E2E third-party: OpenCode Go muse-spark-1.2-contributor

[NOirBRight]

Decision\n\nMaintainer decision: all future real-client E2E is CLI-only. Do not run Desktop/ZCode GUI or manual-evidence cases. Ollama/Ollama Cloud must never be used for E2E again.\n\n## Problem\n\nThe current Windows runner hard-codes Ollama Cloud in four third-party cases and requires an Ollama credential. The dedicated account has exhausted quota (HTTP 402 across Codex CLI/OpenCode/Pi/OMP), and the provider is no longer authorized for E2E. The docs still describe a GUI matrix that is no longer required.\n\n## Scope\n\n- Make CLI-only the only authoritative Windows scope; remove/retire GUI/manual case paths and inputs.\n- Replace hard-coded ollama-cloud provider/model/endpoint/credential names with one maintainer-approved non-Ollama third-party provider contract.\n- Do not read, copy, probe, or invoke Ollama credentials/models in future E2E.\n- Preserve the four official Luna CLI cases and all route/stream/tool/terminal/correlation checks.\n- Update docs/agents/real-client-e2e.md and runner contract tests.\n\n## Acceptance\n\n- [ ] No Ollama/Ollama Cloud identifiers or credential requirements in the active E2E path\n- [ ] No GUI/manual evidence cases in the authoritative Windows gate\n- [ ] Four official CLI cases plus the chosen third-party CLI cases pass\n- [ ] Candidate SHA/build/materializer bindings remain fail-closed\n\n## Needs info\n\nSelect the replacement third-party provider/model and its dedicated credential input. Until selected, do not run third-party real-client E2E.

[NOirBRight]

Maintainer decisions recorded in :\n- Windows real-client E2E is permanently CLI-only. No Desktop/ZCode GUI or manual evidence.\n- Ollama/Ollama Cloud are prohibited. Do not read those credentials or invoke those models.\n\nYoga SHA already proved Official Luna for Codex CLI/OpenCode/Pi/OMP. The four hard-coded Ollama rows failed HTTP 402 (quota exhausted) and must not be retried.\n\nBlocked on maintainer choosing the replacement third-party provider/model and dedicated credential input. Until then, only Official CLI cases + isolated checks.\n

[NOirBRight]

Maintainer decisions recorded in docs/agents/real-client-e2e.md:

• Windows real-client E2E is permanently CLI-only. No Desktop/ZCode GUI or manual evidence.
• Ollama/Ollama Cloud are prohibited. Do not read those credentials or invoke those models.

Yoga SHA f3c16fee7686defa6da61fe99af8b7050d0d9f6f already proved Official Luna for Codex CLI/OpenCode/Pi/OMP. The four hard-coded Ollama rows failed HTTP 402 (quota exhausted) and must not be retried.

Blocked on maintainer choosing the replacement third-party provider/model and dedicated credential input. Until then, only Official CLI cases plus isolated managed-client-config checks.

[NOirBRight]

Maintainer selected the replacement third-party: OpenCode Go Muse Spark 1.2 Contributor.

Contract now coded in the runner/docs/tests:

• provider_id: opencode-go
• diagnostic_provider_id: opencode_go
• Gateway model: opencode-go/muse-spark-1.2-contributor
• client selector: codexhub-opencode-go/muse-spark-1.2-contributor
• Codex CLI selector: opencode-go/muse-spark-1.2-contributor
• endpoint: /v1/providers/opencode-go/responses
• credential file: isolated/credentials/opencode-go.json
• schema: codexhub.real-client-opencode-go.v1
• env: OPENCODE_API_KEY
• -ThirdPartyModel: codexhub-opencode-go/muse-spark-1.2-contributor

CLI cases: codex-cli-opencode-go, opencode-opencode-go, pi-opencode-go, omp-opencode-go.
GUI remains retired. Ollama remains prohibited.

Still blocked on Yoga: no OPENCODE_API_KEY in user/machine/process env, and no dedicated opencode-go.json yet. Place a dedicated (not host-user) key into a fresh run root as isolated/credentials/opencode-go.json before the next live CLI-only run.

