Aeon Update

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

• mode (sync | report, default sync) — sync opens a PR with the framework changes; report computes the delta and notifies, mutating nothing (dry run).
• repo=owner/name — override the upstream source repo (else auto-resolved from this instance's parent, falling back to aeonfun/aeon).
• reset=<sha|fork-point> — force the stored baseline to <sha> (or the merge-base with upstream) before running. Recovery / backfill lever.

Empty ⇒ sync from the auto-resolved upstream. Examples: `` · report · repo=aeonfun/aeon · reset=fork-point.

Today is ${today}. This is the fleet's downstream updater - the counterpart to fork-fleet. fork-fleet looks outward from the parent to find work in the forks worth pulling up; this skill runs inside an instance and pulls the parent's shipped framework changes down - new skills, script/harness fixes, workflow and doc updates - and lands them as a reviewable PR. It is how an instance stays current with aeonfun/aeon without a hand-run rsync-overlay rebase.

Operating principles

• PR, never push to main. Every framework change ships as one reviewable PR. The operator merges.
• Never clobber operator config. aeon.yml, STRATEGY.md, soul/, memory/, output/, .mcp.json and the git-derived catalogs are instance-owned. Upstream changes to them are surfaced for manual review in the PR body, never written into the tree.
• 3-way, not blind overwrite. A framework file the operator has already customized is auto-merged when its local edits and upstream's edits touch disjoint regions (a real git merge-file 3-way, S6); it is only listed as a conflict for a human when the same lines changed on both sides. This is what lets a hand-narrowed workflow keep receiving unrelated upstream fixes without a manual merge every run.
• Silent when in sync. Baseline == upstream HEAD ⇒ nothing to do, no notification.
• The baseline is the watermark, and it advances by merge. The new baseline SHA is written into the PR branch, so the watermark only moves when the operator merges the PR. Unresolved conflicts are tracked separately so advancing the baseline can never silently drop them.

Steps

S0. Bootstrap + load state

mkdir -p memory/topics
[ -f memory/topics/aeon-update-state.json ] || echo '{"baseline_sha":null,"upstream":null,"last_run":null,"last_pr":null,"pending_conflicts":[]}' > memory/topics/aeon-update-state.json

Read memory/MEMORY.md for context and scan the last ~3 days of memory/logs/ - drop anything already reported so a repeat run isn't re-sent. Read the state file:

• BASELINE = .baseline_sha (the upstream commit this instance was last synced to).
• PENDING = .pending_conflicts (files surfaced as conflicts in a prior run, not yet resolved).

S1. Parse ${var}

• MODE = report if the token report (or dry) is present, else sync.
• REPO_OVERRIDE = value of a repo=owner/name token, if any.
• RESET = value of a reset= token, if any (fork-point or a 7-40 char SHA).

S2. Resolve the upstream source

SELF=$(gh repo view --json nameWithOwner -q .nameWithOwner)
UPSTREAM="${REPO_OVERRIDE:-$(gh api "repos/${SELF}" --jq '.parent.full_name // empty')}"
[ -z "$UPSTREAM" ] && UPSTREAM="aeonfun/aeon"

If UPSTREAM == SELF, this instance is canon - there is nothing upstream to pull. Write status AEON_UPDATE_IS_UPSTREAM to memory/logs/${today}.md, send no notification, and stop.

UP_DEFAULT=$(gh api "repos/${UPSTREAM}" --jq '.default_branch')
HEAD_SHA=$(gh api "repos/${UPSTREAM}/commits/${UP_DEFAULT}" --jq '.sha')

S3. Establish / reset the baseline

• reset=fork-point → BASELINE=$(gh api "repos/${UPSTREAM}/compare/${HEAD_SHA}...$(git rev-parse HEAD)" --jq '.merge_base_commit.sha'). reset=<sha> → BASELINE=<sha>. Persist immediately to state, then continue.
• First run (BASELINE null and no reset): a fresh instance already carries all of canon from fork time, so there is no delta to apply - just anchor the watermark. Set baseline_sha = HEAD_SHA, write state, log AEON_UPDATE_BASELINE_SET, send a one-line notify (baseline initialized at <head7>; future runs sync from here), and stop. (To backfill everything since the fork point instead, re-run with reset=fork-point.)
• BASELINE == HEAD_SHA: in sync. Re-verify PENDING (S8) in case a prior conflict is now resolved, update state, log AEON_UPDATE_IN_SYNC, notify nothing, stop.

