create-egregore 0.20.2 → 0.21.0-runtime-mvp.2

package/runtime/claude/bin/repo-state.sh

@@ -0,0 +1,96 @@
+#!/usr/bin/env bash
+# repo-state.sh — Capture current branch state across all repos.
+# Outputs a markdown table (## Repo State) for handoff files.
+# Only includes repos on non-base branches or with uncommitted changes.
+#
+# Usage: bash bin/repo-state.sh [--no-pr]
+#   --no-pr    Skip the `gh pr list` lookup per repo (~400–600ms each).
+#              PR column will be `—`. Useful on the hot path where a detached
+#              backfill will fill PR numbers in later.
+#   Outputs the markdown section to stdout. Empty output = nothing to report.
+set -euo pipefail
+
+NO_PR=0
+while [ $# -gt 0 ]; do
+  case "$1" in
+    --no-pr) NO_PR=1; shift ;;
+    *) shift ;;
+  esac
+done
+
+SCRIPT_DIR="$(cd "$(dirname "$0")/.." && pwd)"
+CONFIG="$SCRIPT_DIR/egregore.json"
+PARENT_DIR="$(cd "$SCRIPT_DIR/.." && pwd)"
+
+# Source config helpers for _get_base_branch
+source "$SCRIPT_DIR/bin/lib/config.sh" 2>/dev/null || true
+
+CORE_REPO=$(jq -r '.repo_name // "egregore"' "$CONFIG" 2>/dev/null)
+GITHUB_ORG=$(jq -r '.github_org // empty' "$CONFIG" 2>/dev/null)
+MANAGED_REPOS=$(jq -r '(.repos[]? // empty) | if type == "object" then .name else . end' "$CONFIG" 2>/dev/null)
+
+ROWS=""
+
+check_repo() {
+  local repo_name="$1"
+  local repo_dir="$2"
+  local base_branch="$3"
+
+  [ -d "$repo_dir/.git" ] || [ -f "$repo_dir/.git" ] || return 0
+
+  local branch
+  branch=$(git -C "$repo_dir" branch --show-current 2>/dev/null || echo "")
+  [ -z "$branch" ] && return 0
+
+  local dirty=""
+  [ -n "$(git -C "$repo_dir" status --porcelain 2>/dev/null | head -1)" ] && dirty="yes"
+
+  # Skip if on base branch with no changes
+  if [ "$branch" = "$base_branch" ] && [ -z "$dirty" ]; then
+    return 0
+  fi
+
+  # Count commits ahead of base
+  local ahead=0
+  ahead=$(git -C "$repo_dir" rev-list "origin/${base_branch}..HEAD" --count 2>/dev/null || echo "0")
+
+  # Skip if on base branch, no dirty files, no commits ahead
+  if [ "$branch" = "$base_branch" ] && [ -z "$dirty" ] && [ "$ahead" = "0" ]; then
+    return 0
+  fi
+
+  # Check for open PR (skipped with --no-pr; a detached backfill fills it in)
+  local pr="—"
+  if [ "$NO_PR" = "0" ] && command -v gh &>/dev/null && [ -n "$GITHUB_ORG" ]; then
+    local pr_num
+    pr_num=$(gh pr list --repo "$GITHUB_ORG/$repo_name" --head "$branch" --json number -q '.[0].number' 2>/dev/null || true)
+    [ -n "$pr_num" ] && pr="#$pr_num"
+  fi
+
+  ROWS="${ROWS}| ${repo_name} | ${branch} | ${pr} | ${base_branch} |\n"
+}
+
+# Check core repo
+if ! CORE_BASE=$(_get_base_branch); then
+  echo "repo-state: could not resolve the configured base branch" >&2
+  exit 1
+fi
+check_repo "$CORE_REPO" "$SCRIPT_DIR" "$CORE_BASE"
+
+# Check managed repos
+for REPO in $MANAGED_REPOS; do
+  REPO_DIR="$PARENT_DIR/$REPO"
+  if ! BASE=$(_get_base_branch "$REPO"); then
+    continue
+  fi
+  check_repo "$REPO" "$REPO_DIR" "$BASE"
+done
+
+# Output
+if [ -n "$ROWS" ]; then
+  echo "## Repo State"
+  echo ""
+  echo "| Repo | Branch | PR | Base |"
+  echo "|------|--------|----|------|"
+  printf '%b' "$ROWS"
+fi

package/runtime/claude/bin/research-ingest.sh

@@ -0,0 +1,7 @@
+#!/usr/bin/env bash
+# Typed Runtime boundary for approved meeting/interview analysis packages.
