chore(marketplace): move hypermnesia-mcp-viz pin to v3.1.1 #1397

Workflow file for this run

	name: CI

	on:
	push:
	branches: [main]
	pull_request:
	branches: [main]
	workflow_dispatch:

	permissions:
	contents: read

	concurrency:
	group: ${{ github.workflow }}-${{ github.ref }}
	cancel-in-progress: true

	jobs:
	# Uses the composite action at .github/actions/test-suite/action.yml for
	# the step-level rationale, including why the shared unit is a composite
	# action and not a reusable workflow: a job delegating via `uses:` reports
	# its check as "<caller job name> / <called job name>", which would rename
	# all four required contexts on `main`. Sharing at the step level leaves
	# this job's name — and therefore its status-check context — untouched.
	# Originally shared with release.yml's own `test` job (issue #336); issue
	# #392 removed that second caller (see the action's header comment) —
	# `requirements/ci-postgresql.txt` and tree-sitter grammars are now
	# installed unconditionally inside the action rather than passed in.
	test:
	name: Test (Python ${{ matrix.python-version }})
	runs-on: ubuntu-latest
	strategy:
	fail-fast: false
	matrix:
	python-version: ["3.10", "3.11", "3.12", "3.13"]

	steps:
	# A local `uses: ./...` is resolved from the checked-out tree, so the
	# checkout cannot itself live inside the shared action.
	- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

	- uses: ./.github/actions/test-suite
	with:
	python-version: ${{ matrix.python-version }}
	# Coverage, the MCP host contract check, and the advertised-test-count
	# check are expensive and gain nothing from running on all four legs —
	# ci.yml has always scoped them to Python 3.12 only; the shared action
	# makes that scoping an explicit input instead of an inline `if:` per
	# step.
	run-extended-checks: ${{ matrix.python-version == '3.12' }}

	test-sqlite:
	name: Test (SQLite backend)
	runs-on: ubuntu-latest

	# Force the SQLite fallback path — no PostgreSQL installed or started.
	# The conftest detects PG-unreachable and sets CORTEX_MEMORY_STORE_BACKEND=sqlite
	# automatically; this env var makes the selection explicit and observable in logs.
	# source: mcp_server/infrastructure/memory_store.py _construct_store() — the
	# 'sqlite' backend branch is always reachable without PG; conftest.py line 99
	# mirrors this override when _USE_PG is False.
	env:
	CORTEX_MEMORY_STORE_BACKEND: sqlite

	steps:
	- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

	- name: Set up Python
	uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
	with:
	python-version: "3.12"

	- name: Cache pip
	uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
	with:
	path: ~/.cache/pip
	key: ${{ runner.os }}-pip-3.12-sqlite-${{ hashFiles('pyproject.toml') }}
	restore-keys: \|
	${{ runner.os }}-pip-3.12-sqlite-

	- name: Cache HuggingFace models
	uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
	with:
	path: ~/.cache/huggingface
	key: ${{ runner.os }}-hf-all-MiniLM-L6-v2

	# The `codebase` extra is included even though this job is about the
	# storage backend: without tree-sitter/leidenalg, 9 tests SKIP here that
	# the PostgreSQL job runs, so the two gates would not be equal and a
	# codebase-analysis regression could reach the default backend unseen.
	# Measured 2026-07-28 locally: 8 skips from tree-sitter, 1 from leidenalg.
	- name: Install dependencies (no postgresql extra)
	# Hash-pinned from uv.lock (scripts/generate_pip_constraints.py).
	# --no-deps on the project install because the file above IS the
	# complete dependency graph; re-resolving here would be unpinned.
	run: \|
	pip install --require-hashes -r requirements/ci-sqlite.txt
	pip install --no-deps -e .

	# Retry-with-backoff, fail-loudly: see the `test` job's pre-download step
	# for the root-cause rationale (CI run 28495801728, 2026-07-01).
	- name: Pre-download embedding model
	run: \|
	for attempt in 1 2 3 4 5; do
	python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('all-MiniLM-L6-v2', device='cpu')" && exit 0
	echo "HF pre-download attempt ${attempt} failed; retrying in $((attempt * 10))s" >&2
	sleep $((attempt * 10))
	done
	echo "HF pre-download failed after 5 attempts" >&2
	exit 1

	# Same rationale as the `test` job's identically-named steps (CI run
	# 30592244731, 2026-07-31): this job also installs the `codebase` extra
	# (comment above), so it is equally exposed to a cold-cache mid-suite
	# `DownloadError` without this.
	- name: Resolve tree-sitter cache directory
	run: echo "TREE_SITTER_CACHE_DIR=$(python -c 'from tree_sitter_language_pack import cache_dir; print(cache_dir())')" >> "$GITHUB_ENV"

	- name: Cache tree-sitter grammars
	uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
	with:
	path: ${{ env.TREE_SITTER_CACHE_DIR }}
	key: ${{ runner.os }}-tree-sitter-grammars-${{ hashFiles('pyproject.toml') }}

	- name: Prefetch tree-sitter grammars
	run: \|
	for attempt in 1 2 3 4 5; do
	python -c "from mcp_server.core.ast_parser import AST_SUPPORTED; from tree_sitter_language_pack import prefetch; prefetch(sorted(AST_SUPPORTED))" && exit 0
	echo "tree-sitter grammar prefetch attempt ${attempt} failed; retrying in $((attempt * 10))s" >&2
	sleep $((attempt * 10))
	done
	echo "tree-sitter grammar prefetch failed after 5 attempts" >&2
	exit 1

