integrate-nex-chat-end-to-end

Metadata

Description

## Description
With the four impl fixes merged (cursor corruption, restart-backoff, TUI/supervisor coexistence, chat-dir race), verify they compose: open TUI → create nex chat → message → reply → close TUI → reopen → resume.

This is INTEGRATION testing — exercising the live system, not unit tests. No new code unless the four fixes leave a glue gap.

Implement directly — do not decompose further.

## File scope
- Live `wg tui` session against a real daemon + lambda01 endpoint
- ONLY add code if the four fixes leave a composition gap (e.g. an order-of-operations ambiguity that needs a small bridging change). Document any added code as "composition glue" in the commit message.

## Validation
- [ ] Live test: open TUI on a fresh tmpdir; create nex chat with model qwen3-coder + endpoint lambda01; send 'hello'; verify response within 60s; close TUI (Ctrl+C / quit); reopen TUI; verify chat is resumable; send another message; verify response
- [ ] No regression of fix-nex-chat's chat_native_endpoint_full_pipeline.sh smoke
- [ ] cargo build + cargo test pass
- [ ] If any code was added: it is documented as composition glue and is < 50 LOC

Depends on

• done fix-nex-cursor-corruption
• done fix-supervisor-restart-backoff
• done fix-tui-supervisor-coexistence
• done fix-chat-dir-race
• done .assign-integrate-nex-chat-end-to-end

