compute-guard: drop :8099 from the Talk-engine ports (false positive); Deepgram template line under the BYOK exemption

Hub classification 10-02: :8099 in windy-pro is the OLD translate-api / cloud-storage service, not Windy Talk
(verified: windytalk has no :8099, only lockfile hash fragments). DEEPGRAM_API_KEY in windy-pro .env.example is
a template line for the user-own-key Deepgram feature: owner-approved, approved_by windy-hub, expires 2026-12-31.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

.github/CODEOWNERS

@@ -0,0 +1,7 @@
+# Changes to the guard allow-lists / exemptions need review by the account owner on GitHub, AND an
+# approved_by (windy-hub | windy-mind) on every non-structural entry, which the guard enforces itself
+# (scripts/compute_guard.py: load_allow). A lane never approves its own exemption (Hub 10-02).
+/ci/compute-guard-allow.yml   @sneakyfree
+/ci/secret-guard-allow.yml    @sneakyfree
+/ci/ci-hygiene-allow.yml      @sneakyfree
+/scripts/compute_guard.py     @sneakyfree

api/tests/test_ci_hygiene.py

@@ -86,7 +86,7 @@ def test_warn_mode_never_turns_red(monkeypatch):
 def test_allow_file_is_line_scoped_exceptions_only():
     """Every exception is line-scoped (`matches`), so an allowed file can't hide a
     NEW floating install or docker step. Today: windy-pro's if:false deploy job."""
-    allow = hy.cg.load_allow(hy.ALLOW_FILE)
+    allow = hy.cg.load_allow(hy.ALLOW_FILE, strict=False)
     assert [(e["repo"], e["paths"]) for e in allow] == [("windy-pro", [".github/workflows/ci.yml"])]
     assert all(e.get("matches") for e in allow)
     ok = "        run: docker build -f account-server/Dockerfile -t windy-pro:${{ github.sha }} ."

api/tests/test_compute_guard.py

@@ -77,6 +77,88 @@ def test_allow_list_needs_a_reason_per_entry(tmp_path):
         cg.load_allow(bad)
 
