refactor: rename Sentinel to toolgauge everywhere user-facing

.github/workflows/ci.yml

@@ -84,7 +84,7 @@ jobs:
       # and runs. Locally a missing toolchain skips those tests; here it must fail
       # them, or "the build scans the example on every commit" would quietly stop
       # being true.
-      SENTINEL_REQUIRE_GO: "1"
+      TOOLGAUGE_REQUIRE_GO: "1"
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-go@v5

.github/workflows/conformance.yml

@@ -67,7 +67,7 @@ jobs:
       # Direction 1: the broker must PASS.
       - name: scan the broker
         run: |
-          uv run sentinel scan \
+          uv run toolgauge scan \
             --endpoint http://localhost:8080/mcp \
             --token "$SCAN_TOKEN" \
             --gate must --format sarif --out broker.sarif --no-color \
@@ -84,15 +84,15 @@ jobs:
       # Without this the job proves only that the harness can say yes.
       - name: the non-conformant fixture must fail the gate
         run: |
-          uv run sentinel fixture serve --profile nonconformant &
+          uv run toolgauge fixture serve --profile nonconformant &
           for _ in $(seq 1 30); do
             curl -sf -o /dev/null http://127.0.0.1:9000/mcp \
               -H 'Mcp-Method: server/discover' \
               -d '{"jsonrpc":"2.0","id":1,"method":"server/discover"}' && break
             sleep 1
           done
 
-          if uv run sentinel scan --endpoint http://127.0.0.1:9000/mcp --gate must --no-color; then
+          if uv run toolgauge scan --endpoint http://127.0.0.1:9000/mcp --gate must --no-color; then
             echo "::error::the non-conformant fixture PASSED the MUST gate — the harness is broken"
             exit 1
           fi
@@ -102,21 +102,21 @@ jobs:
       # "the rules work" from "the rules reject everything".
       - name: the conformant fixture must pass the gate
         run: |
-          uv run sentinel fixture serve --profile conformant &
+          uv run toolgauge fixture serve --profile conformant &
           for _ in $(seq 1 30); do
             curl -sf -o /dev/null http://127.0.0.1:9001/mcp \
               -H 'Mcp-Method: server/discover' \
               -d '{"jsonrpc":"2.0","id":1,"method":"server/discover"}' && break
             sleep 1
           done
-          uv run sentinel scan --endpoint http://127.0.0.1:9001/mcp --gate must --no-color
+          uv run toolgauge scan --endpoint http://127.0.0.1:9001/mcp --gate must --no-color
 
       # The deprecation inventory, in both directions too. It never fails a
       # gate — a deprecated feature that still works is a migration to plan, not
       # a defect — so the assertion is on the CONTENT rather than the exit code.
       - name: the broker depends on no deprecated feature
         run: |
-          uv run sentinel deprecations \
+          uv run toolgauge deprecations \
             --endpoint http://localhost:8080/mcp \
             --token "$SCAN_TOKEN" --format json --out broker-deprecations.json --no-color
           IN_USE=$(python -c "import json,sys; d=json.load(open('broker-deprecations.json')); print(sum(1 for f in d['features'] if f['inUse']))")
@@ -128,7 +128,7 @@ jobs:
 
       - name: the fixture's deprecations are all found
         run: |
-          uv run sentinel deprecations \
+          uv run toolgauge deprecations \
             --endpoint http://127.0.0.1:9000/mcp \
             --format json --out fixture-deprecations.json --no-color
           python - <<'PY'

.github/workflows/registry-scan.yml

@@ -43,7 +43,7 @@ jobs:
         run: |
           args=(registry-scan --out-dir scan --concurrency 8 --timeout 10)
           if [ -n "$LIMIT" ]; then args+=(--limit "$LIMIT"); fi
-          sentinel "${args[@]}"
+          toolgauge "${args[@]}"
           cat scan/summary.md >> "$GITHUB_STEP_SUMMARY"
       - uses: actions/upload-artifact@v4
         with:

