feat(pipeline): Golden Chain for Cloud Dev — tri pipeline for agent swarm #140

Context

Cloud Dev agent swarm works: issue → container → Claude Code → PR. But the pipeline needs hardening. Currently manual spawn via Railway API, no feedback loop, agents can't verify their own code.

Golden Chain: tri pipeline for creating agents

Phase 1: tri cloud pipeline (автоматизация)

tri cloud pipeline <issue-number>
  1. Read issue from GitHub
  2. Spawn container on Railway
  3. Monitor status via heartbeats
  4. If agent STUCK >10min → kill + respawn
  5. If agent DONE → verify PR (zig build locally)  
  6. If PR passes → auto-merge
  7. Cleanup container

Phase 2: Self-healing loop

• Agent fails → auto-create fix-issue with review findings
• Fix-issue auto-spawns fix-agent
• Fix-agent pushes to same branch → PR auto-updates
• Max 3 retry loops per original issue

Phase 3: Improvements needed

Entrypoint fixes

• Advisory self-review (zig build removed — too heavy for Railway)
• Informative Telegram messages (issue title + detail)
• Fix heartbeat subshell bug (doesn't see parent vars)
• Add set -o pipefail for proper exit code capture
• Telegram streaming of Claude Code output (PR feat(cloud): Stream all container logs to Telegram in realtime #133)
• HTML escape for Telegram messages
• Proper exit after "staying alive" phase (don't restart)

Orchestrator fixes

• tri cloud spawn should work from CLI (not just Railway API)
• tri cloud status — real-time dashboard in terminal
• tri cloud pipeline — full Golden Chain automation
• Service reuse (hit 25 service/day limit — need to recycle)
• Container resource limits (memory, CPU, disk)

CI/CD fixes

• zig build in container: needs -Dci=true flag or lighter build
• Pre-cache zig-cache in Docker image for faster builds
• Agent should git checkout existing branch (not create new) for fix-issues

Monitoring

• JSONL event persistence (PR feat(cloud): JSONL event persistence + deduplication + JSON history format (#124) #129)
• ACI protocol (PR feat(cloud): structured ACI protocol (#126) #132)
• Dashboard UI (web page with live container status)
• Cost tracking per agent (Railway API billing)
• Agent solve rate metrics (success/fail/retry counts)

Agent intelligence

• Agent reads CLAUDE.md and follows project conventions
• Agent can run zig build -Dci=true for lighter build check
• Agent self-reviews diff before pushing
• Agent creates meaningful commit messages (not generic)
• Multi-file awareness (reads related files before editing)

Architecture

tri cloud pipeline <N>
  └─ tri cloud spawn <N>           # Railway API
      └─ agent-entrypoint.sh       # Container boot
          └─ claude -p "..."        # Claude Code via z.ai
              └─ git push + gh pr   # Result
  └─ tri cloud monitor <N>         # Watch heartbeats
  └─ tri cloud verify <N>          # Check PR locally
  └─ tri cloud merge <N>           # Auto-merge if clean
  └─ tri cloud cleanup <N>         # Destroy container

Exit criteria

• tri cloud pipeline 999 spawns agent, monitors, verifies, merges, cleans up
• Zero human intervention for simple issues
• Max 3 auto-retries before escalating to human