0.1.8 Beta6: Complete the GLM-5.2 Ollama Chat and V2 CLI path

Parent

Outcome

Run the real Codex CLI through the complete Beta6 text, standard-tool, and Collaboration V2 contracts on the maintained Ollama Cloud GLM-5.2 Chat Completions endpoint.

Acceptance criteria

Text reaches the CLI progressively with bounded buffering/backpressure.
Multi-turn text, default reasoning, and a standard function call/result lifecycle complete with exact identity.
A real read -> target-scoped apply_patch -> verify loop completes.
Collaboration V2 completes spawn, canonical task identity/status, message, follow-up, wait/result, list/interrupt, child result delivery, parent completion, streaming/history, and same-Home restart/readback.
Runtime fields and V2 items are preserved, consumed, reversibly adapted, or rejected by value according to 0.1.8 Beta6: Generate the runtime-derived Chat and Collaboration V2 conversion matrix #66.
Unsupported non-V2 Chat semantics fail visibly without hidden Responses/Official fallback.
Exact Provider/model identity remains unchanged and no cross-Provider execution occurs.
No GLM-specific production conversion branch is introduced.

Blocked by

Authenticated GLM-5.2 Chat qualification passed on source candidate 45e06184f00b5ced9df40069c8fede0423a98840; bounded evidence is in 4d7807c under docs/evidence/authenticated-provider/chat/summary.json.

Real codex-cli 0.147.0 observations: non-stream Chat 200, progressive CLI text, default-reasoning stream observations, exact exec -> reversible apply_patch -> verify, all six Collaboration V2 tools, child result, same-Home Gateway restart/resume with config + AGENTS preservation, Cross-Home rejection before any Gateway request, exact ollama-cloud/glm-5.2 identity, fallback=0, V1=false. The maintained strict native Responses codec is now generically scoped away from Chat attempts (07eeee4), with no GLM-specific conversion branch.

Verification: focused 357 passed + 16 subtests; full Python core 2,458 passed / 90 skipped / 508 subtests, with only 61 known Linux PowerShell/Windows host gaps; quality parse_errors=0.

Final authenticated qualification supersedes the earlier preliminary evidence.

Source candidate: 73e4ed3d444414dedcc1a5902fa9f0a9c848d469
Evidence commit: c82baad (docs/evidence/authenticated-provider/chat/summary.json)

GLM-5.2 passed the real codex-cli 0.147.0 -> Gateway Responses -> ollama-cloud /v1/chat/completions -> Responses path with: non-stream 200; 21 progressive text streams; same-session text resume; explicit default reasoning policy; exact standard/custom call-result identity; target-scoped apply_patch; independently qualified spawn/wait/follow-up/message/list/interrupt turns; canonical task, frozen result shapes, child delivery and history identity; same-Home restart/readback; Cross-Home rejection before request. Exact Provider/model/protocol remained bound, fallback=0, V1=false. No GLM-specific production branch was added; the native Responses codec is scoped away from Chat generically.

Final checks: 366 focused tests + 16 subtests; full Python core 2,458 passed / 90 skipped / 508 subtests with only the 61 known Linux PowerShell/Windows-host gaps; quality parse_errors=0. Final Standards/Spec review found no Spec blocker.