.github/workflows/release.yml

@@ -1,10 +1,10 @@
-# On a GitHub release this publishes the wheel to PyPI as `sentinel-mcp-conformance`,
+# On a GitHub release this publishes the wheel to PyPI as `toolgauge`,
 # pushes a multi-arch container image to GHCR, signs both with cosign (keyless, using
 # this workflow's GitHub OIDC identity, so there is no signing key to leak) and
 # attaches a CycloneDX SBOM of the harness to the release.
 #
 # Verify a release:
-#   cosign verify ghcr.io/patsypppe/sentinel:<version> \
+#   cosign verify ghcr.io/patsypppe/toolgauge:<version> \
 #     --certificate-identity-regexp 'https://github.com/patsypppe/sentinel/.github/workflows/release.yml@.*' \
 #     --certificate-oidc-issuer https://token.actions.githubusercontent.com
 #
@@ -44,8 +44,8 @@ jobs:
         run: |
           uv venv /tmp/smoke --python 3.12
           uv pip install --python /tmp/smoke dist/*.whl
-          /tmp/smoke/bin/sentinel version
-          /tmp/smoke/bin/sentinel-mcp-conformance catalog validate
+          /tmp/smoke/bin/toolgauge version
+          /tmp/smoke/bin/toolgauge catalog validate
       - uses: actions/upload-artifact@v4
         with:
           name: dist
@@ -98,15 +98,15 @@ jobs:
           push: true
           provenance: true
           tags: |
-            ghcr.io/${{ github.repository_owner }}/sentinel:${{ steps.meta.outputs.version }}
-            ghcr.io/${{ github.repository_owner }}/sentinel:latest
+            ghcr.io/${{ github.repository_owner }}/toolgauge:${{ steps.meta.outputs.version }}
+            ghcr.io/${{ github.repository_owner }}/toolgauge:latest
       - name: sign the image by digest
         env:
           DIGEST: ${{ steps.push.outputs.digest }}
-        run: cosign sign --yes "ghcr.io/${{ github.repository_owner }}/sentinel@${DIGEST}"
+        run: cosign sign --yes "ghcr.io/${{ github.repository_owner }}/toolgauge@${DIGEST}"
       - name: the image runs
         run: |
-          docker run --rm "ghcr.io/${{ github.repository_owner }}/sentinel:${{ steps.meta.outputs.version }}" version
+          docker run --rm "ghcr.io/${{ github.repository_owner }}/toolgauge:${{ steps.meta.outputs.version }}" version
 
   attest:
     name: sign the wheel, attach the SBOM
@@ -134,11 +134,11 @@ jobs:
         with:
           path: harness
           format: cyclonedx-json
-          output-file: sentinel-harness.cdx.json
+          output-file: toolgauge-harness.cdx.json
           upload-artifact: false
           upload-release-assets: false
       - name: attach to the release
         env:
           GH_TOKEN: ${{ github.token }}
         run: |
-          gh release upload "$GITHUB_REF_NAME" dist/*.sigstore.json sentinel-harness.cdx.json --clobber
+          gh release upload "$GITHUB_REF_NAME" dist/*.sigstore.json toolgauge-harness.cdx.json --clobber

.pre-commit-hooks.yaml

@@ -4,28 +4,28 @@
 #     - repo: https://github.com/patsypppe/sentinel
 #       rev: v0.4.0
 #       hooks:
-#         - id: sentinel-scan
+#         - id: toolgauge-scan
 #
-# Point it at a server with a sentinel.toml in your repository (scan.endpoint) or
-# the SENTINEL_ENDPOINT environment variable. It runs on pre-push by default: a
+# Point it at a server with a toolgauge.toml in your repository (scan.endpoint) or
+# the TOOLGAUGE_ENDPOINT environment variable. It runs on pre-push by default: a
 # scan makes a few hundred requests and is not something to run on every commit.
 
 # Runs the published image. Needs Docker, nothing else, and cannot drift from the
 # version pinned by `rev`.
-- id: sentinel-scan
-  name: sentinel MCP conformance scan
+- id: toolgauge-scan
+  name: toolgauge MCP conformance scan
   description: Grade a locally-running MCP server against the current MCP spec and fail on a MUST violation.
-  entry: ghcr.io/patsypppe/sentinel:latest scan --gate must --no-color
+  entry: ghcr.io/patsypppe/toolgauge:latest scan --gate must --no-color
   language: docker_image
   pass_filenames: false
   always_run: true
   stages: [pre-push, manual]
 
