feat: Alma/Codex Responses multi-turn tool replay

[starlight02]

Summary

Fixes #9 by adding a process-local tool-call replay cache and tightening Responses ModelInput hygiene so Alma/Codex multi-turn tool results work against Grok CLI (no previous_response_id / item_reference support upstream).

What this PR does

Area	Change
Cache	ToolReplayCache: index completed function_call / custom_tool_call by prev-resp:{id}, optional cache:{prompt_cache_key}, and item:{item.id}
Prepare pipeline	expand item_reference → re-insert missing calls for tool outputs → normalize → prune orphan outputs
ModelInput	drop server-tool history (web_search_call, …); force arguments string; flatten tool output; strip null reasoning fields
Stream	remember calls from output_item.done (item ids immediately; session keys on response.completed, patch empty completed.output)
Safety	only tool calls are cached — never encrypted reasoning (not portable across the credential pool)

Pipeline order

expand item_reference
  → applyToolCallReplay (prev-resp / cache)
  → normalizeInputItem (ModelInput)
  → pruneOrphanToolOutputs

Out of scope

• WebSocket / multi-provider executor changes
• inject build-search defaults
• header fingerprinting / 403 broadening
• cross-account encrypted reasoning replay

Test plan

• go test ./internal/openai ./internal/server
• Unit: previous_response_id restore, item_reference expand, orphan prune, stream index, free-tier client model key, arguments/output harden, server-tool drop, null reasoning strip
• Live: Alma/Codex new session — tool call → tool result → answer without re-requesting the same tool
• Live: payload with web_search_call / reasoning.content: null history does not 422

Notes

Inspired by CLIProxyAPI xAI tool/replay + normalize path; adapted to this repo’s single-process Responses compatibility layer (tool-call subset only).

[gemini-code-assist]

Code Review

This pull request introduces a tool-call replay cache (ToolReplayCache) to support multi-turn continuity for OpenAI Responses clients, allowing them to continue with only previous response IDs or item references and tool outputs. It integrates this cache into the server's streaming and non-streaming response handlers, handles compatibility adjustments (such as flattening tool outputs, stringifying arguments, and stripping null reasoning fields), and adds comprehensive unit tests. The review feedback focuses on performance and type-safety improvements, specifically suggesting that flattenToolOutput return a concrete string instead of any, and recommending a lazy allocation (copy-on-write) pattern in expandItemReferences and pruneOrphanToolOutputs to avoid unnecessary slice allocations when no modifications are made.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[Futureppo]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f18c3405d9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".