feat(taskctl): pipeline agent prompts — rich defaults + external override via opencode.json

[randomm]

Goal

The developer-pipeline and adversarial-pipeline agents used by taskctl need:

1. Rich default prompts — the current stubs (16 lines developer, 8 lines adversarial) are shadows of the full prompts in lievo/prompts/
2. AGENTS.md injection — must be passed as part of the system prompt, same as all other agents
3. External override support — via ~/.config/opencode/opencode.json, same mechanism as developer/ops/etc.

Current vs lievo/ gap

	Current	lievo/ reference
Developer prompt	~16 lines, basic TDD steps	911 lines — full workflow, quality gates, memory protocol, completion format, tool preferences
Adversarial prompt	~8 lines, basic instructions	509 lines — attack vectors checklist, verdict format, re-review workflow, memory protocol

Critical design question: verdict format

lievo/ adversarial prompt expects raw JSON output from the agent:

{ "verdict": "APPROVED", "acceptance_criteria_check": {...}, "critical_issues": [...], ... }

Current pipeline uses taskctl verdict tool calls instead. These are incompatible. Before upgrading prompts, must decide:

• Keep taskctl verdict tool approach (current) and adapt lievo prompt accordingly
• Switch to JSON output parsing (lievo approach)

Required changes to lievo/ prompts

Replace all remory references with:

• vipune search / vipune add (memory)
• colgrep (semantic code search)

Prompt composition design (not yet approved)

Each pipeline agent session should receive:

1. AGENTS.md from project directory
2. Rich default system prompt (from lievo/ as reference, updated)
3. External override from ~/.config/opencode/opencode.json if present

Check how existing agents (developer, ops) handle override mechanism and replicate.

No implementation without explicit design approval.