Home / All Skills / Workflow & Core Tools / continue-codex-work
English (EN)简体中文 (ZH)日本語 (JA)
Workflow & Core Tools · daymade/claude-code-skills

continue-codex-work

Continues interrupted OpenAI Codex work only after read-codex-history verifies the selected rollout identity and complete fork/compaction lineage. Use when the user provides a Codex Session ID, asks to pick up a prior Codex run, says Codex was interrupted, fused, compacted, or stuck, or wants the current Agent to take over without `codex resume`. Restores the original business outcome, unfulfilled requests, user corrections, proven prior assets, current workspace truth, and the next action that directly advances the goal. Do not use when Codex itself natively resumed this same conversation and its prior turns or compaction state are already present in the current context.

L2 Local · Restricted workspaceV1 Commit-PinnedWrite FileDual-pass Audit
Author daymadeGitHub daymade/claude-code-skills ↗Stars 1385License MIT（仓库根 LICENSE：Copyright (c) 2025 daymade；GitHub API spdx MIT）Commit d5c4678cb5
Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

Triage at entry gate: If Codex natively resumed and brought the turns/compaction state of the same session back into the current context, neither this skill nor read-codex-history should be called; proceed directly based on the preserved context and verify the workspace. Use this skill only when a new/different agent context must take over an old rollout or when the user provides a different Session ID.

daymade-claude-code/continue-codex-work/SKILL.md
A restart is not enough to choose the second row. The deciding fact is whether the current conversation is the host-restored continuation of the same Session, not whether a Codex process was restarted.
The entry gate table (native resume vs. external takeover) is the most critical criterion of this skill, preventing redundant session reads and unintended takeovers.

Step 1: Enforce a verified Codex read receipt: Call read-codex-history to obtain a full chrono briefing, which must include prompt-ledger/state-index/rollout sources, exact session_meta.id match, root-to-child fork lineage and inherited byte boundaries, selected user/assistant turns, compaction, end reason, pending calls, and gaps.

daymade-claude-code/continue-codex-work/SKILL.md
Fail closed on a mismatched rollout, missing child rollout, ambiguous physical copy, missing parent, broken byte boundary, or continuation cue with no recoverable parent objective. Never continue from the “closest” file.
The fail-closed criteria are more granular than the Claude version (accounting for Codex's more complex fork/physical copy model).

Step 2: Reconstruct an isomorphic 7-field continuation contract; when a child session contains only 'continue/fork', inherit tasks from the verified parent snapshot—local cues are not standalone objectives.

daymade-claude-code/continue-codex-work/SKILL.md
If the child contains only “继续,” “continue,” or `/fork`, inherit the task from the verified parent snapshot. A local cue is not a standalone goal.
Fork lineage (root-to-child inherited byte boundaries) is a core concept unique to the Codex version.

Step 3: Reconcile with reality: Inspect AGENTS.md/CLAUDE.md in the target project, check cwd/branch/working tree/external states, and verify whether recorded patch/commit/push/download/job/agent outputs actually landed; for transient failures (usage limit/service error), check if the original process recovered on its own before deciding to redo.

daymade-claude-code/continue-codex-work/SKILL.md
rollouts record attempted actions too.
Shares the same philosophy as the Claude version: an attempt is not success, and recovery does not necessarily require redoing work.

Step 4: Execution and gates: Next direct action + feedback loop; ask before branching: 'If this succeeds, which original unfulfilled result becomes smaller?'; prohibit codex resume/--continue; do not overwrite unrelated changes.

daymade-claude-code/continue-codex-work/SKILL.md
Do not run `codex resume` or `codex --continue`.
Together with the Step 5 reporting contract (including Session lineage), this forms a complete closed-loop execution.

2Core Capabilities

01Native resume vs. external takeover determination (Entry gate)
02Verified Codex read receipt (rollout identity + fork/compaction lineage + inherited byte boundaries)
037-field continuation contract + parent snapshot task inheritance (when child session contains only continuation cues)
04Current workspace/external state reconciliation (landing checks for patch/commit/push/download/job)
05Goal-reduction test against drift + prohibition of codex resume + contract-based reporting

3External Dependencies

KindDependency / Endpoint
clicodex（宿主 CLI，禁止 resume/--continue）
skillread-codex-history（同仓兄弟技能，读本地 Codex 会话库）

4Recommended Governance & Evidence

Suggested Controls for Safe Operation
For Agent Runtime
Restrict working directory
For Administrators
Retain operation logs
Evidence & Triggers (File Paths & Excerpts)
  • 本地工作目录: File I/O and artifact output: (Executed by the agent as part of normally authorized current project work—the skill itself contains no write directives; read receipt and contract are mandatory before changes)
Audit Notes: Pure SKILL.md orchestration: no scripts, no network, no credential reading, no custom write implementations; functionality = continues business tasks in the current working tree after obtaining a verified receipt via read-only local session parsing from sibling read-codex-history (project writes are agent's normal authorized work). Caveat: read contents contain full past sessions (privacy sensitive); this skill directory contains only SKILL.md, delegating all parsing to sibling skills.

5Second-Pass Verification

  • [ok] 无脚本/无网络/无凭证 — 目录仅 SKILL.md，无任何命令执行/网络/凭证读取指令
  • [ok] 引用的兄弟技能存在 — read-codex-history 与 continue-claude-code-work 均在 daymade-claude-code/ 下
  • [ok] Entry gate 自洽 — ‘deciding fact is whether current conversation is host-restored continuation of same Session, not whether process restarted’ 前后无矛盾
  • [ok] fail-closed 条件有可操作性 — mismatched rollout/missing child/ambiguous copy/missing parent/broken byte boundary 均为可判定条件
  • [ok] 不夸大能力 — 宣称只做‘external takeover 续接’，原生 resume 场景被显式排除——与 SKILL 内容一致

6Verdict & Scenarios

  • Precisely distinguishes native resume from external takeover, preventing redundant session reads
  • Fail-closed verification of Codex rollout/fork lineage (rejects byte boundary and physical copy ambiguities)
  • Principled handling of transient failures: checks if original process self-recovered before redoing, avoiding duplicated effort
  • Completion unit strictly anchored to business outcomes; reading, parsing, or review passing alone does not count as success
    • Best for: Ideal when a new agent context needs to take over an interrupted/compacted/stuck Codex rollout (user provided Session ID), or when a forked session needs to continue development inheriting tasks from a parent snapshot; emphasizes workflows focused on 'no duplicated work and no drift'.
    Not for: Not suitable for sessions natively resumed by Codex with full context intact (the SKILL mandates proceeding directly); not suitable for Claude Code sessions (use continue-claude-code-work); not suitable for users seeking only history summaries/retrieval (delegated to read-codex-history).
    Install Agent-Ready
    ① Verified Mirror 2026-09-09
    Option A · Direct DownloadDownload continue-codex-work.tar.gz
    sha256: 5f48ef1a57b0b2b3…
    Option B · Copy Guide for Agent
    install.json
    Agents automatically download and verify mirrors from this JSON guide.
    ② Upstream GitHub
    GitHub Repo ↗
    Pinned to commit d5c4678cb5.