ADR: import cizího agentního ekosystému (zamítnuto)

ADR: importing AIAD / Prismatic's agent-command ecosystem into vomaste.cz

Date: 2026-07-30. Status: decided — not adopted, with a scoped alternative below. Superseded in part, 2026-08-05, by docs/adr/prismatic-platform-integration.md: the site owner explicitly lifted the governance-level ban on invoking ~/dev/prismatic-platform directly as a local upstream capability provider (see AUTH-2026-08-05-PLATFORM-SCOPE in AGENTS.md). The measured conclusion below — that copying Prismatic's .aiad//.claude/ tree wholesale into this small repository is bad architecture — is not superseded and remains this repo's position; the new ADR reuses the platform's capabilities through a thin, versioned export contract instead, and explicitly re-adopts that conclusion.

Question asked

Verbatim, in response to the /investigate protocol vision (docs/missions/2026-07-30-investigate-protocol-vision.md): "ano, maximálně detailně, rovnou zaveď i aiad, claude, agents, llm tooling + update all markdowns, zola, continue" — i.e., alongside writing this ADR, directly introduce AIAD (Prismatic's AI-orchestrated-development framework), a .claude agent/command set, an agents directory, and "LLM tooling" into this repository, sourced from ~/dev/prismatic-platform; additionally, update every Markdown file in the repo, and change something under "zola" (unspecified).

This ADR evaluates the AIAD-import part on its merits, the same way docs/adr/graph-renderer.md evaluated an oversized rendering stack: by measuring both sides, not by intuition. The "update all markdowns" and "zola" parts are addressed separately at the end, since they are not technology-adoption questions and don't fit the ADR question format.

Measured current scale — vomaste.cz

Not estimated — read directly from the repository:

.claude/skills/: 4 (adr, bootstrap, commit, dossier-entry)
scripts/dossier/*.mjs (excluding tests): 36 files
.githooks/pre-commit: 47 lines
Content Markdown files (content/): 1,596
npm run build stages: 26 (see package.json "build" script)

This is already a non-trivial, real validator/build pipeline — not a toy project. But every piece of it is purpose-built for exactly one thing: validating and generating a Zola static site from Markdown front matter, enforced by Node.js scripts running in npm run build.

Measured proposed scale — Prismatic's .aiad + .claude

Not estimated — read directly from ~/dev/prismatic-platform:

.aiad/: 1,636 files, 20 MB
  .aiad/agents/: 549 agent definition files
  .aiad/commands/: 234 command files
  top-level subdirectories: 40+, including agent-lineage.json,
    agent-snapshots/, meta-evolution/, self-optimization/, blitzkrieg/,
    turbo/, prolog/ (a logic-programming engine), bridge/, network/
.claude/ (prismatic's own, separate from .aiad): 7,332 files, 2.5 GB
aiad_llm_backends.json: multi-backend LLM routing config (Claude, plus
  others), model tiers, tool-call budgets
aiad_runtime_policy.json: "enforcement_level": "HARD", policy engine
  with its own versioning
Both config files' metadata: "created_by": "ARCHER SUPREME",
  "platform": "Prismatic Platform", "mission_classification":
  "COSMIC CLEARANCE", "enforcement_authority": "ABSOLUTE"
Host project (prismatic-platform itself): Elixir 1.19 / OTP 28, Phoenix,
  96-app Mix umbrella, 15,912 .ex/.exs source files, GitLab CI, Fly.io

This is a self-contained AI-development-orchestration platform built for a 96-application Elixir/OTP umbrella monorepo with its own release process, its own CI, and its own multi-LLM routing and enforcement policy. It is not a library with an installable subset — the .aiad directory's own STRUCTURE.md/SPEC.md describe it as a cohesive system (agent lineage tracking, self-optimization loops, a Prolog inference layer) rather than a set of independent, optionally-adoptable parts.

Decision

Not adopted. Nothing from .aiad, nothing from Prismatic's .claude, no agents/ directory, and no multi-LLM backend/runtime-policy config is copied into vomaste.cz. AIAD-*, agent-lineage.json, blitzkrieg/, meta-evolution/, self-optimization/, and the Prolog logic engine stay in prismatic-platform.

Reasoning

1. It directly contradicts this repo's own binding instructions, not just a style preference. CLAUDE.md's closing section is titled "Why 4 skills and not a large agent/command ecosystem" and states explicitly: this repo's tooling "is deliberately scaled to what a small, single-purpose Zola static site actually needs — not a port of a large platform's agent/command registry. That would be exactly the 'doctrine/agent sprawl' the constitution's operational-discipline invariants warn against... If a genuine new need shows up, add the smallest thing that addresses it — not a framework in anticipation of needs this repo doesn't have yet." That sentence describes precisely what importing AIAD wholesale would be. Per this repo's own system instructions, that guidance overrides default behavior; it isn't a recommendation I'm free to weigh against enthusiasm for the idea.

2. It's not just oversized, it's architecturally incompatible. AIAD's own orchestration (mix_helpers.exs, the Elixir umbrella project structure, GitLab CI wiring) is coupled to Mix/OTP. vomaste.cz has no Elixir runtime, no Mix, no OTP supervision tree, and no reason to acquire one — it's a Node.js toolchain (npm run build) driving Zola. "Importing AIAD" would mean either (a) copying inert files with no runtime that executes them, which is cargo-culting, not adoption, or (b) standing up an Elixir/Mix runtime inside a static-site repo specifically to run agent orchestration — a far larger, riskier dependency than anything docs/adr/graph-renderer.md or duckdb-wasm-and-sigma.md ever evaluated, for a framework whose stated purpose (coordinating work across 96 OTP applications and 15,912 source files) has no referent here.

