Wave 2 features: A2A live view (CEO chime-in + reply budget) and prompter memory (#297)

* feat(a2a): live view — watch fleet conversations, CEO chime-in, reply budget

A2A_MESSAGE_SENT published from A2AService.send (excerpt-capped) and
fanned through the existing /ws/system bridge; CEO-only admin REST for
conversations/messages + a reply route on the publish-bearing send path;
panel /a2a page with live transcript and a composer gated on task-linked
conversations. The matrix gains its one asymmetric rule: CEO may message
anyone, nobody may target the CEO — and agent replies inside a
CEO-opened conversation are hard-budgeted to one per CEO message
(per conversation, per agent), rejected with wait-don't-retry guidance.
Built subagent-driven (Sonnet 5), reviewed; v1 seams documented in the
map delta.

* feat(prompter): intake remembers the task history

Intake spawns now carry a per-project chronological digest of recent
tasks (capped: 15 lines/project, 4000 chars total — ~300-1000 tokens)
merged into the ambient layer, and the interviewer gets a bounded
search_past_tasks tool (one shared implementation behind the grok MCP
tool and the Claude SDK in-process tool) to check precedent
mid-conversation. Informational memory only — the sequencing analyzer
keeps ownership of ordering. Built subagent-driven (Sonnet 5), reviewed;
pre-existing conventions-ambient MegaTask-scope gap flagged, untouched.

---------

Co-authored-by: Renn F <rennf93@users.noreply.github.com>

docs/map/_front.md

@@ -602,3 +602,13 @@ Slices touched: orchestrator (1), tests (1, 2), taskservice + api-routes-schemas
 1. **Five dead comms panel components deleted** (−422 lines; audit-verified zero consumers; MessageComposer/MessageTypeBadge stay).
 2. **e2e scenario 4 (MegaTask umbrella)** — `tests/e2e_smoke/test_megatask_umbrella.py` + arcs helpers (`wire_dependency` via the real sequencing edge, `seed_cell_and_dev`, `set_branch_name`). Proves: sequencing hold (`unmet_dependency` on RS2's i_will_plan), serial root merges to master, umbrella branchless close via ceo-approve (never approve-and-merge).
 3. **PRODUCT FIX: batch root-subtask completion wall** — `_main_pm_complete_guard` (_impl.py ~6708) + `escalate_to_ceo` (task.py ~5321) refused ALL parented tasks; both now consult `is_batch_root_subtask`. Live root-subtasks previously closed only via CEO god-mode. Regression tests in test_choreographer_pm.py + test_task_service_transitions.py. First product bug found BY the harness (subagent-built, Sonnet 5, reviewed).
+
+---
+## Delta 2026-07-03 (2) — A2A live view (branch `feat/wave-2b`, SDD/Sonnet 5, reviewed)
+
+Backend: `EventType.A2A_MESSAGE_SENT` published from `A2AService.send` (excerpt-capped payload), `websocket_bridge` forwarder → `/ws/system` `a2a.message` frames; admin REST (`/a2a/chat/admin/conversations{,/{id}/messages}`, CEO-only) + CEO reply route via the publish-bearing `send` path. `agents_config.can_a2a_direct`: `from ceo` → allowed to anyone (the one asymmetry); `to ceo` stays blocked at creation. **Reply budget** (`_enforce_ceo_reply_budget` in `send_chat_message`): agent msgs ≥ CEO msgs in the conversation ⇒ reject; three-layer enforcement (matrix block → reply-channel resolve → budget). Panel: `/a2a` page + components, `use-a2a-live` hook over `useWebSocket("/system")`, composer gated on task-linked conversations. v1 seams flagged: legacy raw REST chat sends don't publish the live event; agent→CEO reply lookup matches topic-less conversations only; deduped re-send re-publishes an identical frame (idempotent consumer).
+
+---
+## Delta 2026-07-03 (3) — prompter memory (branch `feat/wave-2b`, SDD/Sonnet 5, reviewed)
+
+`TaskService.list_recent_for_project` (recency = coalesce(completed,updated,created)); pure digest builders in `prompter.py` (15 lines/project, 70-char titles, 4000-char total cap); orchestrator `_resolve_history_digest_ambient` (+`_resolve_intake_ambient` merge, best-effort/non-blocking) injected at `_spawn_intake_container` → `_generate_composed_prompt(ambient=…)`; `GET /prompter/live/{id}/search-tasks` (session-liveness = trust boundary, q 2–200, limit ≤10); `query_past_tasks`/`format_search_results` shared by the grok MCP tool AND the Claude SDK in-process tool (full parity, one implementation). FLAGGED pre-existing gap (untouched): `_resolve_conventions_ambient` doesn't cover the MegaTask `project_ids` scope — conventions ambient absent on MegaTask intakes.