[NOirBRight]

Maintainer decisions recorded in docs/agents/real-client-e2e.md:

• Windows real-client E2E is permanently CLI-only. No Desktop/ZCode GUI or manual evidence.
• Ollama/Ollama Cloud are prohibited. Do not read those credentials or invoke those models.

Yoga SHA f3c16fee7686defa6da61fe99af8b7050d0d9f6f already proved Official Luna for Codex CLI/OpenCode/Pi/OMP. The four hard-coded Ollama rows failed HTTP 402 (quota exhausted) and must not be retried.

OpenCode Go credential is now in the Yoga isolated credentials slot with schema codexhub.real-client-opencode-go.v1 (non-empty api_key). The secret is not in this comment.

Next CLI-only run still needs a new output root (copy account+config from a prior isolated, never isolated/work), the updated runner that requires -ThirdPartyModel codexhub-opencode-go/muse-spark-1.2-contributor, pinned -OmpPath, and a Debug portable of the candidate SHA. Yoga checkout is currently on an older branch, so the campaign tree is not what last ran.

[NOirBRight]

Prepared a new Yoga output root cli-only-opencode-go-r1 with copied account+config from the last Official Luna isolated, OpenCode Go credential only (no Ollama file), and no isolated/work.

Still blocked on a Debug portable of the campaign candidate plus the updated CLI-only runner (-ThirdPartyModel codexhub-opencode-go/muse-spark-1.2-contributor, pinned -OmpPath). Yoga D:\\Workstation\\CodexHub is still on an older branch.

[NOirBRight]

Campaign tree is on Yoga worktree D:\\Workstation\\CodexHub-deepmodules at local-only SHA 4994055896e4341ce82c4699c58beb06f4f6042c (not pushed). Debug portable build started detached.

E2E output root cli-only-opencode-go-r1 is already staged with account+config and OpenCode Go credentials (no isolated/work). Run still waits on the portable and then CLI-only -ThirdPartyModel codexhub-opencode-go/muse-spark-1.2-contributor with pinned OMP.

[NOirBRight]

Yoga CLI-only r3 on portable CodexHub_0.1.9-beta.3.13_debug_portable_49940558 (SHA 4994055896e4341ce82c4699c58beb06f4f6042c):

Passed 5/8: all four Official Luna cases, plus codex-cli-opencode-go. Failed 3/8: OpenCode/Pi/OMP third-party (http=200, gateway complete, sentinel/tool counts match). Key is live. r1/r2 were preflight (missing codex.cmd PATH, then ChatGPT AppX bootstrap). Next: confirm model_selected alias codexhub-opencode-go/... vs gateway opencode-go/... and rerun those three cases.

[NOirBRight]

Yoga r4 still 5/8 after the codexhub-opencode-go/... alias (not the failure). r5 case artifacts now record measurement counters.

Failed third-party cases all selected gateway model opencode-go/muse-spark-1.2-contributor (http 200, stream=true, exit 0):

• opencode-opencode-go: tool events 0, terminal events 0
• pi-opencode-go and omp-opencode-go: tool events 1, but 2 assistant_output events (contract requires exactly 1 sentinel output)

Official Luna remains 4/4. codex-cli-opencode-go remains passed. This is live routing plus a Muse Spark client-contract miss, not quota/preflight.

[NOirBRight]

Yoga r6/r7 still 5/8.

Parser updates made OpenCode third-party emit tool+terminal (was 0/0). All three remaining failures are assistant_output=2 with exactly one required sentinel: the extra chunk is non-empty and not the sentinel, so duplicate-sentinel allowance did not help. Official Luna 4/4 and codex-cli-opencode-go still pass. Routing is live; Muse Spark on OpenCode/Pi/OMP adds extra assistant text that the E2E contract correctly rejects.

[NOirBRight]

Local cargo test --locked -- --test-threads=1 (skipping the known flaky diagnostics round-trip): 674 passed, 0 failed, 4 ignored.

Yoga r8 with a stricter "reply with only this exact line" prompt is still 5/8. The three OpenCode-Go failures remain assistant_output=2 with tool+terminal present. Official Luna 4/4 and codex-cli-opencode-go still pass. Stopping further prompt/contract loosening.