Projekt context

Skill valfiguer/Projekt-Skills/skills/projekt-context

Use Projekt as an AI-context store (Serena-style): mirror a codebase's memory/context .md files into a Projekt project's docs (one doc per file, idempotent UPSERT) and load them back as ONE markdown bundle, so Claude Code / Codex / any AI reads repo context cheaply over the API instead of crawling the whole codebase. Use whenever the user wants to save Claude/Serena memories into Projekt, sync a memory directory, or load project context for an agent. Soporta español: memorias, contexto de proyecto, sincronizar memorias, cargar contexto.From its SKILL.md

Install

npx -y skills add valfiguer/Projekt-Skills --skill projekt-context

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

5.2 KB, as published. Nobody here has run it

projekt-context — AI-context store (memories ↔ docs)

⚠ NOT YET PORTED to /api/v1. The base URL is fixed (shares projekt_api.py, now on https://projekt.3xa.es/api/v1), but the endpoint PATHS in context_store.py are still the legacy flat docs ones (/projects/{pid}/docs, /projects/{pid}/docs/{did}?format=markdown) and will 404 on the rewrite. To port: docs live at /organizations/{org}/projects/{proj}/documents (contract paths/documents.yaml; the ?format=markdown round-trip is preserved); confirm with bash ../projekt/scripts/spec_lookup.sh --search documents. Working today: the projekt connect (auth/context) + projekt-issues task create/list.

Keep a codebase's context memories (architecture, gotchas, conventions, runbooks, decisions) as Projekt docs so any AI provider reads them over the API — ?format=markdown, cheap — instead of re-deriving them by crawling the repo and burning tokens. Same idea as Serena's .serena/memories/, but stored in Projekt and shared across providers and machines.

CS="${CLAUDE_SKILL_DIR}/scripts/context_store.py" — use it for every command below.

Prerequisite — connect first

Run the orchestrator's connect steps so the token + org + .projekt-run/context.json exist:

bash "${CLAUDE_SKILL_DIR}/../projekt/scripts/auth_check.sh"
bash "${CLAUDE_SKILL_DIR}/../projekt/scripts/context_sync.sh"

No token? → projekt/references/auth-setup.md. Wrong org (cross-org token 403s)? export the right TREXA_API_TOKEN before connecting.

Commands

1. sync — mirror a memory dir into Projekt docs (idempotent, DRY-RUN first)

UPSERT every *.md in a directory as a doc under a parent (default "Claude Memory"), one doc per file, title = the filename stem (the stable idempotency key). Re-running PATCHes the same titles — never duplicates. The body is sent as Markdown (the server converts it to rich EditorJS blocks — tables, code, callouts); YAML frontmatter is parsed and rendered as a clean lead blockquote (> description _(type)_) instead of being flattened into the doc.

# dry-run (counts files, shows parent + index)
python3 "$CS" sync --project <PID|KEY|name> --memory-dir ~/.claude/.../memory

# apply
python3 "$CS" sync --project <PID|KEY|name> --memory-dir ~/.claude/.../memory --apply

--parent "Claude Memory" — parent doc title the memories nest under (created if missing).
--index MEMORY.md — this file becomes the parent body (the index) instead of a child.
--icon 🧠 — parent icon, on CREATE only.
Title = filename stem, NOT the frontmatter name: slug — keep filenames stable or a rename creates a new doc (the old one lingers; archive it).

2. load — read the store back as one markdown bundle (cheap onboarding)

Concatenate the parent's child docs (each fetched with ?format=markdown) into a single Markdown blob — the one call an agent makes at session start to skip crawling the repo.

python3 "$CS" load --project <PID|KEY|name>                 # → stdout
python3 "$CS" load --project <PID|KEY|name> --out ctx.md    # → file

--parent "Claude Memory" — which tree to bundle.
Output: # <parent> — AI context bundle, the index, then one ## <title> section per memory.

How it pairs with the API

Write: docs accept a markdown field on POST and PATCH (the server runs EditorJsMarkdownConverter::fromMarkdown), so this skill never hand-builds EditorJS JSON.
Read: GET /projects/{pid}/docs/{id}?format=markdown returns compact Markdown (~half the tokens of the raw EditorJS blocks) — what load uses.

Gotchas

Idempotency key is the filename stem. Two files with the same stem collide; a renamed file creates a second doc.
sync writes; load is read-only. sync is dry-run until --apply; load always runs.
Frontmatter is stripped, not stored. Only description + type are surfaced (as a lead blockquote). Put anything you need preserved in the body.
Cross-org / no write permission → 403 on sync; switch token/org. See projekt/references/errors.md.

What it does NOT do

No delete/prune of stale docs (a removed memory file leaves its doc behind — archive it manually).
No bidirectional sync (Projekt → local). One direction: local memories → Projekt docs.
No auth/setup of its own — it reuses the projekt orchestrator's connect + shared API client.

Shared references

projekt/references/endpoints.md (Docs section) · errors.md · auth-setup.md. Shared API client + slim projections live in projekt/scripts/lib/projekt_api.py.