Fork Fleet

${var} — Divergence scope selector; space-separated tokens, order-independent, all optional:

• scope (code | config | both, default both) — which divergence dimension to run.
• repo=owner/name — override the parent repo whose forks are scanned (else auto-resolved).
• fork=owner/name — drill into a single fork (forces code scope; config math needs a fleet).

Empty ⇒ both dimensions over the auto-resolved parent. Examples: `` (both, all forks) · code · config · config repo=octo/aeon · fork=alice/aeon.

Today is ${today}. This is the fleet's divergence monitor. It answers two questions the popularity/liveness skills don't:

1. Code divergence — which active forks are building real work (unique commits, new/modified skills) that's worth pulling back upstream?
2. Config divergence — where does the configured fleet systematically disagree with upstream's enabled / var / model / schedule defaults, so the operator can flip a default the fleet has already voted on?

skill-gap ranks what's popular (top 15 by enabled count). This skill's code branch surfaces per-fork unique work; its config branch surfaces where operators disagree with defaults. If 6 of 8 configured forks enable a skill upstream defaults off, upstream is shipping the wrong default; if 5 of 8 disable a skill upstream defaults on, that skill is noise. Both are peer-learning signals.

Operating principles

• Verdict first, catalog second. The operator reads one line and knows if action is needed.
• Silent when nothing changed. Weekly cadence + a dormant/undivergent fleet = a read-once habit to kill. A clean run notifies nothing.
• Per-fork compare is one call, not three. /compare/{owner}:main...{fork_owner}:main returns ahead/behind/unique commits/files in a single round-trip; the recursive git-tree returns the fork's whole file list in one call.
• Substance ≠ noise. A new skills/*/SKILL.md is worth 100 cron-time edits in aeon.yml. Score accordingly. On the config side, an untouched template fork is not a "vote" — exclude it from divergence math.

Shared setup (all scopes)

S0. Bootstrap + load state

mkdir -p memory/topics
[ -f memory/instances.json ] || echo '{}' > memory/instances.json
[ -f memory/topics/fork-fleet-state.json ] || echo '{"forks":{},"last_run":null}' > memory/topics/fork-fleet-state.json
[ -f memory/topics/fork-digest-state.json ] || echo '{"last_run":null}' > memory/topics/fork-digest-state.json

Read memory/MEMORY.md for high-level context and scan the last ~3 days of memory/logs/ — drop anything already reported so a weekly signal isn't re-sent.

• Read memory/instances.json → the set of repo full_names that are managed instances (tagged separately from organic community forks in the report).
• Read memory/topics/fork-fleet-state.json → prior run's per-fork {pushed_at, ahead_by, default_branch, new_skill_count} keyed by full_name. Used for the code what-changed delta.
• Read memory/topics/fork-digest-state.json → prior config-divergence snapshot (schema in step B8). Used for the config week-over-week delta.

S1. Parse the scope selector

Parse ${var} into tokens:

• SCOPE = code, config, or both (default both if no scope keyword present).
• REPO_OVERRIDE = value of a repo=owner/name token, if any.
• SINGLE_FORK = value of a fork=owner/name token, if any. If SINGLE_FORK is set, force SCOPE=code (single-fork config divergence is meaningless — the config math needs a fleet of ≥2 configured forks).

S2. Resolve the parent/target repo

Resolve PARENT_REPO in priority order:

1. REPO_OVERRIDE if the repo= token was given.
2. Else auto-resolve from this running instance:

   PARENT_REPO=$(gh api repos/$(gh repo view --json nameWithOwner -q .nameWithOwner) --jq '.parent.full_name // .full_name')

3. Else fall back to the first non-comment, non-empty line of memory/watched-repos.md.

If none resolves, write status FORK_DIVERGENCE_NO_TARGET to memory/logs/${today}.md and stop (no notification).