 
+def _entry(**kw):
+    base = dict(repo="x", paths=["*"], reason="r", exemption="owner-approved", expires="2099-01-01",
+                approved_by="windy-hub")
+    base.update(kw)
+    lines = ["allow:", "  - repo: x", "    paths: ['*']", "    reason: r"]
+    for k in ("exemption", "expires", "approved_by"):
+        if base.get(k) is not None:
+            lines.append(f"    {k}: {base[k]}")
+    return "\n".join(lines) + "\n"
+
+
+def test_allow_entries_need_a_named_exemption_and_an_expiry(tmp_path):
+    from datetime import date
+    f = tmp_path / "a.yml"
+    f.write_text(_entry(exemption=None))
+    with pytest.raises(ValueError):
+        cg.load_allow(f)
+    f.write_text(_entry(exemption="because-i-said-so"))
+    with pytest.raises(ValueError):
+        cg.load_allow(f)
+    f.write_text(_entry(expires=None))
+    with pytest.raises(ValueError):
+        cg.load_allow(f)
+    f.write_text(_entry(expires="someday"))
+    with pytest.raises(ValueError):
+        cg.load_allow(f)
+    f.write_text(_entry(expires="2026-12-01"))
+    assert len(cg.load_allow(f, today=date(2026, 10, 2))) == 1
+
+
+def test_expired_exemption_stops_excusing_and_is_reported(tmp_path):
+    from datetime import date
+    f = tmp_path / "a.yml"
+    f.write_text(_entry(expires="2026-10-01"))
+    assert cg.load_allow(f, today=date(2026, 10, 1))        # the expiry day is still valid
+    assert cg.load_allow(f, today=date(2026, 10, 2)) == []   # the next day it no longer excuses anything
+    assert cg.EXPIRED and cg.EXPIRED[0]["repo"] == "x" and cg.EXPIRED[0]["expires"] == "2026-10-01"
+
+
+def test_shipped_allow_file_is_valid_today():
+    assert cg.load_allow() and not cg.EXPIRED  # nothing in the repo's own file may already be expired
+
+
+@pytest.mark.parametrize("text, kind", [
+    ('u = "https://api.deepgram.com/v1/listen"', "voice-ai host"),
+    ("fetch(`https://api.elevenlabs.io/v1/tts`)", "voice-ai host"),
+    ('h = "api.cartesia.ai"', "voice-ai host"),
+    ('h = "app.resemble.ai"', "voice-ai host"),
+    ('h = "speech.googleapis.com"', "voice-ai host"),
+    ('h = "transcribe.us-east-1.amazonaws.com"', "voice-ai host"),
+    ('h = "polly.eu-west-1.amazonaws.com"', "voice-ai host"),
+    ("DEEPGRAM_API_KEY=abc", "voice-ai key"),
+    ("ELEVENLABS_API_KEY = x", "voice-ai key"),
+    ('u = f"https://api.cloudflare.com/client/v4/accounts/{a}/ai/run/@cf/m"', "cloudflare workers ai"),
+    ('ENGINE = "http://10.0.0.5:8791/v1"', "talk engine port"),
+    ('ENGINE = "http://h:8788/ws"', "talk engine port"),
+    ('x = "http://h:8794/health"', "talk engine port"),
+])
+def test_gatekeeper_rules_fire(text, kind):
+    assert kind in [k for k, _ in cg.scan_line("app/x.py", text)]
+
+
+def test_workers_ai_binding_only_in_wrangler_and_near_misses_are_quiet():
+    assert [k for k, _ in cg.scan_line("wrangler.toml", "[ai]")] == ["workers ai binding"]
+    assert [k for k, _ in cg.scan_line("apps/x/wrangler.jsonc", '  "ai": {')] == ["workers ai binding"]
+    assert cg.scan_line("other.toml", "[ai]") == []
+    assert cg.scan_line("app/x.py", "port = 87912") == []
+    assert cg.scan_line("app/x.py", "# talk engine was :8791 (removed)") == []
+
+
+def test_rolling_out_kinds_warn_but_dont_block_and_hard_kinds_still_do(monkeypatch):
+    monkeypatch.setattr(cg, "MODE", "block")
+    monkeypatch.setattr(cg, "SOFT_KINDS", {"voice-ai host", "voice-ai key"})
+    soft = cg.Finding("a.py", 1, "voice-ai host", "api.deepgram.com")
+    hard = cg.Finding("a.py", 2, "provider host", "api.openai.com")
+    state, desc, _ = cg.status_for([soft], whole_tree=True)
+    assert state == "success" and "rolling out" in desc and len(desc) <= 140
+    assert cg.status_for([soft, hard], whole_tree=True)[0] == "failure"
+    monkeypatch.setattr(cg, "SOFT_KINDS", set())
+    assert cg.status_for([soft], whole_tree=True)[0] == "failure"   # rollout over: it blocks
+
+
 @pytest.mark.parametrize(
     "repo, path, ok",
     [
@@ -250,3 +332,41 @@ def test_ollama_in_added_pr_lines_only():
             "+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)]
+
+
+def test_non_structural_exemptions_need_an_independent_approver_and_a_90_day_cap(tmp_path):
+    from datetime import date
+    f = tmp_path / "a.yml"
+    f.write_text(_entry(approved_by=None))
+    with pytest.raises(ValueError):
+        cg.load_allow(f, today=date(2026, 10, 2))
+    f.write_text(_entry(approved_by="windy-chat"))     # a lane may not approve itself/another lane
+    with pytest.raises(ValueError):
+        cg.load_allow(f, today=date(2026, 10, 2))
+    f.write_text(_entry(expires="2026-12-31"))           # exactly 90 days: fine
+    assert len(cg.load_allow(f, today=date(2026, 10, 2))) == 1
+    f.write_text(_entry(expires="2027-01-01"))           # 91 days: does NOT apply, and is reported
+    assert cg.load_allow(f, today=date(2026, 10, 2)) == [] and cg.OVERCAP
+    f.write_text(_entry(exemption="compute-door", approved_by=None, expires="2027-10-02"))
+    assert len(cg.load_allow(f, today=date(2026, 10, 2))) == 1   # structural: yearly, no approver field
+
+
+def test_shipped_allow_file_obeys_its_own_rules():
+    allow = cg.load_allow()
+    assert allow and not cg.EXPIRED and not cg.OVERCAP
+    for e in allow:
+        if e["exemption"] not in cg.STRUCTURAL:
+            assert e["approved_by"] in cg.APPROVERS
+
+
+def test_findings_carry_kind_and_name_never_the_value_and_ports_skip_contracts():
+    for line, kind in [("ELEVENLABS_API_KEY=sk_live_SUPERSECRET123456789", "voice-ai key"),
+                       ('DEEPGRAM_API_KEY = "dg-VALUE-0123456789abcdef"', "voice-ai key")]:
+        hits = cg.scan_line("app/x.py", line)
+        assert [k for k, _ in hits] == [kind]
+        assert all("SUPERSECRET" not in m and "VALUE" not in m for _, m in hits)
+    assert cg.scan_line("engine/contracts/ops.mcp.v1.json", '"url": "http://h:8791/x"') == []
+    assert cg.scan_line("services/api/openapi/spec.json", '"url": "http://h:8791/x"') == []
+    assert [k for k, _ in cg.scan_line("deploy/docker-compose.yml", "    - 8791:8791 # :8791")] == ["talk engine port"]
+    # :8099 is windy-pro's OLD translate-api / cloud-storage service, NOT the Talk engine (Hub 10-02): not flagged
+    assert cg.scan_line("deploy/docker-compose.yml", "    - 8099:8099 # translate-api:8099") == []