-# For machines that already have `sentinel` installed (pip, pipx or uvx).
-- id: sentinel-scan-local
-  name: sentinel MCP conformance scan (installed)
-  description: Same scan, using the sentinel command already on PATH.
-  entry: sentinel scan --gate must --no-color
+# For machines that already have `toolgauge` installed (pip, pipx or uvx).
+- id: toolgauge-scan-local
+  name: toolgauge MCP conformance scan (installed)
+  description: Same scan, using the toolgauge command already on PATH.
+  entry: toolgauge scan --gate must --no-color
   language: system
   pass_filenames: false
   always_run: true

CHANGELOG.md

@@ -2,6 +2,18 @@
 
 ## 0.4.0
 
+### Renamed: Sentinel is now toolgauge
+
+The project was called Sentinel. That name collided in three places: the PyPI package `sentinel`
+installs a top-level `sentinel` module, HashiCorp's Sentinel policy tool installs a `sentinel` command,
+and `mcp-sentinel` is a different MCP testing project. Everything user-facing is now `toolgauge`: the
+PyPI distribution, the import package, the command, the rule namespace (`TOOLGAUGE/...` instead of
+`SENTINEL/...`), the config file (`toolgauge.toml`), environment variables (`TOOLGAUGE_*`), the
+history service (`toolgauge-server`), the Docker image and the GitHub Action. Baselines written by an
+earlier build still load (`sentinelVersion` is read as a fallback); their `SENTINEL/` rule ids no
+longer match, so recreate a baseline after upgrading. The reference server's internal names (database,
+dev headers, `*.sentinel.local`) and the repository URL are unchanged.
+
 The harness stops being a one-shot grader and becomes something a team runs.
 
 ### Measured: the whole MCP registry
@@ -18,11 +30,11 @@ aggregate as an annotation so it can be read without downloading an artifact.
 refined, 7 of 12 after, on a fresh draw; specification findings: 15 of 15 correct). The rule now treats
 enums, paging names, flag booleans and unit-suffixed numbers as self-explanatory and accepts quoted
 parameter names in prose. Headline: 11% of parameters flagged, an estimated 6 to 7% genuinely missing.
-`benchmarks/compare_cisco.py`: Sentinel against Cisco's `mcp-scanner` on a labelled poisoning corpus.
+`benchmarks/compare_cisco.py`: toolgauge against Cisco's `mcp-scanner` on a labelled poisoning corpus.
 
 ### Fixed: the parameter-documentation rule over-reported
 
-`SENTINEL/STYLE/parameters-are-described` counted only schema descriptions. It now also accepts a
+`TOOLGAUGE/STYLE/parameters-are-described` counted only schema descriptions. It now also accepts a
 parameter documented in the tool description (`name: ...`, as SDK docstring `Args:` blocks produce), an
 object whose fields are all described, and self-evident plain-string names (`path`, `url`, ...).
 The benchmark moved from 28% to 17% of parameters as a result.
@@ -32,11 +44,11 @@ The benchmark moved from 28% to 17% of parameters as a result.
 The lint rules only read, so they share one child process instead of starting the server per rule:
 4.4x to 6.7x faster with identical results. `benchmarks/chart.py` draws the 27-server result as an SVG.
 
-### Added: `sentinel lint` and stdio servers
+### Added: `toolgauge lint` and stdio servers
 
-- `sentinel lint --command "npx -y <server>"` runs the manifest rules against a server started as a
+- `toolgauge lint --command "npx -y <server>"` runs the manifest rules against a server started as a
   child process (newline-delimited JSON over pipes), or `--endpoint` for HTTP. Most published MCP
-  servers are stdio programs, which Sentinel could not reach before.
+  servers are stdio programs, which toolgauge could not reach before.
 - `benchmarks/popular_servers.py`: lints 29 published servers; 28 started, 362 tools, 1,349
   parameters, 17% documented nowhere.
 
@@ -49,17 +61,17 @@ session and run, instead of every rule reporting the server unreachable.
 ### Added: benchmarks against servers this project did not write
 
 - `benchmarks/official_negatives.py`: runs the broken fixtures from the official conformance repo and
-  reports which labelled violations Sentinel catches (7 of 7 gradable; one fixture is unusable and one violation has no rule, both listed).
+  reports which labelled violations toolgauge catches (7 of 7 gradable; one fixture is unusable and one violation has no rule, both listed).
 - `benchmarks/sdk_readiness.py`: scans servers built with the official SDKs.
 - Reports now say which protocol generation answered, and a scan of a legacy-only server says so at
   the top instead of presenting a wall of failures as independent bugs.
 - The stub-description threshold is 12 characters, not 20 ("Add two numbers." is a description).
 