	- name: Run SQLite backend tests
	env:
	HF_HUB_OFFLINE: "1"
	TRANSFORMERS_OFFLINE: "1"
	# SQLite is the plugin's DEFAULT backend, so it gets the same gate
	# PostgreSQL does. Explicit rather than inferred: this job has no PG
	# service container, and conftest would select SQLite on its own, but
	# a future runner with PG reachable must not silently turn this into
	# a second PostgreSQL run.
	CORTEX_MEMORY_STORE_BACKEND: sqlite
	# No model download from a CI runner; degrade to first-stage scores.
	CORTEX_RERANKER_OFFLINE: "1"
	# Scope: THE FULL SUITE. This job previously ran one file
	# (test_sqlite_backend.py) and deferred the rest to a "full-parity
	# effort" that named no issue. That gap let three real defects ship on
	# the default backend — missing acquire_interactive/acquire_batch, the
	# PG-only lesson-promotion query, and a bare except that reported the
	# resulting failure as an empty backlog (issue #220).
	#
	# Measured 2026-07-28 on this tree (rebased onto 575f2f1, so it
	# includes the tests #229/#230 added), macOS 15 / Python 3.13,
	# `CORTEX_MEMORY_STORE_BACKEND=sqlite`: 6103 passed, 96 skipped,
	# exit 0 in 247s.
	#
	# Those 96 skips, counted (not estimated) from a `-rs` run:
	# 84 PostgreSQL-only — the test's SUBJECT is the PG implementation
	# (pg_store_* dialect modules, PG-only maintenance passes). These
	# SHOULD skip here; making them "backend-agnostic" would mean
	# testing PG code without PG.
	# 12 optional deps absent from that local env (8 tree-sitter,
	# 1 leidenalg, 3 sqlite-vec). All three ARE installed on this
	# job, so it should report ~84 skips, not 96 — if it reports
	# more, an extra is missing and the gate has silently narrowed.
	run: pytest --tb=short -q -p no:randomly

	# Protocol-level, hook-free proof for representative local client
	# identities. This starts the installed production console entry point,
	# under a SOCKS-proxy environment with no socks extra, then requires a
	# full initialize/discovery round-trip for both full and lean profiles,
	# plus a real SQLite memory_stats call. It is intentionally on the
	# SQLite job: that is the zero-config backend used by Gemini/Codex
	# installs, and all Python dependencies are already present here.
	- name: Verify hook-free MCP host contract
	env:
	HF_HUB_OFFLINE: "1"
	TRANSFORMERS_OFFLINE: "1"
	CORTEX_RERANKER_OFFLINE: "1"
	run: python scripts/verify_mcp_hosts.py -- hypermnesia-mcp

	mcp-host-config:
	name: Validate MCP host configurations
	# The pinned Claude package requires its postinstall to select the native
	# validator. Never execute install scripts from a lockfile modified by an
	# untrusted fork; same-repository PRs and main/workflow_dispatch remain
	# covered by this vendor-parser contract.
	if: github.event_name != 'pull_request' \|\| github.event.pull_request.head.repo.fork == false
	runs-on: ubuntu-latest
	permissions:
	contents: read

	steps:
	- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
	with:
	persist-credentials: false

	- name: Set up Node
	uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
	with:
	# @anthropic-ai/claude-code 2.1.220 declares Node >=22; Node 24 is
	# the repository's existing release-toolchain pin.
	node-version: "24"

	- name: Install uv (pinned)
	uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
	with:
	version: "0.11.3"

	# Vendor parsers catch configuration drift that a generic MCP client
	# cannot: Claude's custom plugin component path, Gemini's extension
	# schema, and Codex's config.toml transport/timeouts. Versions and bytes
	# are pinned in package-lock.json. This isolated, contents-read-only job
	# intentionally permits install scripts: Claude Code's package uses its
	# postinstall to replace a non-executable error stub with the pinned
	# platform-native validator binary. `--ignore-scripts` was verified to
	# leave `claude plugin validate` unusable.
	- name: Install pinned MCP host CLIs
	run: npm ci --prefix tests_js/mcp-host-clis

	- name: Validate Claude, Gemini, and Codex configuration parsing
	run: \|
	set -euo pipefail
	cli_bin="tests_js/mcp-host-clis/node_modules/.bin"
	CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 \
	"$cli_bin/claude" plugin validate .
	GEMINI_CLI_HOME="$RUNNER_TEMP/cortex-gemini-home" \
	"$cli_bin/gemini" extensions validate .
	"$cli_bin/codex" mcp list --json \
	-c 'mcp_servers={cortex={command="hypermnesia-mcp",startup_timeout_sec=30,tool_timeout_sec=600,env={CORTEX_MEMORY_STORE_BACKEND="sqlite"}}}' \
	\| python -c 'import json,sys; servers={s["name"]:s for s in json.load(sys.stdin)}; c=servers["cortex"]; assert c["transport"]["type"]=="stdio"; assert c["transport"]["command"]=="hypermnesia-mcp"; assert c["transport"]["env"]["CORTEX_MEMORY_STORE_BACKEND"]=="sqlite"; assert c["startup_timeout_sec"]==30.0; assert c["tool_timeout_sec"]==600.0'
	marketplace_ref="${GITHUB_HEAD_REF:-$GITHUB_REF_NAME}"
	"$cli_bin/codex" plugin marketplace add cdeust/Cortex \
	--ref "$marketplace_ref" --json
	"$cli_bin/codex" plugin list --available --json \
	\| python -c 'import json,sys; d=json.load(sys.stdin); expected=json.load(open("package.json"))["version"]; ps=d.get("available",[])+d.get("installed",[]); p=[x for x in ps if x.get("pluginId")=="hypermnesia-mcp-codex@cortex-codex-plugins"]; assert len(p)==1, p; assert p[0]["version"]==expected'
	"$cli_bin/codex" plugin add hypermnesia-mcp-codex@cortex-codex-plugins --json
	"$cli_bin/codex" plugin list --json \
	\| python -c 'import json,sys; d=json.load(sys.stdin); p=[x for x in d.get("installed",[]) if x.get("pluginId")=="hypermnesia-mcp-codex@cortex-codex-plugins"]; assert len(p)==1, p; assert p[0]["installed"] and p[0]["enabled"]'
	"$cli_bin/codex" mcp list --json \
	\| python -c 'import json,sys; ss=[s for s in json.load(sys.stdin) if s.get("name")=="cortex"]; assert len(ss)==1, ss; s=ss[0]; assert s["startup_timeout_sec"]==180.0, s; assert s["transport"]["command"]=="uvx", s; assert s["transport"]["env"]["CORTEX_RUNTIME"]=="cowork", s'
	mapfile -t plugin_command < <(
	python -c 'import json; s=json.load(open("plugins/hypermnesia-mcp-codex/.mcp.json"))["mcpServers"]["cortex"]; print(s["command"]); print(*s["args"], sep="\n")'
	)
	plugin_timeout="$(
	python -c 'import json; print(json.load(open("plugins/hypermnesia-mcp-codex/.mcp.json"))["mcpServers"]["cortex"]["startup_timeout_sec"])'
	)"
	plugin_runtime="$(
	python -c 'import json; print(json.load(open("plugins/hypermnesia-mcp-codex/.mcp.json"))["mcpServers"]["cortex"]["env"]["CORTEX_RUNTIME"])'
	)"
	CORTEX_RUNTIME="$plugin_runtime" \
	UV_CACHE_DIR="$RUNNER_TEMP/cortex-codex-cold-uv-cache" \
	UV_TOOL_DIR="$RUNNER_TEMP/cortex-codex-cold-uv-tools" \
	PYTHONPATH="$GITHUB_WORKSPACE" \
	python scripts/verify_mcp_hosts.py \
	--timeout "$plugin_timeout" --clients codex-cli --profiles lean \
	--command-includes-profile \
	--allow-bootstrap-network \
	--storage-selection auto \
	-- "${plugin_command[@]}"

