Merge pull request #2 from sneakyfree/runner-guard

runner-guard: no stranger code on self-hosted runners (PR check + nightly sweep)

api/tests/test_compute_guard.py

@@ -229,3 +229,24 @@ def test_block_mode_never_blocks_grant_owned(monkeypatch):
     assert state == "success" and desc.startswith("⚠ WARN (Grant-owned, not blocking): 1") and f is g
     lane = cg.Finding("a.py", 3, "provider host", "x")
     assert cg.status_for([lane], whole_tree=True, grant=[g])[0] == "failure"
+
+
+def test_veron_ollama_warns_on_added_lines_and_never_blocks(monkeypatch):
+    hits = cg.scan_line("app/llm.py", 'OLLAMA = "http://192.168.1.73:11434/api/generate"')
+    assert [k for k, _ in hits] == ["veron ollama"]
+    assert cg.scan_line("app/llm.py", 'port = 114345') == []                # not the port
+    assert cg.scan_line("app/llm.py", "# was http://x:11434 (removed)") == []  # a comment is not a call
+    f = cg.Finding("app/llm.py", 7, "veron ollama", ":11434")
+    monkeypatch.setattr(cg, "MODE", "block")
+    state, desc, _ = cg.status_for([f], whole_tree=False)
+    assert state == "success" and desc.startswith("⚠ WARN: new Veron Ollama ref app/llm.py:7")
+    assert "Windy Mind" in desc and len(desc) <= 140
+    hard = cg.Finding("app/llm.py", 1, "provider host", "api.openai.com")
+    assert cg.status_for([f, hard], whole_tree=False)[0] == "failure"       # a real violation still blocks
+
+
+def test_ollama_in_added_pr_lines_only():
+    diff = ("+++ b/svc/client.py\n@@ -0,0 +1,2 @@\n+import httpx\n"
+            "+URL = 'http://veron:11434/api/chat'\n")
+    got = cg.parse_added("some-repo", diff, [])
+    assert [(f.kind, f.line) for f in got] == [("veron ollama", 2)]

api/tests/test_runner_guard.py

@@ -0,0 +1,70 @@
+"""runner-guard: workflow shapes that hand a self-hosted runner to strangers."""
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[2]
+_spec = importlib.util.spec_from_file_location("runner_guard", ROOT / "scripts" / "runner_guard.py")
+rg = importlib.util.module_from_spec(_spec)
+sys.modules["runner_guard"] = rg
+_spec.loader.exec_module(rg)
+
+
+def rules(text):
+    return [(r, ln) for _p, ln, r, _m in rg.lint_text("w.yml", text)]
+
+
+def test_pull_request_target_always_fails():
+    assert rules("on: pull_request_target\njobs:\n  a:\n    runs-on: ubuntu-latest\n    steps: []\n")[0][0] == "R1"
+
+
+def test_fork_pr_on_self_hosted_fails_and_points_at_runs_on():
+    wf = "on:\n  pull_request:\njobs:\n  t:\n    runs-on: [self-hosted, linux, x64]\n    steps: []\n"
+    assert rules(wf) == [("R2", 5)]
+
+
+def test_same_repo_gate_or_environment_passes():
+    gated = ("on: [pull_request]\njobs:\n  t:\n    if: github.event.pull_request.head.repo.full_name == github.repository\n"
+             "    runs-on: [self-hosted]\n    steps: []\n")
+    env = "on: [pull_request]\njobs:\n  t:\n    environment: ci\n    runs-on: self-hosted\n    steps: []\n"
+    assert rules(gated) == [] and rules(env) == []
+
+
+def test_outsider_events_on_self_hosted_fail_but_writer_events_pass():
+    wf = "on:\n  issue_comment:\n  workflow_run:\n    workflows: [x]\njobs:\n  t:\n    runs-on: self-hosted\n    steps: []\n"
+    assert sorted(r for r, _ in rules(wf)) == ["R3", "R3"]
+    ok = "on:\n  push:\n    tags: ['v*']\n  workflow_dispatch:\n  schedule:\n    - cron: '0 3 * * *'\njobs:\n  t:\n    runs-on: self-hosted\n    steps: []\n"
+    assert rules(ok) == []
+
+
+def test_expression_runs_on_is_treated_as_self_hosted_and_hosted_runner_is_fine():
+    expr = "on: pull_request\njobs:\n  t:\n    runs-on: ${{ matrix.os }}\n    steps: []\n"
+    hosted = "on: pull_request\njobs:\n  t:\n    runs-on: ubuntu-latest\n    steps: []\n"
+    assert rules(expr) == [("R2", 4)] and rules(hosted) == []
+
+
+def test_broken_yaml_is_a_finding_and_non_workflows_are_ignored():
+    assert rules("on: [push\njobs: {")[0][0] == "R0"
+    assert rules("name: just a file\n") == []
+
+
+def test_pr_mode_posts_each_status_once(monkeypatch, tmp_path):
+    calls = []
+    monkeypatch.setattr(rg, "STATE", str(tmp_path / "s.json"))
+    monkeypatch.setattr(rg, "public_repos", lambda owners: [("o/r", "main")])
+    monkeypatch.setattr(rg, "file_at", lambda *a: "on: push\\njobs:\\n  t:\\n    runs-on: self-hosted\\n    steps: []\\n")
+
+    def fake_gh(*args, check=True):
+        if args[0].startswith("repos/o/r/pulls?"):
+            return "7 abc123 o/r\\n"
+        if args[0].endswith("/files?per_page=100"):
+            return ".github/workflows/ci.yml\\n"
+        calls.append(args[0])
+        return ""
+    monkeypatch.setattr(rg, "gh", fake_gh)
+    rg.cmd_pr(["o"], post=True)
+    rg.cmd_pr(["o"], post=True)
+    assert calls == ["repos/o/r/statuses/abc123"]

