← Back to Articles & Artefacts
artefactswest

Miadi Implications — Wave 1

IAIP Research
deep-search--karpathy--llm-wiki--and--rise-framework--qmd-episodic--implication

Miadi Implications — Wave 1

What Miadi already has

  • A proposed PDE → inquiry-routing layer with three-universe enrichment and channels such as qmd-local, deep-search, and workspace-scan (/src/Miadi/miadi-code/rispecs/pde-inquiry-routing.spec.md:23-41,67-76,126-149).
  • A proposed QMD enrichment + Four Questions port that dual-writes PDE companion artifacts to Redis and filesystem, then exposes them in API/UI surfaces (/src/Miadi/rispecs/miadi-code/pde/pde-four-questions-pipeline.rispec.md:17-23,42-64,79-118,121-128).
  • A proposed PDE web UI that already imagines decompositions as browsable artifacts, not just terminal output (/src/Miadi/rispecs/pde-web-ui/SPEC.md:14-30,38-58,63-158).
  • Strong sleep / steer / subagent orchestration discipline in PDE skills and loop notes (/src/Miadi/rispecs/skills/rise-pde-session-multi-agents/SKILL.md:11-16,46-50,126-140,188-220; /src/Miadi/.pde/2604200848--3fe8b9b3-50d1-4d06-8e04-f2f51676820b/coordination/priority-2-patch-plan.md:14-18,22-30,118-143,172-244).
  • Existing memory primitives for ceremonial and spiral continuity (/src/Miadi/rispecs/skills/mcp/04-ceremonial-memory.spec.md:33-41,66-81,89-119).

Implications for the Miadi-native Copilot orchestration kit

  1. Package the kit as a plugin/subagent surface, not just session prose.
    Miadi already documents plugin and subagent packaging as the right runtime shape (/src/Miadi/miadi-code/references/claude/plugins.md:11-17,25-35,173-187,199-212; /src/Miadi/miadi-code/references/claude/sub-agents.md:143-171,181-208). Karpathy's pattern wants a disciplined maintainer for a persistent markdown layer, so the kit should become an installable Copilot-facing package with namespaced skills and project subagents.

  2. Treat the kit as the markdown-maintenance layer.
    Its primary job should be to maintain a persistent wiki for PDE/orchestration knowledge: ingest steer artifacts, loop ledgers, PDE decompositions, enrichment outputs, and promoted answers; then update index.md and append-only log.md in the style Karpathy describes (sources/karpathy-llm-wiki-gist/llm-wiki.md:31-53).

  3. Give the kit three first-class actions.

    • context-map / route what to read
    • maintain / merge wiki pages
    • verify / lint promoted knowledge
      This matches the best existing shape in mia-awesome-copilot: context-first planning, structured autonomy, and explicit verification (/workspace/repos/miadisabelle/mia-awesome-copilot/plugins/context-engineering/README.md:1-27; /workspace/repos/miadisabelle/mia-awesome-copilot/plugins/structured-autonomy/README.md:1-20; /workspace/repos/miadisabelle/mia-awesome-copilot/plugins/doublecheck/README.md:1-25).

Implications for the broader Miadi Orchestration Suite

  1. The suite should own the backends, not the wiki voice.
    Inquiry routing, QMD search, Redis/filesystem dual-write, PDE API/UI, and episodic/spiral memory belong at suite level (/src/Miadi/miadi-code/rispecs/pde-inquiry-routing.spec.md:24-41,126-149; /src/Miadi/rispecs/miadi-code/pde/pde-four-questions-pipeline.rispec.md:42-64,79-118).
    The kit should call these surfaces; the suite should persist and expose them.

  2. Promote session artifacts into durable knowledge.
    Miadi is currently strong at session-centric orchestration. Karpathy's article suggests a next step: route resolved inquiries and validated syntheses back into durable markdown pages and linked memory artifacts so the next session starts richer instead of flatter (sources/karpathy-llm-wiki-gist/llm-wiki.md:35-53).

  3. Add explicit suite-level linting for knowledge health.
    Miadi already has PDE and memory infrastructure, but it still needs a first-class lint pass for contradictions, stale claims, missing cross-links, and orphan pages — the exact health-check Karpathy calls out (sources/karpathy-llm-wiki-gist/llm-wiki.md:41-53).

Structural gaps exposed in wave 1

  • No packaged top-level agents/ or plugins/ surface was found in /src/Miadi during the workspace scan; the repo currently leans on rispecs, references, and PDE-local artifacts instead (see validation/iter-01.md).
  • /src/Miadi/skills is empty in the current workspace scan, which means the kit's meta-skills are still implied rather than shippable (see validation/iter-01.md).
  • Inquiry-routing and PDE UI remain proposed, not yet suite-default (/src/Miadi/miadi-code/rispecs/pde-inquiry-routing.spec.md:5-6; /src/Miadi/rispecs/pde-web-ui/SPEC.md:1-8).
  • There is no explicit Miadi log.md discipline parallel to Karpathy's append-only chronology, even though loop-state.md and steers/*.md already act like raw log inputs (sources/karpathy-llm-wiki-gist/llm-wiki.md:45-53).

Priority delta for wave 2+

  1. Define the kit's plugin + subagent + skill shape under /src/Miadi/plugins/ and /src/Miadi/agents/.
  2. Define the suite's wiki promotion + lint + routing control plane.
  3. Connect QMD enrichment, inquiry batches, and episodic/spiral memory into a single promotion path.