Skip to content

Phase 2: wire decideLoopContinuation + classifyCompletionState #92

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Pterjudin merged 1 commit into from
Jul 23, 2026
Merged

Phase 2: wire decideLoopContinuation + classifyCompletionState #92

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
93 changes: 65 additions & 28 deletions src/vs/workbench/contrib/cortexide/browser/chatThreadService.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,7 @@ import { CancellationToken } from '../../../../base/common/cancellation.js';
import { ILanguageFeaturesService } from '../../../../editor/common/services/languageFeatures.js';
import { ChatMessage, ChatImageAttachment, ChatPDFAttachment, CheckpointEntry, CodespanLocationLink, StagingSelectionItem, ToolMessage, PlanMessage, PlanStep, StepStatus, ReviewMessage } from '../common/chatThreadServiceTypes.js';
import { selectCompactionWindow } from '../common/compactionPolicy.js';
import { updateConsecutiveToolErrors, computeCompactionOverflowDecision, shouldEscalateModel, computePostEscalationCounters, decideFileReadGate, classifyToolStepOutcome, type ToolMessageType } from '../common/agentLoopDecisions.js';
import { computeCompactionOverflowDecision, shouldEscalateModel, computePostEscalationCounters, decideFileReadGate, classifyToolStepOutcome, decideLoopContinuation, classifyCompletionState, type ToolMessageType } from '../common/agentLoopDecisions.js';
import { isRateLimitErrorMessage } from '../common/providerErrorFormat.js';
import { createSerializer } from '../common/asyncSerializer.js';

Expand Down Expand Up @@ -3507,12 +3507,7 @@ Output ONLY the JSON, no other text. Start with { and end with }.`
while (shouldSendAnotherMessage) {
// CRITICAL: Check for maximum iterations to prevent infinite loops
if (nMessagesSent >= maxAgentIterations) {
// Before giving up: a model that burned the whole step budget without finishing is usually a
// weak/local model spinning. Escalate the task to a more capable model and keep going.
// Phase 2: the escalate-or-stop gate is the pure shouldEscalateModel(); tryEscalateModel still
// performs the switch. shouldCallEscalate==false exactly when fallback is off or the escalation
// budget is spent -- in which case tryEscalateModel would itself return false (its own guard),
// so short-circuiting is behavior-identical and just avoids a no-op await.
// Phase 2: iteration-cap gate via decideLoopContinuation + shouldEscalateModel.
const escDec = shouldEscalateModel({
triggerSite: 'iterCap',
modelFallbackEnabled, escalationCount, MAX_MODEL_ESCALATIONS,
Expand All @@ -3522,18 +3517,28 @@ Output ONLY the JSON, no other text. Start with { and end with }.`
isRateLimitError: false, isNonRetryableError: false,
nAttempts: 0, CHAT_RETRIES,
})
if (escDec.shouldCallEscalate && await tryEscalateModel(`the previous model used all ${maxAgentIterations} steps without finishing`)) {
// Escalation gives the fresh model a clean budget (both loop counters reset). Centralized
// so every escalation site -- including the tool-error sites below -- resets uniformly.
const iterCapDec = decideLoopContinuation({
nMessagesSent,
maxAgentIterations,
consecutiveToolErrors,
maxConsecutiveToolErrors,
lastToolMessageType: null,
toolCallDispatched: false,
awaitingUserApproval: false,
canEscalate: escDec.shouldCallEscalate,
})
if (iterCapDec.action === 'escalate-iter-cap' && await tryEscalateModel(`the previous model used all ${maxAgentIterations} steps without finishing`)) {
const reset = computePostEscalationCounters('iterCap')
nMessagesSent = reset.nMessagesSent
consecutiveToolErrors = reset.consecutiveToolErrors
this._setStreamState(threadId, { isRunning: 'idle', interrupt: 'not_needed' })
continue
}
this._notificationService.warn(`Agent stopped after ${maxAgentIterations} tool iterations.${isLocalModel ? ' Small/local models can struggle with multi-step tool use — try Ask/Normal mode for a direct answer, or use a larger model.' : ''}`)
this._setStreamState(threadId, { isRunning: undefined })
return
if (iterCapDec.action === 'hard-stop-iter-cap' || iterCapDec.action === 'escalate-iter-cap') {
this._notificationService.warn(`Agent stopped after ${maxAgentIterations} tool iterations.${isLocalModel ? ' Small/local models can struggle with multi-step tool use — try Ask/Normal mode for a direct answer, or use a larger model.' : ''}`)
this._setStreamState(threadId, { isRunning: undefined })
return
}
}

