fix(runtime): bound agent output projections

[likun666661]

What changed

• add explicit agent_output views so callers do not receive AgentRun and RuntimeEvent streams together by default
• enforce a 32 KiB default / 128 KiB maximum serialized payload budget across events, runtime events, diagnostics, and artifact references
• expose truncation and budget diagnostics in the tool result
• change the Graph supervisor prompt from max_events: 100 to a bounded 20-event, 32 KiB runtime-event inspection

Why

Issue #1593 reproduced Graph supervisor context overflow after four agent_output calls returned roughly 479–572 KiB each. The tool only bounded event count, not serialized size, and duplicated two event projections into the root session.

This is PR 1 of 4 in the Graph context-overflow repair stack.

Impact

Raw child execution history remains durable and inspectable through explicit views, while normal Graph supervision receives a bounded semantic projection.

Validation

• npm --workspace @maka/runtime run typecheck
• npm --workspace @maka/runtime run build
• targeted SessionManager agent-output tests: 4/4
• Graph/subagent tool tests: 27/27
• git diff --check

The full SessionManager file also exposed an unrelated pre-existing timing-sensitive test failure; the affected agent-output subset passes independently.

[likun666661]

What changed

• add explicit agent_output views so callers do not receive AgentRun and RuntimeEvent streams together by default
• enforce a 32 KiB default / 128 KiB maximum serialized payload budget across events, runtime events, diagnostics, and artifact references
• expose truncation and budget diagnostics in the tool result
• change the Graph supervisor prompt from max_events: 100 to a bounded 20-event, 32 KiB runtime-event inspection

Why

Issue #1593 reproduced Graph supervisor context overflow after four agent_output calls returned roughly 479–572 KiB each. The tool only bounded event count, not serialized size, and duplicated two event projections into the root session.

This is PR 1 of 4 in the Graph context-overflow repair stack.

Impact

Raw child execution history remains durable and inspectable through explicit views, while normal Graph supervision receives a bounded semantic projection.

Validation

• npm --workspace @maka/runtime run typecheck
• npm --workspace @maka/runtime run build
• targeted SessionManager agent-output tests: 4/4
• Graph/subagent tool tests: 27/27
• git diff --check

The full SessionManager file also exposed an unrelated pre-existing timing-sensitive test failure; the affected agent-output subset passes independently.

[likun666661]

What changed

• add explicit agent_output views so callers do not receive AgentRun and RuntimeEvent streams together by default
• enforce a 32 KiB default / 128 KiB maximum serialized payload budget across events, runtime events, diagnostics, and artifact references
• expose truncation and budget diagnostics in the tool result
• change the Graph supervisor prompt from max_events: 100 to a bounded 20-event, 32 KiB runtime-event inspection

Why

Issue #1593 reproduced Graph supervisor context overflow after four agent_output calls returned roughly 479–572 KiB each. The tool only bounded event count, not serialized size, and duplicated two event projections into the root session.

This is PR 1 of 4 in the Graph context-overflow repair stack.

Impact

Raw child execution history remains durable and inspectable through explicit views, while normal Graph supervision receives a bounded semantic projection.

Validation

• npm --workspace @maka/runtime run typecheck
• npm --workspace @maka/runtime run build
• targeted SessionManager agent-output tests: 4/4
• Graph/subagent tool tests: 27/27
• git diff --check

The full SessionManager file also exposed an unrelated pre-existing timing-sensitive test failure; the affected agent-output subset passes independently.

[likun666661]

What changed

• add explicit agent_output views so callers do not receive AgentRun and RuntimeEvent streams together by default
• enforce a 32 KiB default / 128 KiB maximum serialized payload budget across events, runtime events, diagnostics, and artifact references
• expose truncation and budget diagnostics in the tool result
• change the Graph supervisor prompt from max_events: 100 to a bounded 20-event, 32 KiB runtime-event inspection

Why

Issue #1593 reproduced Graph supervisor context overflow after four agent_output calls returned roughly 479–572 KiB each. The tool only bounded event count, not serialized size, and duplicated two event projections into the root session.

This is PR 1 of 4 in the Graph context-overflow repair stack.

Impact

Raw child execution history remains durable and inspectable through explicit views, while normal Graph supervision receives a bounded semantic projection.

Validation

• npm --workspace @maka/runtime run typecheck
• npm --workspace @maka/runtime run build
• targeted SessionManager agent-output tests: 4/4
• Graph/subagent tool tests: 27/27
• git diff --check

The full SessionManager file also exposed an unrelated pre-existing timing-sensitive test failure; the affected agent-output subset passes independently.

