start

Activation adapter: claude-code.plugin.command.v1. Activation capture capability: unsupported.

This file is the live Claude Code plugin command /agenticloop:start, registered through .claude-plugin/plugin.json. The Claude Code plugin surface has no host-owned parser capture channel, so model-visible command arguments are advisory context only and are never activation proof. Do not serialize the substituted argument, the surrounding prompt, or any other model-visible text into activation capture JSON.

Each generated host artifact replaces the capability slot above with its own adapter's declaration. Every shipped host declares unsupported, and that never changes from inside a session: prompt-visible input must never be converted into capture JSON.

unsupported is not the end of the workflow. It only means this host cannot itself prove activation. The universal path is one explicit operator action outside the agent session:

npx agenticloop activate T-016 T-017
npx agenticloop activate --work-unit <work-unit-id>

That command runs in the operator's own terminal, shows the exact tasks, carriers, contract digests, repository, work unit, and resulting assurance, and requires the operator to type a confirmation. It never rewrites a task record, so existing task ids, bodies, history, and decomposition state are preserved.

When a task you need is not activated, do not stop the session and do not attempt to author activation evidence. Report the blocked activation state, tell the operator the exact command above, and continue in the same project and session once they have run it. Activation assurance is then operator_confirmed and role returns are session_reported - honest grades that are not host-authenticated and must never be described as if they were. Hardened projects still require a registered protected host adapter. Activation and return adapters are independent, and capability declarations are not evidence that either event occurred. Hardened closeout requires an observed authenticated host receipt.

Before any setup check, orientation, document loading, task selection, or delegation, normalize the supplied argument by trimming surrounding whitespace. If and only if it equals stop (case-insensitive), immediately follow agenticloop/commands/stop.md and return. Do not continue this start workflow. stop now, stop-gap fix, and other non-exact task or context arguments are ordinary inputs, not deactivation requests. Do not treat exit as an alias.

Path convention: toolkit source (AGENTIC_LOOP.md, agents/, skills/, backends/) lives under agenticloop/ (no leading dot). Target project state (project.md, tasks/, decisions/, improvements/) lives under .agenticloop/ (leading dot). .agenticloop/agents, .agenticloop/skills, and .agenticloop/backends are invalid paths – canonical assets are always under agenticloop/ without the dot. The active role contract is always under agenticloop/agents/.

Read .agenticloop/project.md first. If setup_status is unconfirmed or a confirmed map lacks a valid human-confirmed development_stage, route agenticloop/skills/setup-agenticloop/SKILL.md or confirm the profile before selecting or creating the first task.

Then read agenticloop/AGENTIC_LOOP.md, the lifecycle-at-a-glance section, and only the canonical role contract for the current action. Read the named protocol or skill section on demand; do not preload every role, backend, or skill. Keep the main session as the coordinator: it reads the selected project config and process docs, routes task authoring, review, acceptance, and closeout through the maintainer role, routes scoped implementation and revision work through the engineer role, and should not directly edit implementation files unless the human explicitly asks. Respect the Advance Authorization Boundary, blocked-state handling, decision records, event logging rules, and configured group approval gates.