api/tests/test_secret_guard.py

@@ -146,3 +146,11 @@ def test_warn_kinds_do_not_block(monkeypatch):
     assert sg.status_for([f], True)[0] == "success"
     monkeypatch.setattr(sg, "WARN_KINDS", set())
     assert sg.status_for([f], True)[0] == "failure"
+
+
+def test_pypi_token_shape_hash_only():
+    tok = "pypi-AgE" + "Ab1_-" * 20  # synthetic
+    got = ss.find(f"password = {tok}")
+    assert [k for k, _ in got] == ["pypi token"] and got[0][1] == ss.h8(tok)
+    assert tok not in repr(got)
+    assert ss.find("pypi-AgE-too-short") == []

ci/compute-guard-allow.yml

@@ -45,3 +45,8 @@ allow:
   - repo: windy-git
     paths: ["scripts/compute_guard.py", "ci/compute-guard-allow.yml"]
     reason: "The guard's own pattern list and this file."
+
+  - repo: windy-mind
+    paths: ["*"]
+    matches: [':11434']
+    reason: "Windy Mind IS the compute door (endpoint + key); it may call Ollama. Only the Ollama port is allowed here, any provider host/SDK in Mind still flags."

ci/secret-guard-allow.yml

@@ -57,3 +57,6 @@ allow:
   - repo: windytalk
     hashes: ["c4189d79"]
     reason: "apps/desktop/test/diagnostics.test.ts redaction fixture (hexSecret beside a fake sk-ant token); hash checked against the lockbox 10-01: not present."
+  - repo: windy-agent
+    hashes: ["84e0c0ea"]
+    reason: "tests/test_log_redaction.py:56 Z.ai redaction fixture REPLACED by windy-agent #412 with a synthetic value; hash checked against the lockbox 10-01: not present."

deploy/systemd/windygit-state-backup.service

@@ -0,0 +1,14 @@
+[Unit]
+Description=Windy Git nightly STATE backup (Postgres + Gitea config + repos -> encrypted restic in R2)
+After=network-online.target docker.service
+
+[Service]
+Type=oneshot
+WorkingDirectory=/srv/windygit/src
+# R2 credentials come from the .env; the restic password from /etc/windygit/restic.pass
+# (root 600; the same value lives in the lockbox as RESTIC_WINDYGIT_PASSWORD).
+EnvironmentFile=/srv/windygit/src/.env
+ExecStart=/bin/bash /srv/windygit/src/scripts/backup_state.sh
+Nice=10
+IOSchedulingClass=idle
+TimeoutStartSec=3h

deploy/systemd/windygit-state-backup.service.d/windy-job.conf

@@ -0,0 +1,3 @@
+[Service]
+ExecStart=
+ExecStart=/usr/local/bin/windy-job windygit-state-backup 26h --expect "ok — state backed up" --owner 13 -- /bin/bash /srv/windygit/src/scripts/backup_state.sh

deploy/systemd/windygit-state-backup.timer

@@ -0,0 +1,11 @@
+[Unit]
+Description=Nightly Windy Git state backup
+
+[Timer]
+# 03:07 local, clear of the git-bundle backup at 04:17.
+OnCalendar=*-*-* 03:07:00
+Persistent=true
+RandomizedDelaySec=300
+
+[Install]
+WantedBy=timers.target

deploy/windy0/nightly-runner-guard-sweep.sh

@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+# Nightly (Boss 10-01): runner-guard over the default branch of EVERY public repo in Grant's
+# 5 GitHub accounts (runs on Veron: windy-git scripts/runner_guard.py report). Writes
+# ~/windy-orchestra/RUNNER_GUARD.md (repo, file:line, rule; no file content), appends ONE
+# BOARD line per NEW hit vs the last run, and prints "runner-guard sweep: N hit(s)" last, so
+# the windy-job heartbeat (--expect "runner-guard sweep: 0 hit") goes red while any hit exists.
+set -euo pipefail
+page=~/windy-orchestra/RUNNER_GUARD.md
+state=~/.local/state/runner-guard-sweep.txt
+mkdir -p "$(dirname "$state")"
+out=$(timeout 900 ssh -o BatchMode=yes -o ConnectTimeout=15 ts-veron \
+  'cd /srv/windygit/src && timeout 850 python3 scripts/runner_guard.py report' || true)
+summary=$(grep '^# runner-guard sweep:' <<<"$out" || echo "# runner-guard sweep: ERROR (no summary)")
+hits=$(grep -v '^#' <<<"$out" | grep . || true)
+{
+  echo "# Runner guard: stranger-code paths to self-hosted runners ($(date -u '+%Y-%m-%d %H:%MZ'))"
+  echo "_Nightly; windy-git scripts/runner_guard.py. R1 pull_request_target · R2 fork PR on self-hosted without a same-repo/environment gate · R3 outsider events (issue_comment, workflow_run, ...) on self-hosted · R0 unparseable._"
+  echo; echo "${summary#\# }"; echo
+  echo "| repo | file:line | rule | fix |"; echo "|---|---|---|---|"
+  while IFS=$'\t' read -r repo loc rule msg; do [[ -n $repo ]] && echo "| $repo | $loc | $rule | $msg |"; done <<<"$hits"
+} > "$page"
+new=$(comm -13 <(sort -u "$state" 2>/dev/null || true) <(cut -f1-3 <<<"$hits" | sort -u))
+cut -f1-3 <<<"$hits" | sort -u > "$state"
+if [[ -n "$new" ]]; then
+  n=$(grep -c . <<<"$new")
+  echo "$(date -u +%Y-%m-%dT%H:%MZ) Windy Git: 🚨 runner-guard: $n NEW stranger-code path(s) to a self-hosted runner in public repos; see ~/windy-orchestra/RUNNER_GUARD.md" >> ~/windy-orchestra/BOARD.md
+fi
+echo "${summary#\# }"

