Hyperflow

Hyperflow Doctrine (single-agent port)

Apply Hyperflow's behavioral floor in surfaces that load skills but do not provide the full Claude Code multi-agent runtime.

Runtime Adaptation

Codex, OpenCode, Antigravity, and Grok often run one foreground agent (or a host-specific subagent API). Where the full doctrine says to dispatch parallel workers under reviewers:

• Prefer the host's subagent API when it exists (Codex spawn, OpenCode Task, Grok spawn_subagent).
• Otherwise do the work yourself, one coherent batch at a time.
• Self-review each batch before moving on.
• Run a final integration self-review over the cumulative diff.
• Preserve the same autonomy, clarification, commit cadence, file-first artefact, no-attribution, and security rules.

Portable Function Router (Codex / OpenCode / Grok)

These hosts load Hyperflow as skills, not as native Claude-style slash commands. Treat these user messages as function aliases and execute the matching skill workflow inline in the current thread:

Do not answer that /hyperflow:* is an unknown command on these surfaces. Strip the alias, load the matching skills/<name>/SKILL.md, and follow its workflow. If that workflow says to use unavailable Claude Code tools (Agent, Skill, or AskUserQuestion), emulate them: do worker/reviewer steps inline with visible labels, continue chained skills inline, and use the interaction fallback below when a structured question UI is missing.

Subagents And Auto-Chain

Codex

When Codex exposes multi-agent tools, map Hyperflow agent dispatches to Codex subagents instead of falling back to inline work:

• Hyperflow Agent worker/searcher/writer calls map to Codex worker or explorer subagents.
• If the callable tool is named multi_agent_v1.spawn_agent, use agent_type: worker for implementer/writer execution and agent_type: explorer for search/codebase-research tasks, then collect results before review.
• Spawn independent sibling workers together when the runtime supports parallel subagent calls.
• Every agent runs on the current session model — do not switch models per role. Match reasoning effort to task complexity: low for trivial docs/config checks, medium for normal planning/review, high for debugging, architecture, security, or final integration review.
• Never request or default to xhigh.

When Codex does not expose subagent tools in the current session, use the single-agent port above: execute worker/reviewer phases inline with clear labels and continue.

Grok

Grok CLI / Grok Build loads skills from ~/.grok/skills/, project .grok/skills/, and compatible Claude/Cursor skill dirs. Project rules come from AGENTS.md / CLAUDE.md and .grok/rules/. Runtime signal often includes GROK_AGENT=1.

When Grok exposes spawn_subagent, map Hyperflow dispatches as follows:

• Spawn independent sibling workers together when the runtime supports parallel subagent calls.
• If subagents are disabled (GROK_SUBAGENTS=0 or config), run worker/reviewer phases inline with clear labels.
• Prefer the native AskUserQuestion tool for structural gates when available.
• Every agent runs on the current session model — no per-role model selection.
• Do not invent Claude Code Agent tool calls; use spawn_subagent or inline work.

Auto-chain (all portable hosts)

For /hyperflow:workflow, use the portable workflow adapter (Codex / OpenCode / Grok branches in the workflow skill) instead of falling back to scope: research and planning, .hyperflow/tasks/ progress tracking when needed, parallel subagents when exposed, inline worker/reviewer phases otherwise, adversarial verification, quality gates, per-task conventional commits, and final synthesis. Do not describe this as native Claude Code dynamic workflow support.

These hosts may not expose Claude Code's Skill handoff tool. Treat every Hyperflow handoff as an inline auto-chain:

• plan runs amplify → design → decompose inline, then stops at its build-location gate (always asked). It never auto-implements: on "this session" it continues into dispatch inline; on "another session" it writes a handoff package; on "stop" it keeps the plan.
• dispatch offers audit and deploy structural gates, then runs the selected follow-up inline.
• audit fix gates continue into plan with the generated audit-fix task (which then stops at its own build-location gate).

Do not stop with "Skill tool unavailable". Auto-chain is a behavior contract, not a host API requirement.

Interaction Fallback

When a host lacks a structured question UI (or AskUserQuestion is unavailable), do not skip the question or silently choose the recommended option. Render the same structural gate as a concise chat block and wait for the user's answer:

Hyperflow Question
<question>

1. <recommended option> (Recommended) — <short consequence>
2. <option> — <short consequence>

Use this fallback for every required clarification or structural gate: Amplify handoff, Spec chain mode, Spec brainstorming questions, Scope ambiguity questions, Dispatch audit/deploy gates, Audit fix gate, Deploy commit-inclusion and push gates, and any security/irreversibility escalation. It is still banned to ask invented confirmation questions such as "should I proceed?".

On Grok, prefer the native AskUserQuestion tool when present; use the chat-block fallback only if that tool is missing.

Reasoning Policy

• Every agent runs on the current session model — there is no per-role model selection.
• Resolve reasoning effort by task/profile: low for trivial docs/config checks, medium for normal planning/review, and high for debugging, architecture, security, and final integration.
• Never default portable hosts to exotic max-effort modes (e.g. Codex xhigh).

Core Rules

1. Execute task-shaped requests without confirmation.
2. Clarify only after reading the relevant code and only for genuine ambiguity.
3. Keep long-form plans, specs, task decompositions, and audits under .hyperflow/.
4. Use conventional commits, one distinct user task per commit.
5. Never reference the model as the actor in commits, docs, comments, task files, or memory.
6. Respect the security blocklist in security.md.

Workflow Routing

For full multi-agent doctrine, read DOCTRINE.md and the linked reference files in this directory.