test(acceptance): exercise provider/catalog TUI slash commands through real PTY

[dennisonbertram]

Why this matters

Existing commands /keys and /model are unproven in the real TUI matrix because the fixture lacks configured/missing-provider and /v1/models catalog states. They must be driven through rendered selection/error/success states.

Acceptance contract

Create safe deterministic configured/missing-key and model-catalog fixtures; use 100x30 PTY with keyboard selection/invalid path; assert rendered state, API postcondition and follow-up conversation behavior.

Current architecture and search evidence

/keys cmd/harnesscli/tui/model.go:2091; /model :2068; existing harness provider/model endpoints and catalog fixture seams.

In scope

Create safe deterministic configured/missing-key and model-catalog fixtures; use 100x30 PTY with keyboard selection/invalid path; assert rendered state, API postcondition and follow-up conversation behavior.

Out of scope

All other TUI command groups, cron/callback scenarios, native GUI, and product fixes absent a separately filed live failure.

Work type

Engineering change / feature slice

Change class

Infrastructure, CI, or process change

Protected path

Registry-derived command case -> real harnesscli 100x30 PTY -> rendered TUI state -> isolated harness/API/store or filesystem postcondition -> continued usable conversation -> bounded cleanup.

Cross-surface impact map

Test-only PTY matrix/fixtures and existing read/write API endpoints. Isolated config/workspace/DB paths only. Explicit focus/keyboard/error/confirmation handling, durable probes, no secrets. Native GUI is excluded. No production behavior/schema/deployment change unless a live test demonstrates a separate bug. Required plan, impact map, logs, indexes, focused/race/full regression and hosted CI.

Product and UX contract

Drive exactly the keyboard interaction a user uses; capture success, invalid/error, selection/focus and destructive-confirmation states. A fixture shortage is not an N/A outcome.

Coordination and dependencies

Child work for #1088/#1010. It may build on the non-closing current 24-command matrix foundation, but must use a fresh current-main worktree and a single PR with Closes #N.

Test-first plan

Add red real-PTY fixture case(s) proving the command cannot be credited without rendered and durable evidence, then implement the narrowest deterministic fixture/driver support. Run focused tests, race, ./scripts/test-regression.sh, repeat bounded PTY smoke, and retain artifacts.

Verification plan

Record exact source SHA, terminal frame hashes/transcript, typed keys, API/SSE and relevant DB/filesystem evidence, IDs, cleanup, race/full regression and independent review. Do not claim GUI proof.

Rollout and rollback

Acceptance-only; revert the isolated PR, with no migration or deployment.

Documentation and handoff

Update active plan, impact map, engineering/observational/long-term logs, indexes and #1088 applicability matrix.

Definition of done

• Each named command is driven through a real 100x30 PTY, not reducer-only
• Rendered result and durable postcondition are retained
• Required multi-message/confirmation/invalid behavior is covered
• Focused/race/full regression, independent review and CI pass
• No fixture shortage is recorded as a user-path pass

Process acknowledgement

• I completed current-architecture search and impact analysis and will update scope if evidence changes.