PARENT_NAME="${PARENT_REPO##*/}"
PARENT_OWNER="${PARENT_REPO%%/*}"
PARENT_DEFAULT_BRANCH=$(gh api "repos/${PARENT_REPO}" --jq '.default_branch')

S3. List + classify forks (single call, shared by both branches)

One paginated listing — includes default_branch, archived, disabled, pushed_at, stars, description:

gh api "repos/${PARENT_REPO}/forks" --paginate \
  --jq '[.[] | {full_name, owner: .owner.login, default_branch, pushed_at, pushed_at_epoch: (.pushed_at | fromdateiso8601), stargazers_count, open_issues_count, archived, disabled, description}]'

Skip archived=true or disabled=true. Retain the rest as the total fork population (N_TOTAL). Classify by activity window:

• Active = pushed_at within last 30 days.
• Stale = 30–365 days.
• Dormant = >365 days or never pushed after creation.

The Active set (pushed in the last 30 days) is the shared working set for both branches — this matches the config branch's original 30-day cutoff exactly.

• If SINGLE_FORK is set: filter to that one fork, treat it as active, and skip the classification math.
• If zero active forks (and SINGLE_FORK unset): both branches short-circuit. If there is also no code-side state change (no new forks, none flipped active↔stale vs prior fork-fleet-state.json), write status FORK_DIVERGENCE_QUIET to the log, update the state files' last_run, send no notification, and stop.

Cap active-fork deep processing at 50 per run — if more, rank by pushed_at_epoch desc and trim (log truncated_at=50).

Now dispatch: run Branch A if SCOPE ∈ {code, both}, Branch B if SCOPE ∈ {config, both}.

Branch A — Code divergence (runs when SCOPE ∈ {code, both})

A1. Per-fork compare (one call each)

For each active fork, call cross-repo compare using the fork's own default_branch and full_name (absorbs any repo-rename drift):

gh api "repos/${PARENT_REPO}/compare/${PARENT_OWNER}:${PARENT_DEFAULT_BRANCH}...${FORK_OWNER}:${FORK_DEFAULT_BRANCH}" \
  --jq '{ahead_by, behind_by, status, files: [.files[]? | {filename, status, additions, deletions}], commits: [.commits[]? | {sha: .sha[0:7], msg: .commit.message | split("\n")[0], author: .commit.author.name, date: .commit.author.date}]}'

On 404 (branch missing / fork emptied): mark fork UNREADABLE and continue. On 429: sleep 60s, retry once. On 5xx: sleep 10s, retry once. On persistent fail: mark API_FAIL for that fork.

Cross-repo compare returns unique fork commits (commits) and changed files (up to 300) in one shot — no separate /commits calls needed.

A2. Classify divergence signals per fork

From the files array, tag each fork:

• New skills: files with status=added under skills/*/SKILL.md
• Modified skills: status=modified under skills/*/SKILL.md
• Custom schedule: any change to aeon.yml
• Modified dashboard: any change under apps/dashboard/
• Custom notify: change to notify or notify-jsonrender
• New content: additions under output/articles/ or memory/topics/
• Config changes: changes to CLAUDE.md, .github/, bin/, or root scripts/
• Workflow changes: changes under .github/workflows/

A3. Score each fork (substance-weighted)

score =  10 × (new skill files)
       +  4 × (modified skill files)
       +  2 × min(unique_commits, 15)
       +  3 × (new content files, capped at 5)
       +  2 × (workflow/config files, capped at 3)
       +  1 × (custom-schedule flag)
       +  1 × stargazers

Sort active forks by score descending. Flag any fork with ≥1 new skill file as a PROMOTE candidate; ≥3 unique commits OR ≥1 modified skill as REVIEW; otherwise NOTE.

A4. Deep-read top upstream candidates

For every PROMOTE fork (capped at 5), fetch each unique skill's SKILL.md from the fork's default branch:

gh api "repos/${FORK_FULL_NAME}/contents/${SKILL_PATH}?ref=${FORK_DEFAULT_BRANCH}" --jq '.content' | base64 -d

On failure fall back to the file-tree listing and note "could not read content". Synthesize each unique skill into a 1–2 sentence description of what it does. Do not deep-read REVIEW or NOTE forks (output stays actionable).

