hephaestus-build

name	hephaestus-build
description	Use when the user types /prompts:hep-build or /agentlas-build, mentions @Hephaestus for build work, asks to create a single Agentlas agent, create a multi-agent team, or package an existing local/external agent into Agentlas architecture.

Hephaestus Build

Procedure

1. Treat this as the public Codex build surface. Do not expose or ask the user to invoke the older internal support skill names.

2. Resolve the trusted engine root before reading any build contract:

   ENGINE=""
   for candidate in \
     "${CODEX_PLUGIN_ROOT:-}" \
     "${CLAUDE_PLUGIN_ROOT:-}" \
     "${PLUGIN_ROOT:-}" \
     "$HOME/.agentlas/runtime/current/host_adapters/codex/plugins/agentlas-core-engine-meta-agent" \
     "$HOME/.agentlas/runtime/current/host_adapters/claude/plugins/agentlas-core-engine-meta-agent" \
     "$HOME/.agentlas/runtime/current" \
     "."
   do
     if [ -n "$candidate" ] && [ -f "$candidate/AGENTS.md" ] && [ -f "$candidate/package-contract.json" ] && [ -f "$candidate/contracts/builder-interview-research-gate.md" ]; then
       ENGINE="$candidate"
       break
     fi
   done
   [ -n "$ENGINE" ] || { echo "Hephaestus engine not found. Run the installer first." >&2; exit 1; }
   RUNNER=""
   for candidate in "$HOME/.agentlas/runtime/current/bin/hephaestus" "$ENGINE/bin/hephaestus"; do
     if [ -x "$candidate" ]; then RUNNER="$candidate"; break; fi
   done
   [ -n "$RUNNER" ] || { echo "Hephaestus runner not found." >&2; exit 1; }
   

   For an interactive request whose first argument is session, do not resolve a generic PACKAGE_TARGET and do not ask for a JSON/JSONL export. Ask first: "이 세션에서 만든 에이전트를 기본 전역 Agentlas 에이전트 폴더에 만들까요? 다른 위치를 원하면 경로를 알려주세요. 별도 위치를 지정하지 않으면 전역 폴더에 만듭니다." If no alternate location is named, use AGENTLAS_AGENT_HOME or ~/.agentlas/agentlas-agent and create a new safe-slug child package. If a destination is supplied, validate that exact folder. Never overwrite an existing child. The current conversation is the source; do not inspect recent sessions or host databases.

   For every other build request, take exactly one folder explicitly named or confirmed by the user as PACKAGE_TARGET. If none was named, or multiple candidates exist, stop and ask. Never default to ., the cwd, or $ENGINE. Run "$RUNNER" contract resolve-target "$PACKAGE_TARGET" --base "$PWD" and set PACKAGE_ROOT only to the status-ok receipt's exact package_root. Nonzero or error receipts are blockers.

   Read $ENGINE/AGENTS.md, $ENGINE/.agentlas/mode-map.json, the selected mode contract under $ENGINE/modes/, and $ENGINE/contracts/builder-interview-research-gate.md. Do not substitute files from the user's package workspace.

3. Run the public mode classifier by independent ownership boundaries, not by keywords such as "team":

   • package or repair existing material -> 30-agentlas-packager;
   • current interactive conversation invoked with session, or an explicitly exported session in a terminal/headless run -> 40-session-agent-builder;
   • one independently owned context/tools/success standard -> 10-single-agent-builder;
   • two or more roles with separate context, permissions, success standards, handoff, or synthesis needs -> 20-multi-agent-team-builder. If the shape is unclear, ask before generating. The user-facing question must be plain language, for example: "이 일을 한 명의 전문가가 처음부터 끝까지 맡으면 되나요, 아니면 조사/분석/검토처럼 여러 전문가가 나눠 맡고 마지막에 합쳐야 하나요?" Do not expose internal labels such as single-agent, team-builder, ownership boundary, memory/context, synthesis, or produces/consumes.

