Bound repeated empty tool_search misses for external models #159

Part of #156.

Problem

On the external-model compatibility path, an exact tool_search query can return tools=[] repeatedly while the same search function remains visible in replayed history. The model may resample the identical unavailable callback query without a terminal policy, amplifying requests, tokens, latency, and cost without creating a callable host tool.

Outcome

Bound identical empty tool-search misses in the CodexHub adapter and return one sanitized, machine-classifiable unavailable terminal result. A different query and every non-empty result remain unaffected.

Scope

• Track identical exact-name search misses within one request/turn history.
• Permit at most two identical empty searches or two additional model samples.
• On the bound, emit a deterministic unavailable result and prevent a third identical search sample.
• Preserve successful discovery, distinct queries, call IDs, history, SSE terminalization, and existing Official/third-party tool behavior.
• Record only sanitized query classification/count/status telemetry; do not record prompts, tool arguments, callback addresses, or Task identifiers.

Non-goals

• Do not inject a callback schema without a real host handler.
• Do not implement Codex Host sidebar Worker registration, result receipts, or model-binding readback.
• Do not use GPT, Hidden Subagent, codex exec, Background, or Inline work as a substitute for Enable sidebar-visible third-party Worker callbacks and binding readback #156 acceptance.
• Do not change unrelated retry, transport, apply-patch, or Task activity behavior.

Acceptance criteria

• The first exact-name miss returns the existing explicit empty/unavailable search result.
• A second identical miss produces one classified terminal unavailable result and no third model sample for that query in the turn.
• A distinct legitimate query is not suppressed by the identical-query guard.
• A non-empty search result remains unchanged and callable.
• History replay preserves call/result identity and exactly one terminal outcome.
• Sanitized telemetry records the bound without prompts, callback addresses, credentials, paths, or Task identifiers.
• Official, namespace, MCP, collaboration, and strict apply-patch regressions remain green.

Verification

python -m pytest -q tests/test_routing.py -k "tool_search"
python -m pytest -q tests/test_codex_semantic_adapter.py
python -m pytest -q
python scripts/report_quality_gates.py
git diff --check

report_quality_gates.py is report-only.

Expected hotset

• src-python/codex_proxy.py
• src-python/codex_semantic_adapter.py
• tests/test_routing.py
• tests/test_codex_semantic_adapter.py
• one sanitized fixture under tests/fixtures/ if the deterministic replay needs it

Relationships

• Parent/human Visible Worker gate: Enable sidebar-visible third-party Worker callbacks and binding readback #156
• Full collaboration matrix: 0.1.8 Beta2: Define the Codex CLI 0.146 Collaboration V1/V2 schema boundary #64
• Deferred search classification: 0.1.8 Beta3: Complete the generic client tool_search lifecycle #63

Execution contract

Execution-Contract: v2
Verification-Class: strict
Verification-Commands: targeted routing/semantic-adapter tests; full Python suite once; git diff --check; report-only quality gate
Manual-Evidence: none
Architecture-Decision: resolved
Review-Owner: orchestrator