DocsConcepts
Hooks and native features
Hooks are optional. Only Claude Code and Codex run them, and no pipeline depends on them. Below the hooks, a table shows how the blueprint uses native features of Claude Code, Codex and Hermes. Where a feature overlaps something the blueprint does, the table says why the blueprint keeps its own mechanism as the default. Each tool’s support note on GitHub says what is missing or different there:
Hooks in each tool
| Tool | Hooks |
|---|---|
| Claude Code | 10 handlers |
| Codex | 5 handlers, after you trust them in /hooks |
| Antigravity | None |
| Grok Build | None |
| Pi | None |
| Cursor CLI | None |
| Hermes | None |
| Amp | None |
Native features in each tool
The blueprint uses a tool’s native features only as opt-ins: no pipeline depends on a gated or experimental capability. Where a native feature overlaps something the blueprint does, the table says why the blueprint’s own mechanism stays the default.
| Native feature | How the blueprint uses it | Requirements |
|---|---|---|
Claude Code Agent Teams and Codex multi_agent_v2 |
Optional extras for ab-orchestrate on top of the ledger; team work runs without them in every tool |
Switched on by you: CLAUDE_ (interactive sessions only) or multi_agent_v2 = true under [features] |
Claude Code /goal |
A complement for interactive runs; the ship runner and its state file are the guarantee, because a skill cannot invoke /goal and it does not run headless |
None; generally available in the CLI |
Native /loop and ScheduleWakeup (Claude Code) |
Adds scheduled reruns, but /loop stays in one session and does not reset context, stop on repeated failure or notice a run getting worse, so ab-autonomous-loop keeps its own circuit breaker and degradation checks |
None |
Workflow tool, /workflows, ultracode (Claude Code) |
Opt-in for very large autonomous fan-outs; wave orchestration stays the portable default | Paid plans, the API and Bedrock, Vertex or Foundry; in claude -p only behind a Workflow allow rule, auto or bypass mode, or a PreToolUse hook |
| Fast mode (Claude Code) | Opt-in only | Opus 5.5, Opus 5 and Opus 4.8; research preview, pricing subject to change |
| Session limits (Claude Code) | Swarms and research stay within the native limits: since CLI 2.1.224 there is no per-session subagent total, only a concurrency cap and a nesting depth; host-limits.tsv records the other tools’ limits |
20 concurrent subagents by default (CLAUDE_), spawn depth 3 (CLAUDE_), 200 web searches per session |
| Native injection hardening (Claude Code Agent tool) | Adds to the blueprint’s own read and write scanners (Claude Code and Codex hooks) and the data markers helpers wrap around external text, which cover the main-session reads and writes native hardening does not see | None |
Bundled /deep-research workflow (Claude Code) |
Claude Code bundles a web-search workflow of that name; the blueprint’s research swarm is ab-deep-research, so the two names do not collide |
None |
| Codex plugin listing | Codex’s $ picker shows each skill’s display name and one-line description, and Codex offers ab-project-start right after install |
Checked on Codex 0.160.0 |
| Hermes skill security scan | Skills, prompt files and instruction files carry none of the phrases Hermes treats as injection, no instruction to edit the instructions file by name, and no HTML comments; the portability gate checks the same | Runs on every skill Hermes installs or indexes |