Files

T

L'électron rare 066acde0cb docs: add feature maps, audits, research, plans lots 19-26

- Feature maps: agentic intelligence, digital factory, yiacad global/UI/UX/backend
- Audits: consolidation, refactor manifest, exhaustive refonte
- Research: OSS AI-native CAD, digital factory stack, mascarade observability
- Plans: lots 19-26 (mesh tri-repo, UI/UX refonte, yiacad global,
  intelligence agentique, mistral agents/studio, hypnoled, EDA AI tools,
  git EDA platform)
- Operator docs: machine registry, alignment, sync status, provider compat
- Mascarade/Kill_LIFE product contracts and OPS bridge
- PCB AI fab integration map, artifact WMS index
- References directory

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-03-24 00:20:32 +01:00

2.9 KiB

Raw Permalink Blame History

Kill_LIFE consolidation audit (2026-03-22)

Executive summary

Kill_LIFE has a credible 2026 control plane now: spec-first governance, cockpit TUIs, runtime/MCP contracts, YiACAD native surfaces, and a buildable web/ Git EDA scaffold. The main gap is no longer missing surface area. It is drift between what the docs say is canonical and what the code actually uses for intelligence, review, Git state, artifacts, and realtime.

Strengths

spec-first governance is real and still central
intelligence_tui and runtime_ai_gateway.sh provide machine-readable continuity
YiACAD now spans KiCad, FreeCAD, backend service-first hooks, and a web slice
web/ installs cleanly and next build passes
core OSS choices are pragmatic: Excalidraw, KiCanvas, Yjs, BullMQ, KiBot, KiAuto

Weaknesses

lot 22 was drifting toward "done" while the web Git EDA backlog remained outside the intelligence memory
docs were underspecifying required contract fields compared with the actual JSON Schemas
the web/ save/review/PR path is still filesystem- and JSON-demo-first, not Git-first
realtime is transport-first only; Excalidraw is not yet bound to Yjs scene state
worker outputs are not yet surfaced as proper product artifacts

Opportunities

use MCP and service-first tools for review hints, parts search, artifact fetch, and CI triggers
expose queue/worker/realtime health inside the intelligence lane instead of keeping web ops isolated
turn the product review surface into the differentiator: KiCad diff, Excalidraw diff, PCB preview, hints
keep AI as an overlay on top of Git, Yjs and workers rather than as a competing source of truth

Risks

too many overlapping "canonical" sources if plan 22, plan 23 and TUI logic diverge again
a false sense of Git-native collaboration while writes still bypass Git semantics
premature AI orchestration before Git, artifacts and realtime read models are closed
multi-tenant/security debt if /api/project-files/* and GraphQL stay too open

Priority matrix

P0

Raccorder la lane intelligence aux sources web/ et maintenir un backlog vivant TODO 22 + TODO 23 + specs/04_tasks.md.
Remplacer le save/review web purement local par un read model Git minimal et explicite.
Fermer la boucle worker -> artifacts -> GraphQL -> UI avec des URLs web-servables et un etat CI visible.

P1

Binder Excalidraw a Yjs sans perdre la discipline snapshot Git.
Faire remonter le statut queue/worker/realtime dans la memoire intelligence et preparer le pont vers runtime_ai_gateway.sh.
Ajouter une premiere surface review hints / ops summary en lecture seule.

P2

Formaliser le boundary MCP/service-first pour parts, artifacts, CI et review assist.
Revenir sur LangGraph / Agents SDK seulement apres fermeture des P0/P1.
Ouvrir la couche multi-tenant/auth quand le read model Git/CI est reel.

2.9 KiB Raw Permalink Blame History