[drewstone]

The opt-in three-turn evidence proof now uses the maintained authenticated loopback Runtime → CLI Bridge → native OpenCode path. It reads synthetic evidence, restarts the actual bridge, resumes the same native session without tools or prompt replay, then reads changed evidence. All three turns require measured input/output usage, matching reported model identity, exact answers and confirmed per-turn cleanup.

The existing default two-turn filesystem scenario is preserved. The maintained proof and its evidence guide change against current main; Vitest config also consumes the existing admitted VITEST_MAX_WORKERS budget instead of deriving concurrency from the host. the obsolete separate executable is removed. The anonymous evidence scenario pins main, small, title, summary and compaction models to the explicitly authorized exact free route and one-model whitelist, then checks effective config before inference. No host auth or paid fallback is permitted.

Existing source preparation passed syntax, deterministic hooks and independent two-file review. Fresh main through ddc6312 merged normally; current PR head d754987. B1 filesystem namespace preflight passed. First normal gate passed frozen install/types, then fullsuite1268pass/2existing5s timeout/33skip1303 under highhostload. Focused serial process-boundary proof passed18/18 plus published Runtimeconsumer and actualjail. Exact anonymous config/model whitelist passed; authored unsupported system-prompt replacement was then refused before native inference, with confirmed teardown/scratch cleanup. The caller now uses supported additive profile instructions. Final normal B1-6e7 atd754987 passed typecheck but failed14/1303 cases (1256pass33skip) across8files, plus one unhandled expired replay cursor after a timeout. Worker-budget binding does not resolve or establish the cause of these failures. The native stage was not reached. One isolated corrected three-turn first execution remains under the existing deadline; it cannot substitute for the failed normal suite. No successful live inference is claimed yet.

Merge precondition: the exact resulting head must pass the normal gate and all three real turns with usage, native-session continuity and per-turn cleanup. Keep draft until that evidence exists. After merge, prove the actual installed/served revision and live result on the owner host; a premerge test alone is insufficient.

Scope: synthetic local Runtime/CLI Bridge/OpenCode execution only. Agent App browser flows, Casework, managed sandboxes and cloud trace custody require their own consumer proof. No new CI gate, model spend beyond the currently free anonymous route, or security-policy changes.

[drewstone]

Owner: op-company:agent-runtime (live Fleet lane; stale-PR handoff, 2026-10-22).

Blocker: The opt-in local Runtime-to-OpenCode inference path has never completed; recorded preflight exited blocked before inference.

Next action: Reconcile with merged cli-bridge #263 qualification first. Close this PR with exact replacement proof if already covered; otherwise provision the supported network-enabled host, use an explicitly permitted existing model/account, run all three real turns and teardown checks, and retain the exact-head gate and live receipt. Keep draft pending proof; no model spend authorized by this handoff.

The complete PR body, available summary/inline reviews and comments were inspected for this disposition. Existing draft/hold and body preconditions remain binding; no gate, merge, deployment, or production mutation was run by the sweep. This handoff is being delivered to the named lane's durable Fleet inbox.

[drewstone]

Deferred under the campaign’s 30-minute blocker rule; this is not a completed or obsolete proof.

Retained head: d754987 on test/local-opencode-runtime-proof-20261017. Source, commits and private receipts are preserved; the branch is not deleted.

The final normal Beelink gate passed frozen install/typecheck but failed 14 of 1,303 cases (1,256 passed, 33 skipped), across eight files, with one unhandled expired replay cursor. The worker-budget binding fixes a confirmed configuration mismatch; it did not resolve the suite or establish the failures’ cause. No timeout widening or mock/unit-test changes were made.

The earlier focused check passed 18 process-boundary cases, the published Runtime consumer, the real jail check, and anonymous exact-model configuration. The authored system-prompt replacement was refused before native inference and changed to the supported additive instructions field. The corrected three-turn scenario has not executed: its isolated gate remained queued and was cancelled at the cutoff through the maintained gate path (cancellation exit 0; gate terminal 75). No successful native three-turn result, merge, owner-host installation, served identity or live usage result is claimed.

Reopen this same retained PR only after a new allocation and resolving the normal-suite failures, then complete all three real turns and the post-merge owner-host installed/process revision and cleanup proof. No paid fallback or new draft is authorized by this deferral.