Fix agentic-traffic reliability and enable provider prompt caching

[veerareddyvishal144]

Problem

Agent loops through Lynkr were silently degraded in four ways, all observed and measured on ITSMBench / terminal-bench harnesses (pi agent, Azure gpt-5.6-sol):

1. Episodes killed mid-task: upstream 429-retry exhaustion surfaced as a contentless 200 with finish_reason: stop; agent clients read that as "done" and stopped. Killed 6/7 benchmark episodes in one run.
2. System prompt discarded after turn 1: continuations replaced the client agent's instructions with a generic one-liner.
3. Reasoning never forwarded: reasoning_effort was dropped on the Azure Responses path; thinking requests silently ran at minimum effort. max_output_tokens was also unset for gpt-5.x (read from the wrong key).
4. Provider prompt caching broken (0–3% hit rate): six code paths rewrote request bytes every turn — random tool-call fallback IDs (Date.now()+random), the system-prompt swap, conditional reminder-stripping, timestamped tee IDs in tool-result compression, tier-flapping compression thresholds — so prefixes never matched and every token billed at full price.

Fix

• openai-router: guard that aborts the SSE stream (client retries) instead of serving a clean empty completion.
• databricks: forward reasoning.effort; fix max_output_tokens; surface cache_read_input_tokens through both response conversions into telemetry; add prompt_cache_key (session id or first-user-message hash) on gpt-5.x requests.
• Prefix stability: deterministic content-hash tool-call/tee IDs, one constant system prompt per conversation (full prompt, reminders stripped uniformly on every turn), fixed COMPLEX compression threshold.
• config: new TOOL_RESULT_COMPRESSION_ENABLED env knob (was hardcoded).
• nodemon.json: ignore self-written data/db files (dev restart storms).

Measured results

• Provider cache hit rate: 0–3% → 92–95% on agent loops (~5× real cost reduction; ~$5.00/task → ~$0.87/task at reasoning=high on ITSMBench).
• Through-proxy quality on ITSMBench now statistically equal to the same model direct (35.0% vs 35.51% Pass@1).
• No Azure content-filter rejections across ~2,000 continuation requests with the full system prompt.

Behavior changes for review

• Tool-result compression threshold is now fixed at COMPLEX (>2000 chars) instead of following the routed tier — determinism is required for caching; net effect is lighter compression for most traffic.
• Continuations carry the client's full (reminder-stripped) system prompt instead of the generic replacement.

Known issue (follow-up, not addressed here)

Terminal-bench A/B shows Lynkr still costs significant quality on long compute-heavy coding tasks (direct 17/22 vs through-Lynkr 8/21), dominated by agent timeouts from per-turn latency (buffered non-streaming upstream calls). Fix is true streaming passthrough — proposed as a follow-up issue.

🤖 Generated with Claude Code