feat: support Claude Code 'auto' permission mode

[tiann]

Closes #858

What

Adds Claude Code's auto mode (--permission-mode auto) as a first-class HAPI permission mode for claude-flavored sessions. Auto mode runs without routine permission prompts while a classifier blocks anything irreversible, destructive, or aimed outside the environment — a safer alternative to Yolo (bypassPermissions).

How

auto differs from acceptEdits/bypassPermissions in one important way: those are emulated by HAPI inside its canCallTool handler, while auto (like plan) is enforced inside Claude itself by the classifier. That drives the two non-mechanical changes:

• shared/src/modes.ts — add auto to CLAUDE_PERMISSION_MODES and PERMISSION_MODES (label Auto, tone warning). CLI flag parsing, hub validation, runner spawn args, and the web mode dropdown all derive from these lists, so they pick it up automatically.
• cli/src/claude/runClaude.ts — the message-queue batch hash previously only tracked isPlan. It now treats auto the same way, so switching to/from auto mid-session starts a new Claude process with the matching --permission-mode flag (otherwise the switch would silently not take effect until the next respawn).
• cli/src/claude/utils/permissionHandler.ts — allow auto as a plan-exit target mode, matching Claude Code's own "Approve and start in auto mode" plan-approval option.
• web/src/api/client.ts — extend the approvePermission mode union.

Permission escalations in auto mode (classifier blocks, explicit ask rules) still arrive through the existing --permission-prompt-tool stdio → canCallTool path and surface in the web UI as normal — no local auto-approval is added for auto, since Claude's classifier is the decision-maker.

Notes

• Requires Claude Code ≥ 2.1.83 and an auto-mode-eligible account; on older binaries claude rejects the flag at spawn and the error surfaces in the session, same as any unsupported flag.
• auto stays claude-only: codex/gemini/cursor/kimi/opencode keep their own mode lists and validation rejects it for them (covered by the new shared test).
• dontAsk mode is intentionally out of scope.

Testing

• bun typecheck — clean (cli/web/hub)
• bun run test — cli, hub, web, shared all pass (one pre-existing apiMachine.test.ts failure on macOS, reproduces on clean main, unrelated)
• New test: shared/src/modes.test.ts (auto is claude-only + label/tone)

[cloudflare-workers-and-pages]

Deploying hapi with    Cloudflare Pages

Latest commit:	b64905b
Status:	✅  Deploy successful!
Preview URL:	https://b2f4dfd4.hapi-bqd.pages.dev
Branch Preview URL:	https://feat-claude-auto-mode.hapi-bqd.pages.dev

View logs

[github-actions]

Findings

• No high-confidence findings on the added/modified lines.

Summary
Review mode: initial
No correctness, security, regression, data-loss, performance, or maintainability issues found in the reviewed diff. Residual risk: I did not execute the suite; the most relevant remaining gap is no direct test for the Claude plan-exit fake restart targeting auto.

Testing

• Not run (automation review only)

HAPI Bot