[Reasoning retention 2/2] Replay persisted reasoning safely to vLLM

[franciscojavierarceo]

Parent: #151

Depends on: #152

Goal

Replay complete persisted reasoning items through previous_response_id and conversation continuations in original history order while adapting only the vLLM-bound request copy for vLLM's plaintext continuation capabilities.

Preconditions

Land #152 first so streamed first turns persist the same complete reasoning item as non-streamed first turns.

Scope

Persistence and rehydration

• Keep the complete typed reasoning item in stored response/conversation history.
• Preserve its position relative to assistant messages and function calls.
• Rehydrate historical reasoning before the next turn's new input.
• Preserve the original stored item unchanged, including summary, encrypted_content, and status.
• Confirm _agentic_item_kind never appears in serialized upstream JSON.

vLLM request preparation

Prepare only ctx.enriched_request.input (the vLLM-bound copy):

• Replay plaintext reasoning from content[].text.
• Treat encrypted_content: null as valid.
• When plaintext content and non-null encrypted state coexist, preserve the stored item but omit incompatible encrypted state from the vLLM-bound copy.
• If a reasoning item has non-null encrypted content but no plaintext content, return a clear ExecutorError::InvalidRequest before request serialization/network I/O.
• Never substitute reasoning summaries for continuation state.
• Never attempt to decrypt opaque content.
• Apply the same validation before both streaming and non-streaming inference.

Ordering

Verify the second upstream request preserves this sequence when present:

1. historical user input;
2. historical reasoning item;
3. historical assistant message;
4. historical function call;
5. new function call output;
6. new user input.

The typed storage marker must not appear anywhere in that request.

Sensitive logging

• Remove the uncommitted raw request-body, response-body, error-body, and SSE-line trace statements in executor/inference.rs.
• Retain safe structured diagnostics such as byte counts, response IDs, item types, status, URL, and whether authentication is configured.
• Do not add logs containing reasoning text or encrypted content.

Test plan

Extend the existing stateful Responses integration framework rather than creating a parallel harness:

• two-turn streamed continuation records the first turn's complete reasoning item in the second upstream request;
• equivalent two-turn non-streamed continuation;
• streamed and non-streamed second-turn histories are structurally equivalent;
• original ordering across reasoning, assistant message, function call, function call output, and new user input;
• complete summary, encrypted_content, and status remain present after persistence/rehydration even when vLLM does not consume them;
• plaintext reasoning with encrypted_content: null reaches vLLM;
• plaintext reasoning with preserved non-null encrypted state uses the vLLM-compatible enriched copy while storage remains unchanged;
• encrypted-only reasoning fails clearly before any mock upstream request is recorded;
• a summary-only item is not silently used as reasoning continuation state;
• _agentic_item_kind is absent recursively from upstream JSON.

Use the existing mock request recorder and literal hand-checked response fixtures. Do not hand-author replay cassettes.

Likely files

• crates/agentic-server-core/src/executor/engine.rs
• crates/agentic-server-core/src/executor/rehydrate.rs or a focused executor-side vLLM preparation helper
• crates/agentic-server-core/src/executor/inference.rs
• crates/agentic-server-core/src/types/io/input.rs only if a focused typed traversal helper is warranted
• crates/agentic-server-core/tests/stateful_responses_integration.rs
• existing test support helpers

Avoid transport concerns in core wire types and avoid changing storage migrations.

Acceptance criteria

• Streamed and non-streamed continuations send equivalent complete historical reasoning items upstream.
• Plaintext reasoning reaches vLLM in correct history order.
• Stored summary and encrypted state remain lossless even when excluded from vLLM interpretation.
• Encrypted-only reasoning returns a clear 400-class invalid-request error before any vLLM call.
• Summaries are never promoted into continuation reasoning.
• _agentic_item_kind never leaks upstream.
• Normal logs contain no reasoning text, encrypted content, or raw bodies.
• No schema migration or alternate reasoning-text persistence is introduced.

Related work

• [SSE Completion - Issue 2] Prior-turn reasoning leaks <think> into the visible answer #120 concerns reasoning leakage across inference rounds within an active gateway tool loop. Keep this issue scoped to stored continuation replay; full reasoning.context semantics remain follow-up work.
• Persist reliable Responses continuation checkpoints #71 covers broader certified continuation checkpoints.

Verification

cargo test -p agentic-server-core --test stateful_responses_integration
cargo test -p agentic-server-core
cargo clippy -p agentic-server-core --all-targets -- -D warnings
cargo fmt -- --check

Report the exact commands and results.