deploy/windy0/windy-runner-guard-sweep.service

@@ -0,0 +1,7 @@
+[Unit]
+Description=Nightly runner-guard sweep of every PUBLIC repo's workflows (Windy Git lane)
+
+[Service]
+Type=oneshot
+ExecStart=/usr/local/bin/windy-job windy-runner-guard-sweep 26h --expect "runner-guard sweep: 0 hit" --owner 13 -- %h/bin/nightly-runner-guard-sweep.sh
+TimeoutStartSec=1200

deploy/windy0/windy-runner-guard-sweep.timer

@@ -0,0 +1,9 @@
+[Unit]
+Description=Nightly runner-guard sweep
+
+[Timer]
+OnCalendar=*-*-* 09:40:00 UTC
+Persistent=true
+
+[Install]
+WantedBy=timers.target

docker-compose.yml

@@ -32,7 +32,7 @@ services:
 
   gitea:
     # G2.1 — PIN AN EXACT VERSION. Never `latest`. Record it in SUBSTRATE.md.
-    image: docker.io/gitea/gitea:1.24.6
+    image: docker.io/gitea/gitea:1.24.7
     environment:
       GITEA__database__DB_TYPE: postgres
       GITEA__database__HOST: db:5432

docs/CUTOVER-PRIMARY.md

@@ -0,0 +1,31 @@
+# Checklist: making Windy Git the PRIMARY home of one repo (after launch)
+
+Status 2026-10-01: DRAFT, nothing flipped. GitHub stays the source of truth until Grant says otherwise.
+First candidate: `windy-git` itself (public, low blast radius). GitHub becomes the free off-site push-mirror.
+
+## Preconditions (all must be true)
+- [x] Encrypted state backup (Postgres + Gitea config + repos) nightly, restore drill passed cross-host (10-01).
+- [x] Gitea on a patched release (1.24.7); upgrade path = snapshot first (docs/RESTORE-DRILL.md).
+- [ ] Heartbeat `windygit-state-backup` in heartbeats-veron expected list (asked Cloud/Super Admin 10-01).
+- [ ] A missed/failed backup PAGES (heartbeat 26h), tested once by skipping a night in a drill.
+- [ ] Boss/Grant capacity call on Veron (dedicated non-SMR volume for DB + git storage; root disk < 80%).
+- [ ] Post-launch freeze lifted (Hub). No cutover during store review or a launch window.
+
+## Cutover for ONE repo (reversible at every step)
+1. Announce on BOARD; tell every consuming lane (no schema drift rule). Pause the sync for that repo only:
+   remove it from `REPOS` in `scripts/sync_from_github.sh` FIRST (the sync force-overwrites Windy Git).
+2. Verify Windy Git main == GitHub main (sha equal), all branches/tags present, LFS (if any) in R2.
+3. Add a PUSH-mirror on the Windy Git repo -> GitHub (deploy key or scoped token, write on that repo only,
+   interval 8h + on-commit), so GitHub keeps a current copy. Test with a throwaway branch.
+4. Flip the lanes' remote: `origin` = Windy Git (SSH/HTTPS via Windy SSO token), `github` = secondary.
+   Lanes push to Windy Git only; the mirror carries it to GitHub.
+5. CI keeps running here (no change); remove the `pr_status_bridge` mirror-PR duplication for that repo
+   (PRs are now native here; the bridge's GitHub status posting for it can stay for any open GitHub PRs).
+6. Watch 3 days: mirror lag, backup includes the new writes, no push rejected, no lane still pushing to GitHub.
+7. Rollback at any time: re-add the repo to `REPOS` (sync resumes GitHub->Windy Git, GitHub is intact via the
+   push-mirror) and point lanes' remote back. Nothing is destroyed in either direction.
+
+## NOT in scope for a first cutover
+Deploy workflows (stay disabled; deploys remain manual until a separate runner + scoped deploy keys exist),
+credential repos (soul/anima/kit-army-config), windy-pro (importer refuses by name), opening the forge to
+other members (Grant 09-23: registration closed; Hub 10-01: jit false, 4 conditions before any change).

