Add agent team tracing support

[sufjanfana]

Add agent team tracing support

• Add full tracing support for Claude Code agent teams (lead, teammates, subagents) with dedup, lifecycle tracking, and deferred turn emission
• Rewrite token counting from per-line jq loops to a single parse_transcript pipeline with requestId-based dedup
• Correct OpenInference span kind values and attribute names
• Add two new hooks: TeammateIdle and TaskCompleted (11 hooks total)

What changed

Team-aware Turn lifecycle (stop.sh, common.sh, user_prompt_submit.sh, session_end.sh)

• When a team is active, stop.sh defers the Turn span instead of emitting it. close_active_turn() in common.sh handles emission when the next prompt arrives or the session ends.
• last_trace_id / last_trace_span_id are preserved after Turn emission so late-arriving teammate hooks (SubagentStop, TaskCompleted) can find their parent trace.

Agent grouping spans (post_tool_use.sh, subagent_stop.sh)

• post_tool_use.sh lazy-inits an AGENT grouping span per agent_id and re-parents tool spans under it. Handles TeamCreate/TeamDelete detection, shutdown_response grouping, and a race fix that preemptively marks agents active on Agent/SendMessage calls.
• subagent_stop.sh reuses the pre-created grouping span when the agent used tools, or creates one if it didn't. Teammate vs subagent distinction for span naming and start-time estimation. Transcript offset tracking avoids re-parsing for teammates with multiple work periods.
• A sentinel (agent_<id>_shutdown_complete) prevents duplicate grouping spans between the two hooks.

Team dedup (common.sh)

• check_team_dedup(): teammate marks first, lead checks and skips. Prevents duplicate spans when both fire for the same event. Used in post_tool_use and task_completed.

New hooks

• teammate_idle.sh — Caches team_name in session state (no span emitted)
• task_completed.sh — Emits CHAIN span with task metadata, team context, and dedup
• plugin.json — Registers both hooks, updates description to 11 hooks

Token counting rewrite (common.sh, stop.sh, subagent_stop.sh)

• New parse_transcript() replaces per-line while loops with a single jq -rsc pipeline (O(N) jq invocations → O(1))
• Deduplicates streaming token counts by requestId — only the final cumulative count per group is used
• Includes cache_read_input_tokens and cache_creation_input_tokens in totals

Span kind / attribute corrections (breaking)

• Uppercased OpenInference span kinds: "tool" → "TOOL", "chain" → "CHAIN", "LLM" → "AGENT"
• Renamed subagent attributes: subagent.id → agent.id, subagent.type → agent.name
• Added agent.role (lead/teammate/subagent) and team.name to all span types
• Phoenix REST path now reads span_kind from attributes instead of hardcoding "CHAIN"
• Phoenix attribute conversion includes doubleValue

Performance improvements

• get_timestamp_ms: tries date +%s%3N → perl -MTime::HiRes → python3 → date +%s000 (avoids python3 on most systems)
• generate_trace_id / generate_span_id helpers replace repeated inline pipelines
• Batched state reads in post_tool_use.sh (single jq call for 8+ fields)
• del_states() replaces sequential del_state calls
• get_or_set_state() atomic check-then-set in one lock
• init_state idempotency guard
• send_span accepts span name to skip jq parse for logging
• get_state / del_state / inc_state use --arg instead of string interpolation

Other changes

• build_span JSON-escapes span names (handles quotes, backslashes, tabs)
• Output truncation raised from 5KB to 10KB; prompt capture raised from 1KB to 10KB
• session_start.sh hard-resets state file to prevent stale state
• New tool cases in post_tool_use: Agent, SendMessage, TaskCreate/TaskUpdate, ToolSearch
• notification.sh filters idle_prompt notifications
• jq -n → jq -nc across all hooks
• jq error URL updated to platform-agnostic link

Breaking changes

Existing Arize/Phoenix dashboards that filter on openinference.span.kind values ("tool", "chain", "LLM") or subagent.* attributes will need to be updated to the new uppercase/renamed values.

Test plan

• Single-agent session: verify Turn, tool, and subagent spans emit correctly (no regression)
• Team session: verify lead Turn is deferred until next prompt or session end
• Team session: verify teammate tool spans are deduped (no duplicates from lead)
• Team session: verify AGENT grouping spans appear for each teammate
• Verify TeammateIdle caches team_name without emitting a span
• Verify TaskCompleted emits a CHAIN span with correct parent
• Phoenix: verify span_kind reflects actual span kind (TOOL, AGENT, CHAIN) not hardcoded CHAIN
• Dry run mode: ARIZE_DRY_RUN=true prints correct span structure
• Python Agent SDK: lazy init still works (SessionStart doesn't fire)

🤖 Generated with Claude Code