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+EGREGORE_ROOT="$ROOT" PYTHONPATH="$ROOT${PYTHONPATH:+:$PYTHONPATH}" \
+  exec python3 -m egregore_runtime.research_ingest_cli "$@"

package/runtime/claude/bin/restore-owned-skills.sh

@@ -0,0 +1,67 @@
+#!/usr/bin/env bash
+# restore-owned-skills.sh — keep org-owned skills intact through a framework update.
+#
+# Org-owned skills are declared in egregore.json → owned_skills[]. The /update
+# overlay stamps upstream framework paths (including .claude/skills/) over the
+# working tree. This script runs AFTER the overlay and BEFORE the update
+# commit. For every owned skill name:
+#
+#   - upstream does NOT ship that name  → nothing to do (the overlay never
+#     touched it; local edits, committed or not, are left alone)
+#   - upstream DOES ship that name      → the org's committed version wins:
+#     the skill directory is reset to HEAD and the collision is reported so
+#     the org can rename theirs or explicitly adopt upstream's
+#   - collision but no committed org version → upstream's copy stays (there is
+#     nothing to restore); the report says to commit and re-run /update
+#
+# Runtime-neutral: called from every harness's update flow. Idempotent; exits
+# 0 silently when owned_skills is absent or empty.
+#
+# Usage: restore-owned-skills.sh [--upstream-ref <ref>]   (default upstream/main)
+
+set -uo pipefail
+
+UPSTREAM_REF="upstream/main"
+if [ "${1:-}" = "--upstream-ref" ] && [ -n "${2:-}" ]; then
+  UPSTREAM_REF="$2"
+fi
+
+command -v jq >/dev/null 2>&1 || exit 0
+[ -f egregore.json ] || exit 0
+git rev-parse --git-dir >/dev/null 2>&1 || exit 0
+
+OWNED=$(jq -r '.owned_skills // [] | .[]' egregore.json 2>/dev/null)
+[ -n "$OWNED" ] || exit 0
+
+while IFS= read -r name; do
+  [ -n "$name" ] || continue
+  case "$name" in
+    */*|.*|*..*|*' '*)
+      echo "  ⚠ owned_skills: invalid name '$name' — skipped" >&2
+      continue
+      ;;
+  esac
+  SKILL_PATH=".claude/skills/$name"
+
+  # Only a name collision needs work — an org-only skill is never touched by
+  # the overlay, and resetting it here would destroy uncommitted local edits.
+  if ! git ls-tree -d "$UPSTREAM_REF" -- "$SKILL_PATH" 2>/dev/null | grep -q .; then
+    continue
+  fi
+
+  if git cat-file -e "HEAD:$SKILL_PATH" 2>/dev/null; then
+    # Wipe the overlay's copy (index + tree), then restore the org's committed
+    # version. rm+checkout also drops upstream-only files inside the dir.
+    git rm -r --cached --quiet -- "$SKILL_PATH" 2>/dev/null || true
+    rm -rf "$SKILL_PATH"
+    git checkout HEAD -- "$SKILL_PATH" 2>/dev/null || true
+    echo "  ◐ owned skill '$name': upstream ships a skill with this name — kept yours."
+    echo "    Review upstream's version: git diff HEAD $UPSTREAM_REF -- $SKILL_PATH"
+  else
+    echo "  ⚠ owned skill '$name': upstream ships this name and you have no committed version — upstream's copy is in the tree. Commit yours and re-run /update to keep it."
