pal-learn
/pal-learn
Zapisuje zaobserwowany fakt o routingu PAL w mapie modeli pal-router, po weryfikacji u dostawcy, tak by wiedza o routingu nie traciła aktualności.
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 3
Brak sygnałów. Repozytorium zmienione 08.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Zapisuje zaobserwowany fakt o routingu PAL w mapie modeli pal-router, po weryfikacji u dostawcy, tak by wiedza o routingu nie traciła aktualności.
Wyzwalacze z opisu
Oryginalny opis (EN)
Files an observed PAL routing fact back into pal-router's model map so the framework's routing knowledge compounds instead of going stale. Ingests one observation — a model refused a prompt, a model ID stopped being served, a model started returning 402/429, a fallback worked where the primary failed, a model hallucinated on a task class, a rate or context limit turned out to be different than documented — verifies it against the live provider before writing, and enriches the existing entry rather than appending a duplicate. Triggers on: 'gemini refused again', 'that model is gone', 'nemotron
Jak działa
- Why this exists
- Step 0 — is it a routing fact?
- Step 1 — classify the observation
- Step 2 — verify before writing
- Step 3 — deduplicate, then enrich
- Step 4 — file it
- Step 5 — state what changed
- What NOT to file
Podgląd SKILL.md
--- name: pal-learn description: >- Files an observed PAL routing fact back into pal-router's model map so the framework's routing knowledge compounds instead of going stale. Ingests one observation — a model refused a prompt, a model ID stopped being served, a model started returning 402/429, a fallback worked where the primary failed, a model hallucinated on a task class, a rate or context limit turned out to be different than documented — verifies it against the live provider before writing, and enriches the existing entry rather than appending a duplicate. Triggers on: 'gemini refused again', 'that model is gone', 'nemotron made that up', 'update the model map', 'log this routing failure', 'X 402d', 'the fallback worked', a pasted provider error body (402 / 429 / 404 / refusal text), or the end of a session in which any PAL route behaved differently than the map predicted. --- # pal-learn **One-line pitch:** The mechanism that keeps `pal-router`'s model map true. Routing facts decay — models get withdrawn, free tiers change, refusal behaviour shifts — and a stale map sends work to a model that cannot serve it. Adapted from the `pentest-learn` / `pentest-debrief` loop in the pentesting knowledge base (https://github.com/Darxbloo/Pentesting-Agent-new), which files engagement lessons into the skill that owns them. This applies the same discipline to routing. ## Why this exists The map in `skills/pal-router/model-map.md` named `x-ai/grok-4.1-fast` at 2M context as the fallback for three separate routes. OpenRouter had stopped serving it, and the documented replacement was paid-only — so all three routes had a fallback that returned 402. Nothing in the framework noticed, because nothing owned keeping the map current. That is the failure this skill prevents. ## Step 0 — is it a routing fact? File only facts about **how a model behaves as a delegation target**: availability, cost, refusals, limits, context, and output quality per task class. Not prompt technique, not findings, not target data. A one-off network blip is not a routing fact; a reproducible behaviour is. ## Step 1 — classify the observation | Signal | Meaning | Where it goes | |---|---|---| | 402 / insufficient credits | Paid model on an unfunded account. Billing, not routing. | Mark the model **paid** in the map; do not remove it | | 404 / model not found | ID withdrawn by the provider | Replace the ID, or retire the entry and repoint aliases | | 429, `limit_source: upstream_provider_shared_pool` | Transient upstream congestion on a `:free` model | Note as a known flake; do NOT demote the model | | 429, account or key limit | Quota exhausted | Record the real limit and its scope (per key vs per account) | | Refusal text | Safety classifier declined the task class | Record the task class it refuses, and what to route to instead | | Wrong or invented output | Hallucination on a task class | Record the task class to avoid, with the observed failure | | Fallback succeeded | The chain works | Promote it in the failover order | ## Step 2 — verify before writing Never file a provider claim from memory or from a single failed call. Confirm it: - `mcp__pal__listmodels` — what PAL can actually reach right now. - `curl -s https://openrouter.ai/api/v1/models` — whether the ID is still served. - `curl -s -H "Authorization: Bearer $KEY" https://openrouter.ai/api/v1/key` — credits and limits. Treat its `free_model_daily_requests` as eventually-consistent; trust the provider's documented rule over the counter. - Re-run the failing call once. A single failure distinguishes nothing. Record the date you verified it. Every entry in the map is a claim about a moving target. ## Step 3 — deduplicate, then enrich Grep the destination first. If an entry for that model exists, **add to it** — a new failure mode, a corrected limit, a date. Do not append a second entry under a variant name. If the fact contradicts what is written, correct the text rather than stacking a caveat on it. ## Step 4 — file it | Fact | Destination | |---|---| | Per-model behaviour, limits, failure modes | `skills/pal-router/model-map.md`, that model's entry | | Changed preference order for a task | `model-map.md` "Failover order" + `pal-router/SKILL.md` | | A model named in the rendered hook | `setup.sh` routing strings — the hook must not name a model that cannot serve | | A model alias PAL cannot resolve | the `OPENROUTER_MODELS_CONFIG_PATH` / `CUSTOM_MODELS_CONFIG_PATH` registry | | A claim in user-facing docs | `README.md` routing map, `docs/WORKFLOW.md` | A fact that appears in more than one of these must be corrected in all of them; a stale README is as misleading as a stale map. ## Step 5 — state what changed Report the model, the observation, how it was verified, the files touched, and anything left unverified. Never imply a limit was confirmed when it was inferred. ## What NOT to file - Single transient failures with no second observation. - Provider limits copied from documentation without checking them against the account in use. - Model recommendations with no observed evidence behind them. - Target data, findings, or credentials. This skill touches routing knowledge only.
Wersje
W bazie nie ma innych skilli o nazwie pal-learn.
Migawka bazy z 13.11.2026 · wersja paczki 4212b04f6908 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100pal-router
Kieruje podzadania do tańszych lub darmowych modeli PAL, a gdy któryś odmówi, przełącza na zapasowy, by kontekst Claude Opus został na osąd. - 100create-skill
Tworzy nowy osobisty skill w katalogu ~/.pal/skills/ i podłącza go do każdego zainstalowanego agenta. - 100multi-model-delegation
Prowadzi konsultacje projektowe z wieloma modelami przez PAL (kimi, glm, gemini, gpt): burza mózgów nad projektem i godzenie rozbieżnych odpowiedzi. - 100reflect
Diagnozuje, dlaczego zachowanie PAL nie zadziałało zgodnie z oczekiwaniem, śledząc hooki, instrukcje i logikę aż do luki. - 95claude-model-router
Kieruje zadania do najtańszego modelu, który da tę samą jakość, aby oszczędzać limit użycia Claude. - 100model-router
Kieruje każdego podagenta do najtańszego modelu, który da radę, i zapisuje każdą decyzję routingu przed uruchomieniem agenta.