docs/RESTORE-DRILL.md

@@ -0,0 +1,33 @@
+# Restoring Windy Git from the encrypted state backup
+
+What is backed up (`scripts/backup_state.sh`, restic repo `s3:…/windy-git-backups/restic`, tag `windygit-state`):
+both Postgres databases (`gitea`, `windygit`, custom-format dumps + globals), the whole Gitea data root
+(`/srv/windygit/git`: config, jwt, attachments, avatars, templates AND the bare repositories),
+`/srv/windygit/src/.env`, `deploy/runner/.env`, `/etc/cloudflared`, the windygit systemd drop-ins.
+NOT in it: the restic password itself (lockbox `RESTIC_WINDYGIT_PASSWORD`) and the R2 access key
+(scoped token `windy-git-r2-scoped`, lockbox). Never print either: use `lockbox-get KEY FILE`.
+
+## Drill (any machine with restic + docker; proven on Veron 2026-10-01, counts identical)
+    export RESTIC_PASSWORD_FILE=<0600 file from lockbox-get RESTIC_WINDYGIT_PASSWORD>
+    export AWS_ACCESS_KEY_ID=… AWS_SECRET_ACCESS_KEY=…            # from lockbox-get, into env, not echoed
+    export RESTIC_REPOSITORY=s3:https://<R2 account>.r2.cloudflarestorage.com/windy-git-backups/restic
+    restic snapshots --tag windygit-state
+    restic restore latest --tag windygit-state --target /var/tmp/wg-drill
+    docker run -d --name wg-drill-pg -e POSTGRES_PASSWORD=<random> -e POSTGRES_USER=drill postgres:16-alpine
+    for db in gitea windygit; do
+      docker exec wg-drill-pg psql -U drill -d postgres -c "create database $db"
+      docker exec -i wg-drill-pg pg_restore -U drill -d $db --no-owner --no-privileges \
+        < /var/tmp/wg-drill/var/backups/windygit-state/$db.dump
+    done
+    # compare row counts with live (or with the last known): repository, issue, pull_request, "user",
+    # external_login_user, access_token, action_run, action_run_job
+    docker rm -f wg-drill-pg; rm -rf /var/tmp/wg-drill
+
+## Real disaster (Veron lost)
+1. New Linux host with Docker, a Cloudflare tunnel connector, the repo (`git clone` from GitHub: windy-git).
+2. `restic restore latest --tag windygit-state --target /` (puts /srv/windygit/git, the .env files, /etc/cloudflared back).
+3. `docker compose -p windy-git up -d db`, then pg_restore both dumps into it (as above, into the real db names/owner from `.env`).
+4. `docker compose -p windy-git up -d` + `deploy/runner` runners; re-register runners if the token changed.
+5. Verify: `/api/healthz`, Windy SSO login, `git ls-remote`, one CI run. GitHub is still the source of truth for code,
+   so repo content can also be re-synced from there; the database is what only this backup holds.
+Retention: 14 daily / 8 weekly / 6 monthly (prune on Sundays). Integrity: every run does `restic check --read-data-subset=2%`.

scripts/backup_state.sh

@@ -22,7 +22,16 @@ GIT_ROOT="${GIT_DATA_ROOT:-/srv/windygit/git}"
 umask 077
 mkdir -p "$STAGE"; chmod 700 "$STAGE"; rm -f "$STAGE"/*.dump "$STAGE"/globals.sql
 
-restic cat config >/dev/null 2>&1 || { log "initialising restic repo"; restic init >/dev/null; }
+# systemd gives units no $HOME, and restic wants a cache dir: pin one.
+export RESTIC_CACHE_DIR="${RESTIC_CACHE_DIR:-/var/cache/windygit-restic}"; mkdir -p "$RESTIC_CACHE_DIR"
+if ! err=$(restic cat config 2>&1 >/dev/null); then
+  # only a MISSING repo may be initialised; any other error (auth, network, wrong password) must stop here
+  if grep -qiE "does not exist|is there a repository|unable to open config file" <<<"$err"; then
+    log "initialising restic repo"; restic init >/dev/null
+  else
+    log "FATAL: restic cannot open the repository: $(head -c 300 <<<"$err" | tr '\n' ' ')"; exit 1
+  fi
+fi
 
 PGU=$(timeout 30 docker exec "$DB" printenv POSTGRES_USER)
 [[ -n "$PGU" ]] || { log "FATAL: no POSTGRES_USER in $DB"; exit 1; }

scripts/compute_guard.py

@@ -63,6 +63,9 @@ JS_SDKS = (r"@anthropic-ai/sdk|openai|groq-sdk|@google/generative-ai|@google/gen
 