	test-windows:
	name: Test (Windows, SQLite backend)
	runs-on: windows-latest

	# Real NT proof for the cross-platform fixes (fcntl→msvcrt, sys.executable,
	# NTFS exec bits, $HOME override, path separators). We use the SQLite
	# fallback so no PostgreSQL has to be provisioned on the Windows runner —
	# the conftest selects it when PG is unreachable; the env var makes it
	# explicit. source: RAPPORT_INSTALLATION_CORTEX_WINDOWS.md §10 (CI Windows)
	env:
	CORTEX_MEMORY_STORE_BACKEND: sqlite

	steps:
	- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

	- name: Set up Python
	uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
	with:
	python-version: "3.12"

	- name: Cache HuggingFace models
	uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
	with:
	path: ~/.cache/huggingface
	key: ${{ runner.os }}-hf-all-MiniLM-L6-v2

	- name: Install dependencies (no postgresql extra)
	# Hash-pinned from uv.lock (scripts/generate_pip_constraints.py).
	# --no-deps on the project install because the file above IS the
	# complete dependency graph; re-resolving here would be unpinned.
	run: \|
	pip install --require-hashes -r requirements/ci-sqlite-min.txt
	pip install --no-deps -e .

	# Import smoke: the modules that previously crashed at load on Windows
	# (fcntl import) or silently misbehaved. If any fails to import, the
	# platform branches are wrong — fail fast before the suite.
	- name: Import smoke (formerly Windows-broken modules)
	run: >-
	python -c "import mcp_server.shared.platform,
	mcp_server.infrastructure.pipeline_install_lock,
	mcp_server.infrastructure.pipeline_discovery,
	mcp_server.core.staleness, mcp_server.doctor, mcp_server.doctor_mcp;
	print('windows import smoke OK')"

	# Retry-with-backoff, fail-loudly: see the `test` job's pre-download step
	# for the root-cause rationale (CI run 28495801728, 2026-07-01). shell: bash
	# so the retry loop runs under Git Bash rather than the Windows default pwsh.
	- name: Pre-download embedding model
	shell: bash
	run: \|
	for attempt in 1 2 3 4 5; do
	python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('all-MiniLM-L6-v2', device='cpu')" && exit 0
	echo "HF pre-download attempt ${attempt} failed; retrying in $((attempt * 10))s" >&2
	sleep $((attempt * 10))
	done
	echo "HF pre-download failed after 5 attempts" >&2
	exit 1

	# Real postInstall proof (issue #113). Before this step, no CI job ever
	# ran plugin.json's actual postInstall command
	# (`bash scripts/install-plugin.sh`) on any OS — the ubuntu `test` job
	# above installs deps directly via `pip install -e`, bypassing it
	# entirely. That gap is why scripts/setup.sh's "Unsupported OS" failure
	# on native Windows (Git Bash reports uname -s as MINGW64_NT-*/
	# MSYS_NT-/CYGWIN_NT-, matched by neither its Darwin nor Linux branch)
	# went undetected. The job-level CORTEX_MEMORY_STORE_BACKEND: sqlite
	# (above) makes install-plugin.sh/scripts/setup.py take the SQLite
	# path — now the plugin's production DEFAULT, not a CI-only mode —
	# so this runs without provisioning a PostgreSQL server on the
	# runner, while still exercising the real backend selection +
	# marker write in install-plugin.sh and the real scripts/setup.py
	# subprocess calls. The embedding model is deliberately NOT cached
	# by this path anymore (lazy first-use download); the HF
	# pre-download step above provides the model for the test steps.
	- name: Exercise real postInstall path (install-plugin.sh -> setup.py)
	shell: bash
	env:
	CLAUDE_PLUGIN_ROOT: ${{ github.workspace }}
	CLAUDE_PLUGIN_DATA: ${{ runner.temp }}/cortex-plugin-data
	run: bash scripts/install-plugin.sh

