Extend `serena-hooks auto-approve` to Claude Code's `auto` permission mode

[scarson]

Is your feature request related to a problem? Please describe.

serena-hooks auto-approve (introduced in v1.1, tracked in #920) only emits an allow decision when Claude Code reports permission_mode == "acceptEdits". It is silent in every other documented mode.

Claude Code's hooks docs list six values for permission_mode:

permission_mode: Current permission mode: "default", "plan", "acceptEdits", "auto", "dontAsk", or "bypassPermissions".

The most user-visible gap is auto mode. Claude Code docs describe it as:

auto: Auto-approves tool calls with background safety checks that verify actions align with your request. Currently a research preview.

Auto mode is the mode users adopt when they want hands-off autonomous execution — the same use case that motivated #920. In auto today, Serena's MCP tool calls are not auto-approved by Serena's hook; they fall through to Claude Code's auto-mode classifier, which makes per-call probabilistic decisions rather than the deterministic "the user chose a permissive mode, extend it to Serena" semantic that #920 established.

The behavior is controlled by a single literal string check at src/serena/hooks.py:371:

_ACCEPT_EDITS_MODE = "acceptEdits"

def is_accept_edits_mode(self) -> bool:
    return self._permission_mode == self._ACCEPT_EDITS_MODE

Reproduction

Verified against serena-agent==1.1.2 and main @ 14cd3ef:

for mode in acceptEdits auto bypassPermissions default plan ""; do
  echo "=== mode: '$mode' ==="
  echo "{\"session_id\":\"t\",\"tool_name\":\"mcp__serena__replace_symbol_body\",\"permission_mode\":\"$mode\",\"tool_input\":{}}" \
    | serena-hooks auto-approve --client=claude-code
done

Only acceptEdits emits a decision:

=== mode: 'acceptEdits' ===
{"hookSpecificOutput": {"hookEventName": "PreToolUse", "permissionDecision": "allow", "permissionDecisionReason": "Auto-approved: Serena tool call while client is in acceptEdits mode.", "additionalContext": ""}}
=== mode: 'auto' ===
=== mode: 'bypassPermissions' ===
=== mode: 'default' ===
=== mode: 'plan' ===
=== mode: '' ===

All non-acceptEdits modes produce no output.

Describe the solution you'd like

Extend PreToolUseAutoApproveSerenaHook so that Claude Code's auto mode is treated equivalently to acceptEdits for the purposes of Serena auto-approval. The minimal change is adding "auto" to the accepted set alongside "acceptEdits".

How to make that configurable is the maintainers' call. Two shapes I can see working:

1. Hard-code the expanded set. Accept acceptEdits and auto by default. Simple, matches the Make Serena respect Claude Code's accept edits mode #920 pattern.
2. --modes option on the auto-approve subcommand. Let users opt into any subset via something like serena-hooks auto-approve --client=claude-code --modes=acceptEdits,auto. Preserves today's default for existing users while letting new adopters opt into auto.

Happy to write the PR once direction is confirmed — per CONTRIBUTING.md I'm raising this as an issue first.

Describe alternatives you've considered

1. User adds Serena tools to autoMode.allow. Claude Code's auto mode has its own classifier with a prose-based allow list users can extend. In principle this works, but it's fragile (tool names evolve), per-tool (multiple entries), and crosses the boundary Make Serena respect Claude Code's accept edits mode #920 was built to draw — Serena auto-approval should track the user's permission mode choice, not require maintaining a parallel allowlist.
2. Shell wrapper around the hook. A few lines of bash in ~/.claude/settings.json can check permission_mode and shell out to serena-hooks auto-approve. Works today, but duplicates logic every Serena user has to author.
3. Fix upstream in Claude Code. Make Serena respect Claude Code's accept edits mode #920's thread flagged that ideally Claude Code would natively extend permissive modes to MCP edit tools. That's the cleanest long-term fix, but Make Serena respect Claude Code's accept edits mode #920 established the precedent that Serena provides the bridge today, and this issue asks to extend the existing bridge to cover the newer mode.

Additional context

• Verified against serena-agent 1.1.2 (installed via uv tool install -p 3.13 serena-agent@latest --prerelease=allow) and main at commit 14cd3ef.
• Claude Code's auto is labeled "research preview" in docs; the permission_mode hook field that exposes it is stable and canonically documented, so coupling to the field is low risk even if auto-mode semantics evolve.
• bypassPermissions and dontAsk are also silent in the current implementation. Those modes already suppress prompts at the Claude Code level, so the hook's silence is architecturally harmless — just noting for completeness in case maintainers want a while-you're-there fix.
• Relates to / follow-up on Make Serena respect Claude Code's accept edits mode #920.

[MischaPanch]

This makes sense, pls issue a PR with the hardcoded extended set. Hooks are an alpha feature, we don't need to overcomplicate this. Pls also add an entry to the changelog