feat(tui): remember the last used model across restarts

[dennisonbertram]

Closes #1424

The problem

Pick a model with /model, quit, restart — back on the daemon default. newTUIConfig (cmd/harnesscli/main.go:562) never set Model, so selectedModel started empty every session.

Not a missing store: ~/.config/harnesscli/config.json already persists starred models, gateway, API keys, command history and theme. The model was the conspicuous omission, and it is the one users change most often.

What changed

harnessconfig.Config gains Model, Provider and ReasoningEffort, all omitempty so older files load unchanged and older binaries ignore them. ModelSelectedMsg persists all three through the existing persistConfigField idiom; the constructor applies them right where starred models, gateway, keys and history are already applied.

The three travel together because they are chosen together — a provider or reasoning effort without its model describes nothing, and restoring only the model would leave the status bar agreeing with itself while runs went elsewhere.

Precedence: applying is guarded on cfg.Model == "", so an explicitly requested model always wins. Remembering a preference must never override an instruction. Empty still means "let the daemon choose", so a first run and an unreadable config both behave exactly as today.

The bigger thing this exposed

Making the model persistent turned the TUI test suite into something that mutates the developer's machine. A test run wrote into my real ~/.config/harnesscli/config.json:

"model": "gpt-4.1-mini",
"provider": "openai"

A model no human had chosen. Three tests then failed against it — including one asserting "no model chosen" — and they failed for a reason that had nothing to do with what they test.

Fixed with a package TestMain that redirects HOME once before any test runs, plus a per-test reset of the scratch config in initModel.

Both parts are needed, and the obvious approach does not work:

• t.Setenv in the shared initModel helper panics for any test calling t.Parallel, and several in this package do. Redirecting once in TestMain, before any test starts, is both parallel-safe and race-free.
• The redirect alone is insufficient: one scratch HOME is shared by the whole package, so a test that selects a model still leaks into every test after it. Hence the per-test reset.

I also tried t.Setenv conditionally in the helper; it broke four feedback tests that set their own HOME and assert on files written under it.

Verification

Four tests, red first:

TestRememberedModelAppliedAtStartup   selectedModel = "", want the remembered "gpt-4.1-mini"
TestModelSelectionIsPersisted         stored Model = "", want "claude-sonnet-5"

TestNoRememberedModelLeavesDaemonDefault and TestExplicitModelBeatsRememberedModel are the controls — they passed before the change and must keep passing, so "remembering" cannot be faked by defaulting to something nobody picked, and a preference cannot outrank an instruction.

go test ./cmd/harnesscli/... -race green. My real config after a full race run — no model key, which is the check that the isolation actually holds:

['history_entries', 'starred_models', 'theme']

Live restart is being verified separately and I will post the result here. The unit tests prove persist and restore; they do not prove a real TUI session survives a quit. I will not claim it until I have watched it.

Out of scope

runTUI never receives the -model flag, so harnesscli --tui -model X ignores it today. That is a real defect, recorded in the engineering log, and deserves its own issue. The precedence guard here is written so wiring it needs no further change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WJGxhoFhA8JjkwZFcLGdS5

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[dennisonbertram]

Live restart verified

Driven through a pty against a real daemon on this branch's binaries, with HOME redirected to a scratch directory throughout.

Before any selection — status bar shows the placeholder, no model:

Type /model to select a model  •  Type /help for all commands

After selecting a model via /model:

Model: GPT-4.1

Scratch config after quitting:

{
  "model": "gpt-4.1",
  "provider": "openai"
}

After relaunching, without touching anything:

GPT-4.1

Step 8 matches step 4. The model survives the restart.

Two details worth recording rather than glossing:

reasoning_effort was not written, because GPT-4.1 is not a reasoning model and the picker never offered an effort to choose. That is the omitempty behaving correctly, not a dropped field — but it does mean this run exercised two of the three persisted values, not all three.

The switcher in this build is a flat fuzzy filter across providers, not the two-level provider drill-down I described in the issue. My description was wrong; the behavior under test is unaffected.

Hygiene: the real ~/.config/harnesscli/config.json was re-checked afterwards and still contains only starred_models, history_entries, theme — no model key. That is the check that the TestMain isolation actually holds outside the test suite too.