Claude Code keeps retrying the same (2026)

The Problem

Claude Code is stuck retrying the same failing operation – modifying code slightly and re-running it, each time getting the same or a similar error. Each retry costs 3,000-8,000 tokens (code modification + command execution + error reading). A 5-retry spiral burns 15,000-40,000 tokens in 2-3 minutes. At Opus 4.6 rates ($15/$75 per MTok), that is $0.23-$3.00 wasted on an approach that is not working.

Quick Fix (2 Minutes)

  1. Type “stop” or press Escape to interrupt the current retry loop.
  2. Run /compact to clear the accumulated failed attempts from context.
  3. Restate the task with explicit constraints: “The previous approach of modifying X did not work after 3 attempts. Try a different approach: [describe alternative].”
# Step 1: Interrupt
# Press Escape or type "stop"
# Step 2: Compact away the failed attempts
/compact Keep: the original error and task description. Discard: all failed fix attempts.
# Step 3: Redirect
# "The type error in auth.ts was not fixed by adding null checks.
#  Instead, check if the JWT library version is compatible with Node 22.
#  Read package.json for the jsonwebtoken version."

Why This Happens

Retry spirals occur because Claude Code’s context accumulates evidence that a particular approach is “close to working.” Each slightly different error message reinforces the agent’s commitment to the current strategy. Three factors compound:

  1. Anchoring bias – the accumulated context of 3-4 attempts makes the agent believe it is on the right track, even when the fundamental approach is wrong.
  2. Context pollution – each failed attempt adds 3,000-8,000 tokens of noise to the context, making it harder for the agent to reason clearly about alternatives.
  3. Missing root cause – the agent is treating symptoms (the error message) rather than diagnosing the root cause (wrong library version, missing configuration, etc.).

The token cost compounds because each retry re-sends the entire conversation context, which now includes all previous failed attempts. By retry 5, the context may contain 25,000+ tokens of failed attempt history, costing $0.38+ per turn at Opus rates just to re-read the failures.

The Full Fix

Step 1: Diagnose

Confirm that the agent is in a retry loop by checking the pattern:

# Check the /cost output
/cost
# If token count has jumped 20K+ in the last few exchanges with no progress: retry spiral confirmed
# Look for the pattern in the conversation:
# - Same command being run repeatedly
# - Same error or slight variations of the same error
# - Small code changes between attempts (e.g., adding null checks, changing variable names)

Step 2: Fix

Break the loop and redirect the agent to a fundamentally different approach.

# Compact to remove the failed attempts
/compact Keep: original task, error description. Discard: all fix attempts.
# Redirect with explicit constraints
# Instead of: "Fix the TypeError in auth.ts"
# Say: "The TypeError in auth.ts is NOT a null reference issue (already tried 3 times).
#       Investigate: 1) JWT library version compatibility, 2) Node.js crypto API changes,
#       3) TypeScript strict mode type narrowing. Read package.json first."

Step 3: Prevent

Add retry limits to CLAUDE.md that trigger before spirals begin.

# CLAUDE.md -- retry spiral prevention
## Error Handling Rules
- Maximum 3 attempts on any single error
- After attempt 1: re-read the FULL error message and identify the root cause
- After attempt 2: if the same error recurs, the approach is wrong -- try a different strategy
- After attempt 3: STOP. Report: what was tried, what failed, and 2 alternative approaches
- Never retry: permission errors (403/401), missing file errors, configuration errors
- For build/test errors: read the error output completely before making changes

Cost Recovery

After a retry spiral, minimize the damage:

# 1. Compact to remove the wasted context
/compact
# 2. Verify compaction saved tokens
/cost
# 3. If the task is still needed, restart with a focused prompt
# This fresh approach typically costs 5K-15K tokens vs the 40K+ already wasted
# 4. Consider switching to a cheaper model for the diagnostic phase
# Use Haiku ($0.80/$4 per MTok) to investigate, then Sonnet to implement

The compaction alone saves significant money on remaining turns. If the context was 120K tokens and compaction reduces it to 40K tokens, each subsequent turn saves $1.20 in input at Opus rates.

Prevention Rules for CLAUDE.md

# CLAUDE.md -- copy-paste this section
## Retry Spiral Prevention
- MAX 3 RETRIES on any failing operation
- After each failure: read the COMPLETE error message before retrying
- After 2 failures with the same error: CHANGE APPROACH (do not modify the same code)
- After 3 failures: STOP and report:
  1. What was attempted (3 sentences max)
  2. The exact error each time
  3. Two alternative approaches to try
- Never retry: auth errors, missing file errors, permission errors, syntax errors in config files
- If a test fails 3 times: check if the test itself is wrong, not just the implementation

Which model? → Take the 5-question quiz in our Model Selector.

Estimate tokens → Calculate your usage with our Token Estimator.

Try it: Paste your error into our Error Diagnostic for an instant fix.