-### Added: `sentinel demo` and tool-description lint
+### Added: `toolgauge demo` and tool-description lint
 
-- `sentinel demo`: starts two bundled practice servers (one broken, one correct), scans both and
+- `toolgauge demo`: starts two bundled practice servers (one broken, one correct), scans both and
   prints the verdicts. Works from a plain `uvx` install because the practice servers ship in the wheel.
-- `SENTINEL/STYLE/description-states-purpose` and `SENTINEL/STYLE/parameters-are-described`,
+- `TOOLGAUGE/STYLE/description-states-purpose` and `TOOLGAUGE/STYLE/parameters-are-described`,
   grounded in two 2026 studies of MCP tool descriptions (arXiv 2602.14878, 2602.18914).
 
 ### Fixed: rules that were stricter than the specification
@@ -72,7 +84,7 @@ deprecated and points at its successor.
   `server/discover` from version negotiation, so it failed conformant servers.
 - Demoted to SHOULD (new ids): `discover-reports-server-info`, `list-changed-advertised-truthfully`
   (now INDETERMINATE when the probe cannot tell), `malformed-json-is-parse-error`.
-- Moved out of the spec namespace to `SENTINEL/STYLE/`: `tools-have-descriptions`, and the
+- Moved out of the spec namespace to `TOOLGAUGE/STYLE/`: `tools-have-descriptions`, and the
   `initialize` / `ping` / `logging/setLevel` / `resources/(un)subscribe` removed-method checks. A
   dual-era server MAY keep serving them. The `sampling` and `roots` checks were dropped: they remain
   functional during their deprecation window.
@@ -89,8 +101,8 @@ deprecated and points at its successor.
 
 ### Added: operating it
 
-- **`sentinel.toml`**: endpoint, gate, timeouts, rule toggles, severity overrides, waivers.
-  Precedence is CLI > `SENTINEL_*` > file > `extends` > default, tested cell by cell, and every
+- **`toolgauge.toml`**: endpoint, gate, timeouts, rule toggles, severity overrides, waivers.
+  Precedence is CLI > `TOOLGAUGE_*` > file > `extends` > default, tested cell by cell, and every
   resolved value records its source in the report.
 - **Waivers** with a required `expires_at`. A malformed date is an error; an expired waiver stops
   applying and the finding fails again, labelled.
@@ -115,9 +127,9 @@ deprecated and points at its successor.
 
 `x-mcp-header` constraints (5), `tools-declare-valid-input-schema`, tool-name hygiene (2), tools
 capability agreement, `tampered-requeststate-rejected`, `input-required-carries-requests-or-state`,
-`input-requests-respect-declared-capabilities`, and the beyond-spec `SENTINEL/SECURITY/`
+`input-requests-respect-declared-capabilities`, and the beyond-spec `TOOLGAUGE/SECURITY/`
 `tool-description-injection`, `wildcard-scopes`, `manifest-drift`, `tls-posture` and
-`SENTINEL/OPS/latency-budget`. A third fixture, `partial`, exercises every `NOT_APPLICABLE` branch.
+`TOOLGAUGE/OPS/latency-budget`. A third fixture, `partial`, exercises every `NOT_APPLICABLE` branch.
 
 Every `MCP/` rule cites a heading anchor, checked against the specification source. Three security rules had cited a page that does not exist (`basic/security_best_practices`; the page lives at `docs/2026-07-28/tutorials/security/`) and now point at the real one.
 
@@ -133,7 +145,7 @@ SBOM on each release.
 - **`mcp/`**: the reusable packages extracted from `broker/internal/` into a module of their own,
   with `server.New`, `server.InsecureAnonymous`, an in-memory audit sink, and `mcp/example`, a
   one-tool server the build scans with `--gate must` and `--gate security`.
-- **`fleet/` (`sentinel-server`)**: ingest, history, posture, drift, expiring waivers, an API and a
+- **`fleet/` (`toolgauge-server`)**: ingest, history, posture, drift, expiring waivers, an API and a
   server-rendered dashboard with a strict CSP. Memory or Postgres behind one contract suite.
   `scan --publish` sends a report and never changes the exit code.
 

