Story

Codex CLI: This skill is triggered by description matching. State any arguments in your message. Sub-agent spawning is not supported — perform any review steps inline.

What /story does

/story <parent#> walks the open sub-issues of a parent ("story") ticket in their GitHub-defined order, threading /start and /submit-for-review for each one into a single interactive session. It removes a few routine per-ticket prompts (/start's plan-confirm gate and verified-locally prompt) — the operator already approved the whole story plan at story start.

What /story deliberately does not do: it does not auto-decide on /submit-for-review's needs-attention warnings or must-address blockers. Those remain human-judgment moments. When the review stops, the driver pauses and waits — when the operator finishes, the driver resumes with the next sub-issue.

/story is intentionally a thin coordinator. See #167 for the design discussion and #177 for the deferred path that would automate judgment moments based on observed usage.

Step 1 — Argument check

Required: a single integer (the parent issue number). If missing or non-numeric, abort:

"/story requires a parent issue number. Usage: /story <parent#>"

Step 2 — Fetch and echo the story plan

Find owner/repo if not already known:

gh repo view --json owner,name

Verify the parent and fetch its sub-issues:

gh issue view <parent#> --json number,title,state
python3 CodeCannon/skills/github-agile/scripts/list-sub-issues.py <parent#>

Stop immediately if any of these hold:

• Parent issue is closed → "Issue #<parent#> is closed. /story requires an open parent."
• Sub-issue list is empty → "Issue #<parent#> has no sub-issues. /story requires a parent with linked sub-issues."
• All sub-issues are closed → "All sub-issues of #<parent#> are closed. Story is complete."

Otherwise, echo the plan (filter to state == "open" for the work list; show closed ones for context):

Story #<parent#>:

Open sub-issues (worked in this order):

1. #
2. #

...

Closed sub-issues (already complete): <"none" or list>

Routine per-ticket prompts (plan-confirm at /start and verified-locally at end of /start) are auto-passed under the driver. Review-stage prompts (needs-attention warnings, must-address blockers) still stop and ask — they are the human-judgment moments.

Type go to start, or share concerns first.

Wait for the operator. Proceed only on unconditional go. Treat any other response as discussion — address concerns and re-ask. If the operator abandons, stop with nothing to clean up.

Step 3 — Ticket loop

For each open sub-issue, in the order returned by the API:

3a — Pre-flight escalation check

Read the sub-issue body (already fetched, or re-fetch via gh issue view <sub#> if needed). Stop and ask before invoking /start if any of these triggers fire:

• The acceptance criteria contain phrases like "TBD", "decide later", "?", or otherwise indicate unresolved scope.
• The body mentions a sensitive surface: authentication, authorization, payments, billing, secrets, credentials, production configuration, destructive operations (DROP TABLE, rm -rf, git push --force, mass deletes, schema drops).
• The body or any project-level guidance (e.g. CLAUDE.md) names a domain the operator has flagged for explicit review.

If any trigger fires:

"Sub-issue #<sub#> triggered escalation: .

• proceed — invoke /start for this sub-issue anyway (operator confirms they want it under the driver).
• defer — skip this sub-issue for now and move on to the next.
• stop — exit the driver with a partial summary."

Wait for the operator. Honor the choice exactly.

3b — Invoke /start <sub#> under the driver

Print this preamble line immediately before invoking the skill (this is the signal /start recognizes):

[story-driver: parent=<parent#> ticket=<index> of <open-count>]

Then invoke /start <sub#>. /start recognizes the preamble (see Case B in /start) and auto-passes its plan-confirm gate and verified-locally prompt. All other /start behavior is unchanged — including writing the code.

3c — /submit-for-review runs from /start's auto-proceed

/start under the driver auto-fires /submit-for-review itself (because its verified-locally prompt is auto-approved). The driver does not invoke /submit-for-review separately — it observes the outcome and routes accordingly. The tier routing applies as today:

• clean or informational → auto-merge. The driver continues to the next sub-issue.
• needs-attention → /submit-for-review itself stops and asks the operator (address now / follow up later / accept as-is). The driver waits for the operator to respond through /submit-for-review. When /submit-for-review completes (either by merging or by routing back to coding), the driver resumes — either with the next sub-issue (on merge) or by re-running /submit-for-review after the operator fixes things.
• must-address → /submit-for-review itself stops; the operator fixes the blocker; the driver resumes on the next /submit-for-review invocation.

3d — Per-ticket safety

Track an attempt counter per sub-issue (start at 1; increment on each /start → /submit-for-review pass for the same sub-issue):

• Attempt cap of 2: if a sub-issue would require a third pass, stop and ask:

  "Sub-issue #<sub#> has hit the attempt cap (2). Continue manually or stop the story?"

• No-progress guard: before each attempt past the first, capture git diff origin/dev on the feature branch. After the attempt, if the new diff is byte-identical to the previous attempt's diff, the iteration made no code change. Stop and ask:

  "Sub-issue #<sub#> made no code progress on this attempt. Stop the story or take over manually?"

Both guards exist to prevent unbounded retry loops; neither tries to be clever.

Step 4 — Track follow-up tickets created during the story

During each /submit-for-review run, follow-up tickets may be created (the needs-attention "follow up later" branch and the Step 9 selective branch both do this). Keep a running list of those follow-up issue numbers and their titles for the story summary in Step 5. The list is purely informational — sub-issue closure status is still the source of truth for "what got done."

Step 5 — Story session end

When the loop exits (all open sub-issues processed, or operator chose stop), post a session summary as a comment on the parent issue. Skip the comment silently if nothing actually merged in this session.

Create a temp directory if not already created:

python3 CodeCannon/skills/github-agile/scripts/make-workdir.py

Use the file-writing tool to create <tmpdir>/story_summary.md:

## /story session summary

Worked by the /story driver in one interactive session.

**Merged:**
- #<sub#> <title> → PR #<pr#>
- ...

**Deferred or not addressed:**
- #<sub#> <title> (reason: <escalation / cap / no-progress / operator-stop>)
- ... (or "none")

**Follow-up tickets created during the story:**
- #<f#> <title>
- ... (or "none")

Post via the comment-posting script:

python3 CodeCannon/skills/github-agile/scripts/post-issue-comment.py <parent#> <tmpdir>/story_summary.md

Then close out to the operator:

Story #<parent#> session done. Merged: N. Deferred: M. Open sub-issues remaining: K. If K > 0, run /story <parent#> again to resume from the next open sub-issue.

Hard rules

• /story runs entirely inside the current interactive session. No claude -p, no headless subprocesses, no daemons.
• /story does not write or edit code itself — it sequences /start and /submit-for-review. The code is written inside /start as it normally would be.
• /story does not auto-decide on /submit-for-review's needs-attention or must-address tiers. Those remain human-judgment moments.
• Sub-issue state is the source of truth. The Step 5 session summary is informational — deleting or editing it must not affect resumption. A fresh /story <parent#> invocation always re-derives state from open sub-issues.
• Attempt cap and no-progress guards stop the story rather than retry past the limit.
• If the driver itself is interrupted (operator switches tasks, session ends), no state needs to be saved. The next /story <parent#> resumes by picking up the next open sub-issue.