/source-health

name: source-health
description: >-
  Read-only audit of brief source performance. Scans recent brief frontmatter
  (sources_used, sources_skipped, sources_quiet) for 3+ consecutive quiet or skipped
  streaks, plus a fetch_log-based cost-efficiency analysis. Never modifies feed-sources.md.
  Invoke with /hermitd-feed:source-health [--last N].

Source Health

Read-only audit of brief source performance. Scans recent briefs and flags underperforming sources before they silently waste fetch budget.

When to use

• Manually, to spot-check source health.
• Wired into weekly review to catch slow-building problems.
• After adding a new source, to confirm it's yielding items.

Steps

1. **Read `feed-sources.md`** — extract the active source names (Name column).

2. **Glob briefs** — list `.hermit/briefs/*.md`, sort descending by filename (newest first), take the last **N = 10** (override with `--last N`).

3. **Parse each brief's frontmatter** — extract:

• `sources_used`: sources that contributed ≥1 item.
• `sources_skipped`: sources that failed to fetch (network error, Chrome unavailable, script error).
• `sources_quiet`: sources that fetched cleanly but yielded 0 items.
• `fetch_log`: per-source efficiency array (may be absent in older briefs — treat as missing, not an error).

If a source appears in none of the three arrays, mark it `unknown` for that brief (may predate the schema).

4. **Compute per-source streaks** — walking newest→oldest for each active source:

• `consecutive_quiet`: consecutive briefs the source was in `sources_quiet`.
• `consecutive_skipped`: consecutive briefs the source was in `sources_skipped`.
• Reset the streak when the source appears in `sources_used` or is `unknown`.

5. **Flag threshold violations** — any source where `consecutive_quiet >= 3` OR `consecutive_skipped >= 3`.

6. **Output a health table:**

   ## Source Health — YYYY-MM-DD (last 10 briefs)

   | Source                | Last 10 (newest→oldest)    | Streak       | Status    |
   |-----------------------|----------------------------|--------------|-----------|
   | Source A              | ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓      | —            | healthy   |
   | Source B              | ~ ~ ~ ✓ ~ ~ ~ ✓ ~ ~       | 3 quiet      | ⚠ warn    |
   | Source C              | ✗ ✗ ✗ ✗ ✓ — — — — —      | 4 skipped    | ✗ flag    |

Legend: `✓` contributed · `~` quiet (fetched, nothing relevant) · `✗` skipped (fetch failed) · `—` not yet in sources or brief predates schema.

Status thresholds:

• `healthy` — in `sources_used` in the last 3 briefs.
• `⚠ warn` — quiet 2 consecutive times (not yet flagged, worth watching).
• `✗ flag` — quiet **or** skipped 3+ consecutive times → needs operator review.

7. **Summary line** — after the table:

   Flagged: N source(s) need review. Quiet: M. Skipped: K.

If nothing flagged: `All N sources healthy in the last 10 briefs.`

8. **Recommendations** — one line per flagged source:

• Quiet 3+: "Consider removing or replacing — consuming a fetch slot with zero yield."
• Skipped 3+: "Check source type — may need Chrome (use `x`/`chrome` type) or the URL has changed."

9. **Efficiency Analysis** — reads `fetch_log` from recent frontmatter for per-source cost efficiency. Only runs when at least 3 recent briefs have `fetch_log` data; if fewer, skip entirely (no output).

a. **Load fetch_log data** from step 3. Group by source name; discard briefs where `fetch_log` is absent.

b. **Per-source aggregates** (across all briefs where the source appears in fetch_log):

• `avg_tokens`: mean of `tokens_approx`.
• `avg_yielded`: mean of `items_yielded`.
• `avg_p1`: mean of `items_scored_p1`.
• `yield_ratio`: `avg_yielded / (avg_tokens / 1000)` — items per 1K tokens (higher = more efficient).
• `p1_rate`: `avg_p1 / avg_yielded` when `avg_yielded > 0`, else `—` (dash).
• `briefs_with_data`: count of contributing briefs.

c. **Flag type-upgrade candidates.** Sources where `tokens_approx ≥ 15000` (Chrome-tier cost) AND `yield_ratio < 0.15` → "type-upgrade candidate" (low yield at Chrome cost suggests checking for an RSS feed).

d. **Output efficiency table** — append after the streak table:

      ## Efficiency Analysis (last N briefs with fetch_log)

      | Source      | Type    | Avg tokens | Yield/1K | P1 rate | Verdict          |
      |-------------|---------|------------|----------|---------|------------------|
      | Source A    | rss     | 3K         | 2.3      | 29%     | ✓ efficient      |
      | Source B    | reddit  | 5K         | 1.8      | 33%     | ✓ efficient      |
      | Source C    | chrome  | 20K        | 0.05     | 100%    | ⚠ low yield → RSS? |
      | Source D    | x       | 20K        | 0.0      | —       | ✗ no yield       |

Verdict thresholds:

• `✓ efficient` — yield_ratio ≥ 0.5 (web/rss) or ≥ 0.15 (chrome/x/reddit-chrome).
• `⚠ low yield` — below threshold but still contributing some items.
• `✗ no yield` — avg_yielded < 0.1.
• `→ RSS?` — append to the verdict for type-upgrade candidates.

Summary line:

      Efficiency: N sources efficient · M low-yield · K type-upgrade candidates

Output

Print the tables to the conversation. Do NOT write to any file — this skill is read-only and never modifies `feed-sources.md`. Source removals require operator approval.