Surface new Claude Code JSONL event types (cache misses, queue ops, mode/tool/skill/MCP deltas)

[phiat]

Summary

Claude Code 2.1.150+ emits several JSONL event types we silently dropped. This PR adds parser + renderer coverage for the high-signal ones.

New stream markers:

• ── cache miss: tools_changed, +51k tokens ── — surfaces assistant.diagnostics.cache_miss_reason. Directly explains why per-agent context% spikes unexpectedly (e.g. ToolSearch loading a new tool busts the prompt cache, forcing tens of thousands of tokens to be re-sent).
• ── queued: <prompt> ── / ── dequeued ── — top-level queue-operation events when the user types a follow-up prompt while CC is still working.
• ── plan mode exit ──, ── auto mode ── — attachment subtypes for mode transitions.
• ── tools: +28 ──, ── MCP: +1 ──, ── skills: 50 total ── — deferred_tools_delta / mcp_instructions_delta / skill_listing deltas.

Suppression rules:

• skill_listing.isInitial=true (session-start bulk dump) — dropped.
• *_delta with empty added/removed/readded — dropped.
• queue-operation enqueues whose content is a <task-notification> (TaskStop result re-injection) — dropped; not a user-typed prompt.

Implementation:

• New TypeCacheMiss and TypeSessionEvent (umbrella marker) types in parser.
• AssistantMessage.Diagnostics added; emits an extra TypeCacheMiss item alongside the assistant message that bust the cache.
• parseAttachment extended for the four new attachment subtypes.
• ParseLine extended for top-level queue-operation.
• Renderer adds two early-return branches mirroring the existing TypeCompactMarker / TypePRLink pattern.
• One-line content snippet helper (oneLineSnippet) keeps long queued prompts from breaking the divider layout.

Deferred: permission-mode change detection. CC re-emits this on every prompt, so without per-session dedup it floods the stream. Will land as a separate change since it needs state outside the stateless parser.

Tests: 13 new parser tests covering each handler + each suppression case. Smoke-tested against real 2.1.150 sessions — produces 19–20 session-event markers and 3 cache-miss markers per ~1000-line session, proportionate to actual user activity.

Test plan

• go build ./... && go test ./... && go vet ./... all clean
• Run claude-esp against a fresh CC 2.1.150 session and observe cache-miss markers appear after a ToolSearch / MCP load
• Run against a session with user-queued follow-up prompts; verify queued: markers appear
• Verify session-start does NOT show a giant skill_listing marker
• Verify a long queued paragraph renders as a single-line ellipsised marker

🤖 Generated with Claude Code