code: adopt OpenCode and core-backed cooperative task tools

[erikleblansch]

Owner-selected verification — 40d89016

Adds a real owner CLI and isolated check executor, with failed check output continuing the same native session within the original deadline and permissions. Fixed checks are selected before model execution; native summaries must match actual owner receipts. Cancellation and cleanup are joined. Selected-test success is not general task correctness (taskVerified remains false). 57 focused checks and an actual bwrap failure/pass/isolation/cancellation smoke pass. The editor selection UI remains pending. The real-model trial now selects the same verifier before startup, checks original test bytes before every execution, and retains only closed check status/counts. Original prompt, tests, model/token budget, 12 native approvals, 40-minute total deadline and final independent acceptance criteria are unchanged. The generic verifier now admits ordinary multi-test failures within 4096 raw/8192 encoded bytes, with complete overflow refusal. Ten verifier checks, ten trial checks and seven hosted-contract checks pass. Exact-source run 37076283235 is running; no model-guided repair is yet proven. See docs/OWNER_VERIFICATION.md.

---

Previous source update — d6ec3146

Pins core 845cc84d0d0b766ab1c5227231dbf6c8eaeb8cc3 and fixes the zero-temperature contract: explicit generation-policy handshake, Qwen greedy_v1, request/result binding and refusal of legacy/unsupported cores instead of silent sampling. Ordinary conversation clients retain the legacy handshake and omitted-policy behavior. 100 focused JavaScript checks and 26 Python contracts pass; exact-source run 37073231635 is terminal FAILED. One actual native read completed; the second core response was assistant/stop. No edit or test command ran, the fixture remained unchanged and the independent check failed. Both real core responses report confirmed cleanup. Guest private/runtime cleanup and QEMU/scratch removal passed; the outer-host IPv6-route hash changed for an undetermined reason, so full unchanged outer-host state is not claimed. ZIP SHA-256 64dd8f80faf89d31f2974bb7f1c9cb52a82d14dde8f2d3e24c5c30804b3f32f5; original job log SHA-256 cf13b663b718c00c850a30747150d7131d9da215c9dd5b481e9d4b77e375fb99. The generation-policy contract is fixed; useful model-driven coding remains unproved.

Original run 37066003771 remains failed: one actual completed native read, then assistant/stop, no edit, no passing independent check. The proven sampling mismatch is not a proven explanation for that stop. ZIP SHA-256 2f7d35d5e88dbb1e8110f8cdab6a635fcddd668fea13e6afb4fb67d3b35f55ec; original job log SHA-256 f5dadb5d41e9bf1b5c437a97c3bcd31e04aad2ee1d990bde4657c697158f3e29.

---

OpenCode migration candidate

Replace Codex with pinned OpenCode v1.18.34 (aec0b9a6d8898f68f923aaf08b7306d931fd9d76), connected through actual HTTP/SSE sessions to VOLPAROSSA's typed conversation interface. Core owns model selection, peer scheduling and cancellation. Workspace trust, explicit one-shot edit/command permissions, child-session ownership and cleanup remain intact; no automatic model/runtime downloads or external-provider fallback.

The public delegation tool exposes only an owner-approved exact public snapshot through a single-use proxy. Private workspace history never becomes public task context. The current private executor is local; protected private remote collaboration remains required and incomplete.

Previous diagnostic candidate: 44479640f01de584370b81a18a4e84f7ebe20b0a

Adds closed native tool lifecycle and permission counters, passed through the owner protocol to original trial evidence. Counts distinguish ordinary read/search calls from approved commands and failed/invalid tools, without exporting text, arguments, paths or IDs. Missing historical diagnostics remain null, not zero. Model instructions, budgets, guest task, tool authority and independent correctness checks are unchanged.

Forty focused JavaScript checks and twenty-six Python contract checks pass. An exact-source original-model trial has been requested for this commit/tree 3d548aab229d02c54ef21575dc84de72a4c9a5c0; there is no new live success claim.

Earlier completed original-model trial

Run 37059547905 at e021e5ca52bf8bd23c857ba2917a2ddfe567eb15 is failed. Source-built OpenCode, hosted KVM and pinned Qwen provisioning succeeded. Two actual core responses (one function call and one assistant response) completed and cleaned up; the native turn ended, but the fixture was unchanged and the independent coding test failed. Zero command/edit approvals were recorded. Native read/search tools do not require those approvals, so the old approved_read: 0 does not establish that no read occurred. Original tool details were not retained; the cause remains unknown.

Guest network and private-state cleanup, runtime units, QEMU and scratch cleanup passed. An outer-host IPv6-route hash changed for an undetermined reason: full unchanged outer-host state is not claimed. Original artifact ZIP SHA-256: 20e768b6ee3535564b3d967a1e192402258087fb6b21cae315874f5d68c14314; original job log SHA-256: ebf52ea810aa2bae88cb5a719d05fef44648a03e5f496d26c87aa8f65e9a976f.

Earlier evidence and scope

Actual pinned OpenCode can execute approved disposable-file edits and the public delegation tool with synthetic core replies. These establish native adapter/permissions/cleanup behavior, not model-driven coding quality. The real public-peer trial received eleven terminal worker receipts from two providers but incomplete leaf answers, not successful synthesis. Earlier real-model failures remain failures; no diagnostic or prompt change relabels them.

See docs/OPENCODE.md for original evidence and source provenance. Core public execution remains coordinated with core PR #178. Hosted trial infrastructure is enabled via merged PR #7. This migration supersedes PR #4 without deleting its history.

Draft: actual model-driven edit/test success, complete joined peer answers, native editor completion, protected private network execution, full immune supervision and additional platforms remain unfinished.