Agentic Workflow Roadmap

🔬 Research

Agentic Workflow Roadmap

A portable, progressive-disclosure scaffold and agent-memory system for filesystem-based AI agent workflows. Three tiers: kernel (universal, forkable by anyone), stack (my opinionated toolkit, forkable for similar setups), work (my personal case study). Long-form rationale lives in 01-kernel/PHILOSOPHY.md + 01-kernel/principles/.

Done — First-Iteration Init (2026-04-17)

Structural foundation

Three-tier decomposition — 01-kernel/, 02-stack/, 03-work/, plus 00-meta/ for scaffold-dev tooling
Migration from flat .claude/ + scattered docs/research/ into tier structure (non-destructive copies + deletions of merged skills)
Stratum-tagging convention (1–5) applied to every markdown file via automated injector — 88 files updated
Frontmatter hygiene pass — doubled-frontmatter merge (26 files), CRLF→LF (21), colons-in-values quoted (17)

Philosophy & principles content

01-kernel/PHILOSOPHY.md — 10 invariants as claims about the world (~1.6k words)
10 principle pages with citations — capture→work→output, temperature gradient, skills/agents, progressive disclosure, convention-as-compressed-decision, single canonical addressability, five strata, four channels of context, meta/self-reference, multi-entity design (~9.5k words)
Ultraplan research attribution preserved at 03-work/memory/research/2026-04-17-kernel-and-dressing.md

Stack content

Stack overview with layer diagram + dependency graph
7 layer pages — AI coding, terminal, cross-device, knowledge mgmt, homelab, dev-infra, editor ext
Decision matrix — Claude Code vs Gemini vs Codex, Zellij vs tmux, Tailscale vs alternatives, Obsidian vs alternatives
Patterns — cross-device SSH full walkthrough, image-paste pipeline (Zipline + ShareX + sharex-clip2path)
Profiles migrated — bashrc snippets, keybindings

Work-tier content (mostly personal)

03-work/memory/tool-picks.md — current defaults across all layers
03-work/memory/preferences.md — coding style, review criteria, anti-preferences, vibe-coding philosophy
03-work/memory/project-references.md — 30+ projects grouped by type with “what pattern lives here”
03-work/project-types/astro-starlight-docs.md — canonical pattern for how I build docs sites (with zz- prefix convention)
03-work/rebuild/ — 11-page rebuild-my-machine flow (index + 9 steps + verify + troubleshooting)
03-work/homelab/ — PII-aware placeholder (actual specifics stay gated)

Kernel components (migrated + merged)

13 meta-skills in 01-kernel/skills/ (merged workflow-guide + workflow-meta → workflow-scaffold; kept obsidian-family and proactive-patterns separate after honest review)
10 meta-agents in 01-kernel/agents/ — seacow-scaffolder, skill-writer, agent-writer, workflow-expert, workflow-improver, claude-md-updater, workspace-advisor, improvement-logger, test-improver, knowledge-curator
7 slash commands in 01-kernel/commands/
01-kernel/ARCHITECTURE.md
Hook scripts in 01-kernel/scripts/hooks/
Skeleton templates in 01-kernel/templates/skeleton/ (5 project types migrated)

Site (Astro Starlight)

Meta / scaffold-dev tooling

00-meta/stratum-audit/classify.mjs — 226-file classifier with path + keyword heuristics
00-meta/stratum-audit/add-stratum-frontmatter.mjs — retrofit stratum on source files
00-meta/stratum-audit/pii-scan.mjs — PII gate before any public-visibility change (0 critical after audit)
00-meta/skills-audit.md — rationale for 14→13 skill count (aggressive merge rejected for 2 honest reasons)

Identity

README.md rewrite — vision-first, tier overview, quick tour by audience
CONTRIBUTING.md — selective-acceptance policy, no-AI-attribution rule, style guide
ROADMAP.md (this file)

Status as of init complete: 70 HTML pages build, 39/39 smoke checks pass, 0 PII-critical findings.

Phase dependency map

Phases are goal-oriented, not time-boxed. Some block others; some are parallelizable.

flowchart TD
    P8["Phase 8<br/>Extract kernel + stack repos"]
    P9["Phase 9<br/>Parallelization research"]
    P10["Phase 10<br/>Backend → cyberbaser parity"]
    P11["Phase 11<br/>Content distillation + legacy cleanup"]
    P12["Phase 12<br/>Symlinks + global install"]
    P13["Phase 13<br/>Site polish"]
    P14["Phase 14<br/>Public release"]

    P8 --> P10
    P8 --> P12
    P11 --> P13
    P10 --> P13
    P12 --> P13
    P13 --> P14
    P8 --> P14

    P9 -.->|"independent;<br/>feeds 02-stack content"| P11

    classDef blocking fill:#fff4e6,stroke:#d97706,stroke-width:2px
    classDef gated fill:#fee2e2,stroke:#b91c1c,stroke-width:2px
    class P8 blocking
    class P14 gated

Reading the map: Phase 8 (extraction) is the structural pivot — it gates the multi-repo backend work (P10) and global install (P12). Phase 14 (public release) is the only phase that requires all upstream work plus a clean PII audit. Phase 9 is a parallel research track whose output lands in 02-stack/patterns/.

Phase 8 — Extract kernel + stack to their own repos

Why: kernel is generically useful; stack is fork-worthy for similar setups; keeping them in a monorepo forever conflates audiences and creates maintenance friction.

Phase 9 — Parallelization research + content

Why: user flagged git worktrees + parallel agent patterns as under-researched. Deserves the Ultraplan treatment; output feeds directly into 02-stack/patterns/parallel-agents-worktrees.md.

