sema-seed

Sema Seed Loop

Generative loop for growing the vocabulary. Pull a random word, think about it openly through the sema lens, mint only if a genuine pattern emerges.

Preflight (once per session)

1. Ask the user about oversight. Default to asking confirmation before each mint:

   "Do you want to review each pattern before I mint it, or should I mint autonomously when the four checks pass?"

   Set this for the session and stick with it unless told to change mode.

2. Verify minting is enabled. sema_mint is exposed by default. If it's missing, the deployment has disabled it. Tell the user:

   "Minting isn't available on this server — it's been disabled via SEMA_DISABLE_MINT=true. Unset that env var and restart to re-enable."

3. Verify the active DB is writable. Call sema_use() (no args) to see current. If it's the bundled DB, tell the user to sema build my.db --preset full and sema_use(db_path="my.db") first.

4. Offer the UI. Ask: "Want to open the Sema UI to watch new patterns appear?" If yes, invoke the sema-ui skill.

The run

One word per run, unless the user explicitly asks for more. Most runs end with no mint — that's expected.

Never dismiss a word before reasoning through it. Even words that look obviously trivial, overly abstract, or already covered must go through step 2 first. The reasoning itself is the test — you cannot know whether a word names a missing pattern until you have actually tried to compose existing handles around it. Skipping the thinking and jumping straight to "no, this isn't useful" defeats the purpose of the seed loop.

1. Pull a random word

# macOS has shuf via coreutils; if not, use awk:
awk 'BEGIN{srand()} {print rand() "\t" $0}' /usr/share/dict/words | sort -k1,1 | head -1 | cut -f2-

Or with shuf:

shuf -n 1 /usr/share/dict/words

2. Reason about the word using sema patterns

Output at least 3-4 paragraphs of reasoning where sema handles are woven through the prose as the load-bearing vocabulary. Not a search summary. Not a definition. An actual chain of thought about the word, with handles doing real work in every paragraph.