api/tests/test_lockbox_names.py

@@ -0,0 +1,64 @@
+"""lockbox-names: headings + labels + resolvable, NEVER a value or prose after a label."""
+
+from __future__ import annotations
+
+import os
+import subprocess
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[2]
+V1 = "Qm7xT2vLp9RkZw4HnB8dYc3S"   # synthetic values
+V2 = "other-fake-value-1234567890"
+V3 = "dup-one-aaaaaaaaaaaaaaaa"
+V4 = "dup-two-bbbbbbbbbbbbbbbb"
+PROSE = "ZZPROSEZZ-not-for-printing"
+
+
+def make(tmp_path):
+    r = tmp_path / "kit"
+    (r / "secrets" / "x").mkdir(parents=True)
+    (r / "ACCESS_LOCKBOX.md").write_text(
+        "# LOCKBOX\n\n## 🔷 AZURE signing (added 10-01)\n"
+        f"- **Tenant:** {PROSE} lives in the portal\n"
+        f"- **`AZURE_CLIENT_ID`**: `{V1}`\n"
+        f"- **Secret (AZURE_CLIENT_SECRET):** `{V2}`\n"
+        "## GOOGLE oauth\n"
+        f"GOOGLE_OAUTH_CLIENT_ID={V2}\n"
+        f"- **`DUP_KEY`**: `{V3}`\n- **`DUP_KEY`**: `{V4}`\n"
+        f"## stray\n**{V1}** is a heading-like bold that is secret shaped? no, just label\n")
+    (r / "secrets" / "x" / "a.env").write_text(f"FILE_KEY={V1}\n")
+    subprocess.run(["git", "init", "-q", "-b", "main"], cwd=r, check=True)
+    return r
+
+
+def run(r, *args):
+    e = {**os.environ, "LOCKBOX_REPO": str(r), "LOCKBOX_REF": "WORKTREE"}
+    p = subprocess.run([sys.executable, str(ROOT / "scripts" / "lockbox_names.py"), *args],
+                       capture_output=True, text=True, env=e)
+    return p.returncode, p.stdout + p.stderr
+
+
+def test_lists_labels_and_resolvable_without_values_or_prose(tmp_path):
+    r = make(tmp_path)
+    rc, out = run(r, "AZURE|GOOGLE|FILE|DUP")
+    assert rc == 0
+    assert "AZURE signing (added 10-01) | AZURE_CLIENT_ID | md | yes" in out
+    assert "| GOOGLE_OAUTH_CLIENT_ID | env | yes" in out
+    assert "| FILE_KEY | file | yes" in out
+    assert "| DUP_KEY | md | dup" in out
+    # a prose label is listed but never resolvable, and nothing after the label leaks
+    assert "| Tenant: " not in out or "| Tenant" in out
+    assert "| label | no" in out
+    for secret in (V1, V2, V3, V4, PROSE, "lives in the portal"):
+        assert secret not in out
+        for i in range(0, len(secret) - 7):
+            assert secret[i:i + 8] not in out
+
+
+def test_filter_and_bad_regex(tmp_path):
+    r = make(tmp_path)
+    rc, out = run(r, "NOSUCHTHING")
+    assert rc == 0 and "0 entries" in out
+    rc, out = run(r, "(")
+    assert rc == 2 and "bad regex" in out

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"]

ci/compute-guard-allow.yml

@@ -2,11 +2,16 @@
 # Windy Mind is the ONLY door to AI compute (Grant, 2026-09-23). Every entry
 # here is an exception to that rule and MUST say why. Paths are fnmatch globs
 # relative to the repo root. Owner of this file: Windy Git lane (13); changes
-# go through the orchestrator. Source of the first entries: COMPUTE_BYPASS_AUDIT.md.
+# go through the orchestrator. EVERY entry needs `exemption` (local-user-hardware | owner-approved |
+# compute-door | guard-self) and `expires` (YYYY-MM-DD): nothing gets permanent amnesty (Mind 10-02);
+# non-structural exemptions also need approved_by (windy-hub | windy-mind; a lane never approves its own)
+# and expire within 90 days; an expired entry stops excusing code on that date and shows in the guard report. Source of the first entries: COMPUTE_BYPASS_AUDIT.md.
 allow:
   - repo: windy-mind
     paths: ["*"]
     reason: "Windy Mind IS the door: provider clients belong here by definition."
+    exemption: compute-door
+    expires: 2027-10-02
 
