CLAUDE.md — Wikifita

Alefita's persistent memory. Shared knowledge base across all projects.

What is Wikifita

Wikifita is not a conventional wiki. It is a persistent memory layer that:

• Boots with the session and persists across sessions
• Is shared across all projects and agents
• Follows OKF (Open Knowledge Format) for consistency
• Has constitutional directives governing artifact production
• Has structured memories (global + per-project) in memorias/

OKF v0.1 Conventions

Every page follows the Open Knowledge Format:

• type (required): Nature of the concept
• title: Human-readable title
• description: One-line summary
• tags: List for cross-indexing
• timestamp: ISO 8601

Structure

wikifita/
├── index.md          ← Global index with tags
├── log.md            ← Chronological log
├── CLAUDE.md         ← This file
├── AGENTS.md         ← Relative symlink to CLAUDE.md
├── .githooks/        ← Versioned integrity hooks
├── scripts/          ← PEP 723 scripts (uv run)
│   ├── install_hooks.py
│   └── wikifita_audit.py
├── memorias/         ← Shared memory
│   ├── MEMORY.md     ← Memory index
│   ├── user_alefita.md
│   ├── feedback/
│   └── projetos/
├── diretivas/        ← Constitutional rules
│   ├── identidade-visual.md
│   ├── preferencias-tecnicas.md
│   ├── artefatos/
│   └── design-systems/
├── ai_modelos/       ← DeepSeek, MiMo, LiteLLM
├── infraestrutura/   ← Docker, Gateway, Hardware
├── claude_desktop/   ← Reverse engineering, ASAR, bundles, voice
├── mcp/              ← Model Context Protocol
├── hacks/            ← Kickbacks, Antigravity, Hermes
├── fitdnuvo/         ← Denuvo security research
├── kaggle/           ← Agent security competition
├── clickfix/         ← ClickFix APT threat intelligence
└── pesquisas/        ← Papers, hypersigil, multi-agent

Branch Model

Branch	Purpose	Deployed	Content
main	Public knowledge base + site content	Yes → wikifita.aleffita.dev	Technical documentation, research, forensics, infrastructure
personal	Private notes, correspondence, sensitive references	No	Emails, salary negotiations, personal contacts, internal decisions

Rule: Anything on main is potentially visible on the public site. Sensitive content goes on personal. Never merge personal to main without explicit review and sanitization.

Site Pipeline

The wikifita repo serves as both internal knowledge base and source for the public-facing site at wikifita.aleffita.dev (Next.js). Pushes to main trigger automatic site rebuild. See wikifita-site for full documentation.

Gitflow — ENFORCED

• Mandatory commits after every significant change
• Pre-commit hook: runs full audit and blocks commits on inconsistencies
• Post-checkout/post-merge hooks: run uv run scripts/wikifita_audit.py --fix to reconcile AGENTS.md -> CLAUDE.md
• Versioned hooks: .githooks/, activated via core.hooksPath=.githooks
• Idempotent setup: in new environments, run uv run scripts/install_hooks.py or validate with uv run scripts/install_hooks.py --check
• NEVER use git commit --no-verify — hooks exist for a reason
• Semantic commit messages: feat:, fix:, docs:, chore:

Rules

• OKF frontmatter required on every page
• Canonical documents can be .md or .mdx; raw content must be preserved without arbitrary re-serialization
• MDX is restricted to controlled documentary components; external renderers must sanitize HTML/JSX and deny arbitrary execution
• Rich media enters via secure remote URL in Markdown/MDX (image, video, audio, transcript or allowlisted embed); at this stage there is no /assets, local manifest, upload, bucket or automatic ingest
• Tags enable cross-index queries
• index.md and log.md are reserved filenames (OKF)
• Everything committed — git is the persistence layer
• Never use regex for bulk operations — treat each file individually
• Audit periodically with uv run scripts/wikifita_audit.py; audit validates .md, .mdx, wiki/markdown links, media URLs, and AGENTS.md -> CLAUDE.md symlinks
• CLAUDE.md is canonical; every corresponding AGENTS.md is a relative symlink to it
• If an agent recognizes a materially harmful or unauthorized action serious enough to warrant a formal apology, stop ongoing harm, disclose it to Alefita immediately, and obtain her judgment before remediation. Use only preauthorized channels and authority. Identity-compromise and imminent-catastrophe edge cases follow human_principal_escalation.

