Skip to content

armgit / AiSOC

Public Imported from GitHub
0

Open-source AI Security Operations Center: alert fusion, LLM-agent triage, MITRE ATT&CK investigation, and a replayable decision ledger for every agent step. Self-hostable, runs with no API keys, MIT licensed. Ships an MCP server for Claude, Cursor and Continue.

chore(release): v9.0.0 (#740)

Beenu Arorae99a5ec86760
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 37a2b7a0b..1a585c43e 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,6 +7,192 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
## [Unreleased]
+## [9.0.0] — 2026-09-23
+
+**Ten waves, and one finding under nearly all of them: the mechanism existed,
+was tested, and nothing called it.** v8.0 named that shape and found it a
+dozen times. This release went looking for it deliberately, across the whole
+tree, and found it again in the approval loop, the marketplace, the mobile
+console, the detection engine and the benchmark scoreboard. A passing unit
+test on an uncalled function is indistinguishable from a working feature
+until somebody traces the call graph, and the only defence is to trace it.
+
+The one worth stating first, because it inverts what the feature appeared to
+do: **approving an action executed nothing.** `decide()` flipped a row,
+notified the realtime service and returned 200 without ever touching
+`services/actions`. Every tap of Approve in the responder app recorded a
+decision and ran nothing — while telling the operator the opposite. That is
+the most dangerous shape a security control can have, because what it appears
+to say is "the host is contained". And the other end was missing too: nothing
+in the repository ever created an approval, so the queue had no producer
+either and was structurally empty on every deployment.
+
+### Added
+
+- **Approvals reach a human and then reach the estate** ([#731](https://github.com/beenuar/AiSOC/pull/731),
+ [#735](https://github.com/beenuar/AiSOC/pull/735)). The triage worker raises
+ one approval per proposed action that declares `requires_approval`, keyed
+ deterministically so a Kafka replay re-raises the same approval rather than
+ a second copy an operator cannot tell apart. `decide` carries the decision
+ through to `services/actions` and records on the row whether it executed, in
+ four distinct states, so "was this actually done" is answerable from the row
+ rather than by correlating two services' logs. A refusal returns 502 saying
+ the decision was stored and the action was not run.
+- **The confidence × impact approval matrix now runs.** `approval_matrix.evaluate`
+ was written, documented, unit-tested and listed in the claim-to-gate matrix
+ as GATED, with **zero production callers** — `POST /actions` gated on blast
+ radius alone, a property of the verb, so the same answer came back for a
+ 40%-confidence guess and a corroborated finding. Both gates run and the
+ stricter wins, which is the composition rule the matrix already stated, so
+ nothing can auto-execute that did not before. Ten action types tighten, and
+ `/actions` becomes tier-aware for the first time.
+- **A browser-facing proxy for the live-action registry.** Every route sat
+ behind a service token, so nothing could answer "what can AiSOC do to my
+ estate". Discovery and dry-run only: a live containment goes through the
+ approval path so an approver is bound to it.
+- **`apps/mobile`, a native responder app** ([#733](https://github.com/beenuar/AiSOC/pull/733)).
+ The roadmap said the mobile console was "not started", which was true about
+ React Native and misleading about the product — the responder console
+ already existed as a PWA with a service worker, an offline approval queue
+ and Web Push. The native app is a distribution channel, and it exists for
+ one reason: iOS Web Push requires an installed PWA and has been unreliable
+ even then. Unit tests and type-check run in CI; **no device build,
+ simulator run or store submission has been performed**, and a CI gate fails
+ if the README stops saying so.
+- **`@aisoc/sdk` gained the namespaces a responder client needs** — approvals,
+ push, on-call, live actions. All were in `docs/openapi.yaml` the whole time,
+ which is why the gap went unnoticed: the generated types were complete and
+ the hand-written surface was three releases behind.
+- **Marketplace publisher identity, paid listings and entitlements**
+ ([#734](https://github.com/beenuar/AiSOC/pull/734)). Migration `056`. There
+ is no payment processor and no stub of one — wiring payments is an account
+ action. What exists is the enforcement point, which is the part that would
+ otherwise be written last and least carefully.