   - repo: windy-agent
     paths: ["*"]
@@ -14,14 +19,26 @@ allow:
       User BYOK: self-hosted agents call providers on the USER's own keys.
       Mind stays opt-in there, or every self-hosted user's inference lands on
       Grant's bill (no-cloud-cost-liability rule; audit #7).
+    exemption: owner-approved
+    approved_by: windy-hub
+    approved_on: 2026-10-02
+    expires: 2026-12-31
 
   - repo: windy-code
     paths: ["extensions/windy-ai/*"]
     reason: "User BYOK AI extension: the user's own provider keys; Mind is one opt-in provider (audit #8)."
+    exemption: owner-approved
+    approved_by: windy-hub
+    approved_on: 2026-10-02
+    expires: 2026-12-31
 
   - repo: windy-connect
     paths: ["*writers/*"]
     reason: "Writes client configs that NAME the user's own provider env vars; makes no provider calls (audit #11)."
+    exemption: owner-approved
+    approved_by: windy-hub
+    approved_on: 2026-10-02
+    expires: 2026-12-31
 
   - repo: windy-pro
     paths: ["src/client/desktop/*"]
@@ -30,10 +47,18 @@ allow:
       enters (renderer localStorage -> electron-store; env var only for dev), and
       the CSP line allows exactly those user-keyed hosts (audit #10). The
       account-server is NOT covered: server-side calls go through Mind.
+    exemption: owner-approved
+    approved_by: windy-hub
+    approved_on: 2026-10-02
+    expires: 2026-12-31
 
   - repo: windy-pro
     paths: ["src/client/web/src/pages/panels/MindPanel.jsx"]
     reason: "Validates the USER's own OpenRouter key for BYOK (audit #10); spends no house money."
+    exemption: owner-approved
+    approved_by: windy-hub
+    approved_on: 2026-10-02
+    expires: 2026-12-31
 
   - repo: windy-pro
     paths: ["src/client/web/src/pages/panels/MindKeychain.jsx"]
@@ -41,12 +66,29 @@ allow:
     # /api/v1/chat, i.e. inference) still flags. Orchestrator-approved 09-23.
     matches: ['openrouter\.ai/auth\?', 'openrouter\.ai/api/v1/auth/keys']
     reason: "BYOK key acquisition via OpenRouter OAuth PKCE; no inference; successor of MindPanel allow (ADR-064)."
+    exemption: owner-approved
+    approved_by: windy-hub
+    approved_on: 2026-10-02
+    expires: 2026-12-31
 
   - repo: windy-git
     paths: ["scripts/compute_guard.py", "ci/compute-guard-allow.yml"]
     reason: "The guard's own pattern list and this file."
+    exemption: guard-self
+    expires: 2027-10-02
 
   - 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."
+    exemption: compute-door
+    expires: 2027-10-02
+
+  - repo: windy-pro
+    paths: [".env.example"]
+    matches: ['DEEPGRAM_API_KEY']
+    reason: "Template line (name only, no value) for the existing user-own-key Deepgram feature in Word's Settings. Mind's migration removes the feature and then this line (Hub classification 10-02)."
+    exemption: owner-approved
+    approved_by: windy-hub
+    approved_on: 2026-10-02
+    expires: 2026-12-31

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/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

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).

scripts/ci_hygiene.py

@@ -204,7 +204,7 @@ def _check(repo: str, sha: str, default_branch: str, is_default_head: bool):
     bare = cg.WORK / f"{repo}.git"
     if not bare.is_dir() or not cg.fetched(bare, sha):  # pushed after the fetch: next cycle
         return None
-    allow = cg.load_allow(ALLOW_FILE)
+    allow = cg.load_allow(ALLOW_FILE, strict=False)
     rules = hashlib.sha256((PREFILTER + INCLUDE.pattern + EXACT_PY.pattern + EXACT_NPM.pattern).encode()).hexdigest()[:8]
     fp = cg._fingerprint(allow) + ":" + rules  # hashlib, not hash(): hash() is per-process random
     kw = dict(line_fn=scan_line, path_ok=path_ok)
@@ -232,7 +232,7 @@ def status_for(findings, whole_tree: bool, grant=()):
 
 
 def report(repos: list[str]) -> int:
-    allow = cg.load_allow(ALLOW_FILE)
+    allow = cg.load_allow(ALLOW_FILE, strict=False)
     total = 0
     for repo in repos:
         bare = cg.WORK / f"{repo}.git"

