radar-harvest

What this skill does

/tech-radar fetches ~500 items every morning and the digest keeps ~3. That discarded 99% is exactly the knowledge that should compound, and a line like "GLM-5.2 takes the open coding crown" is a signal to act, not just to report. This skill mines the same raw fetch into two flows:

1. Knowledge growth — the wiki-worthy discards become references/ pages so Servy's knowledge grows every day.
2. Adoption — named tool/model releases Servy could use flow through an adopt / watch / skip decision; STRONG adopts become [capability] backlog items (→ the backlog-util.py bridge → the-builder-room).

It composes existing machinery, it is not a parallel system: input is tech-radar-fetch.py's raw JSON; output lands in the references/ wiki and the evolution-backlog.md bridge. Engine: scripts/radar-harvest.py. Cloud sibling: routines/radar-harvest.md.

Governance (read before running)

• Ingesting wiki pages + filing backlog items = Tier-S (auto). The Ingestion loop already auto-writes the wiki; this is the same risk class.
• Actually WIRING an adopted tool — installing a CLI, editing the multi-brain model roster, adding an MCP — is Tier-X. This skill SUGGESTS it (digest + a [capability] item + a pending.md line); Gera arms the install/edit. Never auto-install or auto-edit the roster. (A doc-only note in references/ai-dev-tools-landscape.md is the one Tier-S exception.)
• SoftServe boundary (references/softserve-ai-usage-policy.md): public/personal data only. Never ingest customer IP, never route it through the multi-brain.

Process

0. Ensure today's fetch exists. If knowledge/radar/<date>-raw.json is missing, run python3 scripts/tech-radar-fetch.py (don't re-fetch if it's already there — same-day dedup).
1. Plan. python3 scripts/radar-harvest.py plan --date <date> → JSON {ingest:[...], adopt:[...]}. This is the scored shortlist, not the whole 500.
2. Ingest (knowledge growth). For each ingest item (respect the cap), follow references/wiki-protocol.md:
   • Read the source/summary; if it's thin, fetch the URL for substance (WebFetch / ctx_fetch_and_index).
   • Write a NEW concept/tool/entity page or fold into the closest existing page (the engine already dampened things the wiki covers — prefer appending a dated note to a stub-worth item). Quality over count: a few solid pages beat 8 stubs.
   • Update references/index.md and append a ## [<date>] ingest | radar-harvest line to references/log.md.
3. Adopt (signal → capability). For each adopt candidate:
   • Log the signal for cross-day accumulation: python3 scripts/radar-harvest.py signal --tool "<name>" --title "<t>" --url "<u>" --source "<s>".
   • python3 scripts/radar-harvest.py score to see which tools are STRONG (≥3 mentions, or ≥2 mentions from ≥2 sources over 7d).
   • For a STRONG candidate, run a quick adopt / watch / skip decision — invoke /ai-apply on "should Servy adopt <tool>?" or spin the-lab for a dossier. On adopt: file python3 scripts/backlog-util.py add --tier T2 --source radar-harvest --desc "[capability] adopt <tool>: <one-line why> — wire via <skill/script/roster note>" (→ bridge → builder). On watch: keep the signal, no backlog item. On skip: nothing.
   • If adoption implies a Tier-X wiring (CLI install / roster edit / MCP), add a pending.md line for Gera to arm — do NOT wire it yourself.
4. Digest. python3 scripts/radar-harvest.py digest --plan <plan.json> and send to Telegram (TG_BOT_TOKEN/TG_CHAT_ID from .env): pages added, adopt candidates + verdicts, what was queued vs what needs Gera to arm.
5. Report in chat: pages ingested (with links), adopt verdicts, capabilities filed, Tier-X items parked for Gera.

Caps & discipline

• Ingest ≤ 8 pages/day, adopt ≤ 5 candidates/day (engine defaults; tune with --ingest-cap/--adopt-cap). Don't flood the wiki — the warden agent lints it.
• Re-runnable and idempotent-ish: re-running the same day appends, the dedup dampener stops re-paging covered topics, and the signal store dedups by tool key over the window.
• "adopt" argument → skip step 2, run only the adoption pass (steps 3-4).