Skip to content

fix(reflection): prevent action loops by reducing max attempts and adding loop detection #114

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
dzianisv merged 2 commits into from
Feb 18, 2026
Merged

fix(reflection): prevent action loops by reducing max attempts and adding loop detection #114

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
57ad843
fix(reflection): prevent action loops by reducing max attempts and ad…
Feb 15, 2026
85d7a44
chore: merge main into issue-80-reflection-loop-prevention
Feb 18, 2026
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
102 changes: 95 additions & 7 deletions reflection-3.test-helpers.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,7 +87,7 @@ export function inferTaskType(text: string): TaskType {
return "other"
}

export function buildSelfAssessmentPrompt(context: TaskContext, agents: string, lastAssistantText?: string): string {
export function buildSelfAssessmentPrompt(context: TaskContext, agents: string, lastAssistantText?: string, attemptCount?: number): string {
const safeContext = {
...context,
detectedSignals: Array.isArray(context.detectedSignals) ? context.detectedSignals : []
Expand All @@ -107,6 +107,11 @@ export function buildSelfAssessmentPrompt(context: TaskContext, agents: string,
? `\n## Agent's Last Response\n${lastAssistantText.slice(0, 4000)}\n`
: ""

const currentAttempt = attemptCount || 0
const attemptSection = currentAttempt > 0
? `\n## Reflection History\n- This is reflection attempt ${currentAttempt + 1}/${MAX_ATTEMPTS} for this task.\n- Previous reflections found the task incomplete.\n- If you are repeating the same actions without progress, set "stuck": true and explain what is blocking you.\n`
: ""

return `SELF-ASSESS REFLECTION-3

You are evaluating an agent's work against workflow requirements.
Expand All @@ -121,7 +126,7 @@ Analyze the task context, the agent's last response, and the tool signals to det

## Tool Commands Run
${safeContext.toolsSummary}
${assistantSection}
${assistantSection}${attemptSection}
${agents ? `## Project Instructions\n${agents.slice(0, 800)}\n\n` : ""}Return JSON only:
{
"task_summary": "...",
Expand All @@ -148,7 +153,9 @@ Rules:
- Direct pushes to main/master are not allowed; require a PR instead.
- Provide a PR URL and CI status when a PR is required.
- If stuck, propose an alternate approach.
- If you need user action (auth, 2FA, credentials), list it in needs_user_action.`
- If you need user action (auth, 2FA, credentials), list it in needs_user_action.
- If you are repeating the same actions (deploy, test, build) without making progress, set "stuck": true.
- Do not retry the same failing approach more than twice — try something different or report stuck.`
}

export function parseSelfAssessmentJson(text: string | null | undefined): SelfAssessment | null {
Expand Down Expand Up @@ -416,13 +423,74 @@ export function getGitHubCopilotModelForRouting(modelSpec: string | null | undef
}

const FEEDBACK_MARKER = "## Reflection-3:"
const MAX_ATTEMPTS = 5
const MAX_ATTEMPTS = 3
const ACTION_LOOP_MIN_COMMANDS = 4
const ACTION_LOOP_REPETITION_THRESHOLD = 0.6

/**
* Detects when the agent is repeating the same commands/actions without progress.
* Unlike detectPlanningLoop (read-heavy without writes), this catches action loops
* where the agent IS making write-like operations but repeating the same ones.
*/
export function detectActionLoop(messages: any[]): {
detected: boolean
repeatedCommands: string[]
totalCommands: number
} {
if (!Array.isArray(messages)) {
return { detected: false, repeatedCommands: [], totalCommands: 0 }
}

const commands: string[] = []
for (const msg of messages) {
if (msg.info?.role !== "assistant") continue
for (const part of msg.parts || []) {
if (part.type !== "tool") continue
const toolName = (part.tool || "").toString().toLowerCase()
const input = part.state?.input || {}

if (toolName === "bash") {
const cmd = (input.command || input.cmd || "").toString().trim()
if (cmd) {
const normalized = cmd.replace(/\s+/g, " ").replace(/\d{10,}/g, "TIMESTAMP").toLowerCase()
commands.push(normalized)
}
} else if (toolName !== "read" && toolName !== "glob" && toolName !== "grep" && toolName !== "todowrite") {
const key = `${toolName}:${JSON.stringify(input).slice(0, 100)}`
commands.push(key)
}
}
}

if (commands.length < ACTION_LOOP_MIN_COMMANDS) {
return { detected: false, repeatedCommands: [], totalCommands: commands.length }
}

