feat(chat): conversational create_skill — brief + confirm + explain (beat 21)

[AVADSA25]

What (Mickael's ask)

"create a skill for X" built silently with no explanation. He wants it conversational like the auto-escalation offer: "Plan → I review → approve → then go… the user should be briefed directly from CODEC in the chat."

Flow

1. "create a skill that tells me the moon phase" → CODEC briefs: "I'll build a skill for: tells me the moon phase. I'll generate it, then stage it in your Skills tab for review — nothing runs until you approve it there. Build it?" + a [Build it] / [Cancel] chip.
2. Vague ("make a skill") → "What should the skill do? e.g. 'a skill that checks the bitcoin price'."
3. [Build it] → POST /api/chat/build_skill runs the real create_skill review-gate flow → conversational outcome: "Done — I generated moon_phase and staged it in your Skills tab. Open Skills to see what it does, then Approve to make it live."

Design

• routes/chat.py: _detect_create_skill_request() intercepts the chat surface before the pre-LLM hijack, so create_skill no longer fires silently; streams the brief + a skill_confirm SSE frame. New POST /api/chat/build_skill. _skill_outcome_message() rephrases the terse result.
• codec_chat.html: renderSkillConfirmChip + Build/Cancel handlers, mirroring renderEscalateChip. No emoji.
• MCP/voice create_skill paths untouched — this conversational layer is chat-only.

Verify

Detection, confirm-decision, outcome message, and the real build path all exercised headlessly (staged a dad_joke skill via /api/chat/build_skill path, then cleaned up). pytest -k "chat or try_skill or escalate or skill_review" → 119 passed. ruff clean. No manifest change (no skills/ edit).

🤖 Generated with Claude Code