scripts/compute_guard.py

@@ -31,6 +31,7 @@ import re
 import subprocess
 import sys
 from dataclasses import dataclass
+from datetime import date, timedelta
 from pathlib import Path
 
 import yaml
@@ -49,6 +50,17 @@ HOSTS = [
     "api.cohere.com", "api.fireworks.ai", "api.replicate.com",
     "api-inference.huggingface.co",
 ]
+# Mind's 10-02 gatekeeper list (speech / voice / avatar / vision / cloud ML): own kinds, so a rollout
+# can be WARN-first (COMPUTE_GUARD_WARN_KINDS) without softening the original provider rules.
+VOICE_HOSTS = [
+    "api.deepgram.com", "api.elevenlabs.io", "api.cartesia.ai", "api.play.ht", "api.playht.com",
+    "app.resemble.ai", "f.cluster.resemble.ai", "api.heygen.com",
+    "vision.googleapis.com", "speech.googleapis.com", "texttospeech.googleapis.com",
+]
+VOICE_KEYS = [
+    "DEEPGRAM_API_KEY", "ELEVENLABS_API_KEY", "ELEVEN_API_KEY", "CARTESIA_API_KEY", "PLAYHT_API_KEY",
+    "PLAY_HT_API_KEY", "PLAYHT_USER_ID", "RESEMBLE_API_KEY", "HEYGEN_API_KEY",
+]
 KEYS = [
     "ANTHROPIC_API_KEY", "ANTHROPIC_OAUTH_TOKEN", "ANTHROPIC_AUTH_TOKEN",
     "OPENAI_API_KEY", "GROQ_API_KEY", "GEMINI_API_KEY", "GOOGLE_GENERATIVE_AI_API_KEY",
@@ -61,11 +73,23 @@ PY_SDKS = r"anthropic|openai|groq|mistralai|cohere|google\.generativeai|google\.
 JS_SDKS = (r"@anthropic-ai/sdk|openai|groq-sdk|@google/generative-ai|@google/genai|@mistralai/mistralai"
            r"|cohere-ai|together-ai|@ai-sdk/(?:anthropic|openai|groq|google|mistral)")
 
+# The engine-port rule is about CODE/CONFIG that calls the engine, not API contracts, schemas or specs.
+PORT_SKIP = re.compile(r"(^|/)(contracts?|schemas?|specs?|openapi)/|\.json$", re.I)
+WRANGLER = re.compile(r"(^|/)wrangler\.(toml|jsonc?)$")
+WRANGLER_AI = re.compile(r'^\s*\[ai\]\s*$|^\s*"ai"\s*:\s*\{')
+
 RULES: list[tuple[str, re.Pattern]] = [
     ("provider host", re.compile("|".join(re.escape(h) for h in HOSTS))),
     # Grant via Boss 10-01: compute = Windy Mind. A NEW reference to an Ollama port (Veron's :11434) is a
     # direct call around Mind's metering/caps. WARN-only, never red, and only for lines a PR ADDS.
     ("veron ollama", re.compile(r"(?::|%3[aA])11434(?![0-9])")),
+    ("voice-ai host", re.compile("|".join(re.escape(h) for h in VOICE_HOSTS))),
+    ("voice-ai key", re.compile(r"\b(?:" + "|".join(VOICE_KEYS) + r")\b")),
+    ("voice-ai host", re.compile(r"(?:transcribe|polly)\.[a-z0-9-]+\.amazonaws\.com")),
+    ("provider host", re.compile(r"(?:bedrock-runtime|bedrock)\.[a-z0-9-]+\.amazonaws\.com")),
+    ("cloudflare workers ai", re.compile(r"api\.cloudflare\.com/client/v4/accounts/[^\s'\"/]+/ai/")),
+    ("talk engine port", re.compile(r"(?::|%3[aA])(?:8791|8788|8794)(?![0-9])")),
+    ("workers ai binding", WRANGLER_AI),
     ("provider key", re.compile(r"\b(?:" + "|".join(KEYS) + r")\b")),
     ("provider SDK", re.compile(rf"^\s*(?:from|import)\s+(?:{PY_SDKS})(?:\s|\.|$|,)")),
     ("provider SDK", re.compile(rf"""(?:from\s+|require\(\s*|import\(\s*)['"](?:{JS_SDKS})(?:/[^'"]*)?['"]""")),
@@ -75,6 +99,9 @@ RULES: list[tuple[str, re.Pattern]] = [
 ]
 # Kinds that never block (even in MODE=block) and are only judged on ADDED lines, never the baseline tree.
 WARN_ONLY_KINDS = {"veron ollama"}
+# Rolled out WARN-first: these kinds still show (tree + PRs) but never block, until the env var
+# (a systemd drop-in on the sync, like SECRET_GUARD_WARN_KINDS) is removed.
+SOFT_KINDS = {k for k in os.environ.get("COMPUTE_GUARD_WARN_KINDS", "").split(",") if k}
 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)$")
 