 RULES: list[tuple[str, re.Pattern]] = [
     ("provider host", re.compile("|".join(re.escape(h) for h in HOSTS))),
+    # Grant via Boss 10-01: compute = Windy Mind. A NEW reference to an Ollama port (Veron's :11434) is a
+    # direct call around Mind's metering/caps. WARN-only, never red, and only for lines a PR ADDS.
+    ("veron ollama", re.compile(r"(?::|%3[aA])11434(?![0-9])")),
     ("provider key", re.compile(r"\b(?:" + "|".join(KEYS) + r")\b")),
     ("provider SDK", re.compile(rf"^\s*(?:from|import)\s+(?:{PY_SDKS})(?:\s|\.|$|,)")),
     ("provider SDK", re.compile(rf"""(?:from\s+|require\(\s*|import\(\s*)['"](?:{JS_SDKS})(?:/[^'"]*)?['"]""")),
@@ -70,6 +73,9 @@ RULES: list[tuple[str, re.Pattern]] = [
     ("provider SDK dep", re.compile(rf'''^\s*"(?:{JS_SDKS})"\s*:''')),
     ("provider SDK dep", re.compile(rf'''^\s*["']?(?:{PY_SDKS.replace(chr(92) + ".", "-")})(?:\[[^\]]*\])?\s*(?:[<>=~!]=?|["',]|$)''')),
 ]
+# Kinds that never block (even in MODE=block) and are only judged on ADDED lines, never the baseline tree.
+WARN_ONLY_KINDS = {"veron ollama"}
+OLLAMA_MSG = "compute = Windy Mind (endpoint + key); do not call Veron's Ollama directly"
 DEP_FILES = re.compile(r"(^|/)(package\.json|requirements[^/]*\.txt|pyproject\.toml|setup\.cfg|Pipfile)$")
 
 # Never scanned: tests, docs, lockfiles, vendored/built code, CI config.
@@ -253,7 +259,8 @@ def check(repo: str, sha: str, default_branch: str, is_default_head: bool) -> li
     allow = load_allow()
     fp = _fingerprint(allow)
     if is_default_head:
-        return cached_scan(f"tree:{repo}:{sha}:{fp}", lambda: scan_tree(repo, bare, sha, allow))
+        return cached_scan(f"tree:{repo}:{sha}:{fp}",
+                           lambda: [f for f in scan_tree(repo, bare, sha, allow) if f.kind not in WARN_ONLY_KINDS])
     return cached_scan(
         f"pr:{repo}:{sha}:{fp}",
         lambda: scan_added(repo, bare, f"refs/heads/{default_branch}", sha, allow),
@@ -268,6 +275,11 @@ def status_for(findings: list[Finding], whole_tree: bool,
     Grant-owned code (ci/grant-owned.yml): always WARN, never red (orchestrator
     09-23: his desktop work is never blocked by us)."""
     scope = "in tree" if whole_tree else "added"
+    soft = [f for f in findings if f.kind in WARN_ONLY_KINDS]
+    findings = [f for f in findings if f.kind not in WARN_ONLY_KINDS]
+    if not findings and not grant and soft:
+        f = soft[0]
+        return "success", f"⚠ WARN: new Veron Ollama ref {f.path}:{f.line}. {OLLAMA_MSG}"[:140], f
     if not findings and grant:
         g, n = grant[0], len(grant)
         desc = (f"⚠ WARN (Grant-owned, not blocking): {n} direct AI-provider use{'s' if n > 1 else ''} "

scripts/drill_cross_host.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+# Cross-host restore drill on Windy 0: ONLY lockbox + R2, nothing from Veron. Values never printed.
+# Usage: bash drill_cross_host.sh   (needs lockbox keys RESTIC_WINDYGIT_PASSWORD, WINDYGIT_R2_ACCESS_KEY_ID, WINDYGIT_R2_SECRET_ACCESS_KEY, WINDYGIT_R2_ENDPOINT)
+set -euo pipefail
+umask 077; W=$(mktemp -d ~/.cache/wg-xdrill.XXXXXX)
+trap 'docker rm -f wg-xdrill-pg >/dev/null 2>&1 || true; rm -rf "$W"' EXIT
+lockbox-get RESTIC_WINDYGIT_PASSWORD "$W/pw" >/dev/null
+lockbox-get WINDYGIT_R2_ACCESS_KEY_ID "$W/ak" >/dev/null; lockbox-get WINDYGIT_R2_SECRET_ACCESS_KEY "$W/sk" >/dev/null; lockbox-get WINDYGIT_R2_ENDPOINT "$W/ep" >/dev/null
+export RESTIC_PASSWORD_FILE="$W/pw" AWS_ACCESS_KEY_ID="$(cat "$W/ak")" AWS_SECRET_ACCESS_KEY="$(cat "$W/sk")"
+export RESTIC_REPOSITORY="s3:$(cat "$W/ep")/windy-git-backups/restic"
+restic snapshots --tag windygit-state --compact | tail -3
+restic restore latest --tag windygit-state --target "$W/r" --include /var/backups/windygit-state --include /srv/windygit/git/gitea/conf --quiet
+ls -l "$W/r/var/backups/windygit-state" | awk 'NR>1{print $5, $NF}'
+docker run -d --name wg-xdrill-pg -e POSTGRES_PASSWORD="$(python3 -c 'import secrets;print(secrets.token_hex(12))')" -e POSTGRES_USER=drill postgres:16-alpine >/dev/null
+for i in $(seq 1 30); do docker exec wg-xdrill-pg pg_isready -U drill >/dev/null 2>&1 && break; sleep 2; done
+for db in gitea windygit; do
+  docker exec wg-xdrill-pg psql -U drill -d postgres -qc "create database $db"
+  docker exec -i wg-xdrill-pg pg_restore -U drill -d $db --no-owner --no-privileges < "$W/r/var/backups/windygit-state/$db.dump" 2>&1 | grep -v "already exists" | head -2 || true
+done
+for t in repository issue pull_request '"user"' external_login_user action_run action_run_job; do
+  echo "$t restored=$(docker exec wg-xdrill-pg psql -U drill -d gitea -Atc "select count(*) from $t")"
+done
+echo "cross-host drill OK (cleaned up)"

