Audit Codex session duplicates and retry chains with AgentMeasure
Inspect supported Codex rollout logs locally for duplicate records, retry amplification, and repeated tool failures, preserving unknown evidence and comparing before-and-after snapshots.
npx skills add agentskillexchange/skills --skill audit-codex-session-duplicates-and-retry-chains-with-agentmeasure
Use AgentMeasure Healthcheck when a Codex operator needs to investigate repeated tool failures, inflated execution counts or retry amplification in local session logs. Select the session directory and time window, generate a local report, inspect evidence-backed findings and distinguish logical operations from repeated attempts. The workflow analyzes existing logs; it does not drive the coding agent or modify a project.
What this skill actually does
With Python 3.9 or later, create a virtual environment and install python -m pip install agentmeasure==0.7.2. Start with agentmeasure demo --no-history to inspect synthetic findings. For real sessions, use agentmeasure check --dir /path/to/codex/sessions --json report.json --no-history, choosing a directory and report destination you control. The documented default window is seven days; use --all or explicit date filters only when a broader scan is intended. Read the terminal and HTML report, identify the repeated record IDs or command failures, and propose a specific next diagnostic action.
The checks identify duplicate records, failed-then-repeated command chains and runs of repeated tool errors. Missing outcome or token evidence remains UNPROVABLE. Exit zero means a report was produced, including reports with findings; it does not certify a clean session. To assess a prompt or tool change, save a before snapshot with --save-snapshot before.json, repeat after the change using the same scope, then run agentmeasure compare before.json after.json. Disclose format, version or window mismatches rather than interpreting every delta as an improvement.
The approved boundary is local forensic review of supported Codex rollout logs. The detailed upstream support table verifies Desktop-originated Codex rollouts and marks CLI-originated logs as expected compatible but unverified; older logs may leave retry or failure checks UNPROVABLE. The root README and detailed support table differ about Claude support, so this entry makes no verified Claude Code claim. Codex is the framework mapping because its supported rollout format is central to invocation. The separate provider SDK, metering standard, consulting services, invoice disputes, live quota monitoring and claims of vendor overbilling are outside this workflow.
Keep full reports and snapshots local because they contain file paths and evidence metadata. Preview a reduced summary with agentmeasure share report.json, inspect it, and export or share only if the user explicitly chooses. The tool documents local processing and a whitelist for summaries; no private logs were opened, software installed, reports generated or security review performed during metadata verification. Healthcheck commands, evidence rules and runtime limitations.