Real Prometheus validation — 2026-09-28

Nine SDK runs completed against real native subscription accounts on the user-authorized Prometheus host. Every request travelled from a separately installed client on the application computer, through verified OpenSSH, to a loopback AgenticDriver host on Prometheus. No mock provider, replacement model server or canned response was used.

The host and JavaScript client used published @agenticdriver/sdk@0.2.0-alpha.5. Python used the checksum-verified 0.2.0a5 GitHub wheel in a fresh environment; PyPI organization approval remains pending. Go installed v0.2.0-alpha.5 through the ordinary module proxy without a replace directive. Rust installed the exact 0.2.0-alpha.5 crate from crates.io in a fresh application.

Host and explicit selections#

• Prometheus: Fedora Linux x64, kernel 7.0.12-201.fc44.x86_64, Node 24.16.0.
• Codex: qualified CLI 0.157.0, gpt-6-luna, medium reasoning, existing ChatGPT Pro native session.
• Claude: qualified CLI 2.1.282, claude-haiku-4-5-20251001, existing Claude Pro native session.
• The qualified runtimes were installed in a separate user-owned directory and checked against their previously recorded SHA-256 digests. The normal Codex 0.154.0 and Claude 2.1.268 installations were not replaced.
• The separate host bound only 127.0.0.1:17435. The application-side SSH tunnel also bound only loopback. Existing native sign-in stores stayed on Prometheus; CI containers received no account credentials.
• The configuration imposed no model allowlist override. Only the two exact models above were run. Latest refresh reported nine Codex entries and fourteen Claude entries. The first Claude inspection reported eight entries; later inspections returned fourteen. Those catalogs remain reported metadata, not live qualification of every entry.

Actual application prompts and results#

The prompt source contains the complete inputs. Brandstorm prompts use AgenticDriver's product brief. Literature prompts use short, attributed notes from Lewis et al. and Liu et al.. Workspace prompts ask for an email draft from the actual alpha.5 release handoff. No email was sent and no private mailbox, manuscript or library was accessed.

The actual answers were reviewed. The brand results gave distinct directions, taglines, colour pairs and audience rationales. Literature answers cited both supplied papers and distinguished reported findings from proposed experiments. Email drafts retained pending PyPI approval and did not claim confirmed app adoption. Claude's JavaScript email draft blurred discovery and execution in one action item, despite distinguishing them in its closing note; it needs editorial review. Model completion is not a guarantee of factual or editorial quality.

Codex advertised textStreaming: false and delivered final text in one delta per run. The three JavaScript Claude runs delivered 79, 103 and 42 text deltas. Metadata was refreshed after each of those runs started, and all three completed without cancellation. SDK retries were limited to one attempt. No generation deadline or inactivity timeout was set.

Connection and settings checks#

The real host accepted one-use invitations for separate execution and management profiles. Execution access covered only the two selected instances. Management had no execution grants. A further short-lived connection proved one-provider scoping, rejection of an already-consumed invitation (INVITATION_REJECTED) and rejection of requests after revocation (UNAUTHORIZED). That grant was revoked.

The execution profile could not call provider management (FORBIDDEN). Through the management profile, a Codex display-name edit persisted and read back, a stale revision was rejected (CONFIG_CONFLICT), and the original name was restored. No account identity, native path or model policy was changed by that edit. Both provider catalogs refreshed successfully over the SSH route.

Reported usage and evidence#

These are native CLI reports, including native runtime overhead. Claude's API-equivalent estimate is not an actual subscription charge. Neither adapter reported an actual billed cost. The host's existing JSONL usage sink recorded all nine completed runs; this separate host does not yet forward usage to the shared Usagestat ingestion service.

The machine-readable receipt records each run ID, timing, selection, usage and response checksum, plus checksums of private full receipts. Full responses, prompts and private connection profiles are retained under the local validation state with restrictive permissions. Public evidence omits email/name, token values and native account identifiers. JavaScript syntax, Go/Rust compilation and the documentation build are checked separately. Legacy mock-based suites are not claimed as validation of this work and were not rerun.

Scope and reuse#

The remote host is agenticdriver-remote-testing.service in Prometheus's user systemd. It can be started explicitly with ssh prometheus systemctl --user start agenticdriver-remote-testing.service. The application computer's current tunnel is the transient agenticdriver-prometheus-test-tunnel.service; use the SSH recipe when creating a new route. Neither unit was enabled for automatic startup.

The test execution connection has a six-hour lifetime; the management connection has one hour. Issue a new scoped invitation for later work or for another app; do not reuse a validation credential as a production app grant. The host's separate static operator credential stays private on Prometheus.

This evidence covers the selected accounts, runtimes, models and SSH route at the recorded time. It does not qualify other models, API accounts, verified HTTPS, container account isolation, PDF extraction, vector retrieval or any consumer's complete UI/workflow. The literature input is supplied context, not a vector search. App-owned end-to-end checks and Linux account-container qualification remain separate tasks.