Suppress duplicate same-response side-effect tool calls

[Komzpa]

Summary

• suppress duplicate named side-effect tool calls in the same upstream response when canonical arguments match but the model emits a new call_id
• keep existing replay-burst behavior across response IDs and still allow intentional repeats after an intervening side-effect boundary
• update HTTP/WebSocket coverage so duplicate side-effect calls produce a retryable stream_incomplete instead of silently executing twice

Tests

• uv run ruff format tests/unit/test_proxy_tool_call_dedupe.py tests/unit/test_proxy_utils.py tests/integration/test_proxy_websocket_responses.py
• uv run ruff check app/modules/proxy/tool_call_dedupe.py tests/unit/test_proxy_tool_call_dedupe.py tests/unit/test_proxy_utils.py tests/integration/test_proxy_websocket_responses.py
• uv run pytest -q tests/unit/test_proxy_utils.py::test_stream_responses_suppresses_same_response_http_tool_calls_with_distinct_call_ids tests/unit/test_proxy_utils.py::test_process_upstream_websocket_text_suppresses_same_response_distinct_tool_call_ids tests/unit/test_proxy_tool_call_dedupe.py tests/unit/test_proxy_http_bridge.py -k 'duplicate or tool_call or replay' tests/integration/test_proxy_websocket_responses.py -k 'duplicate or tool_call or replay'

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[Komzpa]

@codex review

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[Soju06]

The flat-call part of this is a real spec-compliance fix: openspec/specs/responses-api-compat/spec.md ("Same-response side-effect tool-call replays are suppressed") already mandates that a changed call_id alone must not make a same-response replay distinct, and main currently violates that for sequential flat calls. But as written the new check overreaches into identity-scoped calls:

1. P1 — collapses distinct namespaced/code-mode call identities. The new same-response check (app/modules/proxy/tool_call_dedupe.py:140) runs before identity_scoped_call is computed at line 149, and same_response_argument_key deliberately zeroes the call_id slot. So two collaboration.spawn_agent custom_tool_calls in one response with identical input and different call_ids — the exact scenario the "Namespaced side-effect replay dedupe preserves call identity" requirement pins down ("Calls with different namespaces or different nonblank call IDs MUST remain distinct, even when their names and canonical arguments match"; scenario "Distinct namespaced spawns use identical arguments") — now suppress the second spawn and convert the whole turn into a stream_incomplete failure. Same for code-mode exec with identical source and distinct call_ids (2026-07-12 harden delta: "without collapsing distinct call identities"). Since the client retry will regenerate the same two spawns, this is a deterministic retry loop for legit multi-agent v2 parallel spawns. CI is green only because downstream same-response distinct-call_id coverage exists for flat calls, not namespaced ones. Fix: move the check below line 149 and gate it on not identity_scoped_call (identity-scoped calls already get cross-response protection via retained call_id), and add a downstream test mirroring the spec scenario.

2. P3 — burst clear also drops response-scoped parallel_tool_use keys. _clear_downstream_side_effect_burst_keys (line 202) pops every response-scoped key with a None call_id, which includes the per-tool_use keys written by _mark_duplicate_parallel_tool_call_downstream_event. On the response-scoped WS path, parallel([exec X]) -> exec Y -> parallel([exec X]) in one response previously trimmed the second X; now it executes twice. If the intervening-boundary reset is meant to apply to parallel batches too, fine, but it should be deliberate.

3. P3 — the intervening-boundary carve-out itself is new semantics. The canonical requirement says forward only the first event, with no boundary exception; the burst reset is a reasonable refinement but deserves a small openspec delta modifying that requirement rather than shipping as silent drift (that drift is how main ended up out of spec here in the first place).