
from codexgo19
Recover the previous actionable Codex request for the current workspace after a crash, compaction, or context loss, so work can resume quickly and accurately.
CodexGo restores the last actionable Codex request for the current workspace when a session is interrupted by compaction, a crash, or other context loss. It reads the saved session snapshot and surfaces a concise recovery summary: the resolved request to continue, the literal last user message, supporting context, and ambiguity hints. This lets the agent or user decide to resume, re-run, or request clarification without guessing.
Use CodexGo at the start of a fresh session when the user explicitly asks to recover the prior Codex context (for example by running codexgo or asking to "recover the last Codex session"). Do NOT auto-trigger during an active conversation from vague prompts like "continue" — it is designed for explicit recovery scenarios.
scripts/codexgo.py — command-line runner that emits a JSON summary of the recovered request.resolved_request, literal_last_user_message, last_conversation_content, and other recovery metadata.Run the provided script (python scripts/codexgo.py --cwd "$PWD" --format json) to produce a structured JSON describing what to resume and why. Review the needs_more_context and ambiguity_hints fields before executing or continuing. If the user requests execution, briefly restate the recovered request and proceed.
CLI-friendly assistants and code-focused agents that run Python scripts and handle structured JSON outputs (Codex-style assistants, Copilot-like agents, or CLI-based agent runners).
CodexGo recovers the previous actionable Codex session request after a crash or compaction. The script is well-written Python with dataclass-based data models, proper CLI argument parsing, and sophisticated heuristic logic for classifying and resolving user messages (including bilingual Chinese/English support). It reads from the local Codex SQLite state database and rollout JSONL files. Script failed in test only because no Codex state DB exists in this environment — not a code quality issue.
Clean, well-structured skill. No security concerns — subprocess calls use list form (no shell injection), reads only local files, no network calls, no credentials. The heuristic classification logic is extensive and thoughtful but somewhat hard to verify without tests.