fix(a2a): wait for final replies before resolving RPCs

[davidrobertson]

What does this PR do?

Fixes two integration bugs in the A2A plugin branch from #41711:

1. Inbound A2A RPCs now wait for the final gateway reply before resolving the blocked JSON-RPC request.
2. A2AAdapter.connect() accepts adapter-agnostic gateway reconnect kwargs such as is_reconnect.

The important behavior change is in A2AAdapter.send(): only sends marked with metadata["notify"] == True satisfy the pending A2A reply future. Interim/status/editable sends are acknowledged locally but do not complete the RPC.

This is needed because #41711 routes inbound A2A tasks into the live Hermes gateway session, following the design direction from #11025. That is the right architecture, but the normal gateway path can emit progress/status/steering messages before the final answer. Without this guard, A2A callers can receive a compaction banner or steering notice instead of the actual task result.

Related contributor work:

• Feature: A2A (Agent-to-Agent) Protocol Support — Remote Agent Discovery, Communication & Interoperability #514 — umbrella A2A support request
• Feature Request: Remote Agent Connection / Multi-Instance Delegation #689 — remote Hermes-to-Hermes delegation use case
• feat: add A2A (Agent-to-Agent) protocol integration #4135 — earlier A2A client/server/orchestration implementation by @Ridwannurudeen
• feat: add A2A (Agent-to-Agent) protocol support #11025 — earlier live-session A2A implementation by @iamagenius00
• feat(a2a): consolidated Agent-to-Agent protocol plugin (closes #514) #41711 — active consolidated plugin implementation by @teknium1
• fix(a2a): 3 bug fixes + userContext OBO for caller identity #45996 — related A2A bug bundle by @shivasymbl
• fix(a2a): honor top-level contextId for multi-turn memory #53756 — focused contextId multi-turn fix by @knoal
• fix(a2a): hold lock across fut.done/set_result in send() to fix race #53757 — focused send() lock race fix by @knoal

This PR is intentionally based on hermes/hermes-8d223d48, the head branch for #41711, because main does not contain the A2A plugin files yet.

Related Issue

Addresses #56433 when #41711 lands.

Also related: #514, #689, #56434, #56435.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• plugins/platforms/a2a/adapter.py
  • Changed connect() to accept **_kwargs for gateway lifecycle compatibility.
  • Changed send() to ignore non-final sends while a context has a pending A2A reply future.
  • Preserved the pending-lock protection around fut.done() / fut.set_result() so this remains compatible with the race fix from fix(a2a): hold lock across fut.done/set_result in send() to fix race #53757.
• tests/plugins/test_a2a_plugin.py
  • Added regression coverage showing interim/editable sends do not resolve the pending reply future.
  • Updated the live inbound round-trip mock to send the same metadata={"notify": True} final marker the gateway uses.
  • Added coverage for connect(is_reconnect=True).

How to Test

1. Check out this branch against the feat(a2a): consolidated Agent-to-Agent protocol plugin (closes #514) #41711 A2A branch.
2. Run:

/home/hermes/.hermes/hermes-agent/venv/bin/python -m pytest -o addopts='' tests/plugins/test_a2a_plugin.py -q

Result from this branch:

.........................................                                [100%]
41 passed in 2.45s

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: Ubuntu 24.04 / Linux 6.8; branch based on feat(a2a): consolidated Agent-to-Agent protocol plugin (closes #514) #41711

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — N/A
• I've updated cli-config.yaml.example if I added/changed config keys — N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide
• I've updated tool descriptions/schemas if I changed tool behavior — N/A

Screenshots / Logs

Focused A2A plugin suite:

/home/hermes/.hermes/hermes-agent/venv/bin/python -m pytest -o addopts='' tests/plugins/test_a2a_plugin.py -q
.........................................                                [100%]
41 passed in 2.45s

[kuangmi-bit]

Small, focused fix. Two issues addressed:

1. RPC reply timing - waiting for the final gateway reply before resolving is correct. Interim/status sends should not prematurely resolve the A2A request. This matches the A2A protocol expectation that the response carries the final state.

2. Reconnect kwargs passthrough - is_reconnect and similar adapter-agnostic kwargs need to flow through to connect(). Good catch for gateway restart scenarios.

80 additions, 2 files, no new deps. LGTM.

[teknium1]

Thanks for the focused stacked fix. The gateway contract supports its approach: final streaming sends receive metadata["notify"] = True in gateway/stream_consumer.py:246, while reconnect plumbing invokes adapters as connect(is_reconnect=...) in gateway/run.py:3500.

Current main has no tracked plugins/platforms/a2a/ implementation, so this behavior cannot be independently reproduced against HEAD; the PR depends on the unmerged #41711 branch whose head is this commit's parent (2c317416).

Automated hermes-sweeper review.

[teknium1]

Merged to main via #77109 with your commit cherry-picked intact (authorship preserved) — the notify-gated final-reply capture is now part of the landed A2A plugin. Thanks @davidrobertson!