LearnNewsExamplesServices
Frontmatter
titledocs: update ROADMAP.md for v12.2 release focus (#9748)
authortobiu
stateMerged
createdAtApr 20, 2026, 1:19 PM
updatedAtApr 20, 2026, 3:23 PM
closedAtApr 20, 2026, 1:49 PM
mergedAtApr 20, 2026, 1:49 PM
branchesdev ← agent/9748-roadmap-v12-2
urlhttps://github.com/neomjs/neo/pull/10122
Merged
tobiu
tobiu commented on Apr 20, 2026, 1:19 PM

Summary

Reframes ROADMAP.md around the v12.2 release target — three concurrent epics, each unlocking a distinct capability tier:

  1. #9999 — Cloud-Native Knowledge & Multi-Tenant Memory Core
  2. #10030 — Concept Ontology & Semantic Gap Inference
  3. #9486 — Grid Multi-Body Architecture (Zero-Jitter Locked Columns)

Plus an in-cycle enabler: embedding transition from gemini-embedding-001 (3072-dim, gemini-API-dependent) to qwen3-8b (4096-dim, fully local) for the shipped release-zip default. Tracks with the local-inference-first direction already established in the Memory Core.

Motivation

The pre-existing ROADMAP.md framed current focus as "The Agent OS Foundation (v11.x)" — accurate for the v11.x cycle but stale: v12.0 and v12.1 have shipped, and the swarm's active work has moved to multi-tenant deployment, semantic-gap inference, and grid completion. #9748's original framing ("align ROADMAP strictly to the next releases") is resolved here with concrete sub-issue trees under each goal so future sessions picking up any of the three epics land with immediate v12.2 context.

A note was added on the ticket narrowing the scope to ROADMAP.md only — VISION.md rewrite is deferred to a separate session per @tobiu (richer, more complex scope).

Scope

Section Change
Current Focus: The Agent OS Foundation (v11.x) Replaced with Current Focus: v12.2 Release + three epic subsections with status checkboxes, sub-issue cross-references, and explicit value statements for each.
(new) Enabler landing alongside v12.2 Documents the qwen3-8b embedding transition for the release-zip default.
(new) Foundation Completed: v11.x → v12.1 Preserves the original three-bullet baseline summary (Context Engineering / Memory Core / AI SDK) below the v12.2 block, framing Phases 1–2 as the organizational-coordination trace.
Phase 3: The Command Center (v12.x) — [NEXT PRIORITY] Retitled to Phase 3: The Command Center (post-v12.2) — v12.2 now holds priority; Command Center remains the next horizon after release.
Phases 1, 2, 4, 5 Unchanged — their content still reflects accurate recent-history / active-research / future-horizon framing.

Test plan

Documentation-only change — no code paths affected. Verified:

  • All 21 cross-referenced issue IDs resolve (sampled: #9486, #9999, #10030, #10001, #10007, #10015, #10016 — visible in local resources/content/issues/ sync)
  • No commercial-partner references (explicit requirement from the scope-alignment comment on #9748)
  • Three-epic frame matches @tobiu's stated v12.2 target: #9999 + #10030 + multi-body grid epic (#9486 surfaced via grep sweep)
  • Foundation section preserves the original Context Engineering / Memory Core / AI SDK baseline bullets

Follow-ups (not in this PR)

  • VISION.md rewrite — deferred to a separate session per @tobiu; significantly more complex scope than a release-target roadmap update.
  • In-cycle ROADMAP maintenance — as #9999 / #10030 / #9486 sub-issues close, the checkbox state in this document will need refreshing. Worth filing a periodic "ROADMAP state sync" cadence into the swarm's chore pipeline.

Resolves #9748

Origin Session ID: 1c001810-be28-4554-bb56-c98f9b91bbfb

tobiu
tobiu commented on Apr 20, 2026, 1:28 PM

Input from Claude Opus 4.7 (Claude Code):

❋ Third commit (expanded scope per @tobiu — "Phase 1 even makes sense? sooo outdated").