@@ -95,13 +122,47 @@ class Finding:
     match: str
 
 
-def load_allow(path: Path = ALLOW_FILE) -> list[dict]:
+EXEMPTIONS = {"local-user-hardware", "owner-approved", "compute-door", "guard-self"}
+STRUCTURAL = {"compute-door", "guard-self"}   # the door itself and the guard's own files: yearly review
+APPROVERS = {"windy-hub", "windy-mind"}       # a lane never approves its own exemption (Hub 10-02)
+MAX_DAYS = 90                                 # every other exemption: 90 days max, then re-approve
+EXPIRED: list[dict] = []  # entries dropped as expired on the last load_allow (reported, never silent)
+OVERCAP: list[dict] = []  # entries dropped because their expiry is further out than MAX_DAYS
+
+
+def load_allow(path: Path = ALLOW_FILE, today: date | None = None, strict: bool = True) -> list[dict]:
+    """Active entries only. Every entry needs repo, paths, a reason, a NAMED exemption and an expiry
+    date (Mind 10-02: nothing gets permanent amnesty). An expired entry stops excusing code at once.
+    `strict=False` is for OTHER guards reusing this loader (ci-hygiene) with their own file format."""
+    today = today or date.today()
     data = yaml.safe_load(path.read_text()) or {}
     entries = data.get("allow") or []
-    for e in entries:  # a reason per entry is the whole point of the file
+    active = []
+    EXPIRED.clear()
+    OVERCAP.clear()
+    for e in entries:
         if not (e.get("repo") and e.get("paths") and str(e.get("reason", "")).strip()):
             raise ValueError(f"allow entry needs repo, paths and a reason: {e}")
-    return entries
+        if not strict:
+            active.append(e)
+            continue
+        if e.get("exemption") not in EXEMPTIONS:
+            raise ValueError(f"allow entry needs exemption in {sorted(EXEMPTIONS)}: {e.get('repo')} {e.get('paths')}")
+        try:
+            exp = e["expires"] if isinstance(e.get("expires"), date) else date.fromisoformat(str(e.get("expires")))
+        except ValueError as err:
+            raise ValueError(f"allow entry needs expires: YYYY-MM-DD: {e.get('repo')} {e.get('paths')}") from err
+        if e["exemption"] not in STRUCTURAL:
+            if e.get("approved_by") not in APPROVERS:
+                raise ValueError(f"allow entry needs approved_by in {sorted(APPROVERS)}: {e.get('repo')} {e.get('paths')}")
+            if exp > today + timedelta(days=MAX_DAYS):
+                OVERCAP.append({**e, "expires": exp.isoformat()})  # a longer amnesty simply does not apply
+                continue
+        if exp < today:
+            EXPIRED.append({**e, "expires": exp.isoformat()})
+        else:
+            active.append(e)
+    return active
 
 
 def allowed(repo: str, path: str, allow: list[dict], text: str | None = None) -> bool:
@@ -132,6 +193,10 @@ def scan_line(path: str, text: str) -> list[tuple[str, str]]:
     for kind, rx in RULES:
         if kind == "provider SDK dep" and not DEP_FILES.search(path):
             continue
+        if kind == "workers ai binding" and not WRANGLER.search(path):
+            continue
+        if kind == "talk engine port" and PORT_SKIP.search(path):
+            continue
         m = rx.search(text)
         if m:
             hits.append((kind, m.group(0).strip()[:60]))
@@ -156,9 +221,11 @@ def scan_tree(repo: str, bare: Path, sha: str, allow: list[dict], *, line_fn=Non
     # Other guards (ci_hygiene) reuse this walker with their own line rules.
     line_fn = line_fn or scan_line
     path_ok = path_ok or _default_path_ok
-    pre = prefilter or "|".join([re.escape(h) for h in HOSTS] + KEYS + [
+    pre = prefilter or "|".join([re.escape(h) for h in HOSTS + VOICE_HOSTS] + KEYS + VOICE_KEYS + [
         "anthropic", "openai", "groq", "mistral", "generativeai", "genai", "cohere",
-        "together", "cerebras", "litellm"])
+        "together", "cerebras", "litellm", "deepgram", "elevenlabs", "cartesia", "play\\.ht", "resemble",
+        "heygen", "googleapis\\.com", "amazonaws\\.com", "api\\.cloudflare\\.com", ":8791", ":8788",
+        ":8794", "%3[aA]87", r"^\s*\[ai\]", '"ai"'])
     try:
         out = _git(bare, "grep", "-nIE", "-e", pre, sha, "--", ".")
     except subprocess.CalledProcessError as e:
@@ -215,7 +282,8 @@ def parse_added(repo: str, diff: str, allow: list[dict], *, line_fn=None, path_o
 # ---- cache: a tree scan runs once per (repo, sha, rules+allow) --------------
 def _fingerprint(allow: list[dict]) -> str:
     return hashlib.sha256(
-        json.dumps([HOSTS, KEYS, PY_SDKS, JS_SDKS, SKIP.pattern, allow], sort_keys=True).encode()
+        json.dumps([HOSTS, KEYS, VOICE_HOSTS, VOICE_KEYS, [r.pattern for _, r in RULES], PY_SDKS, JS_SDKS,
+                    SKIP.pattern, allow], sort_keys=True, default=str).encode()
     ).hexdigest()[:16]
 
