Token efficiency: pm context agenda events re-embed full item payloads already listed in the same response (~35% of brief output) #591

Version: pm 2026.7.18

In pm context (default depth brief), every agenda.events[] entry re-embeds a near-full item payload (item_title, item_type, item_status, item_priority, item_assignee, item_deadline, item_tags) for items that are already fully listed in high_level/low_level of the same response.

Measurement (28-item workspace, toon output)

• Total pm context output: 5794 bytes; agenda section: 2032 bytes (~35%).
• All 6 agenda events referenced items already present in the same response (2 in high_level, 4 in low_level) — every embedded field was a duplicate except at/date/kind.

agenda:
  events:
    - at: "2026-08-01T08:15:37.365Z"
      date: "2026-08-01"
      kind: "deadline"
      item_id: "uxt-jw5v"
      item_title: "Rotate Stripe webhook secrets"   # duplicate of low_level row
      item_type: "Chore"                            # duplicate
      item_status: "open"                           # duplicate
      item_priority: 2                              # duplicate
      item_assignee: "eve"                          # duplicate
      item_deadline: "2026-08-01T08:15:37.365Z"     # duplicate (== at)
      item_tags: ["ops", "security"]                # duplicate

Expected vs actual

• Expected: agenda events reference items by item_id (+ at, kind, and maybe item_title for readability) when the item already appears in the same context response; full embedding only for items not otherwise present.
• Actual: unconditional full embedding — on deadline-heavy workspaces the agenda approaches a second copy of the item lists.

Why it matters for agents

pm context is the command agents are told to run first and most often; every duplicated field is context-window tax paid on each call. A dedup here cuts brief-mode output ~25-30% on realistic workspaces with deadlines, with zero information loss. Sibling of open GH-445 (repeated tags within low-level rows) — that one is within-row redundancy, this is cross-section duplication.