CLAUDE.md

@@ -1,9 +1,9 @@
-# Sentinel — working instructions
+# toolgauge — working instructions
 
 Two products in one repository:
 
 - **`broker/`** — a Go MCP server built natively on the **stateless** MCP `2026-07-28` specification.
-- **`harness/`** — `sentinel`, a Python conformance harness that grades *any* MCP server.
+- **`harness/`** — `toolgauge`, a Python conformance harness that grades *any* MCP server.
 
 The spec is `docs/HANDOFF.md` (SN-HND-001). **Where this repository and the
 [MCP specification](https://modelcontextprotocol.io/specification/2026-07-28/) disagree, the spec wins.**

Dockerfile

@@ -1,4 +1,4 @@
-# The scanner as a container: `docker run ghcr.io/patsypppe/sentinel scan --endpoint ...`.
+# The scanner as a container: `docker run ghcr.io/patsypppe/toolgauge scan --endpoint ...`.
 #
 # Two stages, so the image carries the installed wheel and nothing that built it:
 # no compiler, no pip cache, no source tree. It runs as an unprivileged user, and
@@ -18,11 +18,11 @@ FROM python:3.12-slim
 LABEL org.opencontainers.image.source="https://github.com/patsypppe/sentinel" \
       org.opencontainers.image.description="Conformance harness for MCP 2026-07-28" \
       org.opencontainers.image.licenses="MIT"
-RUN useradd --uid 10001 --no-create-home --shell /usr/sbin/nologin sentinel
+RUN useradd --uid 10001 --no-create-home --shell /usr/sbin/nologin toolgauge
 COPY --from=build /dist/*.whl /tmp/
 RUN pip install --no-cache-dir /tmp/*.whl && rm -f /tmp/*.whl
 USER 10001
 WORKDIR /work
 ENV PYTHONUNBUFFERED=1 PYTHONDONTWRITEBYTECODE=1
-ENTRYPOINT ["sentinel"]
+ENTRYPOINT ["toolgauge"]
 CMD ["--help"]

MEASUREMENTS.md

@@ -61,7 +61,7 @@ Generated in 7.7s.
 
 **Result:** **100%** (52 of 52 seeded violations detected)
 
-**Value:** MCP/MUST 43/43 (100%); MCP/SHOULD 6/6 (100%); SENTINEL/MUST 2/2 (100%); SENTINEL/SHOULD 1/1 (100%)
+**Value:** MCP/MUST 43/43 (100%); MCP/SHOULD 6/6 (100%); TOOLGAUGE/MUST 2/2 (100%); TOOLGAUGE/SHOULD 1/1 (100%)
 
 **Method:** Seeded violations detected ÷ seeded violations. The fixture TAGS each violation in its own source and lists it in `SEEDED_VIOLATIONS`, and a test asserts the two agree — so the denominator describes what the fixture actually does rather than what the scanner happened to find. A scanner supplying its own denominator would be grading its own homework. Reported per severity because a MUST failure and a beyond-spec opinion are not the same claim, and one number over both would let a strong figure hide a weak one.
 

Makefile

@@ -1,6 +1,6 @@
-# Sentinel — two products, one repository.
+# toolgauge — two products, one repository.
 #   broker/   a Go MCP server built natively on MCP 2026-07-28
-#   harness/  `sentinel`, a conformance harness that grades any MCP server
+#   harness/  `toolgauge`, a conformance harness that grades any MCP server
 #
 # Go 1.23+ is required. Homebrew's Go is preferred when present because the
 # system Go on some machines is older than the module's floor.
@@ -84,8 +84,8 @@ demo: ## Run the nine-step demo from docs/HANDOFF.md §13 (needs `make up`)
 
 .PHONY: scan-broker
 scan-broker: ## Scan the running broker; exits 0 when it passes the MUST gate
-	$(UV) run sentinel scan --endpoint http://localhost:8080/mcp --gate must --format text
+	$(UV) run toolgauge scan --endpoint http://localhost:8080/mcp --gate must --format text
 
 .PHONY: scan-fixture
 scan-fixture: ## Scan the non-conformant fixture; expected to exit 1
-	$(UV) run sentinel scan --endpoint http://localhost:9000/mcp --gate must --format text
+	$(UV) run toolgauge scan --endpoint http://localhost:9000/mcp --gate must --format text