fix(compat): apply v1.0 history_length semantics to v0.3 results

[D41910]

A v0.3 agent cannot be asked for an empty history, so the compat transports have to apply the v1.0 history_length semantics to the task they return.

Background

The two protocol versions disagree on history_length = 0:

• v1.0 - 0 truncates the history to zero items; unlimited requires the field to be left unset (apply_history_length in src/a2a/utils/task.py).
• v0.3 - 0 means unlimited. The repo already encodes this rule and tests it in tests/compat/v0_3/test_proto_utils.py::test_message_send_configuration_history_length ("A v0.3 history_length of 0 means unlimited, as in task_query_params"), and proto_utils.FromProto.task_query_params / FromProto.message_send_configuration already normalise 0 to None.

The gap

The compat client transports forward the caller's history_length to the v0.3 agent verbatim and then return whatever comes back:

• rest_transport.get_task sends params['historyLength'] = request.history_length
• jsonrpc_transport.get_task / send_message send the value produced by to_compat_get_task_request / to_compat_send_message_configuration
• grpc_transport.get_task sends history_length=request.history_length

So a v1.0 caller that asks for history_length = 0 (i.e. no history) gets the entire history back from a v0.3 agent - the exact opposite of the request. history_length = N > 0 happens to work only because the v0.3 agent also does the truncation.

This is the outbound direction (v1.0 caller -> v0.3 agent). The inbound direction (v0.3 request -> v1.0 handler) is a separate fix in to_core_get_task_request / to_core_send_message_configuration, already covered by #1312, so this PR deliberately does not touch it.

Fix

Apply the existing apply_history_length helper to the converted task in get_task and send_message of the three compat transports. This is done on the result rather than on the request because v0.3 has no representation for "no history" - the transports ask for everything and then honour the v1.0 contract locally.

caller's history_length	before	after
unset	full history	full history (unchanged)
0	full history	empty history
N > 0	last N	last N (unchanged)

Tests

tests/compat/v0_3/test_history_length_compat.py covers get_task for all three transports with history_length of 0, 1 and unset, plus send_message for REST and JSON-RPC. Reverting the source change fails 8 of the 11 new tests; the 3 "unset keeps history" cases pass both ways and act as the control.

Full suite: uv run pytest -> 2286 passed, 183 skipped, 3 xfailed, 1 xpassed. The 4 errors under tests/integration/cross_version/client_server/ are environmental: they pass in isolation both with and without this change. ruff check and ruff format --check are clean.

AI disclosure

This change was produced with AI assistance and was checked against the specification and the existing tests cited above.

[github-actions]

🧪 Code Coverage (vs main)

⬇️ Download Full Report

	Base	PR	Delta
src/a2a/compat/v0_3/grpc_transport.py	63.64%	65.73%	🟢 +2.10%
src/a2a/compat/v0_3/jsonrpc_transport.py	75.23%	75.68%	🟢 +0.45%
src/a2a/compat/v0_3/proto_utils.py	90.57%	90.94%	🟢 +0.37%
src/a2a/server/cluster/database_event_stream.py	96.94%	92.86%	🔴 -4.08%
Total	93.17%	93.19%	🟢 +0.02%

Generated by coverage-comment.yml