chat-launched-with — Workgraph live mirror

Metadata

Status	done
Assigned	`agent-954`
Agent identity	`3184716484e6f0ea08bb13539daf07686ee79d440505f1fdf2de0357707034c3`
Model	`claude:opus`
Created	2026-04-28T21:23:33.076743389+00:00
Started	2026-04-28T21:51:41.583465222+00:00
Completed	2026-04-28T22:10:54.264805796+00:00
Tags	`eval-scheduled`
Eval score	0.91
└ blocking impact	0.92
└ completeness	0.92
└ coordination overhead	0.88
└ correctness	0.95
└ downstream usability	0.90
└ efficiency	0.85
└ intent fidelity	0.81
└ style adherence	0.90

Description

User reports: launched a new chat configured for codex, got a claude session. Model selection on chat creation is being ignored (or overridden) somewhere in the chat-creation pipeline.

Verbatim: 'lol. i just tried to launch a chat with codex and got... claude. lol lol.'

Likely suspects:

The new chat dialog / wg service create-chat (or whatever the entry point is) doesn't pass the chosen model spec to the spawn
Project-level coordinator.model (claude:opus in this user's config) silently wins over the per-chat selection
Per CLAUDE.md, agency tasks ignore project cascade and pin to claude:haiku — chat agents may have analogous logic incorrectly applied

This is distinct from the codex worker-task handler bug (already fixed in b6c786cef). That was about prompt-stdin wiring for wg add worker tasks. This bug is in the chat-spawn code path, which is a different surface.

Investigation order

Repro: launch a fresh chat, explicitly pick codex / gpt-5.5 in whatever the UI offers. Confirm via wg agents --alive what handler actually spawned.
Find the chat-spawn code path (likely src/commands/service/coordinator_agent.rs, src/commands/service/mod.rs, or the IPC layer between TUI and dispatcher). Trace where the model selection flows from UI → spawn args → executor wrapper.
Identify where the override happens (config cascade vs hardcoded vs IPC drop) and fix it.
Add a regression test: spawning a chat with explicit model spec must result in a worker subprocess invoked with that model.

Validation

Failing test first: spawning a chat agent with model = codex:gpt-5.5 results in a codex CLI subprocess (mock the subprocess, assert on argv)
Implementation makes the test pass
Live smoke: open TUI, create a new chat picking codex, confirm wg agents --alive shows codex executor
Live smoke: same with claude — assert no regression
Live smoke: existing chats spawned pre-fix continue to work on whatever they were on
cargo build + cargo test pass

## Description

User reports: launched a new chat configured for codex, got a claude session. Model selection on chat creation is being ignored (or overridden) somewhere in the chat-creation pipeline.

Verbatim: 'lol. i just tried to launch a chat with codex and got... claude. lol lol.'

Likely suspects:
1. The `new chat` dialog / `wg service create-chat` (or whatever the entry point is) doesn't pass the chosen model spec to the spawn
2. Project-level `coordinator.model` (claude:opus in this user's config) silently wins over the per-chat selection
3. Per CLAUDE.md, agency tasks ignore project cascade and pin to claude:haiku — chat agents may have analogous logic incorrectly applied

This is distinct from the codex worker-task handler bug (already fixed in `b6c786cef`). That was about prompt-stdin wiring for `wg add` worker tasks. This bug is in the chat-spawn code path, which is a different surface.

### Investigation order
1. Repro: launch a fresh chat, explicitly pick codex / gpt-5.5 in whatever the UI offers. Confirm via `wg agents --alive` what handler actually spawned.
2. Find the chat-spawn code path (likely `src/commands/service/coordinator_agent.rs`, `src/commands/service/mod.rs`, or the IPC layer between TUI and dispatcher). Trace where the model selection flows from UI → spawn args → executor wrapper.
3. Identify where the override happens (config cascade vs hardcoded vs IPC drop) and fix it.
4. Add a regression test: spawning a chat with explicit model spec must result in a worker subprocess invoked with that model.

## Validation

- [ ] Failing test first: spawning a chat agent with `model = codex:gpt-5.5` results in a codex CLI subprocess (mock the subprocess, assert on argv)
- [ ] Implementation makes the test pass
- [ ] Live smoke: open TUI, create a new chat picking codex, confirm `wg agents --alive` shows codex executor
- [ ] Live smoke: same with claude — assert no regression
- [ ] Live smoke: existing chats spawned pre-fix continue to work on whatever they were on
- [ ] cargo build + cargo test pass

Depends on

done .assign-chat-launched-with

Required by

(none)

Log

2026-04-28T21:23:33.067629199+00:00 Task paused
2026-04-28T21:24:04.235025253+00:00 Task published
2026-04-28T21:25:34.131344992+00:00 Spawned by coordinator --executor claude --model opus
2026-04-28T21:26:13.610219984+00:00 Starting investigation: looking for chat-spawn code path where model selection might be dropped
2026-04-28T21:34:34.796637209+00:00 Root cause: maybe_auto_enable_chat_pty (state.rs:12454) uses config.coordinator.effective_executor() — IGNORES per-coordinator CoordinatorState.executor_override/model_override. Even when chat created with --executor codex, TUI auto-spawns claude binary.
2026-04-28T21:46:37.756091638+00:00 Reconciliation: task recovered from orphaned state (agent: agent-937)
2026-04-28T21:46:47.775018668+00:00 Spawned by coordinator --executor claude --model opus
2026-04-28T21:47:35.044504439+00:00 Resuming after agent-947 spawn — reviewing prior WIP. Fix touches state.rs (resolve_chat_pty_executor_and_model + auto_enable_chat_pty integration + per-handler model arg + WG_MODEL propagation) and spawn_task.rs (reads WG_MODEL alongside WG_EXECUTOR_TYPE). Three regression tests in state.rs, one in spawn_task.rs.
2026-04-28T21:51:30.854696554+00:00 Task unclaimed: agent 'agent-947' (PID 638541) process exited
2026-04-28T21:51:41.583469200+00:00 Spawned by coordinator --executor claude --model opus
2026-04-28T21:58:11.785889403+00:00 Build: clean. New regression tests pass: chat_launched_with_codex_uses_codex_executor, chat_with_no_overrides_uses_global_default, chat_with_only_model_override_keeps_global_executor, spawn_task_propagates_wg_model_env_var. Pre-existing failures in spawn_task::tests (role_override_wins, coordinator_task_gets_coordinator_role, etc) are unrelated — they fail on main without my changes (verified by stash-then-test; stash dropped cleanly).
2026-04-28T22:10:03.449297358+00:00 Live smoke verified end-to-end: wg chat create --executor codex --model codex:gpt-5 against claude:opus default → CoordinatorState gets overrides → wg spawn-task with daemon-set WG_EXECUTOR_TYPE/WG_MODEL → 'wg codex-handler --chat .chat-0 -m codex:gpt-5' (NOT claude). Same env vars cleared → falls back to global default (claude/opus). Added grow-only smoke scenario chat_launched_with_codex_uses_codex.sh.
2026-04-28T22:10:38.746277545+00:00 Committed: d3b6315b6 — pushed to remote
2026-04-28T22:10:54.264819111+00:00 Task pending eval (agent reported done; awaiting `.evaluate-*` to score)
2026-04-28T22:14:13.279066442+00:00 PendingEval → Done (evaluator passed; downstream unblocks)