Scheduled collect → enrich → digest
This skill runs a prebuilt aevatar workflow (scheduled_collect_enrich_digest) that turns
one connected source into a ranked, AI-summarized digest. It is the native aevatar
re-expression of the common "data scraper agent" pattern — collect → enrich → store —
with the engine's own constraints respected (see Scope below).
A run does three things, in order:
- Collect — one
nyxid_proxycall runs a Tavily web search (tavily-search-chrono-ai) for the brief. - Enrich — one batched
llm_callscores (0–100) and one-line-summarizes every item against the run brief. - Digest — the ranked result is frozen as the workflow output and is watchable in the observatory.
There is no per-run database write in v1 — the run is the record. Delivery and persistence are optional extensions (see Scope).
Protocol (follow in order)
- Confirm the source is connected. This instance is wired to the nyxid connector
tavily-search-chrono-ai(Tavily web search, ChronoAI org-shared). Callnyxid_services{"action":"list"}and confirm that slug is present andallowed. If it is missing, stop — connect Tavily (or retarget the workflow to another source) before running. - The prompt is the search query. There are no placeholders to fill for the default (Tavily)
source —
collectalready searchestavily-search-chrono-ai. Compose the run prompt as a single clean query line: the topic plus any recency window ("aevatar workflow updates, last 7 days"). It is both the Tavily query and the ranking brief, and — because the engine has no clock — the only place a date/window can enter. (To retarget another source, edit thecollectbody +extract_itemspath in the YAML.) - Run it. Call
aevatar_start_workflowwithworkflow_id: "scheduled_collect_enrich_digest"andinputs.promptset to the query/brief. A cleanrun_finishedwith norun_errormeans it executed end to end; then scan the digest output for error-shaped content. - Report honestly. Return the ranked digest and the observatory link. State that you verified it ran structurally; do not claim the scoring is objectively correct.
Scheduling
The engine does not schedule itself. "Recurring" means an external trigger (a scheduled agent
or any cron caller) invokes aevatar_start_workflow on an interval, each time passing a fresh brief
with the current window. Set that trigger up separately; this skill is the thing it calls.
Inputs
| Input | Where | Purpose |
|---|---|---|
prompt |
inputs.prompt |
The search query + relevance criteria + any date/window. Required — it is the Tavily query, the ranking brief, and the only clock. |
| source (slug/path/array-path) | hardcoded in the YAML | Wired to tavily-search-chrono-ai → POST /search → results[]. Edit the collect body + extract_items path to retarget. |
Output
A single ranked digest (text), highest score first, one line per item:
<score> — <title> — <one-line why it matters> — <url>. Also persisted as the run record in the observatory.
Scope (what this draft deliberately does and does not do)
- Does: single-source collect, one batched enrich pass, deterministic extraction, ranked digest, observatory visibility.
- Does not (v1):
- Cross-run dedup / "only new since last time". That needs state that outlives a run, which in aevatar belongs to an actor / read-model, not a stateless workflow. Do not fake it with in-workflow memory.
- Feedback learning (ECC's "gets smarter from your decisions"). Same reason — persistent preference state is an actor concern, a future companion skill, not this pipeline.
- Multi-source fan-out. Add sources by composing collect steps (or a
parallel/foreachover a source list) — kept to one source here for a clean first run. - Delivery / storage to an external store. Append a final
tool_call(e.g. post the digest to a chat connector, or write to a store the user has) once the digest path is verified.
See DESIGN.md for the full ECC → aevatar mapping, connector dependencies, and verification checklist.