S4. Compare baseline → HEAD

gh api "repos/${UPSTREAM}/compare/${BASELINE}...${HEAD_SHA}" --jq '{
  ahead: .ahead_by, behind: .behind_by, status,
  commits: [.commits[]? | {sha: .sha[0:7], msg: (.commit.message | split("\n")[0]), date: .commit.author.date}],
  files:   [.files[]?   | {filename, status, previous_filename, additions, deletions}]
}' > /tmp/aeon-update-compare.json

Error handling:

• 404 / status: "diverged" with no merge base (baseline not an ancestor of HEAD - history rewrite or unrelated repo): stop with AEON_UPDATE_BASELINE_UNREACHABLE; notify the operator to re-run with reset=fork-point or reset=<sha>.
• Cross-repo compare returns at most 300 files; if .files looks truncated, note files_truncated=true in the report - the operator can run again after merging to pick up the remainder.

S5. Partition changed files

Classify every entry in .files by path. A file is OPERATOR-owned (surfaced, never auto-written) if its path matches any of:

aeon.yml            STRATEGY.md         soul/**             memory/**
output/**           .mcp.json           .env*               aeon.db
skills.lock         eyebrowlock.json    catalog/*.json      .claude/**  (except .claude/skills/aeon/**)
apps/dashboard/outputs/**

Everything else is OWNED (a candidate for auto-apply): skills/**, scripts/**, bin/**, harness-adapter/**, .github/**, apps/** (except apps/dashboard/outputs/**), CLAUDE.md, AGENTS.md, docs/**, .github/README.md, LICENSE, CHANGELOG.md, .gitignore, eyebrow.policy.json, and tracked root helpers (aeon, ...).

catalog/*.json and eyebrowlock.json are OPERATOR-owned here only so they are never blindly copied - they are regenerated from the synced sources in S7, which is the correct way to reconcile them.

S6. 3-way classify each OWNED file

Set up a workspace and, for each OWNED file f, fetch upstream's HEAD and BASELINE blobs:

WORK=$(mktemp -d)
fetch() { gh api "repos/${UPSTREAM}/contents/$1?ref=$2" --jq '.content' 2>/dev/null | base64 -d; }   # $1=path $2=ref
h() { sha256sum 2>/dev/null | cut -d' ' -f1; }

Decide f's disposition from its status and a content 3-way (local-current vs upstream@BASELINE vs upstream@HEAD):

Never CLEAN-DELETE a skills/<name>/ directory whose <name> is not present in upstream's tree - fork-only skills are operator work and are structurally untouched (upstream's compare can only reference paths that exist upstream).

Also never CLEAN-DELETE a skills/<name>/ directory if <name> is currently enabled: true in the operator's aeon.yml (grep -E "^ ${name}: *\{[^}]*enabled: true" aeon.yml) - upstream retiring a skill the operator has actively scheduled is exactly the case validate-config.js's skill-refs check exists to catch, but only after the PR is merged; nothing in the PR review itself would otherwise flag it. Downgrade this case to CONFLICT (reason: enabled-skill-removed-upstream) instead of deleting - the directory stays, and S9 surfaces it as its own loud PR-body section rather than folding it into "Applied cleanly" or the generic conflict list.