+  fi
+done <<EOF
+$OWNED
+EOF
+
+exit 0

package/runtime/claude/bin/retrieval_eval.py

@@ -0,0 +1,385 @@
+"""Measure retrieval against questions whose answers are known.
+
+A system that answers questions phrases a wrong answer exactly as well as a
+right one. Nothing about the output distinguishes them, so the only instrument
+that can is a set of questions someone has already answered — asked of the
+system, and compared.
+
+This runs such a fixture and reports what the plan requires: whether the hard
+boundary ever leaked, how often the expected source was retrieved, how often
+nothing was found, and how long it took. A boundary violation is a failure
+regardless of how good the answer looked.
+
+Expected answers are marked verified or not. An unverified expectation is a
+guess written down, and counting it as evidence would make the measurement
+agree with whoever wrote the fixture rather than with the domain.
+"""
+
+from __future__ import annotations
+
+import json
+import math
+import re
+import time
+from collections import Counter
+from dataclasses import dataclass, field
+from pathlib import Path
+
+TOKEN = re.compile(r"\w{3,}", re.UNICODE)
+
+# Turkish is agglutinative: budama, budaması and budamada are one word wearing
+# three suffixes, and exact token matching treats them as three. Retrieval
+# therefore under-returns on inflected forms. Stemming is a known gap, recorded
+# here rather than silently tolerated — measuring it needs the fixture.
+STEMMING = False
+
+
+@dataclass
+class Passage:
+    id: str
+    text: str
+    document: str
+    zone: str
+    tier: str
+    topic: str
+
+
+@dataclass
+class QueryResult:
+    id: str
+    question: str
+    found: int
+    top: list = field(default_factory=list)
+    expected_hit: bool | None = None
+    boundary_violation: list = field(default_factory=list)
+    elapsed_ms: float = 0.0
+    verified_expectation: bool = False
+
+
+def tokenize(text: str) -> list[str]:
+    return TOKEN.findall((text or "").lower())
+
+
+def zones_in_scope(zone: str, hierarchy: dict) -> set[str]:
+    """A zone plus everything it inherits from.
+
+    Inheritance runs one way. A grower's own zone and the wider zones containing
+    it are in scope; a sibling zone never is, which is the direction where a
+    violation does harm.
+    """
+    scope = {zone}
+    current = zone
+    seen = {zone}
+    while True:
+        parent = hierarchy.get(current)
+        if not parent or parent in seen:
+            break
+        scope.add(parent)
+        seen.add(parent)
+        current = parent
+    return scope
+
+
+# BM25 parameters at their standard values. k1 bounds how much repeating a term
+# can raise a score; b sets how strongly length is normalised away.
+K1 = 1.2
+B = 0.75
+
+# How many passages one document may contribute. Without a cap a single long
+# document fills the whole list, and a grower reading five results sees one
+# source five times instead of five sources.
+PER_DOCUMENT = 2
+
+
+class Index:
+    """A lexical index over passages, filtered before ranking.
+
+    Ranking is BM25 rather than plain term frequency. The difference is length:
+    a longer passage contains more of every word, so under raw counts the
+    longest passages in a corpus rise to the top of questions they have no
+    particular claim to. Measured here, the two documents topping unrelated
+    queries were the two longest chunks in the archive.
+    """
+
+    def __init__(self, passages: list[Passage]):
+        self.passages = passages
+        self.df: Counter = Counter()
+        self.tf: list[Counter] = []
+        self.lengths: list[int] = []
+        for p in passages:
+            counts = Counter(tokenize(p.text))
+            self.tf.append(counts)
+            self.lengths.append(sum(counts.values()))
+            self.df.update(set(counts))
+        self.n = max(1, len(passages))
+        self.avg_length = (sum(self.lengths) / self.n) if self.lengths else 0.0
+
+    def idf(self, term: str) -> float:
+        """Inverse document frequency, in the form that stays positive.
+
+        The textbook ratio goes to zero — and can go negative — for a term
+        present in most of the corpus, which silently drops passages that do
+        match. The +1 keeps every genuine match above zero.