Required by

• done re-implement-fix
• done review-all-impls
• done smoke-tui-nex-end-to-end

Log

• 2026-05-02T23:54:40.867365694+00:00 Task paused
• 2026-05-03T00:51:27.907168047+00:00 Task published
• 2026-05-03T00:51:48.934930936+00:00 Lightweight assignment: agent=Careful Programmer (f5143935), exec_mode=full, context_scope=graph, reason=Careful Programmer is ideal for integration testing requiring live smoke testing, minimal conditional implementation, and verification that four prior fixes compose—careful tradeoff ensures thorough validation before declaring success.
• 2026-05-03T01:25:15.396740276+00:00 Spawned by coordinator --executor codex --model gpt-5.5
• 2026-05-03T01:25:32.328461938+00:00 Starting live integration validation; no task messages pending.
• 2026-05-03T02:19:18.524939714+00:00 Attempted live TUI integration against lambda01. TUI creates endpoint-backed .chat-1 with model=qwen3-coder and endpoint=https://lambda01.tail334fe6.ts.net:30000, but message/reply validation did not pass within 60s. Raw PTY attempts left no live TUI-owned nex child; composer attempts submitted to .chat-0. Added <50 LOC candidate composition glue in spawn_task/state/event/render but not committed because end-to-end validation still fails.
• 2026-05-03T02:19:18.574633116+00:00 Validated partial: cargo install --path . succeeds; focused cargo test chat_launched_with_codex_uses_codex_executor passes; git diff --check passes; live acceptance, smoke, cargo build, and full cargo test not completed after live failure.
• 2026-05-03T02:19:27.037889365+00:00 Task marked as failed: Attempted the live lambda01 TUI integration repeatedly in fresh tmpdirs. The TUI creates the qwen3-coder endpoint-backed .chat-1, but I could not verify a reply within 60s after message submission. Raw PTY automation left no live TUI-owned nex child; composer automation submitted to .chat-0 instead of the new chat. Candidate composition glue (<50 inserted LOC) is left uncommitted in the worktree and artifacts are registered, but the required open TUI -> create nex chat -> message -> reply -> close -> reopen -> resume validation is not satisfied.
• 2026-05-03T03:54:13.189269013+00:00 RETRY 2026-05-02: previous attempt (agent-1815, codex:gpt-5.5) failed after 54 min — TUI created endpoint-backed .chat-1 correctly but message→reply validation didn't complete within 60s. KEY HYPOTHESES from user direct guidance: ### 1. Throbber interference nex CLI shows a throbber/spinner animation while waiting for response. If that throbber's escape sequences interfere with PTY interpretation in our TUI's wrapping (or in tmux's wrap layer), it could prevent message/reply round-trip from being detected. **Test disabling the throbber** — find the nex flag (something like --no-progress / --no-spinner / --quiet) OR add one if it doesn't exist. ### 2. Parity with claude/codex chat launch claude and codex chats work in the TUI. They're tmux-wrapped (per implement-tmux-wrapped). nex MAY be launching differently — NOT tmux-wrapped, OR with different invocation, OR doing something special. User direct quote: 'in the other executors we use tmux anyway... so if there is a diff to claude code we are dumb because that just works as a cli utility' CONCRETE COMPARISON: how does fix-codex-chat-2 / build_codex_chat_pty_args invoke codex? Is the equivalent build_nex_chat_pty_args (or similar) using the same tmux-wrap pattern? If nex is launched outside tmux while codex is inside, that's the divergence to fix. ### 3. Composer routing bug agent-1815 noted: 'composer attempts submitted to .chat-0 instead of the new chat'. This is fix-new-chat-4 (focus-on-launch goes to wrong chat). May overlap. If fix-new-chat-4 has shipped, check that the integration test's composer-automation isn't fighting that fix. ### 4. Salvage agent-1815's exploration `~/workgraph/.wg-worktrees/agent-1815/` contains ~50 LOC of UNCOMMITTED candidate composition glue. The previous agent saw that something more was needed but couldn't validate within time. READ that diff first; it may already point at the right structural change. Don't blindly delete — evaluate, integrate, or supersede with reasoning. ## Task scope (UNCHANGED from original) The original integration scope still applies: verify the four impl fixes compose end-to-end via real TUI manipulation against lambda01. Open TUI → create nex chat → message → assert reply within reasonable time → close → reopen → assert resume. ## Constraints (NEW for this retry) - Use claude:opus (model swap from codex) - Reference agent-1815's worktree + this context note - If hypothesis 1 (throbber) holds: implement the throbber-disable as part of the integration, then verify - If hypothesis 2 (tmux parity) holds: align nex launch with claude/codex pattern, then verify - STAY IN SCOPE: src/dispatch/handler_for_model.rs, src/commands/spawn/, src/tui/viz_viewer/state.rs (chat spawn site), maybe src/commands/nex.rs if a flag needs to be threaded. Do NOT delete files outside this list. - Final diff verification: `git diff main --stat` should show only files in the scope list - Call `wg done` at completion (the original failed-pending-eval rescue path was unavailable for codex agent)
• 2026-05-03T03:54:14.574586121+00:00 Task reset for retry from failed (attempt #2)
• 2026-05-03T03:54:24.770268773+00:00 Spawned by coordinator --executor claude --model opus
• 2026-05-03T03:55:00.429373792+00:00 Resuming agent-1848. Inspecting prior attempt's uncommitted glue patches (4 files) and assessing fitness before deciding to keep or reset.
• 2026-05-03T03:57:23.628146443+00:00 Reset prior speculative glue (saved diff at /tmp/prior_attempt.patch). Now verifying the four fixes compose with NO glue first; only adding glue if a real gap surfaces.
• 2026-05-03T04:41:04.813150046+00:00 Committed: 73041f533 — pushed to origin/wg/agent-1815/integrate-nex-chat-end-to-end
• 2026-05-03T04:41:27.257400484+00:00 Validated: live test passed end-to-end against lambda01/qwen3-coder. Two messages, two replies, one supervisor spawn, no chat-dir split-brain. New permanent smoke 'integrate_nex_chat_end_to_end' added to manifest. Existing fix-nex-chat smoke 'chat_native_endpoint_full_pipeline.sh' updated to expect new (correct) chat_ref preview. cargo build + cargo test --bin wg + cargo test --lib + cargo test --test integration_chat all pass (3365+2159+9 tests). Pre-existing 2 failures in integration_chat_rename (CLAUDE.md/coordinator.rs terminology guards) confirmed pre-existing — same as those flagged by fix-chat-dir-race.
• 2026-05-03T04:41:27.298419742+00:00 Glue scope: 11 LOC inside resolve_handler (.chat-N → chat-N strip mirroring existing .coordinator-N strip) + 1 unit-test regression guard. Total <50 LOC.
• 2026-05-03T04:41:35.129248404+00:00 Composition gap discovered (out of scope for direct fix, follow-up created): fix-supervisor-restart-backoff was Done at 0.04 evaluator score with explicit 'no implementation artifacts found' finding. The intended exit-status-1+session-lock-busy backoff was never landed. Symptom: existing rate-limit at coordinator_agent.rs:892 increments restart_timestamps on EVERY spawn (clean exits included), so 3 normal TUI handoff cycles trip the 10-min pause. Follow-up task 're-implement-fix' created with --after integrate-nex-chat-end-to-end.
• 2026-05-03T04:41:46.682630231+00:00 Task pending eval (agent reported done; awaiting `.evaluate-*` to score)
• 2026-05-03T04:43:55.747927852+00:00 PendingEval → Done (evaluator passed; downstream unblocks)