3-way content merge (OWNED files only). A modified file that reached the otherwise row means the operator diverged from BASELINE and upstream changed the file too. Do not give up on it - most of the time the two sets of edits are in different parts of the file (e.g. the operator narrowed the workflow's env: secrets block while upstream bumped a timeout and a retry loop elsewhere), and a real 3-way merge combines both losslessly. Attempt it before declaring a conflict:

fetch "$f" "$BASELINE"  > "$WORK/base"    # upstream@BASELINE (the common ancestor)
fetch "$f" "$HEAD_SHA"  > "$WORK/head"    # upstream@HEAD (what to bring in)
cp "$f" "$WORK/local"                      # operator's current copy (ours)
if git merge-file -p --diff3 "$WORK/local" "$WORK/base" "$WORK/head" > "$WORK/merged.$$" 2>/dev/null; then
  disposition=CLEAN-MERGE   # exit 0 = disjoint hunks; the merged file carries BOTH edits - write it in S7
else
  disposition=CONFLICT      # exit >0 = the same lines changed on both sides; surface for a human as before
fi

git merge-file exits 0 only when the merge is clean (operator and upstream touched disjoint regions); the merged output preserves the operator's customization AND applies upstream's change. A non-zero exit means a genuine overlap - keep it a CONFLICT and list it with the upstream diff (S9). Only OWNED files are ever 3-way-merged; OPERATOR-owned paths are always surfaced, never written. A merged file gets the same S7 YAML/JSON parse-check as any written file - if the merge produced something that no longer parses, abort that file back to CONFLICT rather than committing it.

S7. Apply CLEAN changes on a branch (sync mode only)

If MODE == report, skip to S9. Otherwise:

BR="aeon-update/sync-$(echo "$HEAD_SHA" | cut -c1-7)"
git checkout -b "$BR"

Write every CLEAN-ADD / CLEAN-UPDATE (mkdir -p "$(dirname f)" then write the HEAD blob to f), write every CLEAN-MERGE (the merged file $WORK/merged.$$ from S6, over the existing f), and git rm every CLEAN-DELETE. CONFLICT and OPERATOR files are not touched - they go in the PR body only. (CLEAN-MERGE counts as a clean apply for the "nothing CLEAN applied" test below.)

If any skills/** path was applied, regenerate the derived catalogs from the synced sources (never copy them from upstream):

bin/generate-skills-json && bin/generate-packs-json && bin/generate-skill-icons
node scripts/gen-agents-md.js || true

Refresh the eyebrow integrity lock for any NEWLY-ADDED skill. ci-skill-integrity fails a PR when a present skill's skills/<slug>/SKILL.md has no "discoveredFrom": "skills/<slug>/SKILL.md" entry in eyebrowlock.json. That entry is produced only by the eyebrow binary, which is not preinstalled in this run - so a CLEAN-ADD of a new skill would otherwise land the PR CI-red. Fetch the binary at the exact version ci-skill-integrity.yml pins, parsed from that workflow's alexverify/eyebrow/action@<sha> # vX.Y.Z line so it can never drift from CI - a hardcoded version writes a lock that CI's (newer) eyebrow verify then rejects as drift, which is the recurring cause of red sync PRs. Verify the downloaded tarball against the release's own checksums.txt before running it (the same checksum-verified install the action does), then rescan:

EYEBROW_OK=0
EB=$(command -v eyebrow || true)
if [ -z "$EB" ]; then
  # Use the SAME eyebrow version ci-skill-integrity.yml pins, parsed from the
  # workflow's `alexverify/eyebrow/action@<sha> # vX.Y.Z` comment. A hardcoded
  # version silently drifts from CI (the lock this writes with an older binary
  # then fails CI's newer `eyebrow verify` as drift - the recurring sync red);
  # deriving it self-heals across action bumps. Fall back to v0.4.2 if unparsable.
  EBV=$(grep -oE 'alexverify/eyebrow/action@[0-9a-f]+ *# *v[0-9]+\.[0-9]+\.[0-9]+' .github/workflows/ci-skill-integrity.yml | grep -oE 'v[0-9]+\.[0-9]+\.[0-9]+' | head -1)
  EBV=${EBV:-v0.4.2}
  # Linux runner (ubuntu-latest) assumed; unknown arch => skip to the fail-safe.
  case "$(uname -m)" in
    x86_64)        A=amd64 ;;
    aarch64|arm64) A=arm64 ;;
    *)             A= ;;
  esac
  TB="eyebrow_${EBV#v}_linux_${A}.tar.gz"
  # Download the tarball + the release's checksums.txt and verify the tarball
  # against it BEFORE extract or exec (mirrors the action's checksum-verified
  # install). A tampered or moved asset fails the check, so we do NOT run it and
  # fall through to the fail-safe. --ignore-missing checks only the asset present.
  if [ -n "$A" ] && gh release download "$EBV" -R alexverify/eyebrow -p "$TB" -p checksums.txt -D "$WORK/eb" 2>/dev/null; then
    if (cd "$WORK/eb" && shasum -a 256 --check --ignore-missing --strict checksums.txt >/dev/null 2>&1); then
      tar xzf "$WORK/eb/$TB" -C "$WORK/eb" 2>/dev/null \
        && EB=$(find "$WORK/eb" -type f -name eyebrow | head -1) && chmod +x "$EB" 2>/dev/null || true
    else
      echo "::warning::eyebrow $TB failed checksums.txt verification - not executing"; EB=
    fi
  fi
fi
# Run with a SCRUBBED env (allowlist PATH+HOME only). eyebrow scan is a local
# file-hasher - it needs no secrets and no network - so denying it the run's
# secret env (GH_GLOBAL + provider/notify keys) means even a bad binary that
# slipped the checksum cannot read or exfiltrate them. If the scan fails, EYEBROW_OK
# stays 0 and the fail-safe below covers it.
[ -n "$EB" ] && env -i PATH="$PATH" HOME="$HOME" "$EB" scan --path . --lockfile eyebrowlock.json 2>/dev/null && EYEBROW_OK=1

Fail-safe - guarantees a green PR without the binary. If EYEBROW_OK is still 0 (binary unavailable or scan failed) and this run has any CLEAN-ADD of a skills/** SKILL.md, do not ship a skill the lock cannot cover: revert each such new skill from the branch (git rm -r --cached skills/<slug> + drop it from the working tree) and re-classify it as CONFLICT with reason needs-eyebrowlock-scan. S9 surfaces it with the exact operator command (eyebrow scan --path . --lockfile eyebrowlock.json then commit). A CLEAN-UPDATE / CLEAN-MERGE of an existing skill needs no rescan - eyebrow verify allows content drift (it fails only on a new egress host or a new CRITICAL), and the skill already has a lock entry. This trades auto-installing a brand-new upstream skill (rare) for never landing a red PR; the skill still arrives, just as a one-line manual step in the PR.

Then validate config:

node scripts/validate-config.js aeon.yml || echo "validate-config flagged (may be pre-existing drift; note, do not abort on it)"

If nothing CLEAN applied (every upstream change was CONFLICT or OPERATOR): open no PR. Fold all changes into the report, log AEON_UPDATE_MANUAL_ONLY, and go to S10 (notify the operator that the sync needs a manual merge). If a file we wrote breaks YAML/JSON parsing, abort: git checkout . && git checkout ${UP_DEFAULT} && git branch -D "$BR", exit AEON_UPDATE_VALIDATION_FAILED, notify with the failing file.

