feat: headless api mode

[AryanNandanwar]

Refs #810

Description

Defer ink and @/app loading in cli.tsx until the interactive TUI branch so --acp (and other non-TUI paths) no longer pay the Ink/App module-graph cost at startup. Extract a framework-free parseContextLimit() for --context-max, add ACP startup metrics to the quality benchmark, refresh the baseline, and correct ACP vs legacy --vscode docs.

Related to #810 — first step toward Headless API Mode. --acp already runs the Ink-free ACP server; this PR removes the remaining CLI-level cost of resolving Ink/App before that branch is taken.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation update

Changeset

• Added a changeset (pnpm changeset) describing this change for the changelog

Docs-only or internal chores need no changeset (or run pnpm changeset --empty to note that intentionally).

Testing

Automated Tests

• New features include passing tests in .spec.ts/tsx files
• All existing tests pass (pnpm test:all completes successfully)
• Tests cover both success and error scenarios

Focused coverage used for this change:

• pnpm run test:ava source/cli.spec.ts source/app/utils/app-util.spec.ts source/app/utils/handlers/context-max-handler.spec.ts source/acp/acp-capabilities.spec.ts
• pnpm run test:types && pnpm run test:format && pnpm run test:lint
• pnpm run build && pnpm run test:benchmark (confirms acp_module_count / acp_boot_ms_approx)

Manual Testing

• Tested with Ollama
• Tested with OpenRouter
• Tested with OpenAI-compatible API
• Tested MCP integration (if applicable)

Manual smoke checks:

• node dist/cli.js --json --acp run "hi" rejects with --json cannot be combined with --acp
• node dist/cli.js --acp --provider … --model … starts the ACP server (no Ink TUI)
• ACP module-URL capture shows no dist/app/App.js / dist/app/index.js

Checklist

• Code follows project style guidelines
• Self-review completed
• Documentation updated (if needed)
• No breaking changes (or clearly documented)
• Appropriate logging added using structured logging (see CONTRIBUTING.md)

[will-lamerton]

Thanks for this @AryanNandanwar , the core change is right and the win is bigger than the description suggests. I measured the module graph before/after on the built dist:

• --acp: 612 -> 427 modules
• --plain run: 597 -> 411 modules
• copilot login: 593 -> 11 modules

So the eager Promise.all at the top of main() was costing every non-TUI path, not just ACP. Worth calling that out in the description.

A few things before this can go in:

1. pnpm test:knip fails, which is a required check. parseContextLimit at source/app/utils/handlers/context-max-handler.ts:15 is now an unused re-export once app-util.ts points at the new module. Deleting that export line should do it.

2. Please drop the benchmarks/baseline.json changes from this PR. audit_high_vulns: 2 → 17 is a security signal that shouldn't get absorbed by a startup-cost change, and the wall-clock numbers (interactive_boot_ms_approx 606 → 1415, first_render_ms_approx 549 → 1127) are machine-dependent, so committing them makes the baseline less useful for everyone else. The stale counters are worth refreshing, just as a separate chore PR.

3. This doesn't fully bypass Ink. source/custom-tools/formatter.tsx imports ink directly and is pulled in by the tool registry, so --acp and --plain still load all 54 Ink modules plus React. Could you switch Closes #810 to Refs #810 and reword the changeset to patch? Splitting the formatters out of tool registration is the follow-up that would actually close it, happy to open an issue for that.

Minor: a spec next to the new source/utils/parse-context-limit.ts would be good, right now it's only covered transitively through app-util.spec.ts.

pnpm test:types and the focused specs both pass on my end.

[AryanNandanwar]

Hey @will-lamerton , I've made the required changes. Changed the PR Description to Refs instead of closes.
I'd be happy to work on the Ink import in formatter as well