Release Policy¶
Sentinel follows semantic versioning for repository-owned public behavior:
- Patch: compatible fixes, documentation corrections, and evidence clarifications.
- Minor: backward-compatible skills, routes, fields, or integrations.
- Major: incompatible plugin interfaces, artifact schemas, commands, or behavior contracts after
1.0.0.
Before 1.0.0, a minor release may make an incompatible security or evidence-integrity correction only when the changelog names the exact affected interface, provides a migration, and the compatibility page marks the break. Silent schema reinterpretation is never allowed.
Upstream Roboflow behavior is not versioned by this project. A release must name any tested upstream assumptions rather than imply they are frozen.
Release gate¶
Before a human publishes a release:
- reconcile versions in all plugin manifests and generated records,
- run strict documentation, link, eval, hook, and pre-commit checks,
- validate supported host manifests and exact install commands,
- review permissions, secrets, paid actions, and data-flow disclosures,
- verify every changed headline claim against committed evidence,
- list mocked, private, single-run, and pending evidence,
- update the changelog and compatibility notes,
- inspect the built package or plugin from a clean environment,
- re-run live routing (
make eval-trigger-live) on the current route set, at least two repeats and on both hosts where available, and record the summary; a release that adds, removes, or rewrites a skill trigger cannot reuse an older run, - triage open outcome reports (
outcome-reportlabel) and note any that change a support tier, - refresh the upstream Roboflow skill roster (
evals/platform-routing/upstream-skills.json: names, source commit, date) from a currentroboflow/computer-vision-skillsclone, then runSENTINEL_UPSTREAM_SKILLS_DIR=<clone>/skills make eval-platform-routingand update the delegation map for any drift, - download Roboflow's current published tool catalog, regenerate the gate policy with
python scripts/build_gate_policy.py --catalog <file> --release <release tag> --as-of <date> --expect-sha256 <digest>, review thehooks/gate-policy.jsondiff (new tools, changed read-only flags, every override reason), then runSENTINEL_TOOL_CATALOG=<file> SENTINEL_UPSTREAM_SKILLS_DIR=<clone>/skills make eval-upstream, which fails when either source is missing or has drifted, - have the maintainer run paid live dual-plugin routing once per release on a clean Claude Code host profile with both plugins installed:
python3 evals/trigger-live/run_live.py --host claude --dual-cases --with-plugin-dir <path to a clone of the official Roboflow plugin> [--model sonnet|opus], at least two repeats of the six prompts; record per-group counts, thefront_door_collisionslist and collision count, and pack results fromevals/trigger-live/runs/withpython3 scripts/publish_run.py pack --suite trigger-live --run <results file> --set <set-id>; a release changing a skill trigger or the upstream roster cannot reuse an older run; list a missing run as pending evidence in step 6, never claim it, - have the maintainer record one paid live session per release climbing the baseline ladder from rung 0 to rung 3 on one public licensed dataset; pack rung 0 with
python3 scripts/package_golden_run.py pack ...followingexamples/golden-run/README.md, and commit the packed golden run plus a folder containing the redacted project ledger (.vision-delivery/ledger.jsonl), per-rung scorer output, and session transcript; rungs 1 to 3 have no packer yet, so assemble that folder by hand and review it; list a missing run as pending evidence in step 6, never claim it.
Remote publication, tags, releases, marketplace changes, and branch protection are human-owned operations. Their existence must not be claimed until verified on the remote service.
GitHub Pages deployment additionally requires a maintainer to select GitHub Actions as the Pages source in the repository settings. Required checks and branch protection must be configured on the remote repository before they can be described as enforced.
Evidence policy¶
A release note may call a route "proven" only when the exact claim has a versioned fixture or study, reproducible command, result, environment, sample size, and material limitations. A workflow with instructions but without comparable live evidence is "guided." Platform-specific claims remain delegated to current official Roboflow sources.
Security fixes may be prepared privately and disclosed after users have a practical remediation path. See the repository security policy.