// CRITICAL: Check stream state first - if execution was interrupted/aborted, stop immediately
Expand Down Expand Up @@ -4810,7 +4815,19 @@ Output ONLY the JSON, no other text. Start with { and end with }.`
const mcpTool = mcpTools?.find(t => t.name === toolCall.name)

const { awaitingUserApproval, interrupted, completionSignaled } = await this._runToolCall(threadId, toolCall.name, toolCall.id, mcpTool?.mcpServerName, { preapproved: false, unvalidatedToolParams: toolCall.rawParams }, isLocalModel, chatMode, isCapableLocalModelFlag)
if (interrupted) {

const toolCompletion = classifyCompletionState({
toolCall: { name: toolCall.name },
completionSignaled,
interrupted,
awaitingUserApproval,
fileReadLimitExceeded: false,
readFileLimitReached: false,
synthFired: false,
synthCompletionSignaled: false,
synthInterrupted: false,
})
if (toolCompletion.action === 'terminate_interrupted') {
this._setStreamState(threadId, undefined)
if (activePlanTracking?.currentStep) {
const updatedStep = this._markStepCompletedInternal(threadId, activePlanTracking.currentStep, false, 'Interrupted by user')
Expand All @@ -4824,7 +4841,7 @@ Output ONLY the JSON, no other text. Start with { and end with }.`
return
}

if (completionSignaled) {
if (toolCompletion.action === 'terminate_completion') {
this._setStreamState(threadId, { isRunning: undefined })
return
}
Expand All @@ -4841,23 +4858,43 @@ Output ONLY the JSON, no other text. Start with { and end with }.`
const lastToolMessageType: ToolMessageType | null = lastToolMsg?.role === 'tool'
? (lastToolMsg as ToolMessage<ToolName>).type
: null
const errDec = updateConsecutiveToolErrors(consecutiveToolErrors, lastToolMessageType, maxConsecutiveToolErrors, /* escalationAvailable */ false)
consecutiveToolErrors = errDec.nextConsecutiveToolErrors
if (errDec.action === 'halt') {
// Before giving up: this is the single most common local-model failure mode (invents tool
// names, writes empty files, never converges). Escalate to a more capable model and let it
// recover the SAME task — it sees the failed attempts in history and corrects.
if (await tryEscalateModel(`the previous model failed ${consecutiveToolErrors} tool calls in a row`)) {
// Match the iter-cap escalation: the fresh model gets a clean budget -- reset BOTH loop
// counters (previously only consecutiveToolErrors reset, so the new model inherited a spent
// iteration budget and could hit the iteration cap before finishing).
const loopLastType = (lastToolMessageType === 'success' || lastToolMessageType === 'tool_error' || lastToolMessageType === 'invalid_params')
? lastToolMessageType
: null
const toolErrEscDec = shouldEscalateModel({
triggerSite: 'toolErrorCap',
modelFallbackEnabled, escalationCount, MAX_MODEL_ESCALATIONS,
nMessagesSent, maxAgentIterations,
consecutiveToolErrors, maxConsecutiveToolErrors,
isAutoMode: false, autoFallbackExhausted: true,
isRateLimitError: false, isNonRetryableError: false,
nAttempts: 0, CHAT_RETRIES,
})
const postToolDec = decideLoopContinuation({
nMessagesSent,
maxAgentIterations,
consecutiveToolErrors,
maxConsecutiveToolErrors,
lastToolMessageType: loopLastType,
toolCallDispatched: true,
awaitingUserApproval,
canEscalate: toolErrEscDec.shouldCallEscalate,
})
const toolErrorCountForMessage = postToolDec.action === 'escalate-tool-errors'
? maxConsecutiveToolErrors
: postToolDec.nextConsecutiveToolErrors
consecutiveToolErrors = postToolDec.nextConsecutiveToolErrors
if (postToolDec.action === 'escalate-tool-errors') {
if (await tryEscalateModel(`the previous model failed ${toolErrorCountForMessage} tool calls in a row`)) {
const reset = computePostEscalationCounters('toolErrorCap')
nMessagesSent = reset.nMessagesSent
consecutiveToolErrors = reset.consecutiveToolErrors
this._setStreamState(threadId, { isRunning: 'idle', interrupt: 'not_needed' })
continue
}
this._addMessageToThread(threadId, { role: 'assistant', displayContent: `Stopped after ${consecutiveToolErrors} failed tool calls in a row.${isLocalModel ? ' Small/local models can struggle with multi-step tool use — try Ask/Normal mode for a direct answer, or a larger model.' : ''}`, reasoning: '', anthropicReasoning: null })
}
if (postToolDec.action === 'hard-stop-tool-errors' || postToolDec.action === 'escalate-tool-errors') {
this._addMessageToThread(threadId, { role: 'assistant', displayContent: `Stopped after ${toolErrorCountForMessage} failed tool calls in a row.${isLocalModel ? ' Small/local models can struggle with multi-step tool use — try Ask/Normal mode for a direct answer, or use a larger model.' : ''}`, reasoning: '', anthropicReasoning: null })
this._setStreamState(threadId, { isRunning: undefined })
this._addUserCheckpoint({ threadId })
return
Expand Down Expand Up @@ -4925,8 +4962,8 @@ Output ONLY the JSON, no other text. Start with { and end with }.`
}
}

if (awaitingUserApproval) { isRunningWhenEnd = 'awaiting_user' }
else { shouldSendAnotherMessage = true }
if (postToolDec.action === 'await-user') { isRunningWhenEnd = postToolDec.isRunningWhenEnd ?? 'awaiting_user' }
else if (postToolDec.action === 'continue') { shouldSendAnotherMessage = true }

this._setStreamState(threadId, { isRunning: 'idle', interrupt: 'not_needed' }) // just decorative, for clarity
}
Expand Down
Loading