test: preserve multi-tool assistant turns in E2E showcase

[gnanirahulnutakki]

Context

CodeRabbit review of PR #365 found that the cloud-gated E2E showcase constructs invalid conversation history when Ollama returns more than one tool call in a single assistant turn. This defect is pre-existing on dev at 8af2340965a28894129a7300611f8dc210023776; PR #365 only formatted the block and did not change its AST or behavior.

Exact evidence

Pre-existing source:

• test_multi_turn_conversation, original lines 708-772
• The malformed construction is specifically original lines 744-766.

CodeRabbit's exact finding:

Update the message construction around the multi-call loop to append the original assistant tool-call message only once, outside the loop. Keep one corresponding tool message per call inside the loop, using each call’s tool_name, while preserving the existing evaluation behavior.

When resp.message.tool_calls contains N calls, the current loop appends N synthetic assistant messages, each containing only one call. A valid transcript has one assistant turn containing the original ordered set of N tool calls, followed by N corresponding tool-result messages. The current malformed history can make a later model turn reject, reinterpret, or lose the relationship among parallel calls.

The affected class is guarded by @ollama_required, so ordinary CI and cloud-disabled local runs skip this path. That allowed the defect to remain invisible even though the showcase reports multi-turn coverage when credentials are present.

Expected sequence

For an assistant response containing calls [call_a, call_b], append exactly:

1. one assistant message preserving the original assistant content and ordered tool_calls=[call_a, call_b];
2. one tool message for call_a, using call_a.function.name and its corresponding result;
3. one tool message for call_b, using call_b.function.name and its corresponding result.

Continue to send one /evaluate request per tool call and preserve the existing evaluation counter/decision behavior.

Acceptance criteria

• The original assistant tool-call turn is appended once, outside the per-call result loop.
• All tool calls from that turn remain attached to the single assistant message in original order.
• Exactly one matching tool-result message is appended per call, in the same order, with the correct tool name and result.
• /evaluate is still invoked exactly once per call with that call's name and parsed arguments.
• Add a credential-free regression using a fake response with at least two tool calls; assert the exact assistant/tool message sequence presented to the next model turn.
• The regression fails against dev at 8af2340965a28894129a7300611f8dc210023776.
• The existing single-call behavior remains valid.
• The cloud-disabled showcase retains its expected skips; an optional credentialed smoke may be recorded separately without making CI depend on an external model.
• Ruff 0.13.0 and repository pre-commit hooks pass.

Related: #365, #361

[gnanirahulnutakki]

Implementation is ready in PR #373 at signed/DCO commit d280fc6dfb720dd838de720bd6fc651476224263.

Evidence:

• old exact-dev behavior failed the new two-call regression by producing two synthetic assistant turns with legacy name result fields;
• the fix appends the original assistant message exactly once, preserves ordered calls, emits one tool_name result per call, and evaluates each call exactly once with its own parsed arguments;
• rejected follow-up requests now propagate and cannot become a false showcase success;
• credential-free targeted result: 21 passed, 10 expected cloud-disabled skips;
• Ruff 0.13, format, pre-commit security hooks, quick repository checks, and diff checks pass;
• independent review is clear;
• hosted CodeRabbit and CI are running.

Primary contract: https://docs.ollama.com/capabilities/tool-calling

Adjacent gaps were decomposed into #374 and #375 instead of broadening this issue. The five pre-existing pytest warnings remain tracked by #371.

[gnanirahulnutakki]

#367 closeout — merged and exact-dev green

PR #373 merged into dev as 424735afd813540ae74ae6e33f6e22b4ac521536 from final reviewed, signed, DCO head 18954b63db3b239552cccbf7fa0b33766c218fcc. The reviewed and merge trees are identical at 2e32243be87182b9e5d721cac55b41ba35c6a38e; GitHub reports the merge signature as valid.

Final head evidence:

• 36 successful checks, 6 intentional release-only skips, 0 failures/pending;
• 0 open CodeQL alerts and 0 unresolved review threads;
• local targeted suite: 29 passed, 10 unchanged credential-disabled Ollama skips.

CodeRabbit found one valid fail-open mapping defect. The signed remediation now maps explicit PERMIT to ok, explicit DENY to denied, and absent/unusable/non-200/unknown evidence to unknown. CodeRabbit confirmed the semantics and resolved its only thread.

All seven exact-SHA post-merge dev workflows passed:

• tests: https://github.com/ArdurAI/ardur/actions/runs/29981071944
• CodeQL: https://github.com/ArdurAI/ardur/actions/runs/29981071955
• OCI proxy: https://github.com/ArdurAI/ardur/actions/runs/29981071985
• Python package: https://github.com/ArdurAI/ardur/actions/runs/29981071952
• Hugo: https://github.com/ArdurAI/ardur/actions/runs/29981071917
• format/schema validation: https://github.com/ArdurAI/ardur/actions/runs/29981071974
• secret scan: https://github.com/ArdurAI/ardur/actions/runs/29981071959

Adjacent scope is decomposed as #374 for the five remaining legacy Ollama transcript builders and #375 for credentialed fail-closed CI. #375 is parked on dependency, credentials, cost, and required-gate policy. Next safe focused lane is #371. Cluster qualification is separately parked on interactive Duplo JIT auth; when resumed, every workload will use exactly nodeSelector.tenantname: duploservices-saasops3.