const counts = new Map<string, number>()
for (const cmd of commands) {
counts.set(cmd, (counts.get(cmd) || 0) + 1)
}

const repeatedCommands: string[] = []
let repeatedCount = 0
for (const [cmd, count] of counts) {
if (count >= 3) {
repeatedCommands.push(cmd)
repeatedCount += count
}
}

const detected = repeatedCommands.length > 0 && repeatedCount / commands.length >= ACTION_LOOP_REPETITION_THRESHOLD

return { detected, repeatedCommands, totalCommands: commands.length }
}

export function buildEscalatingFeedback(
attemptCount: number,
severity: string,
verdict: { feedback?: string; missing?: string[]; next_actions?: string[] } | undefined | null,
isPlanningLoop: boolean
isPlanningLoop: boolean,
isActionLoop?: boolean
): string {
const safeVerdict = verdict ?? {}
const missingItems = Array.isArray(safeVerdict.missing) ? safeVerdict.missing : []
Expand All @@ -445,6 +513,19 @@ Pick the FIRST item from your existing todo list and implement it. Open a file w
Start coding NOW. No more planning.`
}

if (isActionLoop) {
return `${FEEDBACK_MARKER} STOP: Action Loop Detected (attempt ${attemptCount}/${MAX_ATTEMPTS})

You are repeating the same commands without making progress. Running the same deploy/test/build cycle again will produce the same result.

STOP and do ONE of these:
1. If the same test/eval keeps failing, analyze the failure output and fix the root cause before re-running.
2. If you cannot fix the root cause, explain what is blocking you and ask the user for help.
3. Try a completely different approach (e.g., test locally instead of via deployment).

Do NOT re-run the same command hoping for a different result.`
}

if (attemptCount <= 2) {
const missing = missingItems.length
? `\n### Missing\n${missingItems.map((m) => `- ${m}`).join("\n")}`
Expand All @@ -463,11 +544,18 @@ Please address these issues and continue.`
const missingBrief = missingItems.length
? `Still missing: ${missingItems.slice(0, 3).join(", ")}.`
: ""
return `${FEEDBACK_MARKER} Still Incomplete (attempt ${attemptCount}/${MAX_ATTEMPTS})
return `${FEEDBACK_MARKER} Final Attempt (${attemptCount}/${MAX_ATTEMPTS})

${missingBrief}

You have been asked ${attemptCount} times to complete this task. Stop re-reading files or re-planning. Focus on the specific items above and implement them now. If something is blocking you, say what it is clearly.`
You have been asked ${attemptCount} times to complete this task. This is your LAST chance before reflection stops.

If you cannot complete the remaining items:
- Explain clearly what is blocking you
- Set needs_user_action if you need user help
- Try a different approach instead of repeating the same steps

Do NOT re-read files or re-plan. Either implement the fix now or explain why you cannot.`
}

export function shouldApplyPlanningLoop(taskType: TaskType, loopDetected: boolean): boolean {
Expand Down
115 changes: 106 additions & 9 deletions reflection-3.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,7 @@ async function reportError(err: unknown, context?: Record<string, string>): Prom

const SELF_ASSESSMENT_MARKER = "## Reflection-3 Self-Assessment"
const FEEDBACK_MARKER = "## Reflection-3:"
const MAX_ATTEMPTS = 5
const MAX_ATTEMPTS = 3

const JUDGE_BLOCKED_PATTERNS = [
/\bhaiku\b/i,
Expand All @@ -36,6 +36,8 @@ const JUDGE_BLOCKED_PATTERNS = [

const PLANNING_LOOP_MIN_TOOL_CALLS = 8
const PLANNING_LOOP_WRITE_RATIO_THRESHOLD = 0.1
const ACTION_LOOP_MIN_COMMANDS = 4
const ACTION_LOOP_REPETITION_THRESHOLD = 0.6

type TaskType = "coding" | "docs" | "research" | "ops" | "other"
type AgentMode = "plan" | "build" | "unknown"
Expand Down Expand Up @@ -276,11 +278,76 @@ export function shouldApplyPlanningLoop(taskType: TaskType, loopDetected: boolea
return taskType === "coding"
}