Deep research session on parallel AI-agent workflows
- Git worktrees — how to use, gotchas, integration with Claude Code’s per-directory session model
- OpenCode oh-my-opencode background agents — whether to adopt, ban-risk status in 2026
- Multi-agent coordination patterns (beyond what inter-agent-messaging skill documents)
- Watching: Conductor extension, other parallel-agent tooling
Pattern doc: 02-stack/patterns/parallel-agents-worktrees.md — the canonical how-to
Update 02-stack/decisions/index.md with parallel-work decision rows
Update 03-work/memory/preferences.md with my parallel-work flow

Phase 10 — Backend alignment to cyberbaser

Why: as the scaffolded sites mature, the “full-featured Starlight” backend stack (Nova theme + site-graph + blog/changelog + announcements + tags + image-zoom) is worth adopting across all three agentic-workflow sites. MVP uses Flexoki for shipping speed; Nova is the target.

Install + configure Nova theme on agentic-workflow site (swap from Flexoki)
Add starlight-announcement for phase status banners
Add starlight-blog as changelog prefix
Add starlight-tags with proper tags.yml config
Align PageTitle.astro with cyberbaser’s conventions (custom badges + dates)
Once extracted, apply same backend to agentic-kernel + agentic-stack sites
Migration is pure config + plugin swap; content (markdown + frontmatter) unchanged

Phase 11 — Content distillation + legacy cleanup

Why: original docs/, research/, examples/, knowledge-base/, tutorials/, System/, templates/ still live at repo root. Sync script no longer pulls from them. They’re raw source material for distillation, not content.

Audit each legacy folder’s markdown — is content methodology (stratum 1–4) or knowledge work (stratum 5 / exits)?
Distill methodology bits into appropriate tier pages (principle pages, patterns, stack layer content, work memory)
Exit actual knowledge work to user’s external Obsidian vault
Delete legacy folders once content is extracted
Run full smoke pass after each folder’s cleanup

Phase 12 — Symlinks, global install, kernel consumption

Why: Claude Code needs to see skills/agents at ~/.claude/skills/ and ~/.claude/agents/. Currently, copies exist in project-local .claude/. The install.sh script should symlink (or copy) kernel content into the global location.

kernel/scripts/install.sh finalized — symlinks ~/.claude/skills/ → $KERNEL_REPO/skills/, same for agents, commands, scripts
Validate Windows fallback (Windows git doesn’t default to core.symlinks=true)
install.sh prints clear “what was installed” summary
Tested on a fresh VM via the rebuild flow (03-work/rebuild/05-install-kernel.md)
In-repo .claude/ at root becomes either a symlink or a documented cache of kernel content

Phase 13 — Site polish

Brand CSS — distinct palette from cyberbaser (not Nova default; project-specific)
Custom SVG hero for homepage (optional)
Site graph tuned — principles section should visibly cluster
Search tuning (Pagefind) — boost principle pages
Mobile navigation pass — real phone testing
Accessibility audit — axe or similar
Edit-link config verified against actual repo path
starlight-announcement configured for current project phase
Status badges on all pages — cross-check stratum: frontmatter is correct

Phase 14 — Public release

Final PII audit clean (0 critical, 0 high-findings-without-tolerance)
LICENSE finalized for each repo
CONTRIBUTING.md tuned per repo (kernel = open-source-friendly; stack = selective; workflow = mostly personal)
README final pass per repo
Switch GitHub repos to public — one at a time, starting with agentic-kernel
Enable GitHub Pages on each
Cross-linking on all sites verified
Announcement (blog post, README banner) explaining the three-repo structure

Watching / deferred

OpenCode ecosystem

If sst/opencode session bugs are fixed (issues #4378, #3551, #4557), revisit as a Claude Code alternative.
oh-my-opencode native background-agent support makes it attractive for parallel work if reliability improves.

AI coding CLI wrappers

OpenCode, OpenClaw, T3Code currently avoided for ban-risk + security reasons. Revisit 2027 if an ecosystem entrant proves 2+ years clean operation with provider cooperation.

Knowledge graph integration

Neo4j + Ollama + MegaMem — Obsidian → Neo4j bridge for semantic search. Pre-beta; watching.
Openclast (my project) — browser-based Obsidian with CRDT sync. Becomes an alternative for multi-device vault editing when it matures.

Enterprise scale

If the scaffold ever needs to serve a team, identity layer becomes load-bearing. Authelia or Keycloak ACLs for per-user access to different tiers.
Enterprise MCP gateways (microsoft/mcp-gateway) — aware but not adopted for personal use.

Tooling experiments

Tests beyond smoke: lightweight Playwright suite for interaction patterns (lightbox, search, navigation) if pain points emerge.
Auto-update-attention hook — cron-run script that reads mtime and updates status: hot | warm | cold | frozen in frontmatter automatically.

Ideas backlog

Visual workflow designer for agentic pipelines
Agent performance metrics (which agent + which skill combo gets things done)
Skill effectiveness tracking — auto-detect skills loaded but never referenced in output
Cross-project skill learning — pattern mining across projects
Workflow recording + replay for reproducible agent runs
Version control for agent/skill evolution (semver per skill?)
sharex-clip2path for Mac/Linux equivalents
Obsidian plugin that reads this scaffold’s principles + patterns as in-vault reference

Design principles for this roadmap

Done section is dense + specific. Checkboxes + short descriptions. No weasel words about “completed initial setup” — list what.
Phases are goal-oriented, not time-boxed. “Extract kernel” is a phase; “Q2 2026 work” is not.
Every phase answers: what, why, what’s the exit criterion.
Deferred != dead. Watching + ideas backlog stay as explicit parking.
Updated every significant merge. The README + site link here; the Done section should grow.

Last updated: 2026-04-17