A5. Compute week-over-week delta (code)

Compare the current active-fork set to prior fork-fleet-state.json:

• NEW_FORK: full_name absent from prior state
• NEW_ACTIVE: was stale/dormant, now active
• WENT_STALE: was active, now stale/dormant
• NEW_SKILLS: active in both snapshots, new_skill_count increased
• GONE: archived / deleted since prior run

A6. Pick the code verdict

One line. Priority order:

1. NEW UPSTREAM CANDIDATE: {fork} — if ≥1 PROMOTE fork has ≥1 new skill not present in prior state
2. ACTIVE FLEET: {N} forks building — if ≥3 PROMOTE+REVIEW combined
3. FLEET STIRRING: {N} new active — if ≥2 NEW_FORK or NEW_ACTIVE
4. HOLDING PATTERN: {N} active, no new work — active forks present but nothing crossed REVIEW
5. DORMANT: no active forks — shouldn't reach notify (S3 gates it); included for the log-only path

A7. Build the code-divergence article part

Assemble this block (it becomes Part 1 of the combined article in the final section):

## What changed this week
- **New forks**: [list or "none"]
- **Went active**: [list or "none"]
- **New skills landed**: [fork → skill names, or "none"]
- **Went stale**: [list or "none"]
- **Archived/deleted**: [list or "none"]
(Omit the entire section if every bucket is empty.)

## PROMOTE — upstream contribution candidates

### {fork_full_name} — score N [MANAGED | COMMUNITY]
**Activity:** last pushed YYYY-MM-DD · stars N · +N/-M commits vs upstream
**Unique skills:**
- `skills/foo/SKILL.md` — {one-line synthesis of what it does, from deep-read}
- `skills/bar/SKILL.md` — {synthesis}

**Why promote:** {1–2 sentence take — what this skill does that upstream lacks, and whether it's generalizable}
**Suggested action:** Open a PR cherry-picking `skills/foo/` (or reach out to {owner} to upstream themselves).

(Repeat for each PROMOTE fork, capped at 5. If PROMOTE is empty: "No upstream candidates this week.")

## REVIEW — worth a look

| Fork | Score | Ahead | New/Modified | Notable |
|------|-------|-------|--------------|---------|
| owner/repo | N | +N/-M | 0/2 | dashboard rewrite, custom notify |

(Omit if empty.)

## NOTE — low divergence

Terse one-liner per fork: `owner/repo (+N/-M, schedule tweak only)`. Collapse into a count if >5 entries. Omit if empty.

## Fleet vs community

| Category | Count |
|----------|-------|
| Managed instances | N |
| Community forks | N |
| Stale (30-365d) | N |
| Dormant (>365d) | N |

## Code source status
`forks_list=ok|fail · compare_ok=N/M · deep_read=N/M · rate_limit_retries=N · unreadable=N`

If PROMOTE has >5 forks, keep only the top 5 by score; list the rest in REVIEW.

A8. Update code state

Write memory/topics/fork-fleet-state.json:

{
  "last_run": "${today}",
  "last_status": "FORK_FLEET_OK",
  "parent_repo": "owner/repo",
  "forks": {
    "owner/repo": {
      "pushed_at": "YYYY-MM-DD...",
      "default_branch": "main",
      "ahead_by": N,
      "behind_by": N,
      "new_skill_count": N,
      "score": N,
      "tier": "PROMOTE|REVIEW|NOTE|UNREADABLE|API_FAIL",
      "unique_skills": ["skills/foo/SKILL.md", "..."]
    }
  }
}

A9. Set the code branch status

Branch B — Config divergence (runs when SCOPE ∈ {config, both})

B1. Snapshot upstream defaults

Read this running instance's local aeon.yml once. Build (these are baselines — never mutated):

• UPSTREAM_DEFAULTS: dict {skill_name -> {enabled: bool, model: str|null, var: str, schedule: str|null}} for every skill entry under skills:.
• UPSTREAM_SKILLS: set of skill directory names from skills/ (use ls skills/).
• UPSTREAM_TAGS: dict {skill_name -> [tags]} parsed from each skills/<name>/SKILL.md frontmatter (best-effort; missing frontmatter → []).

