frontier-digest

Files (2)

• SKILL.md
• SKILL_detail.md

SKILL.md (first lines)

# frontier-digest — FH Frontier Insight Collection & Synthesis

> Automatically collects the latest AI/agent/harness engineering trends from HackerNews, arxiv, launch channels, GeekNews and vendor changelogs,
> then compresses and synthesizes them into insights directly relevant to forge-harness skills and structure.

## Triggers

```
/frontier-digest                # immediate execution
/frontier-digest --save         # save result to file
"what's on the frontier"
"this week's AI trends"
"harness trends"
"latest agent developments"
"any ideas for improving the harness?"
```

---

## Step 0. API Environment Detection

This is the trend-scan specialization (rung 3) of the **Deep-Research Capability Ladder**
(`../../../../knowledge/shared/harness-core/deep_research_capability_ladder.md`). The ladder owns
the **cross-rung routing** (when a task is trend-scan at all vs general research); the `Priority:`
block below is frontier-digest's own **internal engine resolution** for the HN/arxiv case (API-key
vs WebSearch) — a narrower detection, not a re-definition of the cross-rung ladder.

```
Priority:
0. /deep-research built-in available (check live session skill list)
   → use it as the collection+verification engine (staged source gathering,
     cross-checking, cited synthesis) — Tier-0 route, no API key needed
1. ANTHROPIC_API_KEY environment variable → Claude Sonnet
2. Neither → WebSearch mode (raw data only, no synthesis)
```

Report detection result in one line. Example: `🔑 Using Claude Sonnet` / `🔑 Engine: /deep-research (built-in)`

On first run without an API key, output the environment setup guide (format in §Output-Formats).

---

## Step 0.5. Operator Intake (speculative-interview arm — walled channels)

On **cadence-triggered** runs (7d), ask the operator one line before collecting:

> *"이번 주에 본 벽 뒤 소스(YouTube·LinkedIn·X 등 기계가 못 닿는 링크/요약)가 있으면 던져 주세요 — triage해서 기록합니다. 없으면 그냥 진행할게요."*

- Operator may skip — zero pressure; the autonomous arms (Step 1) run regardless.
- Why: walled channels return 403 to machine fetch — **the operator is the only wide-net sensor for them**. This arm turns ad-hoc link-throwing into a scheduled intake (manual-validated n=2).
- 🟥 **YouTube is half-walled, not walled (corrected 2026-10-29).** A channel's *upload list* is machine-readable: the channel Atom feed, or `yt-dlp --flat-playlist` when that feed answers 404 (measured 2026-10-29: 404 for every channel tried, a well-known control channel included; the same feeds answered again a few hours later — the outage is intermittent, so the fallback stays). Listing new conference talks is the Step 1 `slow-leg` row's job; only *reading* a talk needs the video path below. LinkedIn and X stay operator-only.
- Received sources route to the sister-asset/triage flow with its **lightweight path**: C-tier (territory already covered) = one-paragraph entry only; full cross-audit reserved for A/B-tier. Partial wall-bypass is allowed first: try WebSearch + secondary sources before declaring unfetchable.
- **Video sources (local/laptop only — cloud VMs typically 403 on video hosts)**: probe by **capability, not engine name** via the Sidecar Engine Resolution Protocol. Tier 1 = natively multimodal CLI ingesting the URL directly (⚠️ direct `gemini` CLI EOL 2026-06-18 — probe `agy` or the Gemini API after); never route video to a coding-agent CLI without native video access (it burns tokens and recovers only metadata). Tier 3 = `yt-dlp` transcript fallback, **conditional** — probe the environment first, never assume it works. Unresolvable → operator summary remains the path.

> **Detail**: See `SKILL_detail.md §Video-Harvest` — full tier ladder, probe commands, EOL/router-shell notes — read when a video source needs harvesting.

---

## Step 1. Data Collection

Collect from the sources below (bash/endpoint per source in §Collection-Bash). The backticked `id` is the channel identity the recall lane (`scripts/test_frontier_digest_recall.sh`) checks — keep it when editing a row:

| Source | Method | Cap |
|---|---|---|
| HackerNews `hn` | Algolia API `search_by_date` (date-sorted, never `/search`), score > 30, keyword-relevant | 15 items |
| arxiv `arxiv` | export API, latest by submittedDate | 6 items |
| TLDR AI `tldr-ai` | RSS, title + link | 5 items |
| The Batch (deeplearning.ai) `the-batch` | HTML scraping, `aria-label` + `/the-batch/issue-N` href (the JSON `slug` shape died 2026-09-11) | 5 items, or `UNMEASURED` |
| GeekNews (news.hada.io) `geeknews` | Atom feed (`/rss/news` path), title + link, AI/agent/LLM-relevant | 5 items |
| Launch YC `launch-yc` | official page `ycombinator.com/launches`, newest launches, AI/agent/dev-tool only | 5 items, or `UNMEASURED` |
| Product Hunt `product-hunt` | official Atom feeds, categories `artificial-intelligence` and `developer-tools` | 5 items, or `UNMEASURED` |
| Show HN `show-hn` | Algolia `search_by_date` with `tags=show_hn`, points floor lower than the main HN leg | 8 items, or `UNMEASURED` |
| Vendor changelogs/blogs `vendor-changelogs` | Anthropic, OpenAI, Google DeepMind/Gemini, Meta AI, Cursor — RSS where one exists, else the official news/changelog page | 1-2 per vendor, new since last digest, or `UNMEASURED` |
| Conference channels & tech blogs `slow-leg` | **weekly, not daily**: run only when `python3 scripts/digest_slow_leg.py due` exits 0, then `python3 scripts/digest_slow_leg.py run`. Sources come from an operator-local, gitignored list (`$FH_DIGEST_SLOW_SOURCES`, default `tracks/_meta/digest_slow_sources.tsv`) — which companies someone follows can identify their employer, so no source is named in this public file. Lists new posts and talk uploads only; summarizing a talk goes through `video-ingest` | everything new since last run. `due` exit **3** = not due this week (skip the row); **any other non-zero** (e.g. 2 `STATE-INVALID`) = `UNMEASURED`, never not-due. `run` exit 0 = all read · 4 = some source unread (the listing is still valid; that source waits for next week, so hand it in via Step 0.5) · **any other non-zero** (2 `SOURCES-*`/`STATE-INVALID`, 1 a crash) → `UNMEASURED`. Do not rerun to "retry" — a rerun only appends. Copy **every** source block into the digest, «새 항목 0» lines included (read-and-empty ≠ not read); `UNREACHABLE(…)` · `PARSE-ZERO` · `PARSE-ERROR` lines go in as is and are never `0`. The listing is also saved to `tracks/_meta/digest_slow_<date>.md`. **Never retry a refused feed with a different client identity** — it goes to Step 0.5 as an operator hand-in |

Report progress: `📡 HN 15 items · arxiv 5 items · TLDR 5 items · Batch 5 items · GeekNews 5 items · LaunchYC n · PH n · ShowHN n · Vendors n collected` (+ `· Slow n (+k unread)` on a week the slow leg is due — n = new items listed, k = sources not read; `· Slow not-due` on `due` exit 3; a leg that could not be read is `UNMEASURED`, never `0`)

> **Detail**: See `SKILL_detail.md §Collection-Bash` — curl commands per source with parsing notes — read when executing Step 1.

**Knowledge-currency augment (on-demand, not weekly-scanned)**: distinct from the trend feeds above, these are fetched when a specific version/fact/spec needs confirming against the model's knowledge cutoff — `docs.anthropic.com` (official Claude/API docs · model IDs · pricing), arxiv (already a feed, also a currency target by ID), and Zenodo (research deposits incl. FH's own). Pull these into context for currency rather than scanning them for trends; in egress-restricted environments (e.g. a company network) route the fetch through the approved bridge (n8n), not a direct call.

---

## Step 2. Synthesis

- **With Anthropic API**: run the synthesis prompt (full prompt in §Synthesis-Prompt) — outputs This Week's Frontier Highlights (max 3) / FH Immediate Application Candidates (2-3) / Warning Signals, within 400 characters, no preamble.
- **WebSearch mode (no API key)**: search directly with the WebSearch tool, then synthesize in context (queries in §Synthesis-Prompt).

> **Detail**: See `SKILL_detail.md §Synthesis-Prompt` — full API prompt with FH context block, WebSearch queries — read when executing Step 2.

---

## Step 3. Output

Print the synthesis result in the conversation (format in §Output-Formats): engine line + Highlights + Immediate Application Candidates + Warning Signals + collection stats.

**With `--save` flag**: save to **`{FH}/tracks/_meta/frontier_digest_{YYYY_MM_DD}.md`** — underscores in
the date, matching `date +%Y_%m_%d`. Fallback when no FH install is resolvable:
`~/.claude/forge-harness/tracks/_meta/` → cwd `tracks/_meta/`, same filename either way. After saving:
`✅ Saved: {path}`

⚠️ **This path is load-bearing, not cosmetic.** The cadence detector in `CLAUDE.md §Cadence Rules`
globs exactly `tracks/_meta/frontier_digest_*.md` to decide whether the 7-day proposal is overdue, so
a digest written anywhere else — or with hyphens instead of underscores — is **invisible to the
cadence check forever**, and the skill silently looks never-run. The hub's production runner
(`scripts/frontier_digest_daily.sh` — **hub-local, not distributed in the npm package**: it is half of
a launchd pair and spends CLI calls per run, so an installed copy does not have it and does not need
it — this skill's save path stands alone) already writes this exact path; the previously documented
`digests/frontier_{today}.md` matched neither, and measured 2026-08-11 it had produced **0** files
against 53 real digests in `tracks/_meta/`. Keep this path, the runner, and the cadence glob in sync
— changing one alone re-opens the same hole.

> **Detail**: See `SKILL_detail.md §Output-Formats` — conversation output template, save path priority, fh_signal file format, env setup guide — read when executing Steps 3–4.

---

## Step 4. Chaining — Improvement Suggestion Connection

Immediately after Step 3 output, if **"FH Immediate Application Candidates"** has 1+ items, automatically suggest one of the following:

From chrono-meta/forge-harness, MIT.

More from chrono-meta/forge-harness