basd-coding-dispatch

Use this skill to keep mobile-controlled coding work inside a disciplined Hermes/OpenClaw workflow.

The dispatcher owns the gates. Codex, Claude, and other validated tools are worker backends selected by routing policy, not the center of the workflow.

Core Loop

1. Classify the request.

• Identify whether it is a question, tiny fix, feature, refactor, debugging task, review, release task, or integration task.
• Identify risk: user-facing behavior, security, data loss, public docs, dependency changes, install paths, workflow changes, or provider claims.
• Select the quality profile and edge-case pack. Tiny fixes still receive lightweight independent review.

1. Load or declare the Superpowers activation map.

• Start with using-superpowers when available.
• Map the work to the relevant phases: brainstorming, using-git-worktrees, writing-plans, executing-plans, subagent-driven-development, dispatching-parallel-agents, test-driven-development, systematic-debugging, requesting-code-review, receiving-code-review, verification-before-completion, finishing-a-development-branch, and writing-skills.
• If a required skill/reference is unavailable, state the compensation gates instead of implying it loaded.

1. Select the worker-routing shape.

• Choose one implementer when the task is narrow.
• Split implementation and review when the task is risky or broad.
• Use executing-plans for 1-2 approved tasks, subagent-driven-development for 3+ owned task slices, and dispatching-parallel-agents only for independent domains.
• Keep provider mechanics in references/ and integration docs.

1. Run the Brainstorming Gate when required.

• Use it before implementation planning for broad, ambiguous, user-facing, UI/design, architecture, workflow, public-doc, or public-claim work.
• Exact low-risk tiny fixes stay compact; use a micro-brainstorm only when ambiguity exists.
• Do not let the implementation plan become the first place requirements are invented.

1. Run the spec gate when required.

• Produce a concise spec before code when requirements are ambiguous, user-facing, broad, public, or risky.
• Wait for approval when the user asked for gates or when the scope is not clear.

1. Run the implementation-plan gate.

• Name files, behavior, TDD applicability, expected RED failures, execution mode, review packet, review-intake policy, verification commands, rollback, and branch closeout boundaries.
• Do not write feature code before the plan is approved when plan approval is required.
• Use references/plan-linter.md as a human checklist for feature, risky, public, cross-module, or ambiguous work.

1. Isolate and implement.

• Use a worktree after implementation-plan approval when feature work needs isolation.
• Keep changes scoped to the approved plan.
• Preserve unrelated user changes.
• Avoid worker lock-in unless the integration requires it.

1. Review independently and run review intake.

• Use a separate reviewer or subagent when available for meaningful changes.
• Use lightweight independent review even for tiny fixes.
• Build the reviewer packet from references/review-packets.md.
• Reconcile findings with receiving-code-review: read, restate, verify, evaluate, disposition, then fix only accepted in-scope findings.
• Preserve valid bonus findings as backlog/input instead of discarding them.

1. Verify before completion.

• Run the smallest commands that prove the work, then broader checks when risk justifies them.
• Include Superpowers evidence, or explicit compensation with external gates when a required skill/reference was unavailable.
• Report commands, status, skipped checks, and residual risk.

1. Finish the branch explicitly.

• Use finishing-a-development-branch after verification to present closeout options.
• Do not push, PR, merge, publish, release, deploy, delete branches, or discard work unless the human explicitly approves.
• Update skills or process memory with writing-skills discipline only when useful and in scope.

A2 run manifest / gate ledger is intentionally not part of the v0.1 public skill requirements.

References

Load only the reference needed for the task:

• references/quality-gates.md for gate criteria.
• references/quality-profiles.md for risk-tiered defaults, review levels, and test escalation.
• references/edge-case-packs.md for short recurring-miss checklists.
• references/review-packets.md for independent-review inputs.
• references/superpowers-integration.md for Superpowers phase map, evidence, and compensation.
• references/prompt-templates.md for dispatch, review, review-intake, closeout, and final-report prompt shapes.
• references/plan-linter.md for implementation-plan checklisting.
• references/provider-command-recipes.md for Hermes/OpenClaw install and worker command patterns.
• references/session-topology.md for single-worker, split-worker, and subagent execution shapes.
• references/review-orchestration.md for independent review and review intake.
• references/subagent-skill-bundles.md for assigning focused subagent work.