Initial commit: Hermes Agent Skills collection
This commit is contained in:
Debian
@@ -0,0 +1,28 @@
|
||||
# Session 20250621 — Essensplaner Cronjob Refactoring
|
||||
|
||||
## Context
|
||||
Rechnungen-Organizer Cronjob war hart getroffen: `RuntimeError: Response truncated due to output length limit`. Anlass für komplette Essensplaner-Pipeline-Überarbeitung.
|
||||
|
||||
## Learnings
|
||||
|
||||
### 1. Output Truncation in Agent Cronjobs
|
||||
Bei `Rechnungen-Organizer` gab der Agent riesige Markdown-Blöcke aus → harter Cutoff.
|
||||
**Fix:** Im Prompt explizit begrenzen: `max 20 lines`, `compact summary only`, `no explanations`.
|
||||
**Ergebnis:** Neuer `nutrition-plan-generator-v4.py` mit kompakter 7-Zeilen-Tabellen-Ausgabe statt Voll-Plan.
|
||||
|
||||
### 2. Telegram Delivery Targets
|
||||
User nannte `telegram:Essensplaner` als Ziel. System kannte nur `telegram:Schoen Inc`.
|
||||
**Fix:** `@Essensplaner` ist ein Bot-Name INNERHALB der Gruppe, kein `deliver`-Target. Cronjobs müssen `telegram:Schoen Inc` verwenden.
|
||||
**Workflow:** Bei Zweifel vorher `send_message(action='list')` oder existierende Jobs inspizieren.
|
||||
|
||||
### 3. Navigation mit Tabulator
|
||||
Anstatt `exec` auf kompletten Prompt-Outputs: Arbeitsverzeichnis wechseln → Verzeichnisinhalt tabellarisch anzeigen → relevante Dateien nach Bedarf per `cat` laden.
|
||||
Vermeidet: a) Output-Truncation b) Unlesbarkeit bei ~5000 Einträgen c) Wartezeit.
|
||||
|
||||
### 4. Model + Provider als Paar setzen
|
||||
Bei jedem Update eines Agent-Jobs: `model` und `provider` zusammmen. Einzelnes `model` reicht nicht.
|
||||
Beide Cronjobs (Interview + Plan-Generator) jetzt auf `moonshotai/kimi-k2.6` / `noris`.
|
||||
|
||||
### 5. Referentielle Labels statt Zahlen-IDs
|
||||
Link zwischen Interview (Mittwoch) und Plan-Generator (Freitag) über `current_interview.json`.
|
||||
Bei Livedaten durch den Agent: JSON als Arbeitsdokument, nicht Memory.
|
||||
@@ -0,0 +1,10 @@
|
||||
## Session Summary (2026-06-17)
|
||||
|
||||
- Issue: Cronjob **Rechnungen‑Organizer** failed with `RuntimeError: No LLM provider configured`.
|
||||
- Resolution steps:
|
||||
1. Verified current Hermes config (`hermes config`) – model/provider set to `noris`/`vllm/gpt-oss-120b`.
|
||||
2. Updated the cronjob to use the same model/provider via `cronjob action=update`.
|
||||
3. Confirmed the job now has `model: vllm/gpt-oss-120b` and `provider: noris`.
|
||||
- Additional improvement: Added a new pitfall entry to the **cronjob** skill covering the "No LLM provider configured" error and the required `hermes model` / `hermes config set` steps.
|
||||
|
||||
**Takeaway:** Always ensure a cronjob’s model/provider aligns with the global Hermes configuration, or explicitly set them on the job.
|
||||
@@ -0,0 +1,91 @@
|
||||
# Chefkoch Scraper Session: v4 First Production Deploy
|
||||
|
||||
## Session Date: 2026-06-18
|
||||
## Trigger: User requested continuous scraping to 25,000 recipes
|
||||
## Final State: Scraper running as nohup daemon (PID 80297) + watchdog cronjob
|
||||
|
||||
---
|
||||
|
||||
## Problem Discovery Log (Failure Chain)
|
||||
|
||||
### Failure 1: LLM-based cronjob fails on provider mismatch
|
||||
- Cronjob `chefkoch-real-recipes-refill` (job_id: c77127dd9130)
|
||||
- 06:01: `Error 400: model is required` — no model configured on job
|
||||
- Fix attempted: set `model="moonshotai/kimi-k2.6"` — BUT omitted `provider`
|
||||
- 06:17: `RuntimeError: No Anthropic credentials found` — system defaulted to anthropic provider
