Continue Codex Work
Continue means finish the still-unfulfilled business task. Reading history is a mandatory evidence phase, not the deliverable.
Entry gate: external takeover, not native resume
| Current situation | Action |
|---|---|
| Codex itself natively resumed this same conversation, and the restored turns or compaction state are already in the current context | Do not invoke this Skill or read-codex-history. Continue directly from the retained context, then verify current workspace and external state before changing anything. |
| A new/different Agent context must take over an earlier Codex rollout, or the user identifies another Session ID whose continuity is not already present here | Use this Skill. Step 1's verified history receipt remains mandatory. |
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.
Step 1: Obtain a verified Codex read receipt
Invoke daymade-claude-code:read-codex-history for the exact Session ID and request
the full chronological Session evidence briefing. If the ID is unknown, use that
Skill's inventory or bounded search first.
The receipt must state:
- prompt-ledger, state-index, and rollout sources used;
- selected
session_meta.idand whether it exactly matches the requested ID; - root-to-child fork lineage and exact inherited byte boundaries;
- selected and inherited chronological user/assistant turns;
- compacted context, latest plan, end reason, unresolved calls, files, and gaps.
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.
Step 2: Rebuild the continuation contract
Before choosing the next action or reporting a stage complete, fill this internal contract from the verified read receipt and the project's current stage source. Use the latest human request when no project stage source exists.
| Field | Evidence required |
|---|---|
| Original business outcome | Earliest still-governing human request across inherited and selected timelines |
| Current explicit request | Latest human request that is not only a continuation cue |
| Current delivery stage / non-goals | Latest human correction narrowing what to deliver now and what to defer; do not promote the original long-term ambition back above it |
| Already completed | Independent completion evidence, not Agent narration or an assistant final. Bind it to artifact/session identity and input conditions; reuse it while those conditions are unchanged. Read volatile runtime and delivery state fresh. |
| Still unfulfilled | Requested result without completion evidence |
| User corrections / do-not-repeat | Human messages rejecting a route, assumption, or output |
| Proven assets and successful routes | Existing code, documents, Skills, outputs, commands, and prior successful experiments |
| Next direct action | One action whose success visibly reduces the unfulfilled business result |
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.
Step 3: Reconcile with current reality
- Read the current target project's
AGENTS.md/CLAUDE.md. - Confirm cwd, branch, working tree, relevant files, and external state.
- Verify that recorded patches, commits, pushes, downloads, jobs, or Agent outputs actually landed; rollouts record attempted actions too.
- Retrieve and reuse prior successful assets before creating a replacement.
- For transient failures such as usage limits or service errors, verify whether the original process later resumed and finished before duplicating work. For an existing native Goal, use Native Goal reconciliation to distinguish unfinished business, runtime status, supported recovery and observed automatic continuation. An absent recovery tool in one tool menu is not proof that the live host has no recovery control.
- Check what was in flight when the session stopped. A long task may survive
its controller; a start record proves neither completion nor process death.
Read its existing receipt/output and verify the current process before deciding
whether to monitor, finish archival or rerun. If the process is gone and the
result is unconfirmed, give an authorized rerun a detached
result-to-disk channel (e.g. remote
( cmd > log 2>&1 </dev/null; echo $? > rc ) &polled through the rc file — prefersetsid/nohupwhere the target has them, since a pty-backed session can still SIGHUP a bare subshell) so the next interruption cannot take the result again.
In-flight CI example: An existing result.json names commit H, owner O
and a running test phase; its lifecycle marker records PID P. Before starting
another CI, match P's live argv to H and the exact output path, match its cwd
and child test phase to that receipt, and observe log growth or a child completing
the documented timeout fixture. A matching live operation should be monitored
through its original result, archive and cleanup, keeping its owner and accumulated
budget. A vanished controller or an old receipt mtime does not authorize a second
run; PID existence alone does not establish identity. Reconcile any final-head
delta separately and rerun only the affected checks. This process check follows
Step 1 for external takeover; native continuation still uses the entry gate above.
Step 4: Execute the business task
Perform the next direct action and close its feedback loop. Reviews, Skill work, infrastructure cleanup, format polish, and extra safety machinery remain subordinate unless they directly unblock the original result or the user explicitly promoted them.
Before every material branch, ask: “If this succeeds, which unfulfilled result in the current authorized stage becomes smaller?” No answer means the branch is not continuation work. A blocked source, login or provider route blocks only dependent work; execute other necessary authorized work before declaring the whole task blocked.
Verify the result at the user's consumption surface: searchable bytes alone do not prove a readable page, and generated suggestions do not prove saved human decisions. Reuse passed checks for unchanged inputs; expand verification only for new changes, failures or unresolved evidence gaps.
Do not run codex resume or codex --continue. Do not overwrite unrelated changes.
Step 5: Report against the contract
- Recovered goal: exact Session lineage and original business outcome.
- Executed: actions and independent verification.
- Now complete: contract rows actually closed.
- Remaining: unresolved rows and exact blockers.
- Not repeated: relevant rejected routes deliberately avoided.
Context recovery, parser success, a green review, or a finished subprocess is not by itself successful continuation. The user's business result is the completion unit.