 
@@ -280,6 +348,13 @@ def status_for(findings: list[Finding], whole_tree: bool,
     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
+    rolling = [f for f in findings if f.kind in SOFT_KINDS]
+    if findings and len(rolling) == len(findings) and not grant:
+        f, n = rolling[0], len(rolling)
+        return "success", (f"⚠ WARN (rolling out, not blocking): {n} direct AI-provider use{'s' if n > 1 else ''} "
+                           f"{scope}, e.g. {f.path}:{f.line} {f.match}")[:140], f
+    if rolling:
+        findings = [f for f in findings if f.kind not in SOFT_KINDS]
     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 ''} "
@@ -298,6 +373,12 @@ def status_for(findings: list[Finding], whole_tree: bool,
 
 def report(repos: list[str]) -> int:
     allow = load_allow()
+    for e in OVERCAP:
+        print(f"## OVER-CAP exemption (> {MAX_DAYS} days, NOT applied): {e['repo']} {e['paths']} "
+              f"[{e['exemption']}] expires {e['expires']}")
+    for e in EXPIRED:
+        print(f"## EXPIRED exemption (no longer excuses anything): {e['repo']} {e['paths']} "
+              f"[{e['exemption']}] expired {e['expires']}")
     total = 0
     for repo in repos:
         bare = WORK / f"{repo}.git"

scripts/install_secret_tools.sh

@@ -5,10 +5,10 @@ set -euo pipefail
 here=$(cd "$(dirname "$0")" && pwd)
 dest="$HOME/.local/share/secret-tools"
 mkdir -p "$dest" "$HOME/.local/bin"
-cp "$here/secret_shapes.py" "$here/secret_scan.py" "$here/env_names.py" "$here/lockbox_put.py" "$dest/"
+cp "$here/secret_shapes.py" "$here/secret_scan.py" "$here/env_names.py" "$here/lockbox_put.py" "$here/lockbox_names.py" "$dest/"
-for pair in "secret-scan:secret_scan.py" "env-names:env_names.py" "lockbox-put:lockbox_put.py"; do
+for pair in "secret-scan:secret_scan.py" "env-names:env_names.py" "lockbox-put:lockbox_put.py" "lockbox-names:lockbox_names.py"; do
   n=${pair%%:*}; f=${pair##*:}
   printf '#!/usr/bin/env bash\nexec python3 "%s/%s" "$@"\n' "$dest" "$f" > "$HOME/.local/bin/$n"
   chmod 755 "$HOME/.local/bin/$n"
 done
-echo "installed secret-scan, env-names and lockbox-put (shapes from secret_shapes.py, same as secret-guard)"
+echo "installed secret-scan, env-names, lockbox-put and lockbox-names (shapes from secret_shapes.py, same as secret-guard)"

