A transient sampleTable failure during ingest used to take out every
table in a connection: generateTableDescription returned a hardcoded
'Table not found' string into descriptions.ai, and KtxDescriptionGenerator
was constructed without a logger, so the failure left no trail anywhere.
- sampleTable / sampleColumn calls retry 3x with 200/400/800ms backoff,
honouring KtxScanContext.signal via a new KtxAbortedError.
- On retry exhaustion or missing capability, table generation falls back
to a metadata-only prompt built from column name / native type / comment
/ rawDescriptions. The column path follows the same rule -- call the
LLM when any of samples or rawDescriptions are available; skip only
when both are absent.
- Logger is now threaded from KtxScanContext into the generator. Failures
emit structured KtxScanWarning entries (new description_fallback_used
code, plus existing sampling_failed / enrichment_failed /
connector_capability_missing). ktx scan groups warnings by code so a
batch of identical failures collapses to one summary line plus sample.
- Returns null on failure instead of the 'Table not found' sentinel; the
manifest writer's existing guard already skips empty descriptions, so
schema YAML no longer carries misleading text. SCAN_MANAGED_DESCRIPTION_KEYS
already strips stale 'ai' on merge, so existing YAML clears on next run.
Also suppress AI SDK v6 'system in messages' warning: pull system messages
out of KtxMessageBuilder.wrapSimple's output via a new splitKtxSystemMessages
helper and pass them top-level to generateText (preserves cacheControl
providerOptions on the SystemModelMessage). Agent-runner's local
splitSystemPromptMessages dedupes onto the shared helper.
* fix(llm): wire prompt caching through all Anthropic call sites
- page-triage classifier + light-extraction now put the static skill
prompt in `system:` so the per-document caches hit instead of
re-sending boilerplate in the user message every call.
- Description generation builders return `{ system, user }` with
instruction text + word limit moved into the cacheable system.
- Relationship-LLM proposal framing moved to `system:`.
- `KtxMessageBuilder.wrapSimple` skips the history breakpoint for
single-message calls (cache write that could never be reused).
- Gateway backend now sets `anthropic-beta: extended-cache-ttl-2025-04-11`
so 1h TTLs don't silently downgrade to 5m on Gateway routes.
* fix(llm): keep wrapSimple history breakpoint so multi-step agent loops cache
Reverts the wrapSimple `messages.length > 1` guard from the prior commit.
agent-runner uses wrapSimple with a single user message, but generateText
runs a multi-step tool loop inside it — the cache marker on the first user
message is reused by every subsequent step, so it isn't waste.
The release validator (scripts/validate-llm-debug-jsonl.mjs) also requires
a `message-part` marker target in captured debug JSONL.
