shiplog

name	shiplog
description	Recap of everything shipped since the last run - cross-repo PRs, security fixes, star deltas, and X traction, synthesized into a digest article and a ready-to-post shiplog in your voice.
metadata	{"title":"Shiplog","category":"core","var":"","requires":["XAI_API_KEY?","GH_GLOBAL?"],"tags":["content","social"]}

${var} — Optional, space-separated flags:

• since:YYYY-MM-DD — override the window start (default: when this skill last ran).
• days:N — window = last N days.
• dry-run — render to stdout; write no article, no state, no notify.
• owner/repo — narrow GitHub coverage to that one repo.
• any other word — focus/theme filter (a product name or keyword).

Empty = everything shipped since the last run, across all configured repos.

Produce two artifacts:

1. Digest — a themed, human-readable recap of everything that shipped + traction (the article).
2. Shiplog post — a tight, bulleted, ready-to-post version in the operator's voice, every project @-tagged (the notification).

This is cadence-agnostic: the window is always "since the last run" (memory/state/shiplog-last.json), so the aeon.yml schedule alone decides whether this is a daily, weekly, or on-demand recap. One skill, any frequency.

Read STRATEGY.md, memory/MEMORY.md, and the last 7 days of memory/logs/ for context. Read soul/SOUL.md + soul/STYLE.md before writing any output — the shiplog post must sound like the operator, not a changelog bot. If soul/ is empty, use a clear, direct, neutral voice (drop the signature flourishes below).

Config — all derived, nothing hardcoded

operator        = gh api user --jq .login            # the authenticated operator (PR-author search)
product_handles = memory/products.md `handles:` lines (@x)        # product X accounts to read
flagship_repos  = memory/products.md `repos:` tagged (public)     # the star / north-star story
watched_repos   = memory/watched-repos.md, else products.md repos: # everything shipped across
ecosystem_scouts= memory/products.md `scouts:` line (optional)    # recap accounts to scan for features
star_state      = memory/state/shiplog-stars.json    # snapshot for week-over-week star deltas

If neither watched-repos.md nor products.md yields a repo, exit SHIPLOG_NO_REPOS (notify + log, no article). Sections whose config is absent (X handles, scouts) are skipped gracefully, not failed.

Steps

1. Compute the window — "since last run"

STATE="memory/state/shiplog-last.json"
NOW=$(date -u +%Y-%m-%dT%H:%M:%SZ)
TODAY=$(date -u +%Y-%m-%d)
LAST=""
[ -f "$STATE" ] && LAST=$(jq -r '.last_run_at // empty' "$STATE" 2>/dev/null)
SINCE="${LAST:-$(date -u -d '7 days ago' +%Y-%m-%dT%H:%M:%SZ 2>/dev/null || date -u -v-7d +%Y-%m-%dT%H:%M:%SZ)}"
SINCE_DATE="${SINCE%%T*}"

• since:YYYY-MM-DD in ${var} → SINCE = that date at T00:00:00Z; days:N → N days ago. These override the state file.
• Use $SINCE for ALL time filtering — never substitute "since Monday" or other drift-prone shortcuts. The window is [$SINCE, $NOW); state the span ($SINCE_DATE → $TODAY) in the output.
• Idempotency is the state file (step 8 advances it each run, so windows never overlap). No once-per-day lock — a back-to-back re-run just yields an empty window → SHIPLOG_NOTHING_NEW. Write the digest to output/articles/shiplog-${TODAY}.md; if that name exists and there's genuinely new activity since the last run, use output/articles/shiplog-${TODAY}-2.md rather than clobbering.

2. GitHub activity (the bytes)

Cross-repo PR/commit visibility needs the global token — the built-in GITHUB_TOKEN only sees this repo. Prefer GH_GLOBAL when set:

GHT="${GH_GLOBAL:-$GITHUB_TOKEN}"   # gh reads GH_TOKEN from env; falls back to the repo-scoped token
OPERATOR=$(GH_TOKEN="$GHT" gh api user --jq .login 2>/dev/null)

Track success/failure per source in a sources map; on a single endpoint failure log fail and continue — never abort the whole skill.

a) Operator PRs across all repos in the window (grouped by repo + totals):

