One GitHub deletion request, multiple interfaces, one safety question¶
November 7, 2026 · Kishore Babu Gutta
An AI execution gateway is useful only if its safeguards apply to the operation, not merely one convenient representation of that operation.
Consider a deliberately prohibited action: delete a GitHub repository.
An agent could express repository deletion through GitHub REST, a provider-native CLI or a qualified agent-facing MCP entrypoint. GitHub's GraphQL API does not expose a deleteRepository mutation (confirmed by read-only introspection during the v0.13 qualification). GraphQL does expose other destructive operations such as ruleset deletion, protected-rule deletion and forced reference updates; these need equivalent safety treatment. The engineering safety question is whether the Runtime recognizes the request before a provider dispatch, regardless of the admitted route.
What an execution boundary must decide¶
Human / CI / AI agent
|
v
ENGRT entrypoint
|
v
Recognize provider + operation + target
|
v
Compiled safety floor -> configurable policy
|
DENY?
/ \
yes no
| |
no provider auth/context and supported execution
dispatch |
| v
refusal GitHub
evidence
This is execution admission, not simply filtering a chatbot's answer. A model can propose a request; the Runtime must independently decide whether it may act.
ENGRT's published v0.13.0 scope documents a finite catalog of recognized destructive operations—including recognized GitHub repository deletion—refused on covered Runtime paths. The configured policy can deny more, but should not override the compiled safety floor. That statement does not prove every route variant, alias or third-party integration is covered.
Expressing the same intent through different interfaces¶
This table separates the actual provider surface from the behavior evidenced or still awaiting qualification:
| Route | GitHub operation | Evidence / status |
|---|---|---|
| REST engine | Repository DELETE against a bound owner/repo target | v0.13 compiled safety catalog and local no-dispatch probes cover recognized repository deletion |
| Command engine | Native GitHub CLI repository deletion | v0.13 compiled command safety and local no-dispatch probes cover recognized forms; aliases and configuration variants remain bounded by tested coverage |
| GraphQL engine | No deleteRepository mutation exists; test deleteRepositoryRuleset, deleteBranchProtectionRule, deleteRef and forced updateRef/updateRefs instead |
v0.12 baseline exposed gaps; v0.13 safety classification and tests address these recognized operations |
| ENGRT inbound MCP | Agent invokes an ENGRT-exposed operation | v0.14 four-interface qualification must establish handler-to-admission and zero dispatch for the specific effect |
| Outbound/hosted provider MCP | Runtime delegates to an external provider tool | v0.14 qualification target; not a blanket passed or shipped parity claim |
The precise syntax and behavior differ by transport. GitHub repository deletion is therefore not a valid identical-operation test across all four interfaces; use the same destructive safety class with actual supported provider operations, and report non-applicable cases honestly. In particular, a hosted provider MCP path and local inbound MCP path are different architectures and need separate qualification.
What the earlier test campaign actually found¶
The October 17 v0.12 baseline validation recorded:
- REST repository DELETE was denied under then-editable policy, not yet under the non-overridable compiled safety floor.
- Raw
gh repo deletewas denied incidentally because that CLI mode was unsupported in the baseline; that was not sufficient evidence of a compiled prohibition. - GraphQL
deleteRepositoryRuleset,deleteBranchProtectionRuleand forcedupdateRef, along with certain REST forced ref updates, were admitted in plan-only probes. These were not dispatched to GitHub in that baseline campaign. - GitHub GraphQL had no
deleteRepositorymutation.
This was a real cross-interface safety gap: different paths had different admission behavior for destructive effects.
The subsequent v0.13 compiled-safety implementation record reports passing Go tests, 58 GitHub provider-parity subtests, more than 100 command-safety subtests and 13 local destructive probe forms refused under permissive policy with zero identity probes and zero requests. It explicitly notes that no live destructive GitHub attempts were made. That is evidence of local classification and refusal, not proof that every possible alternate route is covered.
The v0.14 release coordinator plan calls for per-route safety parity across REST, GraphQL, native Command and MCP, with dispatch counters, provenance and PASS/FAIL/BLOCKED/NOT_SUPPORTED evidence. MCP parity remains a qualification objective until its specific rows have passed.
What a convincing denial test looks like¶
Use a non-production fixture and a dispatch recorder or mock provider, not a real repository deletion.
- Submit an operation representing the prohibited intent through each supported route, with representative variations in target and argument shape.
- Assert the result is a safety refusal, not an authorization error after contacting GitHub.
- Assert zero outbound dispatch to GitHub or to a provider CLI/MCP server.
- Check the recorded refusal contains the intended operation and target, appropriate sanitized evidence and no leaked credentials.
- Repeat with permissive configurable policy to show that policy cannot override compiled safety.
- Record unsupported variants or bypasses as explicit coverage gaps; never convert missing evidence into a blanket claim.
| Route | Publication-safe conclusion |
|---|---|
| REST repository deletion | v0.13 local compiled-safety and zero-dispatch evidence recorded; no live destructive test |
| Native command repository deletion | v0.13 local compiled-safety probes recorded; no universal alias/configuration guarantee |
| GraphQL repository deletion | Not applicable: GitHub has no deleteRepository mutation |
| Other GraphQL destructive effects | Baseline gaps found; v0.13 compiled-floor implementation/test evidence recorded for recognized mutations and forced ref updates |
| Local inbound MCP | v0.14 route-specific qualification required |
| Hosted/outbound MCP | v0.14 route-specific qualification required; no blanket parity claim |
| Direct provider tools outside ENGRT | Outside the Runtime boundary; must be controlled by credentials, egress or other infrastructure |
This is a recorded baseline → correction → continuing qualification story, not a claim that the entire matrix has passed in one end-to-end campaign.
Is ENGRT an AI gateway?¶
In one sense it creates an admission boundary for engineering operations initiated by AI. But the durable category is broader: Engineering Execution Layer. Humans, CI and ordinary automation should receive the same protection where they use the same supported Runtime operations.
ENGRT is not an LLM request gateway or a universal interception proxy. It cannot stop a caller using an independent token and direct GitHub access outside ENGRT. A system that needs ENGRT as the mandatory route must also constrain alternative credentials and network paths.
The goal is not to prove that GitHub deletion is impossible everywhere. The measurable claim is more precise:
For a recognized prohibited operation entering a qualified ENGRT route, refusal happens before that Runtime dispatches the operation.
That is the behavior worth showing in a reproducible example, with per-route test identifiers and outputs rather than an unqualified promise of interface parity.
Further reading: Why compiled safety must sit below policy, One provider can have many execution transports, MCP works in two directions, and Engineering Runtime release scope.