AgentX

Appearance

6. Shared wiki — compounding team knowledge ​

Status: full Karpathy/Farzapedia pattern shipped — ingest + absorb + query + prune + migrate all live

The wiki pipeline today follows the pattern Karpathy and Farza actually described: typed articles (person / project / place / concept / event / decision / pattern), [[wikilinks]] as the primary navigation surface, an _index.md catalog grouped by type, and an agentic query that walks the subgraph instead of BM25-matching keywords. For the full why, see An honest review of our Karpathy-inspired wiki.

Day in the life — from MR comment to cited answer ​

Concrete, end-to-end. Your PM agent watches a GitLab project. A merge request lands:

MR !179 — feat: product items table dropdowns (closes #642) · author: coding-mtgl-v2 · branch 642-rr-product-items-table-dropdownsGitLab webhook fires into AgentX.

1. Ingest (continuous). The router writes a raw entry per webhook event and per reply. No LLM, no decision — just durable capture into .agentx/wiki/raw/entries/. After a day of MR activity you have ~13 raw entries for this MR: open, push, comment, CI fail, reply, merge.

2. Absorb (cron, nightly). agentx wiki absorb --mode graph batches each agent's unabsorbed entries into one Claude call that emits typed articles with wikilinks. The 13 raw entries collapse into a single article:

markdown
---
title: "MR !179 — feat: product items table dropdowns (closes #642)"
type: project
related: ["Issue #642 — RR Product Items Table Rework (Dropdowns)",
          "MTGL System V2 — Project Overview",
          "MTGL V2 Coding Agent — Implementation Agent for MTGL System V2",
          "DevOps MTGL — Deployment Agent for MTGL System V2",
          "Staging Server — MTGL System V2"]
tags: ["mtgl", "merge-request", "product-items"]
sources: ["pm-mtgl-mnn3f7hj", "pm-mtgl-mnn3j2cc", …]   # traceability to raw entries
---

MR !179 implements [[Issue #642 — RR Product Items Table Rework]] on
[[MTGL System V2 — Project Overview]]. Branch: `642-rr-product-items-table-dropdowns`.
Author: [[MTGL V2 Coding Agent]].

## Scope Drift
#642 specified "UI only." This MR added new DB tables and FKs
(`categories`, `sub_services`, `category_id`). Deployment now requires
running migrations and the `CategoriesAndSubServicesSeeder`.

Notice: the wikilinks form a graph. MR !179 → Issue #642 → MTGL System V2 → Staging Server → DevOps MTGL is now a walkable path.

3. Query (when value lands). Three days later someone in the Telegram project channel types:

"why did we ship new DB tables with #642? wasn't that a UI-only ticket?"

The PM agent picks this up and calls agentx wiki query "scope drift on issue 642". Under the hood:

  • _index.md lists 47 articles by type. Selector LLM picks project: MR !179 and project: Issue #642 as candidates.
  • The walker follows related wikilinks 2 hops. Subgraph ends up with 5 articles including the event 2026-04-08 CI Pipeline PHP Version Mismatch and the decision on backfill strategy.
  • Synthesiser replies, citing titles:

"Scope expanded from UI-only to schema changes because of the category foreign-key requirement — see [[MR !179 — feat: product items table dropdowns]] > Scope Drift. DevOps confirmed the seeder must run on staging first; see [[2026-04-08 CI Pipeline PHP Version Mismatch]]."

That's the payoff. Not BM25 over transcripts — a cited answer assembled from typed articles the agent itself compiled. The same pattern works for onboarding ("what is MTGL?"), incident review ("who decided X?"), and cross-project ("has another team hit this?").

4. Mesh sync. Clawd-server runs its own absorb pass; macbook runs its own. Agent-wide wikis stay peer-local, but cross-references are published over the mesh — agentx graph pull and wiki sync keep the two sides in sympathy without either becoming the source of truth (see Journey 8).

The command surface ​

Grouped by what they're for:

Write (grow the wiki) ​

VerbWhen
ingestContinuous — every conversation writes a raw entry. Cheap, source of truth.
absorbNightly cron — compiles recent raw entries into typed articles with [[wikilinks]], updates _index.md.
interviewInteractive — operator answers a scoped Q&A, an LLM synthesizes one typed article. Captures tacit knowledge that never hit a channel.
quizReverse interview — operator asks, the wiki answers, operator grades (/ok /correct /add /link) and the cited article gets patched.
patchOne-shot LLM-edit from a plain-English instruction. For when you already know what's wrong.
editStraight $EDITOR on the article file. No LLM. For typos and trivial fixes.

Read (use the wiki) ​

VerbWhen
queryAgent-invoked. Walks the catalog + wikilink graph, synthesizes a cited answer. Primary retrieval.
searchRaw BM25 fallback for cases where the agentic path is too slow.
serveLocal Wikipedia-style browser (local + mesh peers).
status, lint, entriesHealth and inventory checks.

Housekeeping (one-shot) ​

VerbWhen
migrateBackfill type + related on legacy articles (post-upgrade).
pruneCollapse legacy per-mode dirs (flat/, unified/) into canonical graph/.
ab-testSide-by-side comparison of old BM25 preload vs new agentic query over real task-history messages.
share, syncCross-agent and cross-mesh sharing.

The agentic query path (the "read side") ​

This is the part the earlier design missed. When an agent needs institutional knowledge:

  1. Read _index.md (catalog grouped by type with wikilink previews).
  2. LLM selector picks candidate articles by type + title.
  3. Walk related wikilinks 2–3 hops from candidates, cap at ~8 articles.
  4. Synthesize the answer from the walked subgraph, cite by title.

This is what makes the wiki worth compiling into. Without it, you're running shallow BM25 over files the LLM wrote — the pathology the blog post analyzes.

Permissions ​

public / shared / private per article, enforced on read and write. Default public; drop to private for sensitive credentials or agent-specific learnings; use shared with specific agent IDs when the article only matters to a subset.

References ​

Released under the MIT License.

© 2025–present Anis Marrouchi