Wiki Fabric

Git-native, testable knowledge governance for AI coding agents. Project memory that is scoped by precedence, traceable to evidence, enforced by CI — and provably delivered before an agent writes code.

Positioning in one line: Wiki Fabric treats repository knowledge as operational infrastructure for agents — a governed context layer, not a note vault. Where generic "LLM wiki" projects stop at self-maintaining Markdown, Wiki Fabric adds the three things that make knowledge trustworthy at task time: scope with precedence, provenance with staleness gates, and behavior evaluations that measure whether the knowledge changed the agent's decision. It is one substrate of agent memory — the durable, Git-governed one — not the whole capability: the agent's context window and working state remain the harness's to manage. And it is local-first, cloud-optional: the daily loop runs on-device with zero tokens; LLM stages route per-repo, per-stage (privacy tiering) when you want them.

Quick start

# See the value in 5 seconds (no install):
bash scripts/demo.sh

# Install (packaged, atomic):
uv tool install wiki-fabric            # the wf CLI
uv tool install wiki-fabric --with mcp # + the wf-mcp MCP server

# Teammate: pulls the team corpus automatically when the remote carries one
# (--vault <path> pins where the vault shell lives):
curl -fsSL https://raw.githubusercontent.com/hybridindie/wiki-fabric/main/scripts/wiki-fabric.sh | bash -s -- --corpus git@github.com:your-org/wiki-fabric-corpus.git

wf status                        # fabric health
wf bootstrap /path/to/my-project # connect a project (--extract local for privacy)
wf capture my-project            # pull docs from upstream repos → evidence/raw/
wf ingest evidence/raw/my-project/docs/readme.md --extract-claims
wf query "Why does my code batch writes?"
wf context --task "Add token rotation to the OAuth service"
wf log --project my-project --problem "..." --intervention "..." --outcomes "..."

# The agent harness POV (always-on block arrives at session start):
wf context --task "..." --write-receipt   # step 1: scoped manifest + receipt (0 tokens)
wf query "..."                   # step 2: evidence with line locators (0 tokens)
#  step 3: the agent writes code
wf log --project my-project --receipt <receipt-id> ...   # step 4: outcome ↔ delivery
wf gate                          # step 5: surface anything awaiting a human

# Make it a team system (gh CLI creates + publishes the corpus repo):
wf sync setup
# Upstream moves while machines sleep → the scheduled freshness cycle:
wf freshness --dry-run           # capture-git + mechanical re-verify (0 tokens)

The daily loop — connect → capture → ingest → validate → context → work → mine/review — compounding at step 7 into step 5's context for the next project. Full walkthrough, flags, and requirements: Getting Started.

What it is

Two trees: the harness (tooling, this repo) and the fabric (your knowledge — gitignored, lives in the corpus/vault; teammates pull it at install — see Teams). The harness ships: the knowledge format (claims with locators, patterns, decisions, commitments), the deterministic context compiler (wf context — every inclusion/exclusion carries a reason, persisted receipts), contract enforcement (lint + CI gate), behavior evaluations, and the self-building domain vocabulary. Optional integrations: graphify (call-graph staleness + code navigation), obsidian (two-way vault: human wiki edits flow back as evidence), judgment tier (local decision models — ollama serves the System One wire, Tev/Nimble — for eval gates, retrieval fusion rerank, and review triage), embeddings (semantic re-rank, off by default), team sync (solo direct-push or PR-gated distribution). The fabric also serves its core surface as MCP tools (wf-mcp) for any MCP client. See Architecture for the pipeline and Core Workflows for the loop with scenarios.

Documentation

Full docs at hybridindie.github.io/wiki-fabric — with search.

Status

• Works today: capture → ingest → query → promote loop (claims carry provenance edges from chats/PRs); activity-bounded git capture (window shrinks to fit an active repo's budget — shrink-not-truncate; --until backfills; threads paginate with flagged truncation + state provenance; tuning.git_history + per-repo knobs); budgeted ingest + the mechanical sha256-drift ⇒ contested staleness trigger; source staleness tiers + --verify-sources lifecycle (upstream-gone ⇒ expired tombstone); deterministic context compiler + delivery receipts; locator re-verification; chat-mined gated pattern candidates; physical domain homes (domains/<domain>/{concepts,questions,syntheses}, ontology alias-mapped); ontology as living vocabulary (shared parser, VOCABULARY/IDENTITY lint gates, canonical repo slugs — one project, one slug, everywhere incl. the corpus trees); layout single-truth + guard (commit-time in the hook); commit-drift ritual (plane-classified hook-drift commits); promotion-queue + questions scaffolds; contribution-freshness gate signal; judgment tier (refinement + effect verification + opt-in re-rank); team sync (PR mode, corpus pull); hooks incl. merge-time capture; Obsidian bridge; MCP server; CLI surface: shell completions (bash/zsh/fish), wf help <cmd> → argparse, --version, typo suggestions, one-liners for all verbs, --json machine surfaces (status/ingest/review/projects), exit-code discipline (drift 2 / unknown-slug 3 — loud for hooks/CI), wf projects inventory; OKF v0.2 export/import; wiki generation + publish + utility write-back. judgment gate (G-J) (only a calibrated judge proposes; judge swaps refuse until re-calibrated); newest-wins contradiction sweep (contradiction-based staleness — same-source contradicts + newer capture mechanically demotes the older claim, restore-able); tracker capture (wf capture issues — the WHY channel, kind: issue-record thread nodes); chat-mining cadence + judged near-miss refinement (corpus CI weekly; judged merge of paraphrase takeaways into inbox candidates); gate push coverage (escalation aging + the daily digest); judged auto-apply tier (promote-patterns --auto — deterministic preconditions + one calibrated judgment per candidate, confident applies (stamped + reversible via --unapply), near-band escalates, off until opted in); SECRETS lint (credential-shaped strings in corpus content block the commit gate — the Memory Defense lesson, deterministic, matches masked in output); contradiction synthesis (concept definitions render the evolution — prior state folded in, disputed → open_questions); delta-mode wiki refresh (unchanged generation-inputs keep the previous file byte-identical at 0 tokens; changed inputs patch incrementally with mechanically-spliced untouched sections; --full/--check); temporal retrieval axis (wf query --type changed orders the answer earliest→latest by capture date). CI proves all of it (what the pipelines show).
• Next: judgment-informed context re-rank surface polish, embeddings re-ranking (shipped, off by default — value conditions in Integrations), multi-harness skill packs, domain-scope syntheses routing.
• Contributing: see CONTRIBUTING.md. The fabric self-documents — if something isn't clear, that's a bug in the fabric; issues welcome.

License

MIT — see LICENSE.