feat(workflows): display workflow agent cohort in statusline

[tmck-code]

Context

The Workflow tool runs multi-agent orchestrations whose subagents are invisible to the statusline. yas globs subagents/*.meta.json, but workflow agents live a directory deeper (subagents/workflows/<runId>/) and their meta.json carries only {"agentType":"workflow-subagent"} — no description, no per-agent type. A user running a workflow sees nothing, even while several agents are actively burning tokens. This change surfaces those agents, grouped by their workflow run, so the statusline reflects what is actually executing.

Spec: openspec/changes/display-workflow-agents/ (new workflow-cohort capability).

Changes

• Detection (claude/yas/info/workflows.py, new): glob subagents/workflows/<runId>/ dirs under the session — a dir exists the instant an agent starts. Opportunistically read the session-level workflows/<runId>.json to enrich a run with workflowName, current phase, and the agentId → label map; fall back to the transcript's first prompt line and the runId when JSON is absent/stale, so detection never depends on JSON liveness.
• Token/activity reuse: every workflow agent reuses the existing _parse_transcript for tokens, activity snippet, and Done detection (end_ts > 0) — no new per-agent reader.
• Liveness: workflow runs get their own ~120s window (separate from the cohort's 30/60s) so a run survives between-phase lulls without flickering.
• Rendering (renderer.py, layout.py, constants.py): each live run renders as a distinct grouped block after the subagent cohort and task row — header (▸ <name> [<phase>]), per-agent rows reusing subagent_row, summary footer (└ N agents · M done · <tok>). Width <80 collapses to header+summary; per-run rows cap at 6 (overflow → summary); at most 2 workflow blocks render concurrently.
• SessionView gains a @cached_property constructing RunningWorkflows.from_session(...) (info/__init__.py).
• CONTEXT.md glossary gains the workflow-run terms; ops/demo.py gains demo coverage; test/test_workflow_cohort.py (new, 22 tests) covers detection, label fallback, liveness/retirement, narrow collapse, and the agent/run caps.

Checklist

• Tests added or updated for new/changed behaviour

Screenshots / recording

Benchmark

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
main	53.2 ± 3.0	49.3	66.8	1.00
PR	54.0 ± 2.7	50.0	66.2	1.01 ± 0.08

System info

Key	Value
OS	Linux debian-work 6.12.90+deb13.1-amd64 #1 SMP PREEMPT_DYNAMIC Debian 6.12.90-2 (2026-05-27) x86_64 GNU/Linux
Claude Code	2.1.170 (Claude Code)
Terminal	TERM=tmux-256color TERM_PROGRAM=tmux SHELL=/bin/bash COLORTERM=truecolor
Locale	LANG=en_AU.UTF-8 LC_ALL=
Python	Python 3.12.7
uv	uv 0.9.17

🤖 Generated with Claude Code