[likun666661]

What changed

• add explicit agent_output views so callers do not receive AgentRun and RuntimeEvent streams together by default
• enforce a 32 KiB default / 128 KiB maximum serialized payload budget across events, runtime events, diagnostics, and artifact references
• expose truncation and budget diagnostics in the tool result
• change the Graph supervisor prompt from max_events: 100 to a bounded 20-event, 32 KiB runtime-event inspection

Why

Issue #1593 reproduced Graph supervisor context overflow after four agent_output calls returned roughly 479–572 KiB each. The tool only bounded event count, not serialized size, and duplicated two event projections into the root session.

This is PR 1 of 4 in the Graph context-overflow repair stack.

Impact

Raw child execution history remains durable and inspectable through explicit views, while normal Graph supervision receives a bounded semantic projection.

Validation

• npm --workspace @maka/runtime run typecheck
• npm --workspace @maka/runtime run build
• targeted SessionManager agent-output tests: 4/4
• Graph/subagent tool tests: 27/27
• git diff --check

The full SessionManager file also exposed an unrelated pre-existing timing-sensitive test failure; the affected agent-output subset passes independently.

[likun666661]

What changed

• add explicit agent_output views so callers do not receive AgentRun and RuntimeEvent streams together by default
• enforce a 32 KiB default / 128 KiB maximum serialized payload budget across events, runtime events, diagnostics, and artifact references
• expose truncation and budget diagnostics in the tool result
• change the Graph supervisor prompt from max_events: 100 to a bounded 20-event, 32 KiB runtime-event inspection

Why

Issue #1593 reproduced Graph supervisor context overflow after four agent_output calls returned roughly 479–572 KiB each. The tool only bounded event count, not serialized size, and duplicated two event projections into the root session.

This is PR 1 of 4 in the Graph context-overflow repair stack.

Impact

Raw child execution history remains durable and inspectable through explicit views, while normal Graph supervision receives a bounded semantic projection.

Validation

• npm --workspace @maka/runtime run typecheck
• npm --workspace @maka/runtime run build
• targeted SessionManager agent-output tests: 4/4
• Graph/subagent tool tests: 27/27
• git diff --check

The full SessionManager file also exposed an unrelated pre-existing timing-sensitive test failure; the affected agent-output subset passes independently.

[likun666661]

What changed

• add explicit agent_output views so callers do not receive AgentRun and RuntimeEvent streams together by default
• enforce a 32 KiB default / 128 KiB maximum serialized payload budget across events, runtime events, diagnostics, and artifact references
• expose truncation and budget diagnostics in the tool result
• change the Graph supervisor prompt from max_events: 100 to a bounded 20-event, 32 KiB runtime-event inspection

Why

Issue #1593 reproduced Graph supervisor context overflow after four agent_output calls returned roughly 479–572 KiB each. The tool only bounded event count, not serialized size, and duplicated two event projections into the root session.

This is PR 1 of 4 in the Graph context-overflow repair stack.

Impact

Raw child execution history remains durable and inspectable through explicit views, while normal Graph supervision receives a bounded semantic projection.

Validation

• npm --workspace @maka/runtime run typecheck
• npm --workspace @maka/runtime run build
• targeted SessionManager agent-output tests: 4/4
• Graph/subagent tool tests: 27/27
• git diff --check

The full SessionManager file also exposed an unrelated pre-existing timing-sensitive test failure; the affected agent-output subset passes independently.

[likun666661]

What changed

• add explicit agent_output views so callers do not receive AgentRun and RuntimeEvent streams together by default
• enforce a 32 KiB default / 128 KiB maximum serialized payload budget across events, runtime events, diagnostics, and artifact references
• expose truncation and budget diagnostics in the tool result
• change the Graph supervisor prompt from max_events: 100 to a bounded 20-event, 32 KiB runtime-event inspection

Why

Issue #1593 reproduced Graph supervisor context overflow after four agent_output calls returned roughly 479–572 KiB each. The tool only bounded event count, not serialized size, and duplicated two event projections into the root session.

This is PR 1 of 4 in the Graph context-overflow repair stack.

Impact

Raw child execution history remains durable and inspectable through explicit views, while normal Graph supervision receives a bounded semantic projection.

Validation

• npm --workspace @maka/runtime run typecheck
• npm --workspace @maka/runtime run build
• targeted SessionManager agent-output tests: 4/4
• Graph/subagent tool tests: 27/27
• git diff --check

The full SessionManager file also exposed an unrelated pre-existing timing-sensitive test failure; the affected agent-output subset passes independently.