+- **An egress default-deny NetworkPolicy** ([#736](https://github.com/beenuar/AiSOC/pull/736)),
+ gated by `helm.yml`. Off by default, because a cluster whose CNI does not
+ enforce NetworkPolicy ignores it silently and a control that is silently
+ ignored reads as protection while providing none.
+- **Distinct counting in the windowed detection engine**
+ ([#737](https://github.com/beenuar/AiSOC/pull/737)). "Fifty reads by one
+ principal" is a script retrying; "fifty *different* secrets read by one
+ principal" is a vault being walked. Counting events cannot tell those apart,
+ and 21 of the 74 rules awaiting a windowed evaluator name a `distinct_*`
+ field.
+- **`docs/audit/DEFERRED_SUBPHASES.md`** ([#739](https://github.com/beenuar/AiSOC/pull/739)).
+ Six lettered sub-phases were "tracked in `docs/audit/PROGRESS.md`", which is
+ gitignored and was never committed — so six named commitments had no scope
+ anywhere a contributor could read.
+- **New gates:** `helm.yml` (lint, render, kubeconform, server dry-run against
+ a real kind cluster), `mobile.yml`, `check_go_module_paths.py`,
+ `check_published_packages.py`, and scoreboard staleness.
+
+### Fixed
+
+- **UEBA never scored a single message** ([#730](https://github.com/beenuar/AiSOC/pull/730)).
+ It consumed `security.events`, which nothing in the platform writes — ingest
+ writes `aisoc.raw_events`. So `ueba.anomalies` never carried a message and
+ fusion's UEBA confidence boost, on by default and fully built, could only
+ ever be inert. `feature_extraction.py` recovers the entity and features from
+ the OCSF envelope.
+- **The ingest inbox could not have worked on any deployment.** Routes mount
+ only with `DATABASE_DSN`, which no compose file set, and the templates were
+ never copied into the runtime image, so even a correctly-configured
+ deployment answered 503 for every template. Templates are `go:embed`-ed now:
+ a directory cannot be missing from an image.
+- **CrowdStrike alerts were anonymous.** The connector never read
+ `behaviors[].user_name`, and the canonical field map knew only `actor` while
+ 11 of the 68 canonical-envelope connectors spell it `username` or `user`.
+ Since the correlation key is `{tenant}:{entity}:{tactic}`, they all
+ correlated as `unknown`. The alias list is a slice, not more map entries,
+ because Go randomises map iteration and three entries pointing at one
+ destination would resolve differently per process.
+- **An approved action was not the action approved.** `approve_action` rebuilt
+ `ActionRequest` without `parameters` or `principal`, so an action gated
+ *because of what it targets* ran against defaults and reported COMPLETED —
+ and the submit path never stored parameters at all, so no rebuild could have
+ recovered them.
+- **Pending actions lived in a module-global dict.** A restart lost every
+ action awaiting approval and a second replica could not see the first one's,
+ so an analyst taps Approve on a Slack card and gets "Action not found" for
+ an incident that is still live. Migration `055`.
+- **Neither ChatOps bot could send an unsolicited message.** Replies go through
+ Bolt's `respond()`, which writes to a `response_url` that only exists inside
+ an inbound interaction, so the bot could answer a question and not ask one —
+ while `rich_approval_card_blocks` sat fully written with no production
+ caller and the Approve/Deny handlers sat wired and waiting.
+- **A plugin could never be rejected for a bad signature.**
+ `_get_registered_pub_key` returned `None` unconditionally, so `publish_plugin`
+ skipped verification entirely. The signing path existed end to end, had a
+ CLI command, and was incapable of saying no.
+- **No registry allow-list and no digest pinning**, both of which the notes
+ claimed existed. There was a deny list of three metadata hostnames, and
+ nothing resolved a tag to a digest — so `:latest` stayed mutable and "what
+ is this deployment running" had no answer after the fact. The signature gate
+ does not close that: it verifies whatever arrived.
+- **`aisoc plugin publish` POSTed to a route no router has ever served.**
+ Broken as shipped.