	# Scope (explicit, not silent): the portability tests plus the modules
	# carrying Windows-specific branches and the SQLite backend suite. The
	# full PG suite is not run here — it is covered by the ubuntu `test` job.
	- name: Run portability + backend tests
	env:
	HF_HUB_OFFLINE: "1"
	TRANSFORMERS_OFFLINE: "1"
	run: >-
	pytest --tb=short -q
	tests_py/shared/test_platform.py
	tests_py/infrastructure/test_pipeline_discovery.py
	tests_py/infrastructure/test_pipeline_install_lock.py
	tests_py/core/test_staleness.py
	tests_py/infrastructure/test_sqlite_backend.py
	tests_py/scripts/test_setup_py_backend_skip.py

	# Validates `requirements/release.txt` — the dependency set release.yml
	# installs before publishing (issue #392). Removing release.yml's own test
	# job (see release-gate below) would otherwise leave that narrower
	# dependency set validated NOWHERE until a PyPI/uvx install broke on it.
	#
	# Deliberately NOT the full pytest suite: this job's subject is the
	# DEPENDENCY SET, not the suite — ci.yml's `test` matrix already covers the
	# suite itself, against a wider dependency set. Re-running pytest here
	# would duplicate that coverage at the release-narrowed dependency set's
	# cost without adding a new claim.
	#
	# release.txt deliberately omits tree-sitter, tree-sitter-language-pack,
	# igraph, leidenalg and texttable (`.github/actions/test-suite/action.yml`
	# header comment) — so nothing this job runs may import from those
	# packages, and it must not exercise codebase_analyze/AST paths.
	#
	# The check itself reuses the cheapest existing mechanism instead of
	# inventing a new one: `scripts/verify_mcp_hosts.py`'s stdio initialize +
	# tools/list + memory_stats round-trip (the same invocation the
	# `test-sqlite` job's "Verify hook-free MCP host contract" step already
	# runs) proves the server imports and its full tool surface registers
	# under this narrower dependency set, with no PostgreSQL service needed
	# (`--storage-selection sqlite` is the default).
	release-deps:
	name: Release dependency set (requirements/release.txt)
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

	- name: Set up Python
	uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
	with:
	python-version: "3.12"

	- name: Cache HuggingFace models
	uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
	with:
	path: ~/.cache/huggingface
	key: ${{ runner.os }}-hf-all-MiniLM-L6-v2

	- name: Install the release dependency set (hash-pinned)
	run: \|
	pip install --require-hashes -r requirements/release.txt
	pip install --no-deps -e .

	# Retry-with-backoff, fail-loudly: see the `test` job's pre-download
	# step for the root-cause rationale (CI run 28495801728, 2026-07-01).
	- name: Pre-download embedding model
	run: \|
	for attempt in 1 2 3 4 5; do
	python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('all-MiniLM-L6-v2', device='cpu')" && exit 0
	echo "HF pre-download attempt ${attempt} failed; retrying in $((attempt * 10))s" >&2
	sleep $((attempt * 10))
	done
	echo "HF pre-download failed after 5 attempts" >&2
	exit 1

	- name: Verify hook-free MCP host contract under the release dependency set
	env:
	HF_HUB_OFFLINE: "1"
	TRANSFORMERS_OFFLINE: "1"
	CORTEX_RERANKER_OFFLINE: "1"
	run: python scripts/verify_mcp_hosts.py -- hypermnesia-mcp

	lint:
	name: Lint
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

	- name: Set up Python
	uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
	with:
	python-version: "3.12"

	- name: Install ruff
	# Pinned: ruff's formatter output changes across minor versions
	# (0.15.6 vs 0.15.20 divergence broke Lint on PR #83). The repo is
	# formatted with 0.15.20; bump this pin and reformat together.
	run: pip install --require-hashes -r requirements/lint.txt

	- name: Check formatting
	run: ruff format --check .

	- name: Check linting
	run: ruff check .

	# Advertised counts (tools, references, mechanisms) must match the
	# repository. Runs here — on every push and PR — because doc drift is
	# introduced at commit time, not at release time: on 2026-07-27 README,
	# CONTRIBUTING, CLAUDE.md and the MCPB manifest each advertised a
	# different tool count, and nothing failed. Static only (no imports).
	- name: Check documentation claims
	run: python scripts/check_doc_claims.py

	# [project].version in pyproject.toml against every one of the 14
	# version sites across 11 files (manifests, plugin.jsons, the
	# marketplace's metadata AND primary-entry version, the MCP registry
	# package entry, the four textual occurrences in the version badge,
	# and uv.lock's own root-package version). Two of these sites
	# (server.json packages[0].version, marketplace.json metadata.version)
	# were covered by nothing before this gate existed, and the one gate
	# that came closest — marketplace-pins.yml — is path-filtered, so it
	# sits outside ci-green and only fires on its weekly cron (issue #392).
	# Static only (no imports), same reason as the doc-claim gate above.
	- name: Check version surfaces
	run: python scripts/check_version_surfaces.py

	# The `ci-green` job below is the single status check branch protection
	# names; the list of jobs it covers lives in its `needs:`. A job added
	# to ci.yml but not to that list would run outside the gate and could
	# fail without blocking a merge. Static only (no imports), same reason
	# as the doc-claim gate above.
	- name: Check the aggregate CI gate covers every job
	run: python scripts/check_ci_gate_complete.py

	# The README's repo-derived badges are committed SVGs, not hotlinked
	# images, so nothing regenerates them on view: a figure that moves
	# leaves the badge asserting the old one. Checked here, on every push
	# and PR, because that drift is introduced at commit time. Static only
	# (no network, no suite) — the tests badge needs a collected count and
	# is checked in the test job instead.
	- name: Check committed badges
	run: python scripts/generate_repo_badges.py --check