/**
* Detects when the agent is repeating the same commands/actions without progress.
* Unlike detectPlanningLoop (read-heavy without writes), this catches action loops
* where the agent IS making write-like operations but repeating the same ones.
* Example: repeatedly re-deploying and re-running the same failing evaluation.
*/
export function detectActionLoop(messages: any[]): {
detected: boolean
repeatedCommands: string[]
totalCommands: number
} {
if (!Array.isArray(messages)) {
return { detected: false, repeatedCommands: [], totalCommands: 0 }
}

const commands: string[] = []
for (const msg of messages) {
if (msg.info?.role !== "assistant") continue
for (const part of msg.parts || []) {
if (part.type !== "tool") continue
const toolName = (part.tool || "").toString().toLowerCase()
const input = part.state?.input || {}

if (toolName === "bash") {
const cmd = (input.command || input.cmd || "").toString().trim()
if (cmd) {
// Normalize: collapse whitespace and remove trailing timestamps/IDs
const normalized = cmd.replace(/\s+/g, " ").replace(/\d{10,}/g, "TIMESTAMP").toLowerCase()
commands.push(normalized)
}
} else if (toolName !== "read" && toolName !== "glob" && toolName !== "grep" && toolName !== "todowrite") {
// Track non-read tool calls by name + key input params
const key = `${toolName}:${JSON.stringify(input).slice(0, 100)}`
commands.push(key)
}
}
}

if (commands.length < ACTION_LOOP_MIN_COMMANDS) {
return { detected: false, repeatedCommands: [], totalCommands: commands.length }
}

// Count occurrences of each command
const counts = new Map<string, number>()
for (const cmd of commands) {
counts.set(cmd, (counts.get(cmd) || 0) + 1)
}

// Find commands repeated 3+ times
const repeatedCommands: string[] = []
let repeatedCount = 0
for (const [cmd, count] of counts) {
if (count >= 3) {
repeatedCommands.push(cmd)
repeatedCount += count
}
}

// Loop detected if repeated commands make up a significant fraction
const detected = repeatedCommands.length > 0 && repeatedCount / commands.length >= ACTION_LOOP_REPETITION_THRESHOLD

return { detected, repeatedCommands, totalCommands: commands.length }
}

export function buildEscalatingFeedback(
attemptCount: number,
severity: string,
verdict: { feedback?: string; missing?: string[]; next_actions?: string[] } | undefined | null,
isPlanningLoop: boolean
isPlanningLoop: boolean,
isActionLoop?: boolean
): string {
const safeVerdict = verdict ?? {}
const missingItems = Array.isArray(safeVerdict.missing) ? safeVerdict.missing : []
Expand All @@ -303,6 +370,19 @@ Pick the FIRST item from your existing todo list and implement it. Open a file w
Start coding NOW. No more planning.`
}

if (isActionLoop) {
return `${FEEDBACK_MARKER} STOP: Action Loop Detected (attempt ${attemptCount}/${MAX_ATTEMPTS})

You are repeating the same commands without making progress. Running the same deploy/test/build cycle again will produce the same result.

STOP and do ONE of these:
1. If the same test/eval keeps failing, analyze the failure output and fix the root cause before re-running.
2. If you cannot fix the root cause, explain what is blocking you and ask the user for help.
3. Try a completely different approach (e.g., test locally instead of via deployment).

Do NOT re-run the same command hoping for a different result.`
}

if (attemptCount <= 2) {
const missing = missingItems.length
? `\n### Missing\n${missingItems.map((m) => `- ${m}`).join("\n")}`
Expand All @@ -321,11 +401,18 @@ Please address these issues and continue.`
const missingBrief = missingItems.length
? `Still missing: ${missingItems.slice(0, 3).join(", ")}.`
: ""
return `${FEEDBACK_MARKER} Still Incomplete (attempt ${attemptCount}/${MAX_ATTEMPTS})
return `${FEEDBACK_MARKER} Final Attempt (${attemptCount}/${MAX_ATTEMPTS})

${missingBrief}

You have been asked ${attemptCount} times to complete this task. Stop re-reading files or re-planning. Focus on the specific items above and implement them now. If something is blocking you, say what it is clearly.`
You have been asked ${attemptCount} times to complete this task. This is your LAST chance before reflection stops.