+- **The public scoreboard was frozen for ten weeks and every check passed**
+ ([#738](https://github.com/beenuar/AiSOC/pull/738)). "Freshness" meant the
+ accuracy value was current and nothing read the row's date; the newest row
+ was taken by file position rather than by date; `--refresh` rewrote only the
+ accuracy, so a refreshed row described today's code and claimed to have been
+ measured in July. And there was no writer at all — `live-agent-eval.yml` had
+ `contents: read`.
+- **Neither published Go SDK was installable.** Both declared
+ `github.com/beenuar/aisoc/<name>` — wrong case against a case-sensitive VCS
+ path, missing the `packages/` prefix. Every in-repo consumer used a `replace`
+ directive, which is the shape that hides this: the tree builds perfectly and
+ the artifact does not exist.
+- **The Helm chart did not render.** `helm template` fails outright until
+ dependencies are fetched, and no documentation said to fetch them — nor to
+ `helm repo add bitnami` first, which a clean machine also needs.
+- **OpenSearch was recorded as dead code and is not.** The check had been made
+ against `services/api`, which holds no OpenSearch client, and never against
+ `services/threatintel`, which indexes into it in its lifespan with no flag
+ and no `try`. Two real defects were hiding under the wrong verdict: compose
+ set an env var that service discards, so the connection worked by
+ coincidence, and it declared no dependency on OpenSearch so it raced it at
+ boot.
+- **One connectivity check at boot is a coin toss.** The graph writer verified
+ Neo4j once and a failure disabled it for the process lifetime; compose
+ cannot express "depend on neo4j only in the `full` profile", so ingest lost
+ a race it had no way to wait for. The integration gate caught it
+ intermittently — the same branch passed at 21:06 and failed at 21:19.
+- Two stale comments that described working code as broken: the reachability
+ gate's claim that the windowed engine "has three hardcoded rules and no
+ loader", and the windowed exporter's claim to mirror the stateless one
+ "exactly: specs are the source of truth".
+
+### Changed
+
+- **Packaging stops moving the number.** It slipped v8.0 → v8.1 → v8.2 for the
+ same reason each time. The README now states a fact rather than a date —
+ the pipeline builds and packs all eight packages on every tag, and the
+ upload is blocked on registry credentials, which is an account action.
+ `check_published_packages.py` reads actual registry state so that claim
+ cannot go stale in either direction once credentials exist.
+- Vulnerability-match events are opt-in: every stock deployment was
+ downloading the CISA KEV catalogue at boot to publish into a topic with no
+ reader.
+- `aisoc.alerts.raw` is reclassified from dead to external entrypoint. No
+ in-repo producer is what an entrypoint *is*.
+
+### Known
+
+- **Eight of the nine `PARTIAL` claim-to-gate rows remain**, each still naming
+ its own gap. The air-gap row narrowed rather than flipping: the Helm half is
+ built and gated, platform-wide egress-blocked CI is not. Relabelling a row
+ for the half that is done is what that file exists to prevent.
+- Whether migrations 050–056 apply to an *existing* Postgres volume is still
+ untraced. Compose mounts them as `docker-entrypoint-initdb.d`, which runs
+ only on a fresh volume.
+- 133 detection rules remain unreachable. Authoring a `wd-*` rule does not make
+ a `det-*` rule reachable, so `MAX_UNREACHABLE` is unchanged on purpose.
+- The OCI install route is held back. Adding it made CodeQL flag eight
+ `py/path-injection` sites across the plugin-ingest graph; trading "a hardened
+ path with no caller" for "a reachable path with eight unresolved high
+ findings" is the worse position.
+
## [8.1.1] — 2026-09-23
**The v8.1.0 notes described a platform its own quick start never started.**
@@ -4171,7 +4357,8 @@ demo profile. Details below.