S8. Advance the baseline + reconcile conflicts (in the branch)

Recompute PENDING: for every CONFLICT file this run plus every prior PENDING entry, keep it only if sha256(local) != sha256(HEAD blob) (still genuinely divergent). Drop the rest (resolved). A file that was CLEAN-MERGE-applied this run is resolved - never carry it as pending (its sha256(local) != sha256(HEAD blob) because it still holds the operator's edits, but the upstream change is now merged in, so the naive test would wrongly keep it forever; a 3-way-merged file is only a CONFLICT again if a future upstream change overlaps the operator's lines). Likewise drop any prior PENDING entry whose file was CLEAN-MERGE- or CLEAN-UPDATE-applied this run. Exception: enabled-skill-removed-upstream entries have no HEAD blob to diff (the path is deleted upstream) - resolve them instead when the skill is no longer enabled: true in the operator's current aeon.yml (they disabled it, so the CLEAN-DELETE rule can now apply next run) or upstream re-adds a path of that name (re-classify as CONFLICT/modified or CLEAN-UPDATE under the normal rules).

Write memory/topics/aeon-update-state.json and commit it with the sync so merging advances the watermark:

{
  "baseline_sha": "${HEAD_SHA}",
  "upstream": "${UPSTREAM}",
  "last_run": "${today}",
  "last_pr": null,
  "applied": { "added": N, "updated": N, "deleted": N },
  "pending_conflicts": [
    { "path": "scripts/foo.sh", "reason": "operator-customized", "upstream_commits": ["abc1234"] }
  ]
}

Advancing baseline_sha to HEAD means clean files never re-notify, while pending_conflicts carries unresolved merges forward independently - so they resurface each run until the operator actually reconciles them, and are also written to the PR body. (In report mode, do not write state - a dry run mutates nothing.)

S9. Commit + open the PR (sync mode; skip in report mode)

git add -A
git commit -F /tmp/aeon-update-commit.txt      # never inline the message with -m (backticks/`$()` in commit text get shell-substituted)
git push -u origin "$BR"
gh pr create --repo "$SELF" --base "$UP_DEFAULT" \
  --title "aeon-update: sync ${N_COMMITS} upstream commits (${BASE7}..${HEAD7})" \
  --body-file /tmp/aeon-update-pr-body.md

/tmp/aeon-update-commit.txt:

aeon-update: sync upstream ${BASE7}..${HEAD7}

${N_COMMITS} upstream commits from ${UPSTREAM}. ${N_APPLIED} files applied cleanly, ${N_CONFLICT} need manual review. Baseline advanced to ${HEAD7}.

PR body (/tmp/aeon-update-pr-body.md) - only include sections that have content:

## Upstream sync: `${UPSTREAM}` `${BASE7}..${HEAD7}`

**${N_COMMITS} commits** ({earliest date} → {latest date}) · **${N_APPLIED} applied** · **${N_CONFLICT} manual** · baseline → `${HEAD7}`.

### Applied cleanly
- **New skills:** `foo`, `bar`   _(regenerated catalogs + agents.md + skill-icons)_
- **Modified skills:** `baz`
- **Scripts / harness:** `scripts/notify.sh`, ...
- **Workflows:** `.github/workflows/...`
- **Auto-merged (3-way):** `.github/workflows/aeon.yml`, ...  _(your local customization kept; upstream's disjoint changes applied - review the merged hunks)_
- **Docs / other:** `docs/...`, `CLAUDE.md`, ...

### ⚠️ Currently-enabled skills removed upstream
For each CONFLICT with reason `enabled-skill-removed-upstream` (S6):
- `verdikta-hunter` — enabled in your `aeon.yml`, deleted upstream in {commit(s)}. **Not deleted here** so nothing breaks. Pick one: keep it as a fork-only skill going forward (nothing else to do), or disable it in `aeon.yml` to match upstream's current default set.