GH_TOKEN="$GHT" gh search prs --author "$OPERATOR" --created ">=$SINCE_DATE" \
  --json number,title,repository,state,createdAt,url --limit 100 \
  --jq 'group_by(.repository.nameWithOwner)[] | {repo: .[0].repository.nameWithOwner, count: length,
         prs: [.[] | {date: .createdAt[0:10], state, number, title}]}'

If 100 rows come back, note the result may be truncated.

b) Flagship headline numbers — for each flagship_repos entry, count commits + merged PRs in the window (the numbers the audience cares about):

for REPO in $FLAGSHIP_REPOS; do
  GH_TOKEN="$GHT" gh api "repos/${REPO}/commits" -X GET -f since="$SINCE" \
    --jq "\"$REPO commits: \" + ([.[] | .sha] | length | tostring)" 2>/dev/null
  GH_TOKEN="$GHT" gh api "repos/${REPO}/pulls" -X GET -f state=closed -f sort=updated -f direction=desc \
    --jq "\"$REPO merged PRs: \" + ([.[] | select(.merged_at != null and .merged_at > \"$SINCE\")] | length | tostring)" 2>/dev/null
done

c) The security flex — PRs the operator landed in repos they do NOT own (the "a project merged a fix from us" candidates). Filter the Step-2a result to external repos whose title matches a security keyword (security|ssrf|cve|credential|sandbox|escape|injection|vuln|redos|xss|toctou|path traversal|prototype pollution|deserial). Merged ones are the marquee story — if it's a named org (not a random fork), that's a headline bullet.

d) Star delta (north-star metric — flagships only, they're public):

mkdir -p memory/state
for REPO in $FLAGSHIP_REPOS; do
  GH_TOKEN="$GHT" gh api "repos/${REPO}" --jq '.stargazers_count'   # current total for $REPO
done

Read the prior snapshot memory/state/shiplog-stars.json (if present): delta = current_total − last_total per repo. After computing, overwrite the snapshot with { "<repo>": {"count": N, "date": "${TODAY}"}, ... }. If no prior snapshot exists, report totals only and note "no baseline yet — deltas start next run." Do NOT fabricate a delta.

3. X activity (direct X.AI curl — primary)

XAI_API_KEY is injected into this skill's environment (declared in requires:) and is the primary way to read X. For each X source below the primary fetch is a direct curl to https://api.x.ai/v1/responses with Authorization: Bearer {XAI_API_KEY}, using Grok's x_search tool. There is no network sandbox blocking this — just make the calls.

Check the key and give the call room first:

[ -n "$XAI_API_KEY" ] && echo KEY_PRESENT || echo KEY_UNSET

x_search runs a live X search and typically takes 30–120s. When you invoke the Bash tool for any curl below, set the tool's timeout to at least 180000 (180s); each curl carries --max-time 150 so it fails cleanly instead of hanging. A slow curl is not a missing key — never treat a timeout as key-unavailable. If KEY_PRESENT (it will be), Path A is required for every X source.

There are two X sources here — the operator handle ($OPERATOR_HANDLE, the operator's own X account from memory/products.md) and the product/project accounts ($PRODUCT_HANDLES). Fetch each independently into its own tmp file so one failing source can't clobber or sink another.

Path A — X.AI API (primary).

Operator posts (SRC=operator):

jq -n --arg h "$OPERATOR_HANDLE" --arg sd "$SINCE_DATE" --arg td "$TODAY" '{model:"grok-4.6", input:[{role:"user", content:("Search X for posts by @" + $h + " between " + $sd + " and " + $td + ". Return each post with full text, date, type (original|reply|RT — an RT text starts with \"RT @\"), exact engagement counts (likes, retweets, replies; 0 if unknown), and the direct link https://x.com/" + $h + "/status/ID. Return chronological.")}], tools:[{type:"x_search"}]}' > /tmp/xai-shiplog-operator-payload.json
HTTP=$(./secretcurl -s -o /tmp/xai-shiplog-operator.json -w '%{http_code}' --max-time 150 -X POST "https://api.x.ai/v1/responses" \
  -H "Content-Type: application/json" -H "Authorization: Bearer {XAI_API_KEY}" -d @/tmp/xai-shiplog-operator-payload.json)
echo "xai http=$HTTP bytes=$(wc -c </tmp/xai-shiplog-operator.json)"