scripts/pr_status_bridge.py

@@ -54,7 +54,7 @@ REPOS = os.environ.get(
     " windy-drops windy-code-web windy-code windy-traveler windy-registry eternitas"
     " windy-translate windytranslate-site windytraveler-site windy-hand"
     " windy-cloud-sites windy-cloud-domains windy-cloud-vps windytalk windy-pro windy-mind"
-    " windy-inbox windy-text windy-call windy-cell",
+    " windy-inbox windy-text windy-call windy-cell windy-hand-site windy-calendar-site",
 ).split()
 
 # Gitea run status -> GitHub status state. `skipped` is deliberately absent: a

scripts/runner_guard.py

@@ -0,0 +1,209 @@
+#!/usr/bin/env python3
+"""Runner guard: no workflow may let a STRANGER's code reach a self-hosted runner (Boss 10-01).
+
+Our self-hosted GitHub runners run on Veron as `github-runner`, which is in the docker group
+(= root on Veron). Until ephemeral containerised runners exist (after launch), the cheap
+guard is to refuse the workflow shapes that hand a self-hosted runner to outsiders:
+
+  R1 pull_request_target            : runs with secrets/write token in the BASE repo context
+  R2 pull_request on self-hosted    : fork PRs run their own code, unless the job is gated to
+                                      same-repo heads (`if:` on head.repo.full_name == github.repository
+                                      or head.repo.fork == false) or an `environment:`
+  R3 issue_comment / workflow_run / issues / discussion* / pull_request_review* / fork / watch
+                                    : anyone can fire these; never on self-hosted without an environment gate
+  (push, tags, schedule, workflow_dispatch, repository_dispatch, workflow_call: writers only, fine)
+
+A job counts as self-hosted when its runs-on names `self-hosted`, or is an expression we
+can't resolve (conservative). Findings carry file:line, never file content beyond that.
+
+  python3 scripts/runner_guard.py lint FILE...              # local files
+  python3 scripts/runner_guard.py report [--owners a,b]     # default branch of every PUBLIC repo
+  python3 scripts/runner_guard.py pr [--owners a,b] [--post]  # open PRs on public repos: changed workflows
+"""
+from __future__ import annotations
+
+import argparse
+import base64
+import json
+import os
+import subprocess
+import sys
+
+import yaml
+
+OWNERS = ["sneakyfree", "VERONTECH", "Windstorm-Institute", "Windstorm-Labs", "Public-Streamer"]
+CTX = "windy-git/runner-guard"
+OUTSIDE = {"issue_comment", "workflow_run", "issues", "discussion", "discussion_comment",
+           "pull_request_review", "pull_request_review_comment", "fork", "watch"}
+SAME_REPO_GATES = ("head.repo.full_name == github.repository", "github.repository == github.event.pull_request.head.repo.full_name",
+                   "head.repo.fork == false", "!github.event.pull_request.head.repo.fork")
+
+
+def _node_map(node):
+    """{key: (value_node, line)} for a YAML mapping node."""
+    if not isinstance(node, yaml.MappingNode):
+        return {}
+    return {k.value: (v, k.start_mark.line + 1) for k, v in node.value if isinstance(k, yaml.ScalarNode)}
+
+
+def _triggers(on_node) -> dict[str, int]:
+    """{event: line}."""
+    if isinstance(on_node, yaml.ScalarNode):
+        return {on_node.value: on_node.start_mark.line + 1}
+    if isinstance(on_node, yaml.SequenceNode):
+        return {n.value: n.start_mark.line + 1 for n in on_node.value if isinstance(n, yaml.ScalarNode)}
+    return {k: line for k, (_v, line) in _node_map(on_node).items()}
+
+
+def _self_hosted(runs_on) -> bool:
+    if runs_on is None:
+        return False
+    text = yaml.serialize(runs_on) if isinstance(runs_on, yaml.Node) else str(runs_on)
+    return "self-hosted" in text or "${{" in text
+
+
+def lint_text(path: str, text: str) -> list[tuple[str, int, str, str]]:
+    """[(path, line, rule, message)]. Unparseable YAML is a finding (it cannot be reviewed)."""
+    try:
+        root = yaml.compose(text)
+    except yaml.YAMLError as e:
+        line = getattr(getattr(e, "problem_mark", None), "line", 0) + 1
+        return [(path, line, "R0", "workflow YAML does not parse; cannot be checked")]
+    top = _node_map(root)
+    if "on" not in top or "jobs" not in top:
+        return []
+    trig = _triggers(top["on"][0])
+    jobs = _node_map(top["jobs"][0])
+    out = []
+    if "pull_request_target" in trig:
+        out.append((path, trig["pull_request_target"], "R1",
+                    "pull_request_target runs fork code with base-repo secrets; not allowed"))
+    for name, (jnode, jline) in jobs.items():
+        j = _node_map(jnode)
+        ro = j.get("runs-on", (None, jline))
+        if not _self_hosted(ro[0]):
+            continue
+        has_env = "environment" in j
+        cond = j["if"][0].value if "if" in j and isinstance(j["if"][0], yaml.ScalarNode) else ""
+        same_repo = any(g in cond.replace("  ", " ") for g in SAME_REPO_GATES)
+        if "pull_request" in trig and not (has_env or same_repo):
+            out.append((path, ro[1], "R2", f"job '{name}' runs fork PR code on a self-hosted runner "
+                        "(gate it: if: github.event.pull_request.head.repo.full_name == github.repository, or an environment)"))
