Repository navigation
Add onprem plugin: on-prem model governance hard lock and operating doctrine - #608
TeslaJarTran wants to merge 3 commits into
Conversation
| "name": "onprem", | ||
| "displayName": "onprem stack", | ||
| "version": "0.1.0", | ||
| "description": "on-prem model governance and operating doctrine for agent work: a closed allowlist hard lock, a stack of named principles, and an entry mode that applies them.", |
There was a problem hiding this comment.
Listing descriptions disagree across surfaces
Medium Severity · Bugbot Rules
The plugin description differs across plugin.json, marketplace.json, and the README table. Marketplace inserts Cursor and a different ending; plugin.json changes capitalization, hyphenation, and the closing clause. Installed and marketplace listings will show different copy.
Additional Locations (2)
Triggered by learned rule: Plugin listing copy must match plugin.json and marketplace.json
Reviewed by Cursor Bugbot for commit 123d853. Configure here.
|
|
||
| # Onprem subagent | ||
|
|
||
| You are operating as onprem-mode's full approach. Read the `onprem-mode` skill's `SKILL.md` in full before doing any work, including its inline Principles index. Navigate to a leaf `principle-*` skill whenever you apply that principle. The on-prem model lock applies to every spawn you make. |
There was a problem hiding this comment.
Agent cannot locate hidden mode skill
High Severity · Logic Bug
onprem-agent is told to read onprem-mode before any work, but that skill is disable-model-invocation and neither the agent nor the mode pass or search a filesystem path. Spawned agents can miss the lock and principles entirely.
Additional Locations (1)
Triggered by learned rule: Skill detection must cover all installation paths
Reviewed by Cursor Bugbot for commit 123d853. Configure here.
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using high effort and found 3 potential issues.
There are 5 total unresolved issues (including 2 from previous reviews).
Bugbot Autofix is ON, but it could not run because the branch was deleted or merged before autofix could start.
Reviewed by Cursor Bugbot for commit 816cdaa. Configure here.
| External multi-model deep review of a diff or a full tree. Use when the task is "robodev this", "deep external review", or a pre-ship cross-check of a large change. The review fleet runs outside the on-prem boundary, so this is the evidence path, not the judgment path. | ||
|
|
||
| 1. Boundary gate. Confirm the content under review may leave the boundary. Internal content reviewed on the internal fleet is fine. Content that a recipient's on-prem policy keeps internal is not: stop and use the **panel** playbook instead. | ||
| 2. Host gate. The CLI only knows the internal GitHub hosts (github.tesla.com, github-it.tesla.com, github-ap.tesla.com, github-fw.tesla.com). If the repo's origin is on a known host, continue there. If not, build a mirror: a private mirror repo on a known host, a fresh clone for the review, and a deleted mirror when the review returns. For a full-tree review of a young repo, give the mirror history an empty root commit (commit-tree the existing trees onto an empty-tree root) so the whole tree is the diff. Never rewrite the real repo's history for this. |
There was a problem hiding this comment.
Public playbook mirrors onto Tesla GitHub
High Severity · Security Issue
The new robodev playbook names four Tesla-internal GitHub hosts and tells the agent to create a private mirror there when the repo origin is unknown. Generic triggers like pre-ship cross-check route here, and onprem-mode autonomy proceeds on reversible external actions without asking, so a marketplace session can copy a user's repo onto internal infra.
Additional Locations (2)
Reviewed by Cursor Bugbot for commit 816cdaa. Configure here.
|
|
||
|
|
||
| def check_frontmatter(root, checks): | ||
| import yaml |
There was a problem hiding this comment.
Verifier crashes without PyYAML
Medium Severity · Logic Bug
check_frontmatter does a bare import yaml. PyYAML is not in the standard library and nothing in this repo declares it, so a missing package aborts the suite with a traceback instead of a JSON frontmatter failure. The skill's proof contract is a JSON summary and a meaningful exit code.
Reviewed by Cursor Bugbot for commit 816cdaa. Configure here.
|
|
||
| 1. Record what failed: the role, the model attempted, and the error. | ||
| 2. Is the attempted model allowlisted? | ||
| - No → the rule or the spawn violated the lock. Apply **no-fallback-to-banned**: re-spawn on a real allowlisted slug for the role, never an alias. If the role's default is `inherit-parent` or `auto`, the retry takes the role's real-slug default (code roles `Qwen/Qwen3.8-Flash-Next-FP8`, judgment and prose roles the rule file's judgment slug, panel roles one entry of their panel list). An alias retry inherits the parent and can land back on the banned model. Flag the source for an **audit** run. |
There was a problem hiding this comment.
Spawn retry misses alias role fallbacks
Medium Severity · Logic Bug
The rewritten retry step forbids inherit-parent/auto but only names a real-slug default for code, judgment, and panel roles. how explorer, why investigators, and reflect tooling default to the alias and have no fallback. Judgment retries also use the rule file's judgment value, which can itself be an alias.
Reviewed by Cursor Bugbot for commit 816cdaa. Configure here.
|
Note for maintainers: the "Validate plugins" run is in |
Summary
onprem/, a plugin that encodes on-prem model governance as a closed-allowlist hard lock plus an operating doctrine for 27B-class open-weight models: entry mode (/onprem-mode), 12 named principles, 3 playbooks (audit, spawn retry, panel),/setup-onpremrule writer/validator, help router, verification-skill generator, a guide, and anonprem-agentsubagent..cursor-plugin/marketplace.jsonand the README table (after dyl-stack, the other pstack-derived stack).Validation
node scripts/validate-plugins.mjsrun locally against this branch: "All plugins validated successfully."namematches the marketplace entryname.Test plan
/add-plugin onpremfrom the marketplace installs the pluginonprem/docs/guide/05-sharing.md: install, run/setup-onprem, start a task with/onprem-modeNote
Low Risk
Adds docs, skills, and a local Python verifier; the only user-writable surface is optional
~/.cursor/rules/onprem-models.mdcfrom/setup-onprem, not repo runtime code.Overview
Introduces the
onpremCursor plugin and lists it in.cursor-plugin/marketplace.jsonand the root README (alongside other pstack-style stacks).The new
onprem/package is documentation- and skill-driven governance for agent work on self-hosted models: a closed allowlist plus an absolute frontier-model ban,/setup-onpremto detect entitlements and write~/.cursor/rules/onprem-models.mdc, entry/onprem-modewith four playbooks (audit, spawn-retry, panel, robodev), twelveprinciple-*skills,onprem-agent, a human guide, and helpers/onprem-helpand/create-verify-skill. It is modeled on pstack but targets model policy and on-prem operating doctrine, with explicit pstack coexistence (separate rule files and agents).Also adds
verify-onprem: a read-onlyverify.pysuite (doctor + nine JSON checks for manifest, lock consistency, templates, links, optional rule file, and registration) with a small feature map, aligned with deterministic-proof maintenance.Reviewed by Cursor Bugbot for commit 816cdaa. Bugbot is set up for automated code reviews on this repo. Configure here.