Product/project accounts (SRC=projects):

jq -n --arg sd "$SINCE_DATE" --arg td "$TODAY" --arg ph "$PRODUCT_HANDLES" '{model:"grok-4.6", input:[{role:"user", content:("Search X for posts between " + $sd + " and " + $td + " from these accounts: " + $ph + ". Focus on launches, announcements, and any brag about a security fix merged into another project. For each: @handle, full text, date, exact engagement counts (likes, retweets, replies; 0 if unknown), and the direct link https://x.com/handle/status/ID. Skip retweets of others.")}], tools:[{type:"x_search"}]}' > /tmp/xai-shiplog-projects-payload.json
HTTP=$(./secretcurl -s -o /tmp/xai-shiplog-projects.json -w '%{http_code}' --max-time 150 -X POST "https://api.x.ai/v1/responses" \
  -H "Content-Type: application/json" -H "Authorization: Bearer {XAI_API_KEY}" -d @/tmp/xai-shiplog-projects-payload.json)
echo "xai http=$HTTP bytes=$(wc -c </tmp/xai-shiplog-projects.json)"

For each source, on HTTP=200 with a non-empty body, parse that source's file with the standard extractor and mark x_source=api:

jq -r '.output[] | select(.type == "message") | .content[] | select(.type == "output_text") | .text' /tmp/xai-shiplog-<SRC>.json

From the operator text, separate original posts from RTs (RT text starts with RT @) — RTs are amplification, not ships. From the projects text, note the bangers (sort by likes/views) — one or two feed the digest's narrative section.

On a real failure, skip that source — never fabricate posts. If a source's curl returns non-200, an empty/unparseable body, or times out, record the true reason for that source and continue with whatever other sources succeeded. Reason codes: key-unset (only if the check above printed KEY_UNSET), http-<code> (non-2xx), empty (200 but no posts parsed), timeout (exceeded --max-time). Never write "XAI_API_KEY unavailable" when the key was set.

Path B — WebFetch last resort (per source, optional). Only if a source's Path A failed for one of the real reasons above: WebFetch that source's public https://x.com/<handle> profile(s) — no auth — and mark that source x_source=webfetch (lower quality; prefer posts inside the window). If every X source fails both paths, set x_source=none and write the GitHub-only shiplog (note the gap) — never abort.

4. Ecosystem + traction sweep (best-effort — skip gracefully)

• Ecosystem mentions — only if ecosystem_scouts (scouts:) is configured. Fetch with the same Path A X.AI curl as Step 3, into its own tmp file (SRC=ecosystem):

  jq -n --arg sd "$SINCE_DATE" --arg td "$TODAY" --arg es "$ECOSYSTEM_SCOUTS" --arg ph "$PRODUCT_HANDLES" '{model:"grok-4.6", input:[{role:"user", content:("Search X between " + $sd + " and " + $td + " for posts from these recap/scout accounts: " + $es + " that mention any of these products: " + $ph + ". Return each mention with @handle, follower_count, full text, date, and the direct link https://x.com/handle/status/ID — recaps, rankings, partner shares.")}], tools:[{type:"x_search"}]}' > /tmp/xai-shiplog-ecosystem-payload.json
  HTTP=$(./secretcurl -s -o /tmp/xai-shiplog-ecosystem.json -w '%{http_code}' --max-time 150 -X POST "https://api.x.ai/v1/responses" \
    -H "Content-Type: application/json" -H "Authorization: Bearer {XAI_API_KEY}" -d @/tmp/xai-shiplog-ecosystem-payload.json)
  echo "xai http=$HTTP bytes=$(wc -c </tmp/xai-shiplog-ecosystem.json)"
  

  On HTTP=200 + non-empty, parse /tmp/xai-shiplog-ecosystem.json with the standard extractor. On a real curl failure (http-<code> / empty / timeout, or key-unset) skip this source with the true reason — never fabricate a mention. Confirm any handle is real before @-mentioning (a wrong tag in a public post is worse than none). Capture follower counts for the flex ("featured by @X (Nk)"). Skip entirely if no scouts: configured.
• Product traction (OpenRouter / x402 / analytics) — only if a source is configured for the product. If you have an app/server id, WebFetch its page; otherwise say "no product-traction sources wired yet" and move on. Keep any number exactly as measured — don't round 79 → ~80.