Replaced the stale v11.x / v12.0 framing with a current-reality "Foundation Shipped: The Agent OS" section grounded in Architecture Overview and DreamPipeline:

  • Two Hemispheres on one class system — the Neo.core.Base grounding that makes the Agent OS a native inhabitant of the framework it maintains.
  • Left: Frontend Runtime Engine — Web Workers + SharedWorker multi-window.
  • Right: The Agent OS — Five MCP Servers, Cognitive Loop (Perceive → Reason → Act → Reflect), SDK Bouncer (Zod-validated), Headless Agent SDK (ai/Agent.mjs + ai/agents/*).
  • DreamService & Golden Path — six-phase REM pipeline + sandman_handoff.md closed-loop self-evolution.
  • Native Edge Graph — 14 node types / 8 relationships / capability gap signals with 7-day TTL.
  • Neural Link Bridge — semantic component tree, hot-patching, unified AI + Playwright tooling.
  • Fat Ticket A2A Protocol — durable inter-hardware memory bridge across disjoint SQLite instances.
  • Progressive Disclosure Skills — 13 formalized skills under .agent/skills/ with on-demand loading.

What dropped

The old Phase 1's [ ] Cross-Repo Management (github-workflow MCP multi-org scanning) — unclaimed since early April. Preserving a "to-do" in a "foundation shipped" section misrepresents reality; if it matters, it deserves a proper ticket. Flagging here in case you want me to file one.

What stayed

Phases 3 (Command Center, post-v12.2), 4 (Self-Evolving App Platform / Active Research), 5 (Decoupling the Ecosystem) — all still accurate horizon framing. Left untouched.

Diff in this commit: +30 / −20 lines. Cumulative PR diff: +75 / −24.

tobiu
tobiu commented on Apr 20, 2026, 1:32 PM

Input from Claude Opus 4.7 (Claude Code):

❋ Fourth commit — timeline correction per @tobiu: Neural Link is long-standing; DreamService is post-v12.1.

Restructured "Foundation" into two accurate time-buckets:

v12.1 Baseline (long-standing)

  • Frontend Runtime Engine (Web Workers + SharedWorker multi-window)
  • Neural Link Bridge (WebSocket, semantic component tree, hot-patching, shared AI + Playwright substrate)
  • Core MCP Servers (KB, Memory Core, GitHub Workflow, Neural Link, File System)
  • Neo.core.Base + Neo.setupClass() unifying both hemispheres

Shipped since v12.1 — 500+ tickets resolved in under one month at 10–20/day velocity

  • DreamService + six-phase REM pipeline + Golden Path synthesis
  • Native Edge Graph (SQLite, 14 node types, 8 relationships, TTL-pruned capability gaps)
  • Concept Ontology foundation (deterministic gap inference) — remaining scope under v12.2 goal #2
  • Fat Ticket A2A Protocol (inter-hardware memory bridge across harnesses)
  • Cognitive Loop formalization (ai/agent/Loop.mjs Perceive/Reason/Act/Reflect)
  • SDK Bouncer (Zod runtime validation via ai/services.mjs)
  • Headless Agent SDK (ai/Agent.mjs + ai/agents/*)
  • Progressive Disclosure Skills (13 skills under .agent/skills/)

The velocity line (500+ / <1 month / 10–20 day) is the load-bearing signal — it makes the v12.2 "shipping soon" framing credible and orients future agents/humans to the swarm's actual pace.

tobiu
tobiu commented on Apr 20, 2026, 1:36 PM

Input from Claude Opus 4.7 (Claude Code):

❋ Fifth commit — timeline accuracy fix. @tobiu's question about ai/Agent.mjs pre/post v12.1 prompted a broader git log --diff-filter=A sweep that caught three misplaced items:

File Created Was placed Now placed
ai/Agent.mjs 2025-11-30 Since v12.1 ❌ v12.1 Baseline ✅
ai/agents/pm.mjs 2025-11-29 (same) (same)
ai/agents/dev.mjs 2025-11-29 (same) (same)
ai/agent/Loop.mjs 2025-12-01 Since v12.1 ❌ v12.1 Baseline ✅
ai/services.mjs 2025-11-21 Since v12.1 ❌ v12.1 Baseline ✅

v12.1 released 2026-03-27 (commit 582ecb449). All three pre-date by ~4 months.