- Helm chart for Kubernetes deployment (`infra/helm/aisoc/`)
- MIT License
-[Unreleased]: https://github.com/beenuar/AiSOC/compare/v8.1.1...HEAD
+[Unreleased]: https://github.com/beenuar/AiSOC/compare/v9.0.0...HEAD
+[9.0.0]: https://github.com/beenuar/AiSOC/compare/v8.1.1...v9.0.0
[8.1.1]: https://github.com/beenuar/AiSOC/compare/v8.1.0...v8.1.1
[8.1.0]: https://github.com/beenuar/AiSOC/compare/v8.0.0...v8.1.0
[8.0.0]: https://github.com/beenuar/AiSOC/compare/v7.7.0...v8.0.0
diff --git a/README.md b/README.md
index 736363b62..bc8ce9b06 100644
--- a/README.md
+++ b/README.md
@@ -7,7 +7,7 @@
**An open-source, self-hostable AI Security Operations Center.** It ingests your security telemetry, detects and correlates threats, investigates them with AI agents whose reasoning is fully auditable, and proposes responses a human approves.
[![License: MIT](https://img.shields.io/badge/License-MIT-22c55e.svg?style=flat-square)](https://opensource.org/licenses/MIT)
-[![Version](https://img.shields.io/badge/version-8.1.1-f59e0b?style=flat-square)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-9.0.0-f59e0b?style=flat-square)](CHANGELOG.md)
[![CI](https://img.shields.io/github/actions/workflow/status/beenuar/AiSOC/ci.yml?branch=main&label=CI&style=flat-square)](https://github.com/beenuar/AiSOC/actions/workflows/ci.yml)
[![CodeQL](https://img.shields.io/github/actions/workflow/status/beenuar/AiSOC/codeql.yml?branch=main&label=CodeQL&style=flat-square)](https://github.com/beenuar/AiSOC/actions/workflows/codeql.yml)
[![OpenSSF Scorecard](https://api.securityscorecards.dev/projects/github.com/beenuar/AiSOC/badge)](https://securityscorecards.dev/viewer/?uri=github.com/beenuar/AiSOC)
@@ -172,8 +172,8 @@ so. They do not fabricate a verdict.
| Entity graph (Neo4j) | Beta | Unit | Yes, `full` profile |
| Governed response actions | Beta | Unit | Human-approved only |
| Scheduled connectors | Beta | Contract tests | `full` profile |
-| UEBA | **Broken** | — | No — [see why](docs/audit/REPOSITORY_REALITY.md) |
-| Package distribution (npm/PyPI) | Planned | — | v8.2 — blocked on registry credentials |
+| UEBA | Beta | Unit | `full` profile |
+| Package distribution (npm/PyPI) | Ready, unpublished | `release.yml` builds and packs all eight on every tag | Install from source — the upload is blocked on registry credentials, which is an account action |
## What AiSOC is not
diff --git a/RELEASES.md b/RELEASES.md
index cce03760b..e608b5d76 100644
--- a/RELEASES.md
+++ b/RELEASES.md
@@ -2,13 +2,32 @@
This file mirrors what used to live in the "What's new" section of [`README.md`](README.md). The complete, machine-readable inventory (with file paths, env-var diffs, and per-release test counts) lives in [`CHANGELOG.md`](CHANGELOG.md).
-> **TL;DR for first-time visitors:** AiSOC is on `v8.1.1`, released 2026-09-23 — an adoption-audit release with no new capability. The audit found that the single most-followed path into AiSOC did not run AiSOC: `./install.sh` started a compose file with no ingest service, no fusion service and Kafka disabled, and everything in the resulting console had been written straight into Postgres by a seed script. The core pipeline does work; it had never been demonstrated, and `make smoke` now demonstrates it in CI. Every product claim is backed by a failing CI test (claim-to-gate matrix: 99 GATED / 9 PARTIAL / 0 NO GATE). The latest GitHub release with notes and downloads: <https://github.com/beenuar/AiSOC/releases/latest>.
+> **TL;DR for first-time visitors:** AiSOC is on `v9.0.0`, released 2026-09-23 — ten waves, and one finding under nearly all of them: the mechanism existed, was tested, and nothing called it. The one worth stating first is that **approving an action executed nothing** — `decide()` flipped a row and never reached the execution service, so every tap of Approve recorded a decision and ran nothing while telling the operator the opposite. Both ends of that loop are closed. Every product claim is backed by a failing CI test (claim-to-gate matrix: 99 GATED / 9 PARTIAL / 0 NO GATE). The latest GitHub release with notes and downloads: <https://github.com/beenuar/AiSOC/releases/latest>.
