Skip to content

ROADMAP.md

Latest commit

 

History

History
30 lines (23 loc) · 2.91 KB

ROADMAP.md

File metadata and controls

30 lines (23 loc) · 2.91 KB
Raw

llama-smith — Roadmap

Matrix-named feature map. Each name does one job. Built on the Brain + Forge + gotcha feed already shipped.

Already shipped (the spine)

  • The Brain — detects infra, decides which smith-* skills a repo needs → .smith/manifest.json. (branch brain)
  • The Forge / Operator — writes each skill grounded in real files, cites sources, never invents. (branch forge)
  • Gotcha feed — glm-5.1:cloud finds the top-3 surprising-but-true facts, short + humanized, in the scan window. (branch gotchas)

Core (the wow-engine)

  • Memory Matrix — mine this repo's own Claude Code + OpenCode session transcripts (~/.claude/projects/<slug>/*.jsonl, OpenCode storage/) for lessons: every "use X not Y", every facepalm, every fix. The depth the file-scan can't reach. Status: not built.
  • Reality Check — opt-in active probes: ssh in, docker compose ps, real row counts, what cron actually ran vs scheduled. Runtime truth, not file guesses. Status: not built.
  • Oracle Loop — correct a forged skill once, Smith reflects and rewrites its own SKILL.md. Self-evolution. Status: not built.

Surface (what you see)

  • The Construct — Matrix HTML dashboard: all repos, skills, gotchas, exposure, bus-factor. Opens in browser. Status: generator written (lib/dashboard.mjs), needs wiring + open.
  • Glitch Radar — rank gotchas by days until it bites (churn + schedule): "the Friday cron misfires on next cutover." Predictive, not descriptive. Status: not built.
  • Blast Radius — for the one fragile point/secret, show what it takes down (dependency fan-in). Visual. Status: not built.

Advanced (power moves)

  • Smith Swarm — the 10-agent ollama launch claude deep-recon, and/or point it at all repos at once → one mega-dashboard ranking the scariest. Maxes the $80 stack. Status: not built.
  • Agent Duel — two models forge the same skill, a third judges/merges. Anti-hallucination by adversarial agreement. Status: not built.
  • Skill Drop — Smith opens a PR adding the smith-* skills, gotchas as the PR description. Ships itself. Status: not built.
  • The Operator — load the forged skills + lessons, then talk to the repo: "how do I deploy?" answered by the repo's resident expert. The endgame. Status: lore exists (the forge persona); the conversational layer not built. (needs your confirm — see below)

One thing to disambiguate

The Operator has two readings: (a) the forge/dispatcher persona we already use for writing skills, or (b) a new conversational layer — load the skills and chat with the repo. I read it as (b), the endgame "talk to the repo." Confirm?

Suggested build order

Core first (they compound the wow): Memory Matrix → Reality Check → Oracle Loop. Surface in parallel (cheap, visual): finish The Construct now. Advanced last (Swarm, Duel, Skill Drop, Operator) once Core proves out.