full-sprint

name	full-sprint
description	Orchestrates a complete sprint cycle autonomously. Each phase runs as a separate claude -p session for context isolation. Use when the sprint-directive.md is ready and the sprint should execute end-to-end. Also use when the user says "run a full sprint", "execute the sprint", "build everything", "autonomous sprint", or "run the cycle".

Full Sprint Execution

You are orchestrating a complete sprint cycle. Each phase runs in an isolated claude -p session for context isolation — the review MUST NOT have the build context (it biases the review).

IMPORTANT: This skill automates the same phase-by-phase workflow as the manual sprint cycle. Quality comes from context isolation, not from running everything in one session. Each phase runs via claude -p --dangerously-skip-permissions so all hooks fire (Stop runs tests, PostToolUse scans secrets, PreToolUse blocks denylisted commands and destructive git ops).

Step 1 — Load context

Read briefs/sprint-directive.md for the sprint scope and type. Read .claude/CLAUDE.md to get the project name and context.

Step 2 — Determine sprint number

Find the highest sprint number in tasks/sprints/ and use N+1. Create the directory: tasks/sprints/sprint-XX/

Step 3 — Execute phases

For each phase, run a separate claude -p session via Bash:

claude -p --dangerously-skip-permissions \
  "You are working on [project name]. \
   Read .claude/CLAUDE.md for project conventions. \
   Read .claude/skills/[skill]/SKILL.md and follow its instructions exactly. \
   Input: [input file path]. \
   Save output to tasks/sprints/sprint-XX/[output file]. \
   Run tests before declaring done."

Why claude -p without --bare: full hook enforcement applies — the Stop hook runs tests, PostToolUse scans for secrets, PreToolUse blocks denylisted commands and destructive git operations. The CLAUDE.md, rules, and skills load automatically, so the agent gets the full project context without us having to inline everything in the prompt.

The previous v3.x recommendation to use --bare for "~10x SDK startup performance" was retired in v4.0 because the trade-offs (hooks skipped, API-only auth that breaks Claude Max OAuth, missing CLAUDE.md/rules auto-load) outweighed the startup gain. The --dangerously-skip-permissions flag alone gives the same non-interactive behavior with all guarantees preserved.

Phase sequence

Phase 1 — Sprint plan:

• Skill: .claude/skills/sprint-plan/SKILL.md
• Input: briefs/sprint-directive.md
• Output: tasks/sprints/sprint-XX/plan.md

Phase 2 — Build:

• Skill: .claude/skills/build/SKILL.md
• Input: tasks/sprints/sprint-XX/plan.md
• Output: tasks/sprints/sprint-XX/build-output.md

Phase 3 — Review:

• Skill: .claude/skills/review/SKILL.md
• Input: tasks/sprints/sprint-XX/build-output.md
• Output: tasks/sprints/sprint-XX/review-output.md
• Note: this phase invokes code-reviewer and qa-tester subagents internally

Phase 4 — Fix:

• Skill: .claude/skills/fix/SKILL.md
• Input: tasks/sprints/sprint-XX/review-output.md
• Output: tasks/sprints/sprint-XX/fix-output.md

Phase 5 — Red team (security sprint type only):

• If sprint type (from sprint-directive.md) is "security":
  • Skill: .claude/skills/red-team/SKILL.md
  • Input: tasks/sprints/sprint-XX/fix-output.md
  • Output: tasks/sprints/sprint-XX/redteam-output.md
• If sprint type is "normal" or "hotfix": skip this phase

Phase 6 — Capture lessons:

• Skill: .claude/skills/capture-lessons/SKILL.md
• Input: all sprint-XX files
• Output: tasks/sprints/sprint-XX/retrospective.md + updated lessons.md, backlog.md

Step 4 — Verify after each phase

After each claude -p call:

1. Verify the expected output file was created
2. If not created → retry ONCE with the same command, then report failure and STOP
3. After review: note if critical issues were found (informational — /fix handles them)
4. After fix: read fix-output.md and check if tests pass
   • If tests fail → report failure and STOP, do not loop

Step 5 — Post-sprint

After all phases complete:

1. Execute /update-briefs to sync shared memory
2. Create a branch and PR with sprint description:

   git checkout -b sprint-XX-[theme]
   git add -A
   git commit -m "feat(sprint-XX): [sprint goal from directive]"
   git push -u origin sprint-XX-[theme]
   

3. Report completion with summary

Channel mode (remote execution)

If invoked via Channel (Telegram/Discord), adapt your output:

• Send progress updates after each phase completes (1-2 lines)
• Send a full summary at the end (5-10 lines)
• End with: "Sprint complete. PR ready for review."
• If a phase fails, notify immediately
• Keep Channel messages concise — the human is likely on a phone

Gotchas

• Each claude -p session is fully isolated. This is intentional — a review without build context is more critical.
• Subagents within each phase (code-reviewer in /review, architect in /sprint-plan) work normally inside claude -p sessions.
• If a phase takes too long or errors out, the sprint stops. Do not attempt infinite retries.
• The --dangerously-skip-permissions flag is for dev/MVP only. For production, configure permissions in settings.json and remove the flag.
• Monitor your own context: if approaching 60%, wrap up and report status.
• The git add -A in Step 5 is acceptable here because the sprint workspace should only contain sprint changes. Review the diff before pushing if in doubt.

STOP. Your deliverable is a completed sprint with all output files and a PR. Do NOT start the next sprint. The SP-PM or human decides what's next.