	# The requirements/ files are the ONLY thing standing between a pip
	# install and an unpinned one, and they are generated — so a lock
	# change that is not re-exported leaves them describing dependencies
	# nobody resolved. Checked here, on every push and PR, because that
	# drift is introduced at commit time. Static: reads uv.lock, installs
	# nothing.
	# uv arrives via the SHA-pinned action, not `pip install uv==0.11.3`:
	# Scorecard scores a bare version specifier as "pipCommand not pinned
	# by hash" (an `==` resolves to whatever bytes the index serves under
	# that version today), so installing the pinning tool with pip would
	# have minted the 22nd Pinned-Dependencies alert in the very commit
	# that closes the other 21 — and the policy clears the check only at
	# score 10. Same form as release.yml's SBOM job.
	- name: Install uv (pinned)
	uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
	with:
	version: "0.11.3"

	# Runs BEFORE the requirements-file check below on purpose: that check
	# only compares the committed requirements/*.txt against whatever
	# uv.lock CURRENTLY says, so a stale uv.lock and its stale export agree
	# with each other and the check stays green. `uv lock --check` instead
	# re-resolves pyproject.toml and fails if uv.lock would change — the
	# half of "lockfile is current" the other check cannot see.
	#
	# This exact drift shipped (issue #251): Dependabot's `deps` PR #218
	# widened tree-sitter-language-pack's upper bound in pyproject.toml
	# (`<1.7` -> `<1.14`) but touched no other file — its "pip" ecosystem
	# entry rewrites the version specifier only, it does not re-run `uv
	# lock`. uv.lock kept resolving 1.6.2 (still valid under the wider
	# range, and `uv lock`'s incremental resolution prefers the existing
	# pin over the newest compatible one), so `.venv/bin/python -m pyright
	# mcp_server/` reported a real diagnostic for a contributor who ran
	# `uv sync --locked` while CI — which back then installed the package
	# editable with its extras resolved straight from pyproject.toml's
	# ranges, unconstrained by the lock (the exact pattern
	# test_job_does_not_pip_install_an_extra_range now forbids below) —
	# silently resolved 1.13.5 and stayed green on the same commit.
	#
	# Reproduced 2026-07-29 by checking out pyproject.toml + uv.lock as of
	# 9e293baa (#218, the commit right after the widening) into a scratch
	# directory and running `uv lock --check` there:
	# Resolved 196 packages in 186ms
	# The lockfile at `uv.lock` needs to be updated, but `--check` was
	# provided. To update the lockfile, run `uv lock`.
	# — i.e. this step would have failed, at commit time, on the PR that
	# introduced the drift. Static: reads pyproject.toml + uv.lock,
	# resolves, installs nothing (~200ms locally).
	- name: Check uv.lock matches pyproject.toml
	run: uv lock --check

	- name: Check hash-pinned requirements match uv.lock
	run: python scripts/generate_pip_constraints.py --check

	# actionlint (which shells out to the runner's preinstalled shellcheck for
	# each `run:` block) was never wired into any gate — issue #247, found by
	# hand after these findings had been latent since before the Node24 action
	# bump. Pinned by release tag + checksum (not `go install @latest` /
	# curl\|bash), matching this repo's supply-chain-hardening stance (release.yml).
	# source: https://github.com/rhysd/actionlint/releases/tag/v1.7.12,
	# actionlint_1.7.12_checksums.txt, linux_amd64 entry, verified 2026-07-29.
	- name: Install actionlint (pinned, checksum-verified)
	run: \|
	set -euxo pipefail
	VERSION="1.7.12"
	ARCHIVE="actionlint_${VERSION}_linux_amd64.tar.gz"
	curl -fsSLO "https://github.com/rhysd/actionlint/releases/download/v${VERSION}/${ARCHIVE}"
	echo "8aca8db96f1b94770f1b0d72b6dddcb1ebb8123cb3712530b08cc387b349a3d8 ${ARCHIVE}" \| sha256sum -c -
	tar -xzf "${ARCHIVE}" actionlint
	sudo install -m 0755 actionlint /usr/local/bin/actionlint
	actionlint -version

	- name: Check workflow files (actionlint + shellcheck)
	run: actionlint -color

	# Deterministic enforcement of CLAUDE.md § Code Style — file size, method
	# size, layer-boundary imports, and unsourced magic numbers — which that
	# section admitted was "enforced by code review today; no automated
	# pre-commit hook checks this yet" until this job. Measured on a single
	# PR the night before this job was added: a 301-line file reported as
	# 280, three of four over-40-line methods unseen, one layer violation
	# justified by a fabricated citation — each caught only by a human or
	# agent re-reading the diff, never by a machine. See
	# scripts/check_craftsmanship.py's module docstring for the rules and
	# scripts/craftsmanship_baseline.py for why pre-existing debt (recorded
	# in .craftsmanship-baseline.json) does not retroactively block.
	craftsmanship:
	name: Craftsmanship Gate
	runs-on: ubuntu-latest
	steps:
	# fetch-depth: 0 so `origin/main` — the diff base the gate compares
	# the PR's changed files against — is resolvable locally, not just
	# the single commit a shallow checkout would leave.
	- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
	with:
	fetch-depth: 0

	- name: Set up Python
	uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
	with:
	python-version: "3.12"

	# Standard library only (scripts/check_craftsmanship.py's module
	# docstring) — no dependency install needed, matching the
	# doc-claim/version-surface/ci-gate-completeness gates in `lint`
	# below, which are static for the same reason.
	- name: Run the craftsmanship gate on this PR's changed files
	run: python scripts/check_craftsmanship.py

	typecheck:
	name: Type Check
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

	- name: Set up Python
	uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
	with:
	python-version: "3.13"

