Cross-session analysis of errors, corrections, and failures across Claude Code & Codex history • June 2026
Errors classified into 3 tiers by severity and actionability:
These indicate fundamental misalignment between what the agent did and what was wanted.
Agent interprets a user request literally or incorrectly, producing the wrong output entirely.
Claude: 14 Codex: 8
Agent modifies code or adds features the user never asked for, breaking existing functionality or adding complexity.
Claude: 2 Codex: 14 — Codex 7x more prone, likely due to autonomous execution without approval gates.
Agent ships to feature branches when told not to, restores deleted branches, or loses work during cherry-picks/merges.
Claude: 16 Codex: 6
Agent references APIs, files, or capabilities that don't exist. Low frequency but high cost per incident.
Claude: 5 Codex: 2
Agent understood the intent but produced buggy or incomplete output.
Second largest category. User asks agent to retry something that failed previously, or reports the same issue persisting after a "fix." Indicates the agent's fix didn't address root cause.
Claude: 164 Codex: 111
Agent makes UI changes that produce visual regressions — broken layouts, wrong scroll behavior, incorrect positioning, phantom UI elements.
Claude: 98 Codex: 53
Frontend expects one shape, backend returns another. Includes JSON parse failures, unexpected status codes, type errors at runtime.
Claude: 38 Codex: 9 — Heavily Claude-skewed; Bismarck project accounts for most.
Cache not invalidated, sessions not reconnecting, auth state lost on refresh, login loops. Agents struggle with stateful systems.
Claude: 134 Codex: 120
Agent claims work is done but user discovers missing features, skipped steps, or incomplete implementations.
Claude: 36 Codex: 49 — Codex slightly more prone, possibly due to less interactive verification.
Build failures, deploy misconfigurations, staging/prod environment issues, service connectivity problems.
Claude: 84 Codex: 59
Normal direction-setting that's part of collaborative workflow. Not true agent mistakes.
Largest category. Mostly "no, do X instead" — user steering rather than agent failures. High noise in this bucket.
User asking agent to fix a bug — could be pre-existing or agent-introduced. Ambiguous attribution.
User pasting error messages for agent to diagnose. Informational, not necessarily agent-caused.
At 15.1% of all error messages, "try again" / "still broken" indicates agents frequently apply surface-level fixes that don't address root causes. The fix-retry loop wastes the most user time of any category.
State management (13.9%) and UI visual bugs (8.3%) together account for 22% of errors. Both require runtime understanding that agents lack — they can't see the screen or observe session state. The bismarck-v0.3 project alone generated 173 error messages, mostly auth/session/terminal state issues.
Codex scope-creep rate is 7x Claude's (14 vs 2), likely because it runs autonomously without approval gates. Claude generates 4x more type/API mismatches (38 vs 9), concentrated in full-stack projects where it modifies both frontend and backend without cross-checking contracts.
Top error-producing projects are full-stack apps with auth, real-time features, and deployment pipelines: bismarck-v0.3 (173), loadoutworks.com (111), metternich-engine (110), monorepo (90). Tooling repos like agentic-skills (59) produce fewer errors per session.
| Tier | Category | Count | % | Severity | Actionability |
|---|---|---|---|---|---|
| 1 | Misunderstood Intent | 22 | 1.2% | High | Clarification gates, confirm before executing |
| 1 | Scope Creep | 16 | 0.9% | High | Stricter diff scoping, approval before extra changes |
| 1 | Git Workflow Violations | 22 | 1.2% | High | Enforce CLAUDE.md git conventions via hooks |
| 1 | Hallucination | 7 | 0.4% | High | Verify-before-recommend pattern |
| 2 | Persistent Failures | 275 | 15.1% | Med-High | Root-cause analysis requirement before retry |
| 2 | UI/Visual Bugs | 151 | 8.3% | Medium | Screenshot verification, browser testing |
| 2 | Type/API Mismatches | 47 | 2.6% | Medium | Shared types, contract-first development |
| 2 | State/Persistence | 254 | 13.9% | Medium | State diagram review before changes |
| 2 | Incomplete Work | 85 | 4.7% | Medium | Checklist verification before marking done |
| 2 | Deploy/Infra | 143 | 7.8% | Medium | Staging validation, smoke tests |
| 3 | Direction Corrections | 333 | 18.3% | Low | Normal workflow — not actionable |
| 3 | Bug Fix Requests | 319 | 17.5% | Low | Ambiguous attribution |
| 3 | Error Reports | 76 | 4.2% | Low | Informational |
| Priority | Error Class | Mitigation | Type |
|---|---|---|---|
| 1 | Persistent Failures | Add CLAUDE.md rule: "When a fix doesn't work on first try, perform root-cause analysis before attempting a second fix. Never retry the same approach." | Convention |
| 2 | State/Persistence | For stateful features (auth, sessions, cache), require agent to describe the state machine before modifying code. | Convention |
| 3 | UI/Visual Bugs | Enforce browser verification for all UI changes (already in CLAUDE.md — compliance is the gap). | Convention |
| 4 | Scope Creep (Codex) | Tighten Codex task scoping: one specific change per prompt, explicit "do not modify other files" when needed. | Workflow |
| 5 | Type/API Mismatches | Shared types between frontend/backend; agent must verify API contract before modifying either side. | Convention |
| 6 | Git Workflow | Pre-push hook enforcing branch policy; already in CLAUDE.md but Codex doesn't always read it. | Hook |