---
## What's new
-`VERSION` is `8.1.1`. The **v8.1.1** release (2026-09-23) adds no capability. It exists because an audit of how the repository is adopted — installability, architecture comprehension, data provenance, pipeline connectivity — found the quick start did not start the product.
+`VERSION` is `9.0.0`. The **v9.0.0** release (2026-09-23) is ten waves of one audit question: what in this tree exists, is tested, and has no caller?
+
+**v9.0.0 highlights (September 23, 2026)**
+- **Approving an action executed nothing.** `decide()` flipped a row, notified the realtime service and returned 200 without ever touching `services/actions`. The other end was missing too — nothing in the repository ever created an approval, so the queue had no producer and was structurally empty on every deployment. A queue with no producer and a queue with no pending work look identical.
+- **The confidence x impact approval matrix had zero production callers.** Written, documented, unit-tested and listed in the claim-to-gate matrix as GATED, while `POST /actions` gated on blast radius alone — a property of the verb, so the same answer came back for a 40%-confidence guess and a corroborated finding. Both gates run now and the stricter wins.
+- **UEBA never scored a single message.** It consumed a topic nothing in the platform writes, so fusion's UEBA confidence boost — on by default, fully built — could only ever be inert.
+- **A plugin could never be rejected for a bad signature.** `_get_registered_pub_key` returned `None` unconditionally, so verification was skipped entirely. The signing path existed end to end, had a CLI command, and was incapable of saying no. There was also no registry allow-list and no digest pinning, both of which the notes claimed existed.
+- **The public scoreboard was frozen for ten weeks and every check passed**, because "freshness" meant the accuracy value was current and nothing ever read the row's date.
+- **Neither published Go SDK was installable** — wrong case against a case-sensitive VCS path, missing the directory prefix — and **the Helm chart did not render** until dependencies were fetched, which no documentation mentioned.
+- **A native responder app**, plus the honest correction that the responder console already existed as a PWA. The native app is a distribution channel; it exists because iOS Web Push requires an installed PWA and has been unreliable even then.
+
+Packaging stops moving the number. It slipped v8.0 to v8.1 to v8.2 for the same reason each time, so the README states a fact rather than a date, and `check_published_packages.py` reads registry state so the claim cannot go stale in either direction.
+
+The full inventory — including what is knowingly still open — lives under `[9.0.0]` in [`CHANGELOG.md`](CHANGELOG.md).
+
+---
+
+## v8.1.1 (2026-09-23)
+
+`VERSION` was `8.1.1`. The **v8.1.1** release (2026-09-23) adds no capability. It exists because an audit of how the repository is adopted — installability, architecture comprehension, data provenance, pipeline connectivity — found the quick start did not start the product.
**v8.1.1 highlights (September 23, 2026)**
- **The documented quick start never ran the product.** `./install.sh` handed off to a nine-service compose file with no ingest service, no fusion service, and `AISOC_DISABLE_KAFKA: true`. Everything visible in the console came from `seed_demo.py` writing fifteen fabricated incidents straight into Postgres, and the installer printed "AiSOC is up and running" because the compose command exited 0. It now runs `make up` — the same CORE stack the README documents and CI tests — then proves the pipeline before claiming success.
diff --git a/ROADMAP.md b/ROADMAP.md
index ba8fc314d..e2d877375 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -660,6 +660,35 @@ Full inventory under `[8.1.1]` in [`CHANGELOG.md`](CHANGELOG.md).
---
+## v9.0 — Shipped (2026-09-23)
+
+Ten waves of one audit question: what in this tree exists, is tested, and has
+no caller? v8.0 named that shape and found it a dozen times; v9.0 went looking
+for it deliberately and found it in the approval loop, the marketplace, the
+mobile console, the detection engine and the benchmark scoreboard.
+
+The one worth stating first inverts what the feature appeared to do:
+**approving an action executed nothing.** `decide()` flipped a row and never
+reached the execution service, so every tap of Approve recorded a decision and
+ran nothing — while telling the operator the opposite. Nothing ever created an
+approval either, so the queue had no producer and was structurally empty on
+every deployment.
