archi-context-fetch

Invocation

• Auto-invoke: Yes, called by 00_system.md Front Pipeline.
• Trigger location: Before user request enters protocol routing.
• Execution context: When subagent supported must execute in independent subagent/independent context; only fallback to NO-SUBAGENT / NO-SKILL inline execution when no subagent.
• Boundary: Only produce structured intermediate artifacts, does not execute protocols, does not modify project files.

Architext Context Fetch

Purpose

Read minimally necessary Architext documents based on Intent Card, produce structured Context Pack. This Skill is "librarian", only responsible for finding files, extracting facts, marking gaps; does not read product source code, does not modify files, does not execute commands, does not make final decisions for protocol.

Inputs

• intent_card: archi-intent-normalizer output.
• protocol_hint: Protocol about to load, e.g. /archi.change or change/edit.
• known_state: Lightweight state known to caller, e.g. project features, task id, current discussion file.
• available_files: Optional. Architext file list already discovered by caller.

Output Contract

Only output one JSON code block, no explanatory body text.

Continuation Contract: This Skill output is only Front Pipeline step 2. After the main agent receives the Context Pack, if missing_or_stale is empty or contains only non-blocking notes, it must continue to load and execute the protocol pointed to by next_context_action / Intent Card; do not treat the Context Pack as the final reply.

risk_flags are later protocol Gate inputs, not Front Pipeline stop markers. Stop to explain to the user only when missing_or_stale explicitly blocks execution, clarification is required, or a safety Gate must be shown first.

{
  "context_pack_version": 1,
  "for_intent": "change.edit",
  "target": {
    "task_id": null,
    "task_dir": null,
    "files": [],
    "refs": []
  },
  "must_read": [],
  "optional_read": [],
  "already_read": [],
  "relevant_facts": [
    {
      "source": "global/roadmap.json",
      "fact": "<fact directly usable for execution>",
      "confidence": 0.0
    }
  ],
  "missing_or_stale": [],
  "risk_flags": [],
  "do_not_read": [],
  "next_context_action": "<continue loading protocol / ask first / fill gap first>"
}