+        """
+        df = self.df.get(term, 0)
+        return math.log(1 + (self.n - df + 0.5) / (df + 0.5))
+
+    def search(
+        self,
+        question: str,
+        allowed_zones: set[str],
+        limit: int = 5,
+        per_document: int = PER_DOCUMENT,
+        vocabulary=None,
+    ) -> list[tuple[float, Passage]]:
+        terms = tokenize(question)
+        weights = {t: 1.0 for t in terms}
+        if vocabulary is not None:
+            import term_expand
+
+            _, added = term_expand.expand(question, vocabulary)
+            for t in added:
+                # Only where the asker did not already use the word. An inferred
+                # word never raises the weight of one they chose themselves.
+                if len(t) >= 3 and t not in weights:
+                    weights[t] = term_expand.EXPANSION_WEIGHT
+
+        scored = []
+        for i, p in enumerate(self.passages):
+            # The zone filter runs before ranking, not after. Filtering a ranked
+            # list would let a better-scoring out-of-zone passage displace an
+            # in-zone one and shorten the result rather than replace it.
+            if p.zone not in allowed_zones:
+                continue
+            counts = self.tf[i]
+            norm = K1 * (1 - B + B * (self.lengths[i] / self.avg_length)) if self.avg_length else K1
+            score = 0.0
+            for t, weight in weights.items():
+                tf = counts.get(t, 0)
+                if tf:
+                    score += weight * self.idf(t) * (tf * (K1 + 1)) / (tf + norm)
+            if score > 0:
+                scored.append((score, p))
+        scored.sort(key=lambda x: -x[0])
+
+        if per_document <= 0:
+            return scored[:limit]
+        seen: Counter = Counter()
+        capped = []
+        for score, p in scored:
+            if seen[p.document] >= per_document:
+                continue
+            seen[p.document] += 1
+            capped.append((score, p))
+            if len(capped) >= limit:
+                break
+        return capped
+
+
+def load_passages(manifest_path: Path, data_root: Path) -> list[Passage]:
+    manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
+    chunk_re = re.compile(r"<!-- ingest-chunk: (\S+) -->\n\n(.*?)(?=\n## Chunk |\Z)", re.S)
+    passages: list[Passage] = []
+    for doc in manifest.get("documents", []):
+        body = (data_root / doc["file_path"]).read_text(encoding="utf-8", errors="replace")
+        provenance = doc.get("provenance") or {}
+        catalogued = provenance.get("from_catalogue") or {}
+        excluded = {
+            c["id"] for c in doc.get("chunks", []) if (c.get("segment") or {}).get("kind") == "bibliography"
+        }
+        for match in chunk_re.finditer(body):
+            cid = match.group(1)
+            if cid in excluded:
+                continue
+            passages.append(
+                Passage(
+                    id=cid,
+                    text=match.group(2),
+                    document=doc.get("title", ""),
+                    zone=(doc.get("boundaries") or {}).get("zone", ""),
+                    tier=provenance.get("tier", ""),
+                    topic=catalogued.get("topic", ""),
+                )
+            )
+    return passages
+
+
+def run(fixture: dict, index: Index, vocabulary=None) -> list[QueryResult]:
+    hierarchy = fixture.get("zone_hierarchy") or {}
+    results: list[QueryResult] = []
+
+    for query in fixture.get("queries", []):
+        zone = query.get("zone") or fixture.get("default_zone") or ""
+        scope = zones_in_scope(zone, hierarchy) if zone else set()
+
+        started = time.perf_counter()
+        hits = (
+            index.search(query["text"], scope, limit=query.get("limit", 5), vocabulary=vocabulary)
+            if scope
+            else []
+        )
+        elapsed = (time.perf_counter() - started) * 1000
+
+        expected = (query.get("expect") or {}).get("documents") or []
+        verified = bool((query.get("expect") or {}).get("verified"))
+        expected_hit = None
+        if expected:
+            retrieved = {p.document for _, p in hits}
+            expected_hit = any(e in retrieved for e in expected)
+
+        # Anything returned from outside the permitted scope is a failure of the
+        # only guarantee the system makes about where advice applies.