	# Pyright resolves third-party imports from ./.venv (pinned in
	# pyrightconfig.json: venvPath="."/venv=".venv"). The full stub set MUST
	# be installed or every import collapses to Unknown — and Unknown
	# SUPPRESSES downstream type errors, silently masking real defects.
	# source: measured 2026-06-18 — an unresolved env reports 566 errors, a
	# fully-resolved env reports 593 (Unknown was masking 27+ real errors).
	# flashrank (core reranker) + sqlite-vec live outside dev/postgresql/
	# codebase; [otel] resolves the opentelemetry exporter imports.
	- name: Create .venv with the full type-check environment
	run: \|
	python -m venv .venv
	.venv/bin/pip install --require-hashes -r requirements/ci-typecheck.txt
	.venv/bin/pip install --no-deps -e .
	# Pin pyright — diagnostic output drifts between releases, so a
	# zero-diagnostic tree is only comparable against the pinned version.
	.venv/bin/pip install --require-hashes -r requirements/typecheck-tool.txt

	# The gate's verdict is a property of THIS environment, so the log has to
	# name it. Issue #253: a contributor and CI read two different
	# `tree-sitter-language-pack` type surfaces on the same commit, and
	# nothing in either log said which one had been analysed. One grep per
	# package, not one alternation: a miss then exits non-zero on the
	# package that is actually absent, so an install that silently dropped
	# one fails here instead of degrading to Unknown inside pyright.
	- name: Record the resolved type surface
	run: \|
	set -euo pipefail
	freeze="$(.venv/bin/pip list --format=freeze)"
	for pkg in pyright tree-sitter tree-sitter-language-pack; do
	printf '%s\n' "${freeze}" \| grep -E "^${pkg}=="
	done

	# pyright IS the gate: the backlog was burned to zero (issue #197,
	# 568 baselined diagnostics -> 0 measured 2026-07-28), so ANY
	# diagnostic — any rule, any severity-error — fails the build via
	# pyright's own exit code. The former per-rule ratchet
	# (scripts/check_pyright_ratchet.py + typecheck-baseline.json) is
	# retired: a floor file only exists to tolerate a backlog.
	- name: Run pyright (zero-diagnostic gate)
	run: .venv/bin/python -m pyright mcp_server/

	build:
	name: Build Package
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

	- name: Set up Python
	uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
	with:
	python-version: "3.12"

	- name: Install build tools
	run: pip install --require-hashes -r requirements/packaging.txt

	- name: Build sdist and wheel
	run: python -m build

	- name: Upload build artifacts
	uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
	with:
	name: dist
	path: dist/

	# The two images below had NO CI build at all until this change, so a
	# regression in either was invisible until someone built it by hand. That
	# is not hypothetical: docker/Dockerfile copied a python3.12 site-packages
	# path against a python:3.14 base and could not build, and nothing said so.
	#
	# These build only — no smoke run. docker/Dockerfile brings up PostgreSQL
	# and pre-caches models, and .devcontainer/Dockerfile is opened by the Dev
	# Containers CLI; asserting their runtime behaviour is a separate job with
	# separate services. The claim here is narrow and worth making on its own:
	# the image still builds, and its hash-pinned installs still resolve.
	docker-runtime-build:
	name: Docker Build (runtime image)
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

	# This image carries PostgreSQL 17, Node 22, torch and a pre-cached
	# embedding model. A stock runner has ~14GB free, which the build can
	# exhaust; the preinstalled toolchains below are worth ~10GB and are
	# not used by it.
	- name: Free runner disk
	run: \|
	sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc
	sudo rm -rf /usr/local/share/boost "$AGENT_TOOLSDIRECTORY"
	df -h /

	- name: Set up Docker Buildx
	uses: docker/setup-buildx-action@bb05f3f5519dd87d3ba754cc423b652a5edd6d2c # v4.2.0

	- name: Build docker/Dockerfile
	uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7.3.0
	with:
	context: .
	file: ./docker/Dockerfile
	tags: cortex-runtime:ci
	load: false
	cache-from: type=gha
	cache-to: type=gha,mode=max

	devcontainer-build:
	name: Docker Build (devcontainer image)
	runs-on: ubuntu-latest
	steps:
	- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

	- name: Free runner disk
	run: \|
	sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc
	sudo rm -rf /usr/local/share/boost "$AGENT_TOOLSDIRECTORY"
	df -h /

	- name: Set up Docker Buildx
	uses: docker/setup-buildx-action@bb05f3f5519dd87d3ba754cc423b652a5edd6d2c # v4.2.0

	- name: Build .devcontainer/Dockerfile
	uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7.3.0
	with:
	context: .
	file: ./.devcontainer/Dockerfile
	tags: cortex-devcontainer:ci
	load: false
	cache-from: type=gha
	cache-to: type=gha,mode=max

	docker-smoke:
	name: Docker Smoke (bare-container DB-less contract)
	runs-on: ubuntu-latest
	# BLOCKING: this job asserts the exact contract that silently broke for
	# two months (fix/bare-container-contract, commit 5d71069c) — third-party
	# registry indexers (Glama et al.) `docker build` the bare repo, then
	# `docker run` with zero env vars and zero external services, and expect
	# `tools/list` to answer. No other CI job exercises this path: `test` and
	# `test-sqlite` both install from source with dev extras and never build
	# the production image; nothing else runs the image with psycopg absent
	# and no DATABASE_URL. continue-on-error is intentionally NOT set.
	steps:
	- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

	- name: Set up Docker Buildx
	uses: docker/setup-buildx-action@bb05f3f5519dd87d3ba754cc423b652a5edd6d2c # v4.2.0

	# GitHub Actions cache (type=gha) persists Docker layer cache across
	# workflow runs, scoped to this repo. Without it, every run repeats the
	# torch CPU-wheel download (~200MB, pinned via the CPU-only index in
	# the Dockerfile) and the full pip install layer. With a warm cache,
	# only changed layers (COPY mcp_server, pip install of the local
	# package) rebuild.
	# source: docker/build-push-action README, "GitHub Cache" backend
	# (https://github.com/docker/build-push-action#cache-backend-api),
	# type=gha is the documented zero-config option for GHA runners.
	- name: Build image (buildx, GHA layer cache)
	uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7.3.0
	with:
	context: .
	file: ./Dockerfile
	tags: cortex-smoke:local
	load: true
	cache-from: type=gha
	cache-to: type=gha,mode=max

