6 +4989,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -4986,6 +4997,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -5144,6 +5156,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -5151,6 +5164,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -5225,6 +5239,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -5232,6 +5247,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -5293,6 +5309,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -5300,6 +5317,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -5410,6 +5428,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -5417,6 +5436,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -6270,6 +6290,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -6277,6 +6298,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -6352,6 +6374,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -6359,6 +6382,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -6435,6 +6459,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -6442,6 +6467,7 @@ AuthorizationObject can represent a "set" of objects, such as: all workspaces in
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -9362,6 +9388,7 @@ CreateWorkspaceRequest provides options for creating a new workspace. Only one o
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -9369,6 +9396,7 @@ CreateWorkspaceRequest provides options for creating a new workspace. Only one o
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],
@@ -9444,6 +9472,7 @@ CreateWorkspaceRequest provides options for creating a new workspace. Only one o
"string"
]
],
"provider_credential": "string",
"provider_executed": true,
"provider_identity": "string",
"provider_metadata": [
@@ -9451,6 +9480,7 @@ CreateWorkspaceRequest provides options for creating a new workspace. Only one o
],
"provider_model": "string",
"reasoning_delta": "string",
"reasoning_epoch": 0,
"result": [
0
],