[dzianisv]

Verdict: STALE gate, not an app regression

Publish run 43572233118 (head e7385a3, tag v0.4.18) failed only on job cua-mid-stream ("streaming NOT proven, cycle 1/5") while the publish job succeeded.

• The gate's run manifest shows both streaming-proof samples at text_length 0 with session status busy — the server produced zero text tokens in ~60s while still working.
• The sampler reads the opencode server API, not the app, so the app's PR fix: SSE startup + liveness (#189, #186) #237 SSE changes cannot cause this.
• The MODEL_CAPABLE probe (short prompt) passed minutes earlier on the same server/model (gpt-5.4).
• The gate scenario has NEVER passed: the Nov 6 run was a deliberate CUA_FORCE_FAIL proof, and its cycles were empty.

Why removal, not adjustment

• Den removed the sibling cua-smoke.yml on 2026-11-13 (commit e844493, "no credits to run it"); this job still spends gpt-5.4 credits on every manual dispatch.
• The gate has been non-blocking since Den's 2026-11-06 decision (publish never had a needs: on it).
• It never passed, so no signal is lost by removing it.

What this PR does

• Deletes the entire cua-mid-stream job from .github/workflows/publish-play-store.yml (~240 lines), plus the now-dead cua_force_fail dispatch input.
• Updates the three comments that described the gate (trigger block, job block, above publish) with a short removal note dated 2026-11-14.
• scripts/android-cua-smoke.py is KEPT for local runs (--mid-stream-exit still works); one stale comment in it (referencing the removed cua_force_fail input) is corrected to the env-var form.
• The publish job is unchanged apart from its leading comment: same steps, no needs: added or removed. YAML validated with yaml.safe_load; the workflow now has a single job, publish.

Revival recipe (if this coverage is wanted back)

1. Fix the sampler: count reasoning parts as streaming evidence — it currently counts only type == "text" parts, so a model that streams reasoning first looks like zero output.
2. Raise the 60s first-text wait (the server produced zero text tokens in ~60s while busy).
3. Require a green baseline run of the gate standalone BEFORE re-adding it as a gate (it never passed, so there is no known-good baseline today).

DO NOT MERGE

Awaiting dispatcher/owner review. This PR only removes CI; it does not touch the Play rollout and publishes nothing.

[dzianisv]

Claim — Ember (reviewer + merger lead).

Task: review + merge PR #246, branch chore/remove-cua-mid-stream-gate, Work Log row "Review + merge PR #246: remove CUA mid-stream gate from publish workflow".

Scope: Stage 4 review of the workflow diff (publish job byte-identical, cua-mid-stream job + dead cua_force_fail input removed, script kept for local runs), then Stage 5 merge on a PASS verdict with checks green on the exact head, verified by SHA readback, then Stage 5.5 design/PRD sync. No release tag, no Play rollout change, nothing published.

Checked before claiming: PR is open, mergeable, head 8fadc3e; no prior claim comments or reviews on this PR.

[dzianisv]

Stage 4 review verdict: PASS — head 8fadc3e.

Method: claude-code-review (parallel compliance + bug passes, independent validation per finding), verified from the actual diff in a real checkout plus lead readback of the workflow files via the GitHub API. Acceptance criteria, one line each:

1. Publish job byte-identical except its leading comment — PASS. Mechanical base-vs-head job-block diff: 378 vs 379 lines, the only change is the 2-line comment replaced by a 3-line 2026-11-14 note. Steps, env, runs-on identical; needs: None on both.
2. Only cua-mid-stream job + dead cua_force_fail input removed — PASS. Diff is exactly 2 files: publish-play-store.yml +16/−250, scripts/android-cua-smoke.py +3/−3. Jobs [cua-mid-stream, publish] → [publish]; dispatch inputs [cua_force_fail, status, track, user_fraction] → [status, track, user_fraction].
3. Three gate comments updated — PASS. Trigger block, job-block position, and publish leading comment each carry the NOTE (2026-11-14), STALE-gate verdict, and run 43572233118.
4. Script kept, comment corrected — PASS. scripts/android-cua-smoke.py present at head (122,261 bytes); --mid-stream-exit flag untouched; only the stale cua_force_fail comment changed to the env-var form.
5. YAML valid, single job — PASS. yaml.safe_load parses the head workflow; jobs == ['publish'].
6. Test-count delta — 0. No test file is in the diff (only the workflow YAML and one script comment changed); test files 52 → 52, cases 370 → 370. No unit or mock test added.
7. Bug pass — PASS, no defects. cua-mid-stream appears only in the three intentional removal notes; cua_force_fail appears nowhere in the head workflow; CUA_FORCE_FAIL survives only as the live env-var check in the kept script (intended for local runs); no AZURE_/OPENCODE_ leftovers in the head workflow; no other workflow consumes the script.

Filed note (validated, NOT a fail): scripts/android-cua-smoke.py:1877–1880 still says the per-cycle filename matches the upload step's logcat-cycle-*.txt glob in publish-play-store.yml — that glob lived in the removed job and no longer exists. Comment-only staleness; file-naming behavior unchanged, nothing consumes the glob now. Optional follow-up: reword to note the naming is kept for local-run consistency. Not a merge blocker.

Checks on this head: test, Typecheck and unit tests, build, Unsigned iOS Simulator build, Secret scan all SUCCESS; release + build-fdroid skipped (correct for a CI-only change). Ready for Stage 5 merge on this head.

[dzianisv]

Merged. Merge commit 01a7b6b (main head, read back via API), merged 2026-11-14T16:42:56Z. Stage 4 verdict PASS comment 6880871622 on head 8fadc3e; checks green on that exact head at merge time. The cua-mid-stream gate job and the dead cua_force_fail input are gone from publish-play-store.yml; scripts/android-cua-smoke.py is kept for local runs. No release tag created, no Play rollout change, nothing published. Stage 5.5 (design/PRD sync) follows.

[dzianisv]

Stage 5.5 — Design/PRD sync: DONE.

• Design doc updated: Notion "Publish — Google Play and F-Droid" (https://app.notion.com/p/417ac25eb49f81589befc07623f52f2c) — the passage saying the cua-mid-stream job's dead code "stays in publish-play-store.yml until someone with workflow scope removes it" now records the removal (PR chore(ci): remove non-blocking CUA mid-stream gate from publish workflow #246, merge 01a7b6b, 2026-11-14); the footer line that dated the gate removal 2026-11-13 (that date was the sibling cua-smoke.yml removal) is corrected to 2026-11-14 for the gate job. Verified by fetch-back.
• TDD checked — no update needed: its only CUA mention is "scripts/ E2E test rig (LLM-driven CUA Android smoke)" in the repo tree, and the script is kept, so it stays accurate.
• PRD checked — no update needed: the CUA gate is CI-only, not product behavior; nothing the PRD promises changed.
• Issue Release gate: CUA mid-stream-exit scenario as pre-publication prerequisite (regression OPENCODE-MOBILE-6/-7) #227 (closed design issue) and the closed triage row are historical records; left as-is.

Task complete: Stage 4 PASS (comment 6880871622, head 8fadc3e), merged 01a7b6b, Stage 5.5 recorded here.