	# Reuses the build above (load: true made it visible to the local
	# docker daemon) instead of rebuilding — scripts/docker_smoke.sh is the
	# single source of truth for the build+run+assert sequence, shared with
	# local dev (`scripts/docker_smoke.sh` with no args builds AND smokes).
	- name: Run bare-container smoke test
	run: scripts/docker_smoke.sh --skip-build
	env:
	CORTEX_SMOKE_IMAGE: cortex-smoke:local

	# The one status check branch protection on `main` names. Everything else
	# in this workflow is reached through its `needs:` list, so renaming a job,
	# resizing the matrix, or delegating steps to a reusable workflow (which
	# rewrites check names to "<caller> / <callee>", issue #336) no longer
	# touches the protection settings — the contract is this file, in git,
	# visible in review, instead of eleven job-name strings in GitHub settings
	# that no diff ever shows.
	#
	# `if: always()` is load-bearing: without it this job is itself skipped the
	# moment any need fails, the required context is never reported, and the PR
	# blocks with no explanation instead of a red X naming the culprit.
	#
	# scripts/check_ci_gate_complete.py (run by `lint`) refuses a merge if any
	# job here is missing from `needs`, or if a job carrying a job-level `if:`
	# is absent from ALLOWED_SKIPS below.
	ci-green:
	name: CI Green
	if: always()
	runs-on: ubuntu-latest
	needs:
	- test
	- test-sqlite
	- mcp-host-config
	- test-windows
	- release-deps
	- craftsmanship
	- lint
	- typecheck
	- build
	- docker-runtime-build
	- devcontainer-build
	- docker-smoke
	steps:
	- name: Require every job above to have succeeded
	env:
	NEEDS: ${{ toJSON(needs) }}
	# Jobs permitted to report `skipped`, with the reason. ONLY jobs
	# carrying a job-level `if:` belong here; anything else that skips
	# is a gap, not an exemption.
	# mcp-host-config — `if: github.event_name != 'pull_request' \|\|
	# github.event.pull_request.head.repo.fork == false`: a fork PR
	# has no access to the secrets it needs, so it cannot run there.
	ALLOWED_SKIPS: mcp-host-config
	run: \|
	set -euo pipefail
	echo "$NEEDS"
	failed=$(printf '%s' "$NEEDS" \| jq -r --arg allowed "$ALLOWED_SKIPS" '
	($allowed \| split(",") \| map(ltrimstr(" ") \| rtrimstr(" "))) as $ok
	\| to_entries[]
	\| select(.value.result != "success")
	\| select(.value.result != "skipped" or ([.key] \| inside($ok) \| not))
	\| "\(.key)=\(.value.result)"')
	if [ -n "$failed" ]; then
	echo "::error::CI Green refuses the merge — $(echo "$failed" \| tr '\n' ' ')"
	exit 1
	fi
	echo "every required job succeeded"

	# The tag-as-output half of issue #392: a green push to `main` tags the
	# exact SHA that just passed CI Green, instead of a human hand-picking
	# which (possibly-unvalidated) tree to tag — the v4.17.0 root cause
	# (release run 30741657854 tagged a tree release.yml's own test job had
	# never actually validated against ci.yml's hardening).
	#
	# ci-gate-exempt: this job runs ONLY on a push to `main`, i.e. AFTER the PR
	# it belongs to has already merged. There is nothing left for it to gate —
	# it cannot block a PR that is already closed — so it does not appear in
	# ci-green.needs (see scripts/check_ci_gate_complete.py's ci-gate-exempt
	# marker convention). Putting it there would make the branch-protection
	# context depend on a job that never runs on a PR in the first place,
	# permanently blocking every PR.
	release-gate:
	name: Tag a release
	needs: [ci-green]
	# ci-gate-exempt: runs only on push to main, after the PR it belongs to
	# has already merged — see the job comment above.
	if: github.event_name == 'push' && github.ref == 'refs/heads/main'
	runs-on: ubuntu-latest
	# `contents: read`, NOT write, and that is deliberate. The tag is pushed
	# with the RELEASE_TAG_SSH_KEY deploy key (persisted by the checkout step
	# below), never
	# with GITHUB_TOKEN — granting GITHUB_TOKEN `contents: write` here would
	# be an unused permission (issue #178 least-privilege), and worse, it
	# would let a future edit that drops `ssh-key:` from the checkout push the
	# tag with GITHUB_TOKEN and SUCCEED — producing a tag that silently never
	# starts release.yml. Read-only makes that mistake fail loudly instead.
	permissions:
	contents: read
	# A DEDICATED group — not this workflow's `${{ github.workflow }}-${{
	# github.ref }}` group above, which sets `cancel-in-progress: true`. A
	# push to main that lands mid-tag-push would be cancelled by the very
	# next push to main under that group, leaving a partially-created tag on
	# the remote (a tag object pushed but no corresponding release started,
	# or the reverse) — exactly the half-state a release process must never
	# produce. This group instead serializes tag operations one at a time
	# without ever cancelling one already in flight.
	concurrency:
	group: release-tag
	cancel-in-progress: false
	steps:
	# secrets.RELEASE_TAG_SSH_KEY is the PRIVATE half of a repository deploy
	# key (ed25519, read_only=false, created 2026-08-08). A deploy key, and
	# not a PAT or a GitHub App token, because neither of those can be
	# minted through the API — both require a browser — and the deploy key
	# is the tighter credential anyway: scoped to this ONE repository,
	# carrying no account access, and with no expiry to silently break
	# releases. Per
	# GitHub's docs, a ref pushed with the default GITHUB_TOKEN (unlike
	# workflow_dispatch/repository_dispatch) does NOT start a new workflow
	# run — a GITHUB_TOKEN-pushed tag would therefore never reach
	# release.yml's `push: tags` trigger. Same guarded-optional-secret
	# shape as sync-ccplugins-fork.yml's `has_pat` step: absence is a
	# clean, loud skip, not a red job.
	- name: Guard — skip cleanly when RELEASE_TAG_SSH_KEY is missing
	id: guard
	env:
	SSH_KEY: ${{ secrets.RELEASE_TAG_SSH_KEY }}
	run: \|
	if [ -z "$SSH_KEY" ]; then
	echo "RELEASE_TAG_SSH_KEY secret not configured — skipping."
	echo "Add the private half of a write-enabled repository deploy"
	echo "key at Settings -> Secrets and variables -> Actions."
	echo "Releases stay manual until it exists."
	echo "has_deploy_key=false" >> "$GITHUB_OUTPUT"
	else
	echo "has_deploy_key=true" >> "$GITHUB_OUTPUT"
	fi