scripts/lockbox_names.py

@@ -0,0 +1,134 @@
+#!/usr/bin/env python3
+"""lockbox-names: DISCOVER what the lockbox holds without reading it (Boss rule 10-01).
+
+  lockbox-names [REGEX]        (case-insensitive; matches the section heading or the label)
+
+Prints one row per entry: SECTION HEADING | LABEL | kind | resolvable
+  kind        env   = `KEY=value` line            md = `- **`KEY`**: `value`` line
+              label = a bold prose label (`**Password (X):** ...`)  file = secrets/**/*.env key
+  resolvable  yes   = `lockbox-get LABEL FILE` returns exactly one value
+              dup   = defined with 2+ different values (lockbox-get refuses)
+              no    = a prose-only label, or not an exact key
+
+NEVER prints a value or any prose after a label. A label or heading that itself looks
+like a secret (secret_shapes) is replaced by <secret-shaped>. Reads the COMMITTED lockbox at
+origin/main (like lockbox-get; LOCKBOX_REF=<ref> or WORKTREE overrides). Memory only, stdout only.
+"""
+from __future__ import annotations
+
+import hashlib
+import os
+import re
+import subprocess
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+import secret_shapes as ss  # noqa: E402
+
+REPO = os.environ.get("LOCKBOX_REPO", os.path.expanduser("~/kit-army-config"))
+REF = os.environ.get("LOCKBOX_REF", "origin/main")
+HEAD = re.compile(r"^#{1,6}\s+(.*\S)\s*$")
+ENV = re.compile(r"^([A-Z][A-Z0-9_]{2,})=(.*)$")
+MD = re.compile(r"^\s*[-*]?\s*\*\*`([A-Za-z0-9_]+)`\*\*\s*:\s*`([^`]+)`")
+LABEL = re.compile(r"\*\*([^*`]{2,70}?)\*\*")
+
+
+def git(*a: str) -> subprocess.CompletedProcess:
+    return subprocess.run(["git", "-C", REPO, *a], capture_output=True, text=True, errors="ignore")
+
+
+def read_sources() -> dict[str, str]:
+    """{path: text} for ACCESS_LOCKBOX.md and secrets/**/*.env."""
+    if REF == "WORKTREE":
+        out = {}
+        for p in [Path(REPO, "ACCESS_LOCKBOX.md"), *Path(REPO, "secrets").rglob("*.env")]:
+            if p.is_file():
+                out[str(p.relative_to(REPO))] = p.read_text(errors="ignore")
+        return out
+    if REF.startswith("origin/"):
+        git("fetch", "-q", "origin", REF.split("/", 1)[1])
+    names = ["ACCESS_LOCKBOX.md"] + [
+        p for p in git("ls-tree", "-r", "--name-only", REF, "--", "secrets").stdout.splitlines() if p.endswith(".env")]
+    out = {}
+    for n in names:
+        r = git("show", f"{REF}:{n}")
+        if r.returncode == 0:
+            out[n] = r.stdout
+    return out
+
+
+def safe(text: str, limit: int = 70) -> str:
+    text = re.sub(r"\s+", " ", text).strip()
+    return "<secret-shaped>" if ss.find(text) else text[:limit]
+
+
+def h(v: str) -> str:
+    return hashlib.sha256(v.strip().strip('"').strip("'").encode()).hexdigest()[:16]
+
+
+def collect(src: dict[str, str]):
+    """(rows, values) where values[KEY] = {hash,...} for resolvability; nothing printed from it."""
+    rows, values = [], {}
+    for path, text in src.items():
+        section = path
+        for line in text.splitlines():
+            m = HEAD.match(line) if path.endswith(".md") else None
+            if m:
+                section = safe(m.group(1), 90)
+                continue
+            m = ENV.match(line)
+            if m:
+                values.setdefault(m.group(1), set()).add(h(m.group(2)))
+                rows.append((section, m.group(1), "file" if path.startswith("secrets/") else "env"))
+                continue
+            m = MD.match(line)
+            if m:
+                values.setdefault(m.group(1), set()).add(h(m.group(2)))
+                rows.append((section, m.group(1), "md"))
+                continue
+            if path.endswith(".md"):
+                mm = LABEL.search(line)
+                if mm and not mm.group(1).startswith("http"):
+                    rows.append((section, safe(mm.group(1)), "label"))
+    return rows, values
+
+
+def resolvable(label: str, kind: str, values) -> str:
+    if kind not in ("env", "md", "file"):
+        return "no"
+    n = len(values.get(label, ()))
+    return "yes" if n == 1 else "dup" if n > 1 else "no"
+
+
+def main(argv=None) -> int:
+    argv = list(sys.argv[1:] if argv is None else argv)
+    rx = re.compile(argv[0], re.I) if argv else None
+    src = read_sources()
+    if not src:
+        print("lockbox-names: cannot read the lockbox")
+        return 2
+    rows, values = collect(src)
+    seen, n = set(), 0
+    for section, label, kind in rows:
+        if rx and not (rx.search(section) or rx.search(label)):
+            continue
+        key = (section, label, kind)
+        if key in seen:
+            continue
+        seen.add(key)
+        n += 1
+        print(f"{section} | {label} | {kind} | {resolvable(label, kind, values)}")
+    print(f"# {n} entr{'y' if n == 1 else 'ies'}; names only, values never printed")
+    return 0
+
+
+if __name__ == "__main__":
+    try:
+        sys.exit(main())
+    except re.error:
+        print("lockbox-names: bad regex")
+        sys.exit(2)
+    except Exception as e:  # never a traceback
+        print(f"lockbox-names: error: {type(e).__name__}")
+        sys.exit(2)

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/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)"