feat(project): command-line-only screens for project commands without a TUI

[notgitika]

What

Every project command without a TUI screen now opens a command-line-only screen instead of tearing the TUI down with an error. It says so, and shows the command's help — usage, options, arguments, and the "Parameter details" that --help appends — read from the same Commander help the CLI prints. Groups (add, export) open their menu, so every resource beneath them is reachable the same way. Esc returns to the parent menu.

The project menu lists screen-backed commands first, then a command line only divider, then the rest:

 ❯ create   create a new AgentCore project
   deploy   deploy the project to AWS
   invoke   invoke a Runtime or Harness from the current project
   build    build the project's deployable artifacts
── command line only ───────────────────────────────────────────
   add      add project resources
   export   convert project resources into editable code you own
   remove   remove a resource from the project
   dev      run the project locally for development
   status   show the status of the project's deployed resources

How

• project Router declares its screen-backed children via the existing supportedTuiCommands(...) (what eval/identity use); compile() propagates it to nested commands.
• RouterScreen partitions on isTuiCommandSupported instead of filtering, behind a showCliOnly prop. Other menus are unchanged.
• New CliOnlyScreen (Layout + KeyValueTable + ScrollView, over command.createHelp()), and one wildcard route agentcore/project/* replacing Root's hand-written list and ProjectCommandNotImplementedScreen.
• KeyValueTable wraps values under themselves and caps the key column at 50%.

dev is CLI-only here; a TUI screen for it is a follow-up.

Formerly stacked on #2172; rebased onto refactor now that it has merged, so this PR is the single feature commit plus one small fix:

• fix(project): satisfy ProjectBackend in the build/deploy screen test fake — buildDeploy.screen.test.tsx on refactor predates resolveProjectResources landing on the ProjectBackend interface, so tsc --noEmit fails on the base branch today. The fake gets the stub the interface requires.

Also resolved the CodeQL js/bad-tag-filter finding on the test regex (it is a terminal-layout assertion, not an HTML filter; made it case-insensitive so the rule is satisfied).

Testing

bun test 2848 pass; tsc, oxlint, prettier clean. Tests read the CLI-only commands off the Commander tree and assert each shows its usage and every option; plus divider order, add → gateway drill-down, scrolling to the parameter details, a small-terminal sweep, and resize-after-scroll.

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[codecov-commenter]

Codecov Report

❌ Patch coverage is 98.77301% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.15%. Comparing base (2e30e1f) to head (24a3bf6).

Files with missing lines	Patch %	Lines
src/components/CliOnlyScreen.tsx	97.97%	2 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff              @@
##           refactor    #2177      +/-   ##
============================================
+ Coverage     97.13%   97.15%   +0.01%     
============================================
  Files           541      542       +1     
  Lines         37618    37726     +108     
============================================
+ Hits          36542    36651     +109     
+ Misses         1076     1075       -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[agentcore-devx-automation]

AgentCore Harness Review

Verdict: Looks good

This is a substantial, well-structured change and I did not find any blocking issues.

Notes I looked into but don't think need changes:

• ConfirmAction immediate-trigger auto-run: the useEffect that fires run() when phase.kind === "idle" && !isPending && !error relies on setPhase({ kind: "running" }) inside run to prevent re-entry. Under React Strict Mode (double-invoked effects) this could theoretically fire run twice before the state update flushes, but Ink doesn't enable Strict Mode, so this is fine in practice.
• useProject re-keys on process.cwd(): because the query key includes from = process.cwd() read on every render, a mid-session process.chdir would produce a new query missing initialData. Screens don't chdir mid-session; the tests explicitly chdir before mount. Reasonable as-is.
• applyProgressEvent returns a fresh array for stray output events: driveProgress will still call onChange with a new (identical) task list, causing an extra setState in React consumers. Minor, not worth changing.
• Telemetry: the referenced src/cli/telemetry/README.md doesn't exist in the tree, and telemetry is wired at the CLI wrapper (withCommandRun) rather than inside handlers/screens. So these new TUI screens are on the same footing as the existing ones — no gap introduced.
• Test approach: tests use real temp directories, createRootHandler to scaffold a project, and stub only ProjectBackend (a legitimate boundary). Monkey-patching core.projectManager.listTargets to inject a failure is a targeted seam, not excessive mocking.

The refactor of runWithProgress into applyProgressEvent / settleProgress / driveProgress is a nice cleanup and the shared TaskList between CLI and TUI progress is a real win.

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[notgitika]

flaky unit test i just reran it and it passed