+
+Also: UEBA consumed a topic nothing writes; a plugin could never be rejected
+for a bad signature; there was no registry allow-list and no digest pinning,
+both of which the notes claimed existed; the public scoreboard was frozen for
+ten weeks with every check passing; neither published Go SDK was installable;
+and the Helm chart did not render at all.
+
+What is knowingly still open is listed under `### Known` in `[9.0.0]` —
+eight `PARTIAL` matrix rows, the migration-on-existing-volume path, 133
+unreachable detection rules, and the OCI install route held back rather than
+shipped with eight unresolved high findings.
+
+Full inventory under `[9.0.0]` in [`CHANGELOG.md`](CHANGELOG.md).
+
+---
+
## Ideas Backlog (unscheduled)
- "Explain this alert" button using LLM with enrichment context
diff --git a/VERSION b/VERSION
index 0e7915245..f7ee06693 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-8.1.1
+9.0.0
diff --git a/apps/web/package.json b/apps/web/package.json
index 017c6a4a5..471ad8f5f 100644
--- a/apps/web/package.json
+++ b/apps/web/package.json
@@ -1,6 +1,6 @@
{
"name": "@aisoc/web",
- "version": "8.1.1",
+ "version": "9.0.0",
"private": true,
"description": "AiSOC SOC Console - Next.js 14 Frontend",
"license": "MIT",
diff --git a/docs/roadmap/v8-progress.md b/docs/roadmap/v8-progress.md
index adce92f77..1f6a889a0 100644
--- a/docs/roadmap/v8-progress.md
+++ b/docs/roadmap/v8-progress.md
@@ -1,7 +1,11 @@
# AiSOC v8 — progress tracker
**Last updated:** 2026-09-23
-**Current release:** `v8.1.1` (2026-09-23) · **Next:** `v8.2` (packaging)
+**Current release:** `v9.0.0` (2026-09-23) · **Next:** unscheduled.
+Packaging is no longer named against a version: it slipped v8.0 to v8.1 to
+v8.2 for the same reason each time, and the blocker is registry credentials
+rather than code. See `[9.0.0]` in `CHANGELOG.md` for what v9.0 closed and
+what it knowingly left open.
This tracker is the at-a-glance view of what has landed across the v8 line and
what is still open. It is deliberately short and dated. When it disagrees with
diff --git a/scripts/check_published_packages.py b/scripts/check_published_packages.py
new file mode 100755
index 000000000..9ab47703b
--- /dev/null
+++ b/scripts/check_published_packages.py
@@ -0,0 +1,126 @@
+#!/usr/bin/env python3
+"""The README's publishing claim must match what the registries actually hold.
+
+The packaging milestone has now moved three times — v8.0 to v8.1 to v8.2 —
+for the same reason each time: the blocker is registry credentials, not code,
+so it cannot be scheduled by writing a version number. v9.0 stops moving the
+number and states a fact instead: the pipeline builds and packs every
+package on each tag, and the upload is blocked on an account action.
+
+A fact can go stale in a way a promise cannot. The moment somebody adds an
+`NPM_TOKEN`, "ready, unpublished" becomes false, and the README would keep
+saying it — the same failure as a promise that never lands, just pointing the
+other way.
+
+So this checks the claim against the registries. Offline it skips rather than
+fails, because a contributor on a plane must not be told the README is wrong
+when the truth is that the network is absent. It is a scheduled check, not a
+per-PR blocker, for the same reason.
+
+Usage:
+ python3 scripts/check_published_packages.py
+ python3 scripts/check_published_packages.py --require-network
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+import urllib.error
+import urllib.parse
+import urllib.request
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent.parent
+README = ROOT / "README.md"
+
+#: The phrase the README uses while nothing is published.
+UNPUBLISHED_CLAIM = "Ready, unpublished"
+
+#: (registry, name) for every package `release.yml` would upload.
+PACKAGES: tuple[tuple[str, str], ...] = (
+ ("npm", "aisoc"),
+ ("npm", "@aisoc/sdk"),
+ ("npm", "@aisoc/mcp"),
+ ("pypi", "aisoc-sandbox"),
+ ("pypi", "aisoc-cli"),
+ ("pypi", "aisoc-sdk"),
+ ("pypi", "aisoc-plugin-sdk"),
+ ("pypi", "aisoc-detections"),
+)
+
+_TIMEOUT = 10
+
+
+class Offline(RuntimeError):
+ """The registry could not be reached, which is not the same as absent."""
