feat(harness): model catalog, connection test, edit matrix in the create/edit wizard (PHNX-2218)

[muqsitnawaz]

Lands the harness create/edit wizard cluster — PHNX-2218 (parent) and its three open subtasks PHNX-2220 / 2221 / 2222 — by filling the typed no-op seams the RUSH-2219 wizard scaffold (harness-wizard.ts) shipped. One worktree, built in ticket order.

What each ticket delivered

• PHNX-2220 — param surface + model catalog. The model step is a catalog pick from the host's own getModelCatalog (same list behind agents models <host>) with a tier/alias hint and an always-present "type a custom model id…" escape hatch; a host with no probeable catalog falls back to free text. The endpoint step is gated to hosts that carry a base-URL slot. New src/commands/harness-hooks.ts implements the pickModel hook (+ pure buildModelChoices), passed into both wizard runs.
• PHNX-2221 — pre-save connection test. New src/lib/harness-connection-test.ts runs a real agents run <name> "say alive in one word" --headless --timeout 60s through the live path (resolveProfileForRun -> resolveProfileEnv -> buildExecEnv -> spawn) and classifies the outcome into pass / auth / endpoint / model / unknown. Behind a confirm (default yes) with --test / --no-test; non-interactive skips unless --test. A failure never blocks silently — on a TTY it offers keep / edit now / delete-and-cancel.
• PHNX-2222 — editability matrix. harnessEditable(host) is sourced entirely from the resolver maps (baseUrlEnvKeyForHost / authEnvKeyForHost / isSelfUpdatingAgent), each disabled field carrying its reason; defaultEditable is now its boolean projection, so the wizard enable/disable can never drift from what a run honors.

Run evidence

Real connection test on the live path (fresh HOME, claude unresolvable -> classified + non-interactive keep+hint):

$ agents harness add corp --host claude --model gpt-x --base-url http://127.0.0.1:1/v1 --test
Harness 'corp' added — claude + gpt-x.
Testing 'corp' — sending "say alive in one word" through agents run…
Run failed (exit 1). (Resolved custom harness 'corp' -> claude)
Kept anyway. Fix and retest with: agents harness edit corp

Endpoint gating (opencode carries no base-URL slot, not asked/dropped):

$ agents harness add spark --host opencode --model meta/muse-spark-1.1 --no-test
Harness 'spark' added — opencode + meta/muse-spark-1.1.

Tests

• src/lib/harness-connection-test.test.ts — classifier maps real provider stderr -> auth / model / endpoint / unknown / pass.
• src/commands/harness-hooks.test.ts — catalog -> select choices (escape hatch, keep-current, tier/alias hints).
• src/commands/harness-wizard.test.ts — extended for harnessEditable.
• Full suite running on a fleet worker.

Docs: cli/docs/profiles.md + CHANGELOG fragment cli/.changelog/next/PHNX-2218.md.

Closes PHNX-2220, PHNX-2221, PHNX-2222; advances PHNX-2218.

Generated with Claude Code

[muqsitnawaz]

Non-author review — code-reviewer subagent (prix-cloud paused per #1767)

Verdict: APPROVE ✅

Reviewed the full diff vs origin/main against cli/CLAUDE.md §Code review conventions and the ticket intent.

Clean on the substance:

• harnessEditable is sourced entirely from the resolver maps (baseUrlEnvKeyForHost / authEnvKeyForHost / isSelfUpdatingAgent); defaultEditable is a pure projection — no lying/hardcoded capability table. Pinned against MANAGED_AGENT_IDS in tests.
• The connection-test classifier is exhaustively tested against real provider error shapes (auth / model / endpoint / unknown / pass).
• Base-URL step fails loud / disables-with-reason at the boundary rather than silently dropping.
• connectionTestStep()'s permanent skip is intentional (superseded by runConnectionTest, wired at both call sites), not a stub.
• Docs (cli/docs/profiles.md) + CHANGELOG (cli/.changelog/next/PHNX-2218.md) in sync with behavior.

First round was CHANGES REQUESTED on two test-coverage gaps (the tri-state --test/--no-test gate and the pickModel wiring). Both resolved in a17520318 by extracting pure exported functions (connectionTestGate, chooseModelFromCatalog, pickModel) with real unit tests, plus a doc callout for the non-interactive --test behavior. All 74 harness tests pass.

Non-blocking residual: pickModel's "host known but catalog empty" branch and catalogVersionFor remain untested — thin wiring already covered indirectly; not worth another round.