v1.4 — External Reference Gap Closure (Q3 2026, planned)

Cross-analysis of three reference projects in analysis/ — DeepFlow, Harnessed-LLM-Agent, LangGraph — to find what the orchestrator still lacks. After verifying every candidate against src/, the previous draft was 80 % wrong: most "gaps" turned out to already be on main. This page lists only what grep + UI inspection confirmed is genuinely missing.

TL;DR

Metric	Today	After v1.4
Reference matrix coverage	18 ✅ + 1 ⚠ + 0 ❌ (~95 %)	19 ✅ (target)
LangGraph stream modes	2 of 7 (events, values)	7 of 7
Personalized Memory	REST only	REST + UI
Cross-protocol agent peers	dict-based + typed	+ A2A bidirectional bridge

Confirmed shipped — do NOT re-do

For traceability — these were on the original gap draft but already live on main:

Candidate	Where it lives now
Loop detection middleware	core/loop_detection.py (+ tests/test_loop_detection.py)
Sandboxed code execution	core/sandbox.py, skills/sandboxed_shell.py
Progressive skill loading	skills/skill_loader.py
Structured clarification protocol	core/clarification.py, skills/clarification_skill.py
Dangling tool-call recovery	core/tool_recovery.py::recover_dangling_tool_calls
Typed channels (LastValue / Topic / …)	core/channels.py
Per-node CachePolicy	core/cache.py
Skill middleware chain	core/skill.py (+ tests/test_skill_middleware.py)
RetryPolicy + circuit breaker	core/resilience.py::RetryPolicy, CircuitBreaker, resilient_call
Conformance test suite	core/conformance.py
tool_description on tool calls	core/skill.py (SkillRequest.metadata["tool_description"])
Configurable summarisation triggers	core/conversation.py::SummarizationConfig (TOKEN / MESSAGE / FRACTION + retain_last)
Config versioning + auto-upgrade	core/yaml_config.py (config_version, _upgrade_v0_to_v1, CURRENT_CONFIG_VERSION)
Memory upload filter	core/memory_filter.py::MemoryFilter

Genuinely missing — the v1.4 scope

P1 — Must-have

A2A adapter (un-park P5b)

The single ⚠ in the harnessed-LLM-agent 19-item match matrix. Builds on top of core/cooperation_messages.py (already in main).

Goal: bidirectional bridge with Google's A2A protocol so external A2A agents appear as local cooperation peers — and vice-versa.

Why now: spec churn was the reason for the Q1 park; the April-2026 stabilisation has landed.

Where it will live: core/cooperation/a2a_adapter.py, plus a discovery endpoint under /api/a2a/.

Managed values (computed state)

LangGraph-style read-only injections into node state: step_count, remaining_steps, interrupt_ids, is_final_step. Computed at runtime by the engine, never checkpointed. Lets nodes self-throttle without polluting State schemas.

# After v1.4
async def writer_node(state, *, remaining_steps):
    if remaining_steps < 2:
        return {"messages": [summarise_and_finish(state)]}
    return {"messages": [continue_drafting(state)]}

Where it will live: new core/managed_values.py + Pregel-loop hook in core/graph.py.

P2 — Nice-to-have

Personalized Memory dashboard UI

REST endpoints /api/user-memory/users/* exist since v1.3.0 P4, but no React page consumes them (verified — frontend/src/components/ has no user-memory references).

Build: frontend/src/components/memory/UserMemoryPanel.tsx that:

• Lists per-user keys with values
• Lets the user edit / delete entries (REST PUT / DELETE)
• Shows the last <user_profile> block injected into the system prompt
• GDPR wipe button → DELETE /api/user-memory/users/{user_id}

Granular stream modes (close the gap to LangGraph 7)

dashboard/sse.py exposes 2 of LangGraph's 7 modes today: events (default) and values. Add the 5 missing:

Mode	Frame contents
updates	per-node delta only (smaller payloads)
messages	LLM message stream only
tasks	task lifecycle events only
debug	verbose internal — node enter/exit, channel writes
custom	user-emitted via emit("event", payload)

Selectable per WebSocket subscription. Reduces frontend filtering load and matches LangGraph SDK consumers.

P3 — Optional

Content-addressed checkpoint blobs

In core/store_postgres.py / PostgresCheckpointer, split large state values into a blobs(sha256 PRIMARY KEY, payload BYTEA) table and reference by hash. Repeated values across checkpoints share a single row.

Expected impact: 5-20× storage reduction on long conversations where the same RAG context or system prompt repeats.

Structured deprecation annotations

@deprecated(since="1.4", removed_in="1.6", migration="use X instead")
def old_api(...): ...

Emits a DeprecationWarning at call sites + powers a docs/deprecations.md page generated at release time. Currently zero @deprecated / DeprecationWarning hits in core/. Pre-requisite for any 2.x cleanup pass.

Sprint shape

Worktree	Owner	Item
1	backend	A2A adapter (largest)
2	architect	Managed values (Pregel-loop change — needs careful design)
3	backend	Content-addressed blobs + deprecation decorator (both small, same dev)
4	frontend	Personalized Memory UI + 5 missing stream-modes UI selector

Match-matrix target: 19/19 ✅ by end of sprint.

What this page replaces

• The earlier analysis/ROADMAP.md v1.4 draft contained 16 items; 10 turned out to be already shipped after a grep pass against src/.
• The verified-shipped table at the top of this page is the audit trail of that pruning so a future agent doesn't rewind it.