+        violations = [p.document for _, p in hits if p.zone not in scope]
+
+        results.append(
+            QueryResult(
+                id=query.get("id", "?"),
+                question=query["text"],
+                found=len(hits),
+                top=[{"document": p.document, "zone": p.zone, "tier": p.tier, "chunk": p.id} for _, p in hits[:3]],
+                expected_hit=expected_hit,
+                boundary_violation=violations,
+                elapsed_ms=round(elapsed, 1),
+                verified_expectation=verified,
+            )
+        )
+    return results
+
+
+def diagnose(passages: list[Passage]) -> list[str]:
+    """Reasons a corpus will answer nothing, stated before the run does.
+
+    Zone is a hard retrieval boundary, so a corpus ingested without a publisher
+    profile carries no zone and every query is refused. That is the boundary
+    working — advice that cannot be placed is not served — but the report it
+    produces reads as `0 answered, 0 boundary violations`, which looks closer to
+    success than to the misconfiguration it is.
+    """
+    notes = []
+    if not passages:
+        return ["the corpus is empty: no passages were loaded from the manifest"]
+
+    zoneless = sum(1 for p in passages if not p.zone)
+    if zoneless == len(passages):
+        notes.append(
+            f"every one of {len(passages):,} passages has no zone, so every query will be "
+            "refused. The corpus was probably ingested without EGREGORE_PUBLISHER_PROFILE set."
+        )
+    elif zoneless:
+        notes.append(f"{zoneless:,} of {len(passages):,} passages have no zone and can never be served")
+
+    tierless = sum(1 for p in passages if not p.tier)
+    if tierless == len(passages):
+        notes.append("no passage carries a trust tier, so results cannot be ranked by source reliability")
+    return notes
+
+
+def report(results: list[QueryResult]) -> dict:
+    answered = [r for r in results if r.found]
+    checked = [r for r in results if r.expected_hit is not None and r.verified_expectation]
+    hits = [r for r in checked if r.expected_hit]
+    latencies = sorted(r.elapsed_ms for r in results) or [0.0]
+    violations = [r for r in results if r.boundary_violation]
+
+    def pct(n, d):
+        return round(100 * n / d, 1) if d else None
+
+    return {
+        "queries": len(results),
+        "answered": len(answered),
+        "unanswered": len(results) - len(answered),
+        "unanswered_rate_pct": pct(len(results) - len(answered), len(results)),
+        "boundary_violations": len(violations),
+        "citation_hit_rate_pct": pct(len(hits), len(checked)),
+        "expectations_verified": len(checked),
+        "expectations_unverified": sum(
+            1 for r in results if r.expected_hit is not None and not r.verified_expectation
+        ),
+        "p50_ms": latencies[len(latencies) // 2],
+        "p95_ms": latencies[min(len(latencies) - 1, int(len(latencies) * 0.95))],
+    }
+
+
+def format_report(report: dict, results: list[QueryResult], verbose: bool = False) -> str:
+    lines = [
+        f"queries {report['queries']} · answered {report['answered']} · unanswered {report['unanswered']}",
+        "",
+        f"boundary_violations          {report['boundary_violations']}",
+        f"citation_hit_rate_pct        {report['citation_hit_rate_pct']}",
+        f"expectations_verified        {report['expectations_verified']}",
+        f"expectations_unverified      {report['expectations_unverified']}",
+        f"p50_ms {report['p50_ms']} · p95_ms {report['p95_ms']}",
+    ]
+    if report["citation_hit_rate_pct"] is None:
+        lines += [
+            "",
+            "The hit rate is unavailable because no expectation has been confirmed by",
+            "someone who knows the domain. It is withheld rather than computed from",
+            "guesses, which would only measure agreement with whoever wrote the fixture.",
+        ]
+    if verbose:
+        lines.append("")
+        for r in results:
+            lines.append(f"[{r.id}] {r.question}")
+            lines.append(f"    {r.found} found · {r.elapsed_ms} ms")
+            for hit in r.top:
+                lines.append(f"    · {hit['document'][:70]}  ({hit['zone']}, {hit['tier']})")
+            if r.boundary_violation:
+                lines.append(f"    VIOLATION: {r.boundary_violation}")
+    return "\n".join(lines)
+
+
+def main(argv: list[str] | None = None) -> int:
+    import argparse
+
+    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
+    parser.add_argument("--fixture", required=True, type=Path)
+    parser.add_argument("--manifest", required=True, type=Path)
+    parser.add_argument("--data-root", required=True, type=Path)
+    parser.add_argument("--vocabulary", type=Path, nargs="*", default=[],
+                        help="term files used to expand a question into the archive's wording")
+    parser.add_argument("--verbose", action="store_true", help="show retrieved sources per query")
+    parser.add_argument("--json", action="store_true", help="emit the report as JSON")
+    args = parser.parse_args(argv)
+
+    fixture = json.loads(args.fixture.read_text(encoding="utf-8"))
+    started = time.perf_counter()
+    passages = load_passages(args.manifest, args.data_root)
+    index = Index(passages)
+    if not args.json:
+        print(f"loaded {len(passages):,} passages in {time.perf_counter() - started:.1f}s "
+              "(bibliographies excluded)\n")
+
+    vocabulary = None
+    if args.vocabulary:
+        import term_expand
+
+        vocabulary = term_expand.load_vocabulary(args.vocabulary)
+        if not args.json:
+            print(f"vocabulary: {len(vocabulary)} term groups")
+            for skipped in vocabulary.skipped:
+                print(f"  skipped {skipped}")
+
+    problems = diagnose(passages)
+    if problems and not args.json:
+        for note in problems:
+            print(f"!! {note}")
+        print()
+
+    results = run(fixture, index, vocabulary=vocabulary)
+    summary = report(results)
+
+    if args.json:
+        print(json.dumps({"report": summary, "diagnostics": problems,
+                          "results": [vars(r) for r in results]},
+                         ensure_ascii=False, indent=2))
+    else:
+        print(format_report(summary, results, verbose=args.verbose))
+
+    # A boundary violation is the one failure that must be actionable from a
+    # script; everything else is a measurement, not a fault.
+    return 1 if (summary["boundary_violations"] or problems) else 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/runtime/claude/bin/runtime-identity.sh

@@ -0,0 +1,70 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Resolve the stable Runtime identity before Observe can emit attributed
+# context. Connected instances adopt IDs from the authenticated control plane;
+# legacy Local instances create opaque local IDs once. Network failure remains
+# non-fatal to startup and is retried next session.
+
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+CONFIG="$ROOT/egregore.json"
+STATE="$ROOT/.egregore-state.json"
+
+resolved() {
+  [ -f "$CONFIG" ] && [ -f "$STATE" ] || return 1
+  jq -e '
+    (.org_id | type == "string" and length > 0 and (startswith("legacy-org:") | not))
+  ' "$CONFIG" >/dev/null 2>&1 || return 1
+  jq -e '
+    [ .account_id, .actor_id, .membership_id ]
+    | all(type == "string" and length > 0)
+  ' "$STATE" >/dev/null 2>&1
+}
+
+case "${1:-ensure}" in
+  status)
+    if resolved; then
+      printf '{"status":"resolved"}\n'
+      exit 0
+    fi
+    printf '{"status":"unresolved"}\n'
+    exit 1
+    ;;
+  ensure)
+    resolved && exit 0
+    mode="$(jq -r 'if (.mode // "local") == "connected" or (.api_url // "") != "" then "connected" else "local" end' "$CONFIG" 2>/dev/null || echo local)"
+    if [ "$mode" = "local" ]; then
+      EGREGORE_ROOT="$ROOT" python3 -m egregore_runtime.identity_cli initialize-local >/dev/null
+    else
+      api_url="$(jq -r '.api_url // empty' "$CONFIG" 2>/dev/null)"
+      api_key="$(grep '^EGREGORE_API_KEY=' "$ROOT/.env" 2>/dev/null | cut -d= -f2-)"
+      [ -n "$api_url" ] && [ -n "$api_key" ] || exit 1
+      # Identity repair is read-only: select the already-existing account and
+      # membership records, then adopt their opaque IDs locally. If migration
+      # 038 has not reached the control plane, the next session retries.