+        for ev in sorted(OUTSIDE & trig.keys()):
+            if not has_env:
+                out.append((path, trig[ev], "R3", f"'{ev}' can be fired by anyone and job '{name}' is self-hosted "
+                            "without an environment gate"))
+    return out
+
+
+# ---------------------------------------------------------------- GitHub side
+def gh(*args: str, check=True) -> str:
+    r = subprocess.run(["gh", "api", *args], capture_output=True, text=True, timeout=60)
+    if check and r.returncode != 0:
+        raise RuntimeError(f"gh api {args[0]} failed")
+    return r.stdout
+
+
+def public_repos(owners) -> list[tuple[str, str]]:
+    out = []
+    for o in owners:
+        txt = gh(f"users/{o}/repos?per_page=100&type=owner", "--paginate",
+                 "--jq", '.[]|select(.private==false and .archived==false)|.full_name+" "+.default_branch', check=False)
+        out += [tuple(row.split()) for row in txt.splitlines() if row.strip()]
+    return out
+
+
+def workflows_at(full: str, ref: str) -> list[tuple[str, str]]:
+    txt = gh(f"repos/{full}/contents/.github/workflows?ref={ref}", "--jq",
+             '.[]|select(.type=="file")|.path', check=False)
+    files = [p for p in txt.splitlines() if p.endswith((".yml", ".yaml"))]
+    return [(p, file_at(full, p, ref)) for p in files]
+
+
+def file_at(full: str, path: str, ref: str) -> str:
+    raw = gh(f"repos/{full}/contents/{path}?ref={ref}", "--jq", ".content", check=False).strip()
+    return base64.b64decode(raw).decode("utf-8", "replace") if raw else ""
+
+
+def cmd_report(owners) -> int:
+    hits = 0
+    repos = public_repos(owners)
+    for full, branch in repos:
+        for path, text in workflows_at(full, branch):
+            for p, line, rule, msg in lint_text(path, text):
+                hits += 1
+                print(f"{full}\t{p}:{line}\t{rule}\t{msg}")
+    print(f"# runner-guard sweep: {hits} hit(s) in {len(repos)} public repos")
+    return 1 if hits else 0
+
+
+STATE = os.environ.get("RUNNER_GUARD_STATE", "/var/lib/windy-git/runner-guard-posted.json")
+
+
+def cmd_pr(owners, post: bool) -> int:
+    # Post each (repo, sha, state, description) ONCE: the sync runs every 5 min and GitHub caps
+    # statuses per sha+context at 1000.
+    try:
+        with open(STATE) as fh:
+            posted = set(json.load(fh))
+    except (OSError, ValueError):
+        posted = set()
+    seen = set()
+    for full, _branch in public_repos(owners):
+        prs = gh(f"repos/{full}/pulls?state=open&per_page=50", "--jq",
+                 '.[]|(.number|tostring)+" "+.head.sha+" "+.head.repo.full_name', check=False)
+        for line in prs.splitlines():
+            num, sha, head_repo = line.split(" ", 2)
+            files = gh(f"repos/{full}/pulls/{num}/files?per_page=100", "--jq",
+                       '.[]|select(.status!="removed")|.filename', check=False).split()
+            wf = [f for f in files if f.startswith(".github/workflows/") and f.endswith((".yml", ".yaml"))]
+            # a fork's own content is read from the head repo at the head sha
+            src = head_repo if head_repo and head_repo != "null" else full
+            found = [h for f in wf for h in lint_text(f, file_at(src, f, sha))]
+            if found:
+                p, ln, rule, msg = found[0]
+                state, desc = "failure", f"BLOCKED: {rule} {p}:{ln}: {msg}"[:140]
+            else:
+                state, desc = "success", ("OK: no workflow changes" if not wf else
+                                          "OK: no stranger-code path to a self-hosted runner")
+            key = f"{full}@{sha}:{state}:{desc}"
+            seen.add(key)
+            if key in posted:
+                continue
+            print(f"{full}#{num}@{sha[:7]} {state} {desc}")
+            if post:
+                gh(f"repos/{full}/statuses/{sha}", "-f", f"state={state}", "-f", f"context={CTX}",
+                   "-f", f"description={desc}", check=False)
+                posted.add(key)
+    if post:  # keep only keys for PRs still open, so the file never grows without bound
+        os.makedirs(os.path.dirname(STATE), exist_ok=True)
+        with open(STATE, "w") as fh:
+            json.dump(sorted(posted & seen), fh)
+    return 0
+
+
+def main(argv=None) -> int:
+    ap = argparse.ArgumentParser(prog="runner_guard")
+    sub = ap.add_subparsers(dest="cmd", required=True)
+    lint_p = sub.add_parser("lint")
+    lint_p.add_argument("files", nargs="+")
+    rep = sub.add_parser("report")
+    rep.add_argument("--owners", default=",".join(OWNERS))
+    prp = sub.add_parser("pr")
+    prp.add_argument("--owners", default="sneakyfree")  # self-hosted runners exist only there
+    prp.add_argument("--post", action="store_true")
+    a = ap.parse_args(argv)
+    if a.cmd == "lint":
+        hits = []
+        for f in a.files:
+            with open(f, errors="replace") as fh:
+                hits += lint_text(f, fh.read())
+        for p_, ln, rule, msg in hits:
+            print(f"{p_}:{ln}\t{rule}\t{msg}")
+        return 1 if hits else 0
+    owners = a.owners.split(",")
+    return cmd_report(owners) if a.cmd == "report" else cmd_pr(owners, a.post)
+
+
+if __name__ == "__main__":
+    sys.exit(main())