### Needs manual review (conflicts - your local copy diverges from upstream)
For each *other* conflict reason (`enabled-skill-removed-upstream` is covered above, don't duplicate it here): what upstream changed and why it wasn't auto-applied.
- `scripts/foo.sh` — you customized this locally; upstream changed it in {commits}. Upstream diff:
  ```diff
  {short upstream base..head diff for the file}

Operator config changed upstream (not auto-applied - reconcile by hand)

• aeon.yml — upstream added skills / changed defaults: {summary}. Merge the new entries you want (keep your enable/schedule/model choices).
• soul/…, STRATEGY.md — {summary, if changed}

Upstream commits

Capture the PR URL; write it back into `last_pr` in the branch's state file (amend the state commit) so the merged watermark records its own PR.

### S10. Log + notify

Append to `memory/logs/${today}.md` under `### aeon-update`: status, `UPSTREAM`, `${BASE7}..${HEAD7}`, applied/conflict counts, PR URL (or `report`/`manual-only`), and `pending_conflicts` count.

**Notify only on signal** - match `soul/` voice if present. Send when there is a PR, a report with changes, or a manual-only situation; stay silent for `IN_SYNC` / `BASELINE_SET`-with-nothing. Keep it ≤4000 chars:

aeon-update — ${today} {verdict: "synced N commits → PR" | "N changes need manual merge" | "report: N commits behind"}

Upstream ${UPSTREAM} is ${AHEAD} commits ahead. Applied ${N_APPLIED} cleanly, ${N_CONFLICT} need review. {Top applied highlight: e.g. "new skill: token-radar; harness fix in run-harness"} {If conflicts: "Manual: aeon.yml (new skills), scripts/foo.sh (local edit)"}

PR: {url} (or "dry run — nothing changed")

Pass `--mute-key "aeon-update:${HEAD7}"` so a muted sync doesn't re-ping for the same upstream HEAD.

## Exit taxonomy

| Code | When | Notify |
|------|------|--------|
| `AEON_UPDATE_OK` | PR opened with ≥1 clean file (conflicts, if any, listed in it) | Yes - PR link |
| `AEON_UPDATE_MANUAL_ONLY` | Upstream changed only operator-owned / conflicting files - no clean apply, no PR | Yes - manual call-out |
| `AEON_UPDATE_REPORT` | `report` mode - delta computed, nothing mutated | Yes - dry-run summary |
| `AEON_UPDATE_IN_SYNC` | Baseline already == upstream HEAD | No (log only) |
| `AEON_UPDATE_BASELINE_SET` | First run - watermark anchored, nothing applied | One-line notify |
| `AEON_UPDATE_IS_UPSTREAM` | This instance is the upstream repo itself | No (log only) |
| `AEON_UPDATE_BASELINE_UNREACHABLE` | Baseline is not an ancestor of HEAD (history rewrite) | Yes - asks for `reset=` |
| `AEON_UPDATE_VALIDATION_FAILED` | An applied file broke YAML/JSON parsing → branch reverted | Yes - failing file |

## Constraints

- **Never** push to `main` or auto-write an OPERATOR-owned path (`aeon.yml`, `soul/`, `memory/`, `STRATEGY.md`, `.mcp.json`, `output/`).
- **Never** delete a fork-only skill, and **never** copy `catalog/*.json` / `eyebrowlock.json` from upstream - regenerate them.
- **Never** advance `baseline_sha` without carrying unresolved conflicts forward in `pending_conflicts`.
- Cross-repo compare caps at 300 files - note `files_truncated=true` and let a follow-up run pick up the rest.
- A clean, in-sync run is **correct**, not a failure - it notifies nothing.

## Network note

Every network call is `gh api`, which authenticates via `GITHUB_TOKEN` automatically - no `curl`, no `./secretcurl`, no `$SECRET` on the command line for the Bash permission layer to refuse, and no secret beyond the default `GITHUB_TOKEN`. There are no irreversible side-effects: the skill's only mutation is a PR against this instance's own repo, which the operator reviews and merges. Retry policy: on `403` with `X-RateLimit-Remaining: 0`, sleep 60s and retry once; on a persistent contents-API failure for one file, mark it `UNREADABLE` in the report and continue with a partial sync rather than aborting the whole run.