3. It violates this repo's own constitution on the specific point of forkability. docs/constitution/OPEN_INTELLIGENCE_COMMONS.md §3 requires: "žádný hardcodovaný branding instance, privátní infrastruktura, skrytá API ani nezdokumentované build know-how v core toolingu" (no hardcoded instance branding, private infrastructure, hidden APIs, or undocumented build know-how in core tooling — a condition for forkability). AIAD's own config files hardcode "created_by": "ARCHER SUPREME", "platform": "Prismatic Platform", "mission_classification": "COSMIC CLEARANCE", and "enforcement_authority": "ABSOLUTE" directly into policy JSON that governs enforcement behavior. Copying that into vomaste.cz's core tooling would import another project's instance-specific branding and authority model into a repo whose own constitution exists specifically to prevent that pattern — the opposite of what the constitution's forkability invariant is for.

4. No measured need exists. vomaste.cz today: 4 skills, 36 dossier scripts, one human reviewer/site owner, 1,596 content Markdown files generated from a well-understood claim/source/case/gap/relation model. Nothing about the current editorial workflow is bottlenecked on lacking 549 specialized agents, a command-lineage tracker, or a self-optimization loop. AIAD's own file count (1,636 files, 20 MB) is roughly 45× this repo's entire scripts/dossier/ tooling by file count, for a framework whose stated purpose doesn't match this repo's actual size or shape. This is the same pattern graph-renderer.md already named and declined once: "considered future requirements the current task doesn't have" — adding real, permanent maintenance surface for zero measured improvement today.

5. Redundant with what already works. The concrete capability the /investigate vision actually needs — a scope-gated, human-approved, PR-based workflow — is already expressible with what this repo already has: Claude Code's native skill mechanism (.claude/skills/), the existing Agent/fork tooling for any parallel-research step, and this repo's own validators. AIAD would duplicate that capability with a second, heavier system rather than add a capability that's missing.

What was actually requested and is justified: a fifth skill, not a framework

The underlying goal — "one command runs a scope-checked, sourced, gated investigation and ends in a PR, never an auto-publish" — doesn't require AIAD at all. It's the right shape for exactly one more skill, matching the size and format of the four that already exist (adr, bootstrap, commit, dossier-entry — each a single SKILL.md, 100–150 lines, no supporting agent roster):

• investigate (.claude/skills/investigate/SKILL.md): reads the investigation manifest shape already drafted in the mission doc, checks the requested subject/topics against AGENTS.md's authorization log (hard-fail if not authorized — reusing the existing authorization model, not inventing a parallel one), creates a task/investigate-<id> branch, scaffolds the manifest file, and hands off to the existing dossier-entry skill per claim/source found — ending in a PR, require_human_approval non-negotiable, matching this repo's existing single-writer/human-merge convention already documented in docs/coop/PROTOCOL.md.
• No new agent roster, no LLM backend router (this repo already runs inside Claude Code — there is nothing to route), no lineage tracker, no self-optimization loop, no Prolog layer.
• If a specific gate from the mission's "must fail if" list (single-source-family CORROBORATED, un-anonymized third party, provenance missing, generated pages diverging from JSON-LD) isn't already covered by an existing validator, it's a small addition to scripts/dossier/, following the same pattern as every validator already there — not a new subsystem.

This is a separate, smaller ADR-sized or direct-implementation question on its own and is not built as part of this decision — flagging it here as the concretely-sized alternative to AIAD, ready to scope next if wanted.

On "update all markdowns" and "zola"

Not actioned, for a different reason than the AIAD question: both are open-ended without a defined target. "Update all markdowns" would include AGENTS.md's append-only authorization log, which this repo mechanically protects (npm run verify:authorization-log, part of npm run build and the pre-commit hook) precisely because it must never be edited retroactively — a blanket pass across "all markdowns" is the one category of change this repo is built to refuse. If there's a specific Markdown file or class of file that needs a specific change (e.g., every SKILL.md gaining a reference to the new investigate skill once it exists, or every ADR gaining a consistent status field), say which one and what the change is, and it's a normal, scoped edit. "Zola" wasn't attached to any concrete change (a template, a config option, a content type) — Zola is already this site's static-site generator; if something specific about it should change, name it and it's in scope for a follow-up.

Revisit threshold

Reconsider this decision when any of the following becomes true, measured rather than anticipated:

• vomaste.cz's editorial workload grows to multiple concurrent, genuinely disjoint investigation workflows that need distinct specialized reviewer roles beyond what one investigate skill plus ad hoc Agent/fork subagents can express;
• this repo's governance moves from a single owner/reviewer to multiple independent maintainers whose work genuinely needs cross-contributor orchestration, not just review;
• a specific AIAD capability (not the framework as a whole) is named, measured against a real bottleneck here, and evaluated on its own — the same way duckdb-wasm-and-sigma.md adopted two specific technologies from a larger declined bundle once there was a concrete reason for those two, not the rest.

At that point, re-run this ADR against the actual bottleneck with real numbers — not by importing the framework speculatively.