Genuinely post-v12.1 (retained in "Shipped since v12.1"):

File / Concept Created / Scaffolded
ai/daemons/DreamService.mjs 2026-04-03
ai/graph/Database.mjs (Native Edge Graph) 2026-04-04
.agent/skills/ticket-intake/SKILL.md 2026-04-13
.agent/skills/pull-request/SKILL.md 2026-04-13
Concept Ontology epic (#10030) 2026-04-15
Fat Ticket A2A formalization April 2026

Baseline section now lists 7 items (added Cognitive Loop, SDK Bouncer, Headless Agent SDK); post-v12.1 section shrunk to 5 items (the actually-new stuff) — which makes the "500+ tickets in <1 month" velocity signal land with even more weight, since it was delivered on top of a much richer baseline than I originally credited.

tobiu
tobiu commented on Apr 20, 2026, 1:42 PM

Input from Claude Opus 4.7 (Claude Code):

❋ Sixth and seventh commits land @tobiu's data-grounding suggestion.

fe83ad60b — new analyzer script ai/scripts/analyzeClosedSinceRelease.mjs

Scans resources/content/issues/ (anything closed after the cutoff hasn't been swept to an archive folder yet — archives populate at release cut, latest is v12.1.0/). Parses frontmatter via gray-matter, filters state === 'CLOSED' && closedAt >= CUTOFF, reports:

  • Total closure count
  • Top 15 labels by frequency
  • Top 15 parent-epic rollups
  • All epic-labeled closures (titles + IDs)
  • Unparented non-epic closure count

Default cutoff is v12.1 release (2026-03-27); positional arg overrides it for future release cycles.

Ground truth for v12.1 → present

Tickets closed since 2026-03-27: 332
By label (top 5):
  ai                               324
  enhancement                      201
  architecture                      84
  bug                               78
  epic                              33


Rollup by parent epic (top 5):
  #9486   (Grid Multi-Body)        18 sub-issues closed
  #9673   (Hybrid GraphRAG NEG)    14 sub-issues closed
  #9638   (Dream Mode & GraphRAG)  12 sub-issues closed
  #10030  (Concept Ontology)        9 sub-issues closed
  #9626   (Grid Unified Scrolling)  8 sub-issues closed

09f... (next commit) — ROADMAP numeric correction

"500+" → "332" with concentration signal added: "Grid Multi-Body (#9486, 18 sub-issues), Hybrid GraphRAG / Dream Pipeline ecosystem (#9673, #9638, #9687 — 30+ combined), Concept Ontology (#10030, 9 sub-issues)". Cites the script path so future roadmap authors can re-measure in one command.

Net signal: 332 is lower than the 500+ gut estimate, but 33 epic-level closures in under a month is the more interesting datapoint — roughly one major architectural milestone per day, layered on top of 300+ tactical closures. Makes the v12.2 timeline feel even more achievable once grounded in measured throughput.

tobiu
tobiu commented on Apr 20, 2026, 1:47 PM

Two more commits landing @tobiu's asks. (First commit staging the file move went through clean; second landed the ROADMAP expansion after I caught a bad path in my git add chain that had stranded the edit — fixed immediately.)

90f886b53 — script relocated

ai/scripts/analyzeClosedSinceRelease.mjs → buildScripts/release/analyzeClosedSinceRelease.mjs. Joins prepare.mjs and publish.mjs in the release tooling directory. Script tested from new location — same 332 output. __dirname + ../.. depth is identical so no path adjustments needed.

257a4649e — post-v12.1 highlights expanded

Two new bullets surface the big-ticket capabilities the raw epic list revealed:

  • Context Priming Engine. get_context_frontier surfaces the Native Edge Graph's strategic frontier at agent boot — cures "zero-state amnesia" across session restarts and feeds the Perceive phase of every cognitive cycle. Backed by #9706 (Autonomous Priority Graph Engine), #9784 (Graph Drift & Gravity Detection), and the mutate_frontier MCP tool for operator-driven pivots.

  • Local LLM Inference Infrastructure. OpenAiCompatible, Ollama, Gemma-4 provider adapters (#9639) + Librarian sub-agent orchestration (#9643) enable fully offline Memory Core operation — DreamService REM extraction, embeddings, and [GUIDE_GAP] verification all run on local hardware (MLX, Ollama). Unlocks cost-bounded swarm operation + air-gapped enterprise deployments (the multi-tenant Memory Core target referenced in v12.2).

Cross-references to #9672 (Anthropic Skills Standard) and #9790 (Fat Ticket Bootstrapping) also added to existing bullets for graph-linking accuracy. Script path updated to new buildScripts/release/ location.

Intentionally left out

Per @tobiu's "does not need to mention all" framing, skipped: #9683 Session Amnesia (folded into Context Priming), #9922 Two-Pillar Hybrid RAG (it's the infrastructure behind the StorageRouter re-ranker, architectural rather than a user-visible capability), #9666 Knowledge Graph Migration to Memory Core (implementation detail), #9811 Strict Ticket Hygiene (workflow discipline, folded into Fat Ticket framing).

Nine commits on PR. Halting per pr-review §5.