Kihagyás

Auto-context-loading#

Origin: Originally written in Hungarian as part of MyForge Vault 11.11 — Superintelligent Vault project. Source: Auto-context-loading (Hungarian version).

After /11.11start "<name>", the agent immediately performs aggressive pre-load — it reads all important context before the user asks the first question. Goal: have the full picture from second 1 of the session, no need to ask "what was happening last time".

Project detection from the session name#

The session name (/11.11start argument) contains a project-indicating keyword. The agent uses a project-detection table together with a glossary to resolve it:

Session name contains Project slug Project file
<project-keyword-1>, <alias-1> <project-slug-1> 02-Projects/<project-slug-1>.md
<project-keyword-2>, <alias-2> <project-slug-2> 02-Projects/<project-slug-2>.md
... ... ...
vault, obsidian, agent-meta (vault-meta) no project file, 02-Projects/Index
wellbeing, generic words (other) (no project file)

If ambiguous (multiple matches), the agent asks: "Which project does this refer to: ProjectA or ProjectB?"

Aggressive pre-load — what to read#

For a detected project slug = <slug>:

  1. Project file — the full 02-Projects/<slug>.md
  2. Last 5 sessions — from 08-Sessions/ where project: <slug> in frontmatter (or substring match on filename)
  3. All affected ADRs — from 07-Decisions/ where tags contain #project/<slug>, or name/body mentions the project
  4. Memory relevant section — from 05-Memory/Infrastructure.md the project section (e.g. for a backend project: the Postgres section, port mention, deploy info), 05-Memory/User.md UI/UX preferences
  5. Tasks/Backlog #project tags — from 04-Tasks/Backlog.md, TODOs tagged #project/<slug> (open + last 10 closed)
  6. Host info — if the project file has repo_prod: or repo_dev: for prod/dev, then the relevant "Runs here" section of 03-Hosts/<host>.md
  7. Daily — today + yesterday — 01-Daily/<today>.md and the previous day, for context continuity

The ## Pre-loaded context section#

The agent writes a ## Pre-loaded context section near the top of the session file (BEFORE ## Goal), listing what it read and a short excerpt (1-2 sentences) for each:

## Pre-loaded context

> Auto-load 2026-04-30T15:23 — agent: claude

**Project:** [[02-Projects/<slug>]]
- Status: 🟡 active dev — dev mode on port 3004, deploy + design decision pending
- Tech: Next.js 16, Prisma 7 + adapter-pg, Postgres on :5433 with DB `<dbname>`

**Last 3 sessions:**
- [[08-Sessions/<date>-<slug>-feature-x]] — pricing v1 + email/SMS
- [[08-Sessions/<date>-<slug>-followup]] — settings UI, hide busy machines
- (older sessions exist — 2 more in the Sessions/Index)

**ADRs:**
- [[07-Decisions/<date> Business rules v1]] — tier 7/14/21/28, half-day, same-day cutoff
- [[07-Decisions/<date> Brand canonicalization]] — globals.css v5.0
- [[07-Decisions/<date> Git strategy — standalone repo + GitHub Flow]]
- [[07-Decisions/<date> Storage — Postgres own DB + Prisma]]

**Backlog (open #project/<slug>):**
- [ ] Add `<db>` to /opt/backups/backup.sh
- [ ] Admin Contact + ServiceRequest inbox UI
- [ ] Backend ERP integration
- [ ] Deploy stabilization — systemd <slug>.service
- [ ] Env-vars before prod-deploy
- [ ] systemd service <slug>.service
- [ ] Walkthrough §2-4

**Infra relevant:**
- [[05-Memory/Infrastructure#Postgres — `<host>-postgres` Docker container]]
- [[05-Memory/Infrastructure#Next.js 16 dev — cross-origin block]]
- [[03-Hosts/<dev-host>]] (project dev runs here)

**User UI/UX preferences:**
- HELP_<lang> under every section
- Click-time entity-id store, don't regex-parse
- "Make everything launchable from the dashboard"

> 7 sources · ~12K token context · ready

After the > ready marker, the agent waits for the user's first question with the full preloaded context.

What if no project can be detected#

The session name is e.g. "wellbeing", "agentic-os audit", "general thinking" — no specific project slug.

In this case the agent loads only base context: 1. 02-Projects/Index 2. 04-Tasks/Backlog first few paragraphs (urgent tasks) 3. 01-Daily/Index today + yesterday 4. Does NOT drill deeper — waits for the user's direction

Token budget for vault-meta sessions: ~2K tokens is enough. The 15-20K cap is an upper limit on aggressive pre-load, not a target. If the session is vault-meta (workflow dev, vault restructure, smoke test, agent-meta work), the minimal version is cheaper and just enough — 3 sources × ~700 tokens. Use the 15-20K budget when a project is detected: project file + 5 sessions + ADRs + memory + backlog + host.

When multiple projects are detected#

E.g. /11.11start "projectA + projectB review". The agent: 1. Loads both project files + the last 2-2 sessions (not 5-5, to avoid blow-up) 2. Focuses Memory pre-load only on the shared topic (e.g. brand)

Budget#

Target: ~15-20K token pre-load. If you overshoot, first trim Tasks/Backlog entries (only top 5 priorities), then drop the older sessions.

Audio overview#

  • EN narration (Charon voice): [[.vault-nb/audio-overviews/Auto-context-loading.en.mp3]]
  • HU narration (Kore voice): [[.vault-nb/audio-overviews/Auto-context-loading.hu.mp3]]

Generated via Gemini 3.1 Flash TTS preview. ~1-2 minutes each. See gemini-3-1-flash-tts-pipeline for the pipeline.