hyperflow-dispatch
Hyperflow execution phase. Use when a task file exists in .hyperflow/tasks/ and the work needs building — verbs like build, implement, add, refactor, "wire up", "run the plan", "execute the task". Works batches sequentially with self-review and per-task commits. In Antigravity there is no sub-agent fan-out — the single agent does each batch itself.
Allowed Tools
No tools specified
Provided by Plugin
hyperflow
Point it at a GitHub issue and get back a reviewed pull request. Hyperflow turns one AI coding session into a structured engineering pipeline: plan sharpens and decomposes the work, dispatch fans out parallel workers, and a domain specialist reviews every step — nothing ships unreviewed. 18 skills, 22 specialist reviewer and investigator agents, adaptive depth so a 5-line fix never triggers a deep run, and persistent per-project memory that lives in your repo and never leaves it. Runs on whatever model your session already uses — zero config, no API keys, no daemon. Works across Claude Code, Codex App/CLI, OpenCode, Grok, Antigravity, and Cursor.
Installation
This skill is included in the hyperflow plugin:
/plugin install hyperflow@claude-code-plugins-plusClick to copy
Instructions
hyperflow-dispatch — execution phase (Antigravity single-agent)
Execute the task file from hyperflow-plan. No sub-agent dispatch and no model tiers in Antigravity — you do each batch yourself, then self-review before committing. Follow the hyperflow doctrine.
Per batch
- Implement every sub-task in the batch (sequentially; they were planned as small, independent units).
- Self-review the batch diff against the level checklist:
- L1 syntax/format/obvious bugs · L2 spec compliance, naming, edge cases · L3 cross-file integration + security (secrets, injection, validation).
- Elevate to L3 when the change touches auth, data, or external input. Fix anything found before committing.
- Quality gates on affected files: run the project's lint, typecheck, and tests. Fix failures (never
--no-verify). - Commit per sub-task — one sub-task = one conventional commit (respect commitlint: lowercase subject, allowed scope). Stage only that sub-task's files; never commit files you didn't change.
- Update the task file's status block (tick the sub-task, bump progress) and append any durable learning to
.hyperflow/memory/.
After all batches
- Final integration self-review over the cumulative diff — catch cross-batch contradictions, scope leaks, and
any/type regressions. - End gate via AskUserQuestion: offer to run
hyperflow-audit(independent review) and/orhyperflow-deploy. Both are binary (no recommended marker). Never auto-push.
Rules
- A
SECURITY_VIOLATIONhalts immediately — no commit, no continue. - If the working tree is dirty with files you didn't create (concurrent work), never stage them; re-check
git statusbefore each commit. - Auto mode completes every sub-task before any summary — no partial "to resume" hand-offs.