If you cannot complete the remaining items:
- Explain clearly what is blocking you
- Set needs_user_action if you need user help
- Try a different approach instead of repeating the same steps

Do NOT re-read files or re-plan. Either implement the fix now or explain why you cannot.`
}

function getLastRelevantUserMessageId(messages: any[]): string | null {
Expand Down Expand Up @@ -931,7 +1018,7 @@ function extractLastAssistantText(messages: any[]): string {
return ""
}

function buildSelfAssessmentPrompt(context: TaskContext, agents: string, lastAssistantText?: string): string {
function buildSelfAssessmentPrompt(context: TaskContext, agents: string, lastAssistantText?: string, attemptCount?: number): string {
const safeContext = {
...context,
detectedSignals: Array.isArray(context.detectedSignals) ? context.detectedSignals : []
Expand All @@ -951,6 +1038,11 @@ function buildSelfAssessmentPrompt(context: TaskContext, agents: string, lastAss
? `\n## Agent's Last Response\n${lastAssistantText.slice(0, 4000)}\n`
: ""

const currentAttempt = attemptCount || 0
const attemptSection = currentAttempt > 0
? `\n## Reflection History\n- This is reflection attempt ${currentAttempt + 1}/${MAX_ATTEMPTS} for this task.\n- Previous reflections found the task incomplete.\n- If you are repeating the same actions without progress, set "stuck": true and explain what is blocking you.\n`
: ""

return `SELF-ASSESS REFLECTION-3

You are evaluating an agent's work against workflow requirements.
Expand All @@ -965,7 +1057,7 @@ Analyze the task context, the agent's last response, and the tool signals to det

## Tool Commands Run
${safeContext.toolsSummary}
${assistantSection}
${assistantSection}${attemptSection}
${agents ? `## Project Instructions\n${agents.slice(0, 800)}\n\n` : ""}Return JSON only:
{
"task_summary": "brief description of what was done",
Expand Down Expand Up @@ -1005,7 +1097,9 @@ Rules:
- Direct pushes to main/master are not allowed; require a PR instead.
- If stuck, propose an alternate approach.
- If you need user action (auth, 2FA, credentials), list it in needs_user_action.
- PLANNING LOOP CHECK: If the task requires code changes (fix, implement, add, create, build, refactor, update) but the "Tool Commands Run" section shows ONLY read operations (read, glob, grep, git log, git status, git diff, webfetch, task/explore) and NO write operations (edit, write, bash with build/test/commit, github_create_pull_request, etc.), then the task is NOT complete. Set status to "in_progress", set stuck to true, and list "Implement the actual code changes" in remaining_work. Analyzing and recommending changes is not the same as making them.`
- PLANNING LOOP CHECK: If the task requires code changes (fix, implement, add, create, build, refactor, update) but the "Tool Commands Run" section shows ONLY read operations (read, glob, grep, git log, git status, git diff, webfetch, task/explore) and NO write operations (edit, write, bash with build/test/commit, github_create_pull_request, etc.), then the task is NOT complete. Set status to "in_progress", set stuck to true, and list "Implement the actual code changes" in remaining_work. Analyzing and recommending changes is not the same as making them.
- If you are repeating the same actions (deploy, test, build) without making progress, set "stuck": true.
- Do not retry the same failing approach more than twice — try something different or report stuck.`
}

function parseSelfAssessmentJson(text: string | null | undefined): SelfAssessment | null {
Expand Down Expand Up @@ -1403,7 +1497,8 @@ export const Reflection3Plugin: Plugin = async ({ client, directory }) => {
const lastAssistantText = extractLastAssistantText(messages)
const customPrompt = await loadReflectionPrompt(directory)
const agents = await getAgentsFile(directory)
const reflectionPrompt = customPrompt || buildSelfAssessmentPrompt(context, agents, lastAssistantText)
const currentAttemptCount = attempts.get(attemptKey) || 0
const reflectionPrompt = customPrompt || buildSelfAssessmentPrompt(context, agents, lastAssistantText, currentAttemptCount)

await showToast(client, directory, "Requesting reflection self-assessment...", "info")
debug("Requesting reflection self-assessment")
Expand Down Expand Up @@ -1606,6 +1701,7 @@ export const Reflection3Plugin: Plugin = async ({ client, directory }) => {

const loopCheck = detectPlanningLoop(preFeedbackMessages || messages)
const usePlanningLoopMessage = shouldApplyPlanningLoop(context.taskType, loopCheck.detected)
const actionLoopCheck = detectActionLoop(preFeedbackMessages || messages)
const feedbackText = buildEscalatingFeedback(
nextAttemptCount,
analysis.severity || "MEDIUM",
Expand All @@ -1614,7 +1710,8 @@ export const Reflection3Plugin: Plugin = async ({ client, directory }) => {
missing: analysis.missing,
next_actions: analysis.nextActions
},
usePlanningLoopMessage
usePlanningLoopMessage,
actionLoopCheck.detected
)

// Apply task-based model routing to feedback injection
Expand Down
Loading
Loading