scripts/secret_shapes.py

@@ -23,13 +23,14 @@ PATTERNS: list[tuple[str, re.Pattern[str]]] = [
     ("twilio sid/api key", re.compile(r"\b(?:AC|SK)[0-9a-f]{32}\b")),
     ("32-hex secret assignment", re.compile(
         r"(?i)\b[a-z0-9_.-]*(?:token|secret|key|password)[a-z0-9_.-]*[\"']?\s*[:=]\s*[\"']?(?P<v>(?<![0-9a-f])[0-9a-f]{32}(?![0-9a-f]))")),
+    ("pypi token", re.compile(r"\bpypi-AgE[A-Za-z0-9_-]{50,}")),
     ("private key block", re.compile(r"-----BEGIN (?:RSA |EC |DSA |OPENSSH |ENCRYPTED )?PRIVATE KEY-----")),
 ]
 
 # git grep -E (POSIX ERE) prefilter: cheap superset of PATTERNS.
 PREFILTER = ("[0-9]{8,10}:[A-Za-z0-9_-]{35}|gh[pousr]_[A-Za-z0-9]{36}|github_pat_|(AKIA|ASIA)[0-9A-Z]{16}"
              "|xox[abprs]-|sk-ant-|sk-[A-Za-z0-9_-]{32}|sk-proj-|[rs]k_live_|AIza[0-9A-Za-z_-]{35}"
-             "|-----BEGIN [A-Z ]*PRIVATE KEY-----|(AC|SK)[0-9a-f]{32}|[0-9a-fA-F]{32}")
+             "|-----BEGIN [A-Z ]*PRIVATE KEY-----|(AC|SK)[0-9a-f]{32}|[0-9a-fA-F]{32}|pypi-AgE")
 
 
 def h8(value: str | bytes) -> str:

scripts/sync_from_github.sh

@@ -38,7 +38,7 @@ FAILED=0
 
 # Repos Windy Git tracks FROM GitHub. Remove a repo from this list at the moment
 # it flips to Windy-Git-first, or the sync will fight its authors and win.
-REPOS="${SYNC_REPOS:-windy-calendar windy-search windy-registry Windy-Clone WindyCloud windy-cloud-sites windy-mind eternitas windy-agent windy-git windy-chat windy-mail windy-connect windy-drops windy-code-web windy-code windy-traveler windy-translate windytranslate-site windytraveler-site windy-hand windy-cloud-domains windy-cloud-vps windytalk windy-pro windy-inbox windy-text windy-call windy-cell}"
+REPOS="${SYNC_REPOS:-windy-calendar windy-search windy-registry Windy-Clone WindyCloud windy-cloud-sites windy-mind eternitas windy-agent windy-git windy-chat windy-mail windy-connect windy-drops windy-code-web windy-code windy-traveler windy-translate windytranslate-site windytraveler-site windy-hand windy-cloud-domains windy-cloud-vps windytalk windy-pro windy-inbox windy-text windy-call windy-cell windy-hand-site windy-calendar-site}"
 
 # Repos whose TAGS must not reach Windy Git. A tag push fires `on: push: tags`
 # workflows; windy-pro's build-electron is a matrix over ubuntu/macos/windows-
@@ -94,6 +94,11 @@ if ! python3 "$(dirname "$0")/pr_status_bridge.py"; then
   log "FAILED pr status bridge"; FAILED=1
 fi
 
+# Runner guard (Boss 10-01): PUBLIC sneakyfree repos have self-hosted GitHub runners on Veron.
+# A PR that changes a workflow so a stranger's code could reach one gets a red
+# windy-git/runner-guard status. Each status is posted once; never fails the sync.
+timeout -k 10 120 python3 "$(dirname "$0")/runner_guard.py" pr --post || log "runner-guard failed or timed out (non-fatal)"
+
 # CI telemetry -> admin.windyword.ai (shapes declared with Windy Telemetry 40).
 # Sends nothing until WINDYGIT_TELEMETRY_TOKEN is set; never fails the sync.
 timeout -k 10 180 python3 "$(dirname "$0")/telemetry_emit.py" || log "telemetry emit failed or timed out (non-fatal)"