| name | continue-loop |
| description | Keep iterating on the next logical step in a project until a clear stop condition is reached. Use when the user says `/continue-loop`, asks to "keep going", "continue iterating", or wants Codex to repeatedly choose the highest-leverage next task, execute it, verify it, and then continue. Good for open-ended cleanup, refactoring, polish, debt burn-down, and project advancement where the next step should be discovered from the current repo state instead of handed over explicitly. |
Continue Loop
Use this skill to keep making progress without waiting for the user to spell out every step. Treat the loop as a sequence of small, verified passes, not a license to thrash.
Workflow
- Establish the bounds.
- Respect any explicit scope, count, stop condition, or quality bar from the user.
- If the user gave no bound, keep going until you hit a real blocker, diminishing returns, or a clean handoff point worth surfacing.
- Initialize or resume the loop state.
- Resolve
SKILL_DIR to the directory containing this SKILL.md.
- Run:
python3 "$SKILL_DIR/scripts/continue_loop.py" start --cwd "$PWD"
- Add
--goal "..." when the user gave a concrete objective worth preserving.
- Reuse the existing state file unless the user clearly wants a fresh loop.
- Ask the core question.
python3 "$SKILL_DIR/scripts/continue_loop.py" next --cwd "$PWD"
- Answer:
What is the next logical step to focus on next?
- Consider only a few concrete candidates. Pick the step that best matches user intent, dependency order, and narrow verification.
- Do the work.
- Read the relevant code, tests, docs, and repo instructions first.
- Implement the chosen step fully enough to verify it.
- Prefer finishing one seam cleanly over opening several half-done branches.
- Verify before claiming progress.
- Run the smallest meaningful proof for the area you changed.
- If the repo requires screenshots for UI work, take them.
- Record the pass.
python3 "$SKILL_DIR/scripts/continue_loop.py" record \
--cwd "$PWD" \
--focus "Short name of the step you just completed" \
--summary "What changed and why" \
--verification "Exact command(s) or checks you ran"
- Add
--next-hint "..." if the current pass revealed an obvious follow-up.
- Use
--status blocked only when you truly cannot proceed safely.
- Decide whether to continue.
- Continue immediately when there is another clear, valuable step.
- Stop when the next step would be speculative, too risky for reasonable assumptions, or requires a user decision.
- When stopping, run:
python3 "$SKILL_DIR/scripts/continue_loop.py" stop \
--cwd "$PWD" \
--reason "Why the loop should stop here"
Choosing The Next Step
Rank candidate steps against:
- dependency order
- user-stated goals
- architectural leverage
- likelihood of unblocking later work
- risk of breaking unrelated behavior
- cost of focused verification
- signal from failing tests, TODOs, logs, profiler output, or unfinished seams
Prefer:
- steps that unblock more work later
- deleting dead paths over adding compatibility or fallback layers
- finishing an in-flight change before hopping elsewhere
- narrow proofs over broad "maybe everything still works" runs
- small passes that can be explained in one or two sentences
Avoid:
- bouncing between unrelated areas
- speculative rewrites without evidence
- analysis-only loops when you can implement
- broad validation churn before the local change is stable
- infinite looping after the work has clearly reached a pause point
Script Notes
- The helper script stores loop state under
$CODEX_HOME/state/continue-loop by default so the repo stays clean.
- Override the location with
--state-file /path/to/state.json when you want an explicit shared log.
- Use
status to inspect the current loop summary:
python3 "$SKILL_DIR/scripts/continue_loop.py" status --cwd "$PWD"
- The state log is there to keep the loop honest. It should help you avoid repeating the same step or forgetting why the last pass stopped.