add prompts/DSH_FIRST_PROMPT.md
This commit is contained in:
@@ -0,0 +1,21 @@
|
|||||||
|
# First prompt for dsh
|
||||||
|
|
||||||
|
Read `/AGENTS.md` and `.agents/skills/storyline/SKILL.md` in the repo root, then
|
||||||
|
work that project's feature backlog. Concretely:
|
||||||
|
|
||||||
|
1. Run the test suite first: `cd prototype && python -m pytest -q`. It must pass
|
||||||
|
before you change anything.
|
||||||
|
2. Implement `cluster.py --serve :8765` (serve `prototype/out/` — index.html and
|
||||||
|
stories.json — on port 8765) if it is not already wired:
|
||||||
|
`python cluster.py --serve :8765`.
|
||||||
|
3. Keep the LLM top-k cap: only `--topk` ranked stories get summarized, and each
|
||||||
|
input is headline + first ~300 chars. Never summarize the long tail.
|
||||||
|
4. Do these and NOTHING else:
|
||||||
|
- Do NOT rebuild or replace FreshRSS. It stays a stock Docker container.
|
||||||
|
- Do NOT add Kafka or any message broker.
|
||||||
|
- Do NOT add a frontend framework. `out/index.html` stays self-contained HTML.
|
||||||
|
- Do NOT change ports: FreshRSS 8280, RSSHub 1200, Hermes 8080, serve 8765.
|
||||||
|
- Story summaries must hit the host Hermes/llama.cpp (`LLM_BASE_URL`
|
||||||
|
default `http://127.0.0.1:8080/v1`, model `hermes`) — never DeepSeek.
|
||||||
|
5. When done: run `pytest -q` once more, confirm still green, and report which
|
||||||
|
backlog items are complete and which remain.
|
||||||
Reference in New Issue
Block a user