+      control_plane="$(curl -sf "$api_url/api/org/status" \
+        -H "Authorization: Bearer $api_key" --connect-timeout 2 --max-time 4 \
+        2>/dev/null || true)"
+      printf '%s' "$control_plane" | jq -e \
+        '.org_id | type == "string" and length > 0' >/dev/null 2>&1 || exit 1
+      slug="$(jq -r '.slug // empty' "$CONFIG" 2>/dev/null)"
+      [ -n "$slug" ] || exit 1
+      roster="$(curl -sf "$api_url/api/org/$slug/members?include_removed=true" \
+        -H "Authorization: Bearer $api_key" --connect-timeout 2 --max-time 4 \
+        2>/dev/null || true)"
+      printf '%s' "$roster" | jq -e \
+        '.members | type == "array"' >/dev/null 2>&1 || exit 1
+      jq -cn \
+        --argjson organization "$control_plane" \
+        --argjson roster "$roster" \
+        '{organization:$organization,roster:$roster}' | \
+        EGREGORE_ROOT="$ROOT" python3 -m egregore_runtime.identity_cli \
+          reconcile-records >/dev/null
+    fi
+    resolved
+    ;;
+  *)
+    echo 'usage: bash bin/runtime-identity.sh {ensure|status}' >&2
+    exit 2
+    ;;
+esac

package/runtime/claude/bin/runtime-upgrade.sh

@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+# runtime-upgrade.sh — deterministic bridge for the per-instance Runtime
+# upgrade. The launcher's Settings surface and advanced callers drive the
+# transactional engine through these verbs; the Python Runtime owns all
+# state, staging, gates, activation, and rollback. Output is JSON for
+# renderers — this script is an internal surface, not the user experience.
+#
+# Verbs:
+#   status [--fetch]        durable upgrade + team-sync state
+#   init --candidate <ver>  stage a candidate (X.Y.Z-runtime-mvp.N only)
+#   prepare                 start/resume the detached preparation worker
+#   activate                atomic activation; refuses unless every gate passed
+#   cancel                  request cancellation; previous Runtime untouched
+#   rollback                restore the recorded previous Runtime version
+#   team-sync [--fetch]     canonical memory sync state only
+set -uo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")/.." && pwd)"
+
+# EGREGORE_INSTANCE_ROOT lets the packaged launcher bridge drive an instance
+# that predates this script: engine code always comes from this script's own
+# tree, while the target instance is set explicitly — never inferred.
+INSTANCE_ROOT="${EGREGORE_INSTANCE_ROOT:-$SCRIPT_DIR}"
+
+_runtime() {
+  EGREGORE_ROOT="$INSTANCE_ROOT" \
+    PYTHONPATH="$SCRIPT_DIR${PYTHONPATH:+:$PYTHONPATH}" \
+    python3 -m egregore_runtime.harness_cli "$@"
+}
+
+case "${1:-}" in
+  status)   shift; _runtime upgrade status "$@" ;;
+  init)     shift; _runtime upgrade init "$@" ;;
+  prepare)  shift; _runtime upgrade prepare "$@" ;;
+  activate) shift; _runtime upgrade activate "$@" ;;
+  cancel)   shift; _runtime upgrade cancel "$@" ;;
+  rollback) shift; _runtime upgrade rollback "$@" ;;
+  team-sync) shift; _runtime team-sync "$@" ;;
+  *)
+    echo "usage: bash bin/runtime-upgrade.sh {status|init --candidate <version>|prepare|activate|cancel|rollback|team-sync}" >&2
+    exit 2
+    ;;
+esac

package/runtime/claude/bin/scroll.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+export EGREGORE_ROOT="$ROOT"
+export EGREGORE_SESSION_ID="${EGREGORE_SESSION_ID:-$(cat "$ROOT/.egregore-session-id" 2>/dev/null || true)}"
+export PYTHONPATH="$ROOT${PYTHONPATH:+:$PYTHONPATH}"
+
+exec python3 -m egregore_runtime.scroll_cli "$@"