B2. Per-fork enumeration (one tree call + one yml fetch each)

Operate over the active-fork set from shared step S3 (already filtered to forks pushed in the last 30 days — the config branch's original cutoff). For each active fork, run one recursive git-tree call to enumerate files (cheaper than per-path contents):

gh api "repos/${FORK_FULL}/git/trees/HEAD?recursive=1" --jq '[.tree[] | select(.type == "blob") | .path]'

Then fetch the fork's aeon.yml only if the tree contains it:

gh api "repos/${FORK_FULL}/contents/aeon.yml?ref=${FORK_DEFAULT_BRANCH}" --jq '.content' | base64 -d

Error handling:

• 404 / 409 (empty repo): mark status: "no_tree", skip aeon.yml extraction, continue.
• 403 with X-RateLimit-Remaining: 0: sleep 60s, retry once. If still failing, mark status: "rate_limited" and continue.
• Tree contains aeon.yml but the contents call 404s: mark status: "yml_unreadable", continue.
• aeon.yml present but YAML parse fails: mark status: "yml_invalid", continue.

For each readable aeon.yml, extract per-skill {enabled, model, var, schedule}. Treat missing keys as inheriting the upstream default (do not count those as overrides).

Detect fork-only skills: directory names matching skills/<name>/SKILL.md in the fork's tree where <name> is NOT in UPSTREAM_SKILLS. Record {fork_full_name, skill_name, path} for each.

B3. Tier each fork

Compute a divergence signal vector vs UPSTREAM_DEFAULTS:

• enabled_diff: count of skills where the fork's enabled differs from upstream
• var_overrides: count of skills with non-empty var: where upstream's was empty (or a different non-empty value)
• model_overrides: count of skills with model: differing from upstream
• schedule_overrides: count of skills with schedule: differing from upstream
• fork_only_skill_count: count from step B2

Tier the fork:

• CONFIGURED: any signal ≥1 (the fork actively diverged)
• TEMPLATE: aeon.yml readable but every signal is 0 — excluded from divergence math
• UNREADABLE: no_tree / no aeon.yml / yml_unreadable / yml_invalid / rate_limited — tracked in the source-status footer

Let N_CONFIGURED = count of forks tiered CONFIGURED. If N_CONFIGURED < 2: the config branch cannot produce meaningful divergence math. Set config status FORK_SKILL_DIGEST_TEMPLATE_FLEET, record active/template/unreadable counts, emit a stub Part 2 noting the conversion rate, and contribute no config notify signal. Skip steps B4–B6.

B4. Aggregate divergence (the core config analysis)

For each skill name in UPSTREAM_SKILLS, compute four dimensions:

Enable divergence:

• forks_enabled_count: number of CONFIGURED forks with enabled: true for this skill
• forks_disabled_count: number of CONFIGURED forks with enabled: false (explicitly set, not inherited)
• upstream_enabled: bool from UPSTREAM_DEFAULTS
• divergence_pct:
  • If upstream enabled: false: forks_enabled_count / N_CONFIGURED (how many disagree by enabling)
  • If upstream enabled: true: forks_disabled_count / N_CONFIGURED (how many disagree by disabling)
• direction: "ENABLE_UPWARD" (upstream off, forks turn on) or "DISABLE_DOWNWARD" (upstream on, forks turn off)

Var divergence:

• var_override_count: number of CONFIGURED forks where var: differs from upstream
• top_var_value: most common non-empty fork value (with count) — only if ≥2 forks share it

Model divergence:

• model_override_count: number of forks with non-null model differing from upstream
• top_model_value: most common fork model (with count) — only if ≥2 forks share it (signals fleet consensus on a cheaper/different model)

Schedule divergence:

• schedule_override_count: number of forks with schedule differing from upstream
• top_schedule_value: most common fork schedule (with count) — only if ≥2 forks share it