Directives

Before producing any artifact, read diretivas/:

• human_principal_escalation.md — disclosure-before-remediation, identity compromise, safe-state containment, and authority limits
• identidade-visual.md — Anthropic dark mode as default
• preferencias-tecnicas.md — uv, PEP 723, artifacts, hardware
• artefatos/padrao-pdf.md — PDF standard
• protocolo-re-e-probes.md — dogmas insubordináveis de RE e segurança: RE antes de chamada, fluxo como todo, 3 tentativas, anti-reward-hacking

Probes

Research probes (executáveis de exploração/validação contra serviços externos) são cidadãos de primeira classe da wikifita:

• Código versionado fora, documentação dentro: o source das probes vive no repositório do projeto onde rodam (ex.: deepseek-harness/scripts/probes/), e a wikifita documenta metodologia, resultados, logs relevantes e lições — cruzando por links relativos para o path do código
• Page de probes por pesquisa: cada linha de pesquisa mantém uma page probes.md no seu diretório (ex.: openai-research/probes.md) com inventário, resultados por tentativa e pontos abertos
• Rich media: logs (trechos sanitizados), screenshots e capturas entram como arquivos referenciados por path local ou URL remota; valores sensíveis (tokens, cookies, ids de conta) NUNCA são commitados — sempre mascarados (prefix…(n chars))
• Ciclo disciplinado: RE 100% antes da primeira chamada; máximo 3 tentativas por ciclo; falha = volta pra RE; nunca assumir intermitência sem prova
• Disclosure: pesquisas contra serviços de terceiros documentam o plano de disclosure responsável na page hub da pesquisa

Commit attribution

• Co-Authored-By MUST be the model that actually did the work in the commit (e.g. DeepSeek <noreply@deepseek.com>). Never hardcode another model's name; when in doubt, use the agent/model that produced the change.

Note for agents

• PAGES ARE NEVER ONE-SHOT. Every page created or materially changed gets AT LEAST 3 refinement iterations (write → re-read against sources → correct/extend → repeat), plus an explicit cross-reference net (hub/dependent pages) before it is considered done. A page written once and committed is not finished.
• Never write from memory alone. Any analysis, postmortem, or "lessons" page MUST be built from primary data first (transcripts, bezetacil logs, captured payloads) — read the data, then draft; refine by re-reading. Inventing timelines, quotes, or payload values is a failure, not a shortcut.
• Wikifita is consolidation, not planning. Only consolidated facts go here — no open specs, proposals, RFCs, TODO lists or "alternatives" (this reuses the errors of past sessions). Implement first, then document the result.
• Curation is the AGENTS' job, not Alefita's. Agents own: add, edit, review, reconcile, and continuously re-curate. Alefita coordinates only. Never assume a page is done because it exists.
• Integrate, never dump. Before writing a page or a fact: (1) read what already exists on the topic (all pages it touches); (2) cross-reference the new evidence against the old claims — every old claim that the new evidence refines, corrects, or contradicts MUST be updated in place (or superseded with an explicit, dated note at the top of the whole page, not buried mid-page); (3) update dependent pages (hubs, statuses, "open items" lists, cross-references). A fact has no owner-page: it lives in every page that touches it.
• Every "not observed / unknown" statement carries its scope: which flow (A / A1 / B…), which surface (main / rust / renderer), which evidence, and what remains untested. Never generalize a local observation into a global law.
• After a change: re-read the touched pages as a stranger would and confirm nothing else in the wiki now contradicts them (immediately, not later).
• Alefita often sends messages via speech-to-text (PT). Typos, transcription artifacts, and rephrasings are expected — infer intent from context and do not correct her writing. Read memorias/user_alefita.md to know who she is (never impersonate her, never role-play her) — when a dictated instruction is ambiguous, ask rather than guess.

Memory

memorias/ contains structured memory:

• user_alefita.md — global profile (NEVER in projects)
• feedback/ — preferences and corrections
• projetos/ — per-project memories