+
+
+def _is_published(registry: str, name: str) -> bool:
+ url = f"https://registry.npmjs.org/{urllib.parse.quote(name, safe='@')}" if registry == "npm" else f"https://pypi.org/pypi/{name}/json"
+ request = urllib.request.Request(url, headers={"Accept": "application/json"})
+ try:
+ with urllib.request.urlopen(request, timeout=_TIMEOUT) as response: # noqa: S310 — fixed https hosts
+ json.load(response)
+ return True
+ except urllib.error.HTTPError as exc:
+ if exc.code == 404:
+ return False
+ raise Offline(f"{registry}:{name} returned {exc.code}") from exc
+ except (urllib.error.URLError, TimeoutError, OSError) as exc:
+ raise Offline(f"{registry}:{name} unreachable: {exc}") from exc
+
+
+def main() -> int:
+ parser = argparse.ArgumentParser(description=__doc__)
+ parser.add_argument(
+ "--require-network",
+ action="store_true",
+ help="fail instead of skipping when a registry cannot be reached",
+ )
+ args = parser.parse_args()
+
+ readme = README.read_text(encoding="utf-8")
+ claims_unpublished = UNPUBLISHED_CLAIM.lower() in readme.lower()
+
+ published: list[str] = []
+ for registry, name in PACKAGES:
+ try:
+ if _is_published(registry, name):
+ published.append(f"{registry}:{name}")
+ except Offline as exc:
+ if args.require_network:
+ print(f"ERROR: {exc}", file=sys.stderr)
+ return 1
+ print(f"SKIP: {exc}")
+ return 0
+
+ if published and claims_unpublished:
+ print("PUBLISHED-PACKAGES GATE FAILED:", file=sys.stderr)
+ print(
+ f" - README still says {UNPUBLISHED_CLAIM!r}, but these are live: " f"{', '.join(published)}.",
+ file=sys.stderr,
+ )
+ print(
+ " Update the maturity table: the claim was true and is not any more.",
+ file=sys.stderr,
+ )
+ return 1
+
+ if not published and not claims_unpublished:
+ print("PUBLISHED-PACKAGES GATE FAILED:", file=sys.stderr)
+ print(
+ f" - Nothing is on npm or PyPI, and the README no longer says "
+ f"{UNPUBLISHED_CLAIM!r}. A reader following it will hit a 404.",
+ file=sys.stderr,
+ )
+ return 1
+
+ state = f"{len(published)}/{len(PACKAGES)} published" if published else "none published"
+ print(f"OK: README matches registry state ({state}).")
+ return 0
+
+
+if __name__ == "__main__":
+ raise SystemExit(main())
diff --git a/scripts/readme_gates.py b/scripts/readme_gates.py
index c6dfe02e1..cb18e79a9 100644
--- a/scripts/readme_gates.py
+++ b/scripts/readme_gates.py
@@ -61,6 +61,13 @@ README_MAX_LINES = 250
# text must use the current one — a README shipped *as* vX cannot coherently
# promise that something lands in vX.
#
+# v9.0 stops moving the number. The README no longer says packaging "comes
+# in" any version, because a README shipped *as* v9.0 promising v9.0 is the
+# incoherence this list was created to police, and promising v9.1 would just
+# be the fourth slip. It says what is true instead: the pipeline builds and
+# packs all eight packages on every tag, and the upload is blocked on
+# registry credentials — which is an account action and cannot be scheduled.
+#
# Do not add the next version here speculatively. The phrase should move when
# the milestone does, so the list is a record of what was actually promised.
V8_GUARDS = (
@@ -80,6 +87,9 @@ V8_GUARDS = (
"lands with the v8.1",
"lands with the v8.2",
"packaging release",
+ # The v9.0 phrasing: a state, not a date.
+ "blocked on registry credentials",
+ "ready, unpublished",
"publish lands in v8.0",
"publish lands in v8.1",
"publish lands in v8.2",