|
||||
|
||||
**Lesson:** Model + provider MUST be set as a PAIR on agent cronjobs. Setting only one creates a mismatch.
|
||||
|
||||
### Failure 2: `terminal(background=true)` silently killed scraper
|
||||
- Attempted to start scraper via `terminal(background=true)`
|
||||
- Result: Exit code 143 (SIGTERM after kill)
|
||||
- Also: `tcsetattr: Inappropriate ioctl for device` — pseudo-tty interferes with Playwright
|
||||
|
||||
**Lesson:** Background process mode is NOT a nohup replacement. Use explicit `nohup` wrapper script.
|
||||
|
||||
### Failure 3: Playwright browser not found
|
||||
- Playwright installed in nutrition-coach venv, but chromium not found
|
||||
- Error: `Executable doesn't exist at ...chrome-headless-shell`
|
||||
- Browser actually installed under `/home/debian/.cache/ms-playwright/` (installed globally)
|
||||
- venv looking elsewhere
|
||||
|
||||
**Lesson:** Set `PLAYWRIGHT_BROWSERS_PATH=/home/debian/.cache/ms-playwright` explicitly. Do NOT rely on auto-detection when multiple Python environments exist.
|
||||
|
||||
---
|
||||
|
||||
## Working Solution (Final)
|
||||
|
||||
### 1. Standalone Scraper Script
|
||||
- `~/.hermes/scripts/chefkoch_scraper_v4.py` — headless, no LLM, asyncio
|
||||
- Reads/writes directly to `nutrition-coach` profile dir
|
||||
- Uses JSON-LD extraction (fast), Playwright only for URL discovery
|
||||
|
||||
### 2. nohup Wrapper Script
|
||||
- `~/.hermes/scripts/start_scraper_v4.sh`
|
||||
- Sets `PLAYWRIGHT_BROWSERS_PATH` before launching
|
||||
- PID tracking in `/tmp/chefkoch_scraper_v4.pid`
|
||||
- Idempotent: checks if already running
|
||||
|
||||
### 3. Watchdog Cronjob (no_agent)
|
||||
- `chefkoch-scraper-watchdog` (job_id: a884c342c49e)
|
||||
- Runs every 10 minutes
|
||||
- Calls wrapper script
|
||||
- `no_agent: true` — zero LLM cost
|
||||
- `deliver: local` — no notification spam
|
||||
|
||||
### 4. Original Cronjob (agent-based) deprecated
|
||||
- Original `chefkoch-real-recipes-refill` still exists but should be paused
|
||||
- The LLM-based approach (checking count + starting scraper) is overkill
|
||||
- Direct nohup + watchdog is cheaper and more reliable
|
||||
|
||||
---
|
||||
|
||||
## Performance Observed
|
||||
|
||||
- Discovery: ~50 URLs per batch from search pages
|
||||
- Extraction: ~60 recipes/minute with single context
|
||||
- 5,000 recipes in ~60–70 minutes
|
||||
- 25,000 recipe ETA: ~7~10 hours continuous
|
||||
|
||||
---
|
||||
|
||||
## Session-Context References
|
||||
|
||||
- `real_recipes.jsonl` — append-only output file
|
||||
- `scraper_state.json` — tracks seed_cycle for resume
|
||||
- `scraper_v4.out` — live log file
|
||||
- `scraper.log` — legacy log from batch run (still there)
|
||||
|
||||
---
|
||||
|
||||
## New Pitfalls Captured in Skills
|
||||
|
||||
1. **multi-source-recipe-scraper / infinite-refill-scraper-v4.md:**
|
||||
- Browser path detection when multiple venvs exist
|
||||
- `nohup` wrapper pattern for daemon mode
|
||||
- `terminal(background=true)` tcsetattr conflict with Playwright
|
||||
- Watchdog cronjob pattern
|
||||
|
||||
2. **cronjob / SKILL.md:**
|
||||
- Model+provider MUST be set as pair on agent jobs
|
||||
- Custom provider mismatch error signature
|
||||
Reference in New Issue
Block a user