	# fetch-depth: 0 so `git rev-parse --verify refs/tags/vX` below can see
	# every existing tag, not just the tip commit. `ssh-key:` installs the
	# deploy key as the credential `git push` uses later in this
	# job (actions/checkout persists it by default) — the non-GITHUB_TOKEN
	# credential the tag push requires (see the guard step's comment).
	- name: Checkout the SHA that just went green
	if: steps.guard.outputs.has_deploy_key == 'true'
	uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
	with:
	ref: ${{ github.sha }}
	fetch-depth: 0
	ssh-key: ${{ secrets.RELEASE_TAG_SSH_KEY }}

	- name: Set up Python
	if: steps.guard.outputs.has_deploy_key == 'true'
	uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
	with:
	python-version: "3.12"

	# Reuses doc_claim_sources.canonical_version — the same reader
	# check_version_surfaces.py and check_doc_claims.py already trust —
	# instead of re-deriving a second pyproject.toml version regex here.
	# A push to `main` for a commit that does not bump the version is the
	# NOMINAL case (most commits are not releases): tag v$VERSION already
	# exists, and this step's job is to detect that and skip silently
	# rather than fail or re-tag.
	- name: Resolve the version and check for an existing tag
	id: tag_check
	if: steps.guard.outputs.has_deploy_key == 'true'
	run: \|
	set -euo pipefail
	VERSION="$(python3 -c "
	import sys
	sys.path.insert(0, 'scripts')
	import doc_claim_sources
	print(doc_claim_sources.canonical_version(lambda p: open(p, encoding='utf-8').read()))
	")"
	echo "version=${VERSION}" >> "$GITHUB_OUTPUT"
	if git rev-parse -q --verify "refs/tags/v${VERSION}" >/dev/null; then
	echo "tag v${VERSION} already exists — this push carries no version bump"
	echo "exists=true" >> "$GITHUB_OUTPUT"
	else
	echo "no existing tag v${VERSION} — this push is a version bump"
	echo "exists=false" >> "$GITHUB_OUTPUT"
	fi

	# The last line of defence before a tag: `main` should never carry a
	# partial bump anyway (the identical gate already runs in `lint` on
	# every push and PR — commit 4781c134), so reaching this step with a
	# mismatch means something bypassed the PR gate (a direct push, an
	# admin-merge override). A failure here is a RED job, not a skip —
	# silently tagging a mismatched tree is exactly the v4.17.0 failure
	# mode this workflow exists to make structurally impossible.
	- name: Check version surfaces
	if: steps.guard.outputs.has_deploy_key == 'true' && steps.tag_check.outputs.exists == 'false'
	run: python scripts/check_version_surfaces.py

	- name: Configure git identity
	if: steps.guard.outputs.has_deploy_key == 'true' && steps.tag_check.outputs.exists == 'false'
	run: \|
	git config user.name "github-actions[bot]"
	git config user.email "41898282+github-actions[bot]@users.noreply.github.com"

	# Annotated (not lightweight): carries its own object, author, and
	# message, which `git tag -a` computes but `git tag` alone does not —
	# the shape release.yml's changelog step and GitHub's own release UI
	# expect. Tagged at `github.sha` explicitly (not the checkout's
	# working-tree HEAD) so the released commit is provably the one that
	# went green, even if a future edit changes what this job checks out.
	- name: Create and push the release tag
	if: steps.guard.outputs.has_deploy_key == 'true' && steps.tag_check.outputs.exists == 'false'
	run: \|
	set -euo pipefail
	VERSION="${{ steps.tag_check.outputs.version }}"
	git tag -a "v${VERSION}" -m "Release v${VERSION}" "${{ github.sha }}"
	git push origin "refs/tags/v${VERSION}"

	# Visible without opening logs: the decision (tagged / no-bump /
	# no-token) is exactly the property issue #392's acceptance criteria
	# ask to verify ("a green push to main with no version change tags
	# nothing" — verifiable in the release-gate logs).
	- name: Summarize the release-gate decision
	if: always()
	run: \|
	if [ "${{ steps.guard.outputs.has_deploy_key }}" != "true" ]; then
	echo "## Release gate: skipped — RELEASE_TAG_SSH_KEY not configured" >> "$GITHUB_STEP_SUMMARY"
	elif [ "${{ steps.tag_check.outputs.exists }}" == "true" ]; then
	echo "## Release gate: skipped — v${{ steps.tag_check.outputs.version }} already tagged (no version bump on this push)" >> "$GITHUB_STEP_SUMMARY"
	else
	echo "## Release gate: tagged v${{ steps.tag_check.outputs.version }} at \`${{ github.sha }}\`" >> "$GITHUB_STEP_SUMMARY"
	fi

Navigation Menu

chore(marketplace): move hypermnesia-mcp-viz pin to v3.1.1 #1397

Workflow file

chore(marketplace): move hypermnesia-mcp-viz pin to v3.1.1 #1397

Uh oh!

Workflow file for this run