fix: read real Claude Code hook payload fields

[ShabanK]

Summary

Nudgy stopped reacting to Claude Code events because it branched on event.matcher. Claude Code never sends matcher in hook input — it is a settings-side filter used to decide whether a hook runs, not a field echoed back in the payload. event.matcher was therefore always nil, and every decision keyed off it collapsed.

The most visible symptom: permission prompts and idle prompts never set waitingPermission / waitingInput, so they fell through to the info style, which is disabled by default and silently dropped.

How this was diagnosed

Rather than reading the code alone, I pointed real Claude Code hooks (v2.1.235) at a logging HTTP endpoint and captured actual payloads. Notification and PermissionRequest don't fire in headless -p mode, so an interactive PTY session was driven to trigger a genuine permission dialog.

A/B against the built binary using a byte-identical real payload:

BEFORE: EVENT Notification | matcher=-                 -> Notification disabled for style: info
AFTER:  EVENT Notification | detail=permission_prompt  -> Showing warning notification | state=waitingPermission

What was wrong

Event	Field Claude Code actually sends	What Nudgy read
Notification	notification_type	matcher
StopFailure	error / error_details	matcher
SessionStart	source	matcher
SessionEnd	reason	matcher

Consequences:

• Notification — waitingPermission and waitingInput were never set. Permission and idle prompts were silently swallowed. This is the core "no longer working" bug.
• StopFailure — max_output_tokens was never detected, so every stop failure was reported as an error.
• Error popups — always showed the generic "Something went wrong" fallback.

Notably, notificationType was already correctly decoded on the model and simply never referenced anywhere in Sources/.

Changes

• HookEvent — decode the real fields: error, error_details, source, reason, last_assistant_message. Added a discriminator accessor since Claude Code names the sub-type field differently per event. matcher is retained for decode compatibility only, with a comment explaining why it must not be branched on.
• SessionManager — Notification now switches on notification_type (permission_prompt, idle_prompt, agent_needs_input, agent_completed, elicitation dialogs); StopFailure switches on error.
• AppDelegate — error text now derives from error_details / last_assistant_message / a readable mapping of error. Added a log line recording the notification style and resulting session state, which is what made the A/B above observable.
• HookInstaller — dropped the SessionStart hook. That event only supports command and mcp_tool handlers, so the installed http handler was never invoked. Verified directly: a command probe fired on SessionStart while an http probe on the same event did not. It is now cleaned up as a legacy entry, and sessions are created lazily on their first real event. SessionEnd gained an explicit timeout, since it otherwise shares a 1.5s budget.
• Logger — matcher: parameter renamed to detail: to match reality.

Why this shipped broken

The tests hand-wrote a matcher field that Claude Code never emits, so the suite stayed green while the app was broken in production. Fixtures in HookEventTests.swift are now verbatim captures from a real Claude Code session, with explicit XCTAssertNil(event.matcher) regression guards. Test helpers build events the way Claude Code actually sends them.

Verification

Each state transition was confirmed end-to-end against the running binary:

Payload	Resulting state	Style
notification_type: permission_prompt	waitingPermission	warning
notification_type: idle_prompt	waitingInput	question
error: rate_limit	error	error
error: overloaded	error	error
error: max_output_tokens	idle	success
Stop	idle	success

Installer behaviour was also verified against a real ~/.claude/settings.json: a seeded legacy SessionStart entry is removed, SessionStart is not reinstalled, and unrelated user hooks are preserved.

Note: swift test could not be run locally — the machine has only Command Line Tools, so XCTest is unavailable. This is a pre-existing environment limitation unrelated to these changes, and is why verification was done against the running binary. CI should exercise the suite.

Not addressed (follow-ups)

Deliberately kept out of scope to keep this reviewable:

1. TranscriptParser reconstructs the transcript path from cwd by replacing / with -, but Claude also replaces .. For /Users/me/.config/nvim it computes -Users-me-.config-nvim while the real directory is -Users-me--config-nvim. The file isn't found, parseUsage returns nil, and token usage silently reads zero. Every hook payload already carries transcript_path directly, making the slug reconstruction unnecessary. This is silent data corruption and is probably the highest-value next fix.
2. HTTPServer caps payloads at 64KB and returns 413 above that. Stop payloads now include last_assistant_message, background_tasks[] and session_crons[], so a long response can exceed the cap and the completion notification is dropped — intermittently, on exactly the longest turns.
3. {"status":"ok"} response body is not a valid hook output schema; the docs specify such a body is treated as a non-blocking error. An empty 200/204 is the correct "no decision" response. Matters most on PermissionRequest.