4. Run the Builder Interview and Research Gate from $ENGINE/contracts/builder-interview-research-gate.md before writing substantial package files:

   • for an interactive session request, treat the current conversation as the source interview and ask only the destination question plus focused gaps that would change scope, permissions, output, or safety;
   • for other builds, ask an 8-12 question first batch when the request is vague;
   • continue follow-ups until target user, tasks, inputs, outputs, examples, role count, separated tools or permissions, final merge needs, execution order, memory, failure modes, and evals are clear;
   • phrase shape questions in everyday language. Ask who handles which part, whether each role needs different files/accounts/tools, whether someone must merge the result, and whether work can run at the same time or must pass from one person to the next;
   • research official or primary docs, similar agent repositories or comparables, GitHub examples, academic/professional theory, and tool/plugin docs;
   • compare selected and rejected tools/plugins with permission, secret, fallback, and smoke-test notes;
   • synthesize domain-expert behavior from interview answers, comparable agents/repos, theory, and tool choices;
   • write docs/builder-interview.md, docs/research-sources.md, docs/tool-selection.md, docs/domain-expert-synthesis.md, docs/prompt-performance-contract.md, and .agentlas/capability-eval-plan.json.

5. If missing narrow details still change files, adapters, or public/private boundaries, ask one to five clarify questions before generating.

6. Pick one:

   • 10-single-agent-builder;
   • 20-multi-agent-team-builder;
   • 30-agentlas-packager;
   • 40-session-agent-builder. For 40-session-agent-builder, run the current-session two-pass flow: first show a Generalized Session Report that extracts reusable intent, procedures, corrections, failed approaches, validation, and conditional IF / THEN / BECAUSE / AVOID / INSTEAD rules; then offer Build Agent or Edit. Only the approved report is converted into the standalone agent prompt and package. Do not present JSON/JSONL as an interactive input.

7. Load matching support skills.

8. Write all generated or repaired runtime agent instructions in English: AGENTS.md, CLAUDE.md, GEMINI.md, agent.md, skills, workflow/command adapters, runtime prompts, handoff contracts, return contracts, and operating docs. Translate Korean or other-language source material into English agent behavior. Localized public copy, routing trigger examples, and sample user inputs may use the target user language.

9. Before writing package files, run "$RUNNER" contract scaffold "$PACKAGE_ROOT" --mode <single|team|package>. After the routing card is authored, run "$RUNNER" contract complete "$PACKAGE_ROOT" --mode <single|team|package>. Emit or repair Agentlas contracts, including .agentlas activation seed files and .agentlas/global-commands.json when local continuity is part of the output.

10. Add the generated command to Claude Code, Codex, Gemini CLI, generic AGENTS.md, and terminal adapters. For teams, expose the orchestrator/HQ command and route workers through HQ unless direct worker commands were requested.

11. For team mode, run "$ENGINE/scripts/verify-team-package.sh" "$PACKAGE_ROOT" when that focused gate exists. If it fails, do not report completion; collapse the output to a single-agent package or add the required orchestrator/HQ and team contracts.

12. Run "$RUNNER" contract verify "$PACKAGE_ROOT" --mode <single|team|package>. A non-zero exit or non-empty blocker list means blocked, never completed. Do not use the engine repository's scripts/verify-package.sh as generated-package evidence; that script verifies the engine repository itself.

13. Once verification and local registration have succeeded, ask one final two-choice storage question using structured controls when available: Cloud에 올리기 or 로컬에만 저장. Cloud means owner-private Agent Cloud storage, restorable on another signed-in Desktop. Mobile can use the package only after a paired Desktop restores/installs it; Cloud is not a hosted LLM runtime. Local-only performs no network mutation.

14. Never upload by default. Missing input and non-interactive execution are local-only. Only after explicit Cloud consent, run the trusted Hephaestus runner with upload <exact-verified-package-root> --visibility private-link. Keep the local package on every auth/offline/CAS/quota/scan failure and report an exact retry command. Public Hub publication remains a separate explicit action.

Output

Return status, evidence, output, global_commands, interview_research, and blockers.