OmniRoute/docs/routing/REASONING_REPLAY.md

title, version, lastUpdated

title	version	lastUpdated
Reasoning Replay Cache	3.8.40	2026-06-28

Reasoning Replay Cache

Source of truth: src/lib/db/reasoningCache.ts, open-sse/services/reasoningCache.ts Last updated: 2026-06-28 — v3.8.40

OmniRoute captures assistant reasoning_content produced by thinking-mode models and replays it transparently on multi-turn requests when the upstream provider requires it. This eliminates the HTTP 400 errors that strict providers raise when a client's conversation history is missing the prior turn's reasoning.

Why This Exists

Several thinking-mode providers reject a follow-up turn unless the previous assistant message includes the original reasoning_content. The upstream returns 400 with messages like:

Param Incorrect: The reasoning_content in the thinking mode must be passed back to the API.

But typical clients (Cursor, Cline, Roo Code, OpenAI SDK) strip reasoning_content from the history they replay. OmniRoute restores it from a server-side cache so the request the upstream sees is consistent. Issue #1628 introduced the hybrid memory/SQLite persistence so the cache survives process restarts.

Architecture

Turn N (assistant generates):
  → response contains reasoning_content + tool_calls
  → if requiresReasoningReplay(provider, model): cacheReasoningFromAssistantMessage()
      writes (memory + DB), keyed by every tool_call.id
  → forward response to client (which may or may not retain reasoning)

Turn N+1 (client sends follow-up):
  → translator detects: requiresReasoningReplay(provider, model) === true
  → for each assistant message with tool_calls and no reasoning_content:
      lookupReasoning(toolCalls[0].id) → memory → DB
      hit  → msg.reasoning_content = cached; recordReplay()
      miss → msg.reasoning_content = "" (legacy fallback for older DeepSeek)
  → upstream sees consistent history → no 400

Capture happens in open-sse/handlers/chatCore.ts (two sites, at the two cacheReasoningFromAssistantMessage call sites). Replay happens in open-sse/translator/index.ts after schema coercion but before dispatch.

Plain (non-tool-call) assistant turns are keyed differently: buildAssistantMessageCacheKey() digests the session scope plus the normalized OpenAI-format transcript up to that turn, because DeepSeek requires the reasoning of every prior turn once tools is present. For Responses-API targets (for example opencode-go/deepseek-v4-flash, routed to /responses) the upstream body carries input, not messages, so translateRequest() (open-sse/translator/index.ts) reports the pivot transcript it digested through a callback option and the capture sites digest that same transcript. The Responses replay pass runs on the OpenAI pivot for every source format, so Anthropic Messages clients (Claude → OpenAI → Responses) are replayed too.

Storage — Hybrid Memory + SQLite

The hot path uses an in-memory Map (LRU-by-creation) backed by a SQLite table for crash recovery and dashboard visibility.

Layer	Implementation	Purpose
Memory	Map in open-sse/services/reasoningCache.ts	Fast lookups, evicts oldest at 200
DB	reasoning_cache table (src/lib/db/)	Persists across restarts, drives stats

Writes go to both. Reads consult memory first, then fall back to DB (DB hits are promoted back into memory). DB failures are non-fatal — the in-memory cache continues to serve the hot path.

Defaults:

• TTL: 2h (TTL_MS = 2 * 60 * 60 * 1000)
• Max memory entries: 200 (MAX_MEMORY_ENTRIES)
• Eviction: oldest createdAt first

Database Schema

Migration: src/lib/db/migrations/033_create_reasoning_cache.sql

CREATE TABLE IF NOT EXISTS reasoning_cache (
  tool_call_id   TEXT PRIMARY KEY,
  provider       TEXT NOT NULL,
  model          TEXT NOT NULL,
  reasoning      TEXT NOT NULL,
  char_count     INTEGER NOT NULL DEFAULT 0,
  created_at     TEXT NOT NULL DEFAULT (datetime('now')),
  expires_at     INTEGER NOT NULL
);

Indexes: expires_at, provider, model, created_at. expires_at is stored as Unix epoch seconds; the SELECT layer normalizes legacy text values via EXPIRES_AT_EPOCH_SQL.

Provider / Model Detection

Replay is enabled when requiresReasoningReplay(provider, model) returns true. The function checks two lists in open-sse/services/reasoningCache.ts.

Provider IDs (exact match, case-insensitive):

• deepseek
• opencode-go
• siliconflow
• nebius
• deepinfra
• sambanova
• fireworks
• together
• kimi-coding
• kimi-coding-apikey
• xiaomi-mimo

Model regex patterns (case-insensitive):

• /deepseek-r1/i
• /deepseek-reasoner/i
• /deepseek-chat/i
• /deepseek[-/]?v4[-.]flash/i and /deepseek[-/]?v4[-.]pro/i (V4 Flash / Pro, optional -free suffix)
• /(deepseek|zen\/deepseek)-v4/i
• /kimi[-/]k\d/i
• /qwq/i
• /qwen.*think/i
• /glm.*think/i
• /^mimo[-.]?v\d/i

Adding a new strict provider/model means appending to one of these lists and writing a unit test asserting replay injection. The PR description should cite the exact upstream 400 string that motivated the change.

REST API

The cache exposes two endpoints under src/app/api/cache/reasoning/route.ts. Both require management authentication (isAuthenticated from @/shared/utils/apiAuth).

Method	Endpoint	Description
GET	/api/cache/reasoning	Stats + paginated entries
GET	/api/cache/reasoning?provider=deepseek&model=...&limit=	Filtered listing (limit clamped to [1, 200])
DELETE	/api/cache/reasoning	Clear everything (memory + DB) and reset hit/miss counts
DELETE	/api/cache/reasoning?provider=deepseek	Clear only entries for one provider
DELETE	/api/cache/reasoning?toolCallId=call_abc	Delete a single entry

GET response shape:

{
  "stats": {
    "memoryEntries": 12,
    "dbEntries": 47,
    "totalEntries": 47,
    "totalChars": 138291,
    "hits": 84,
    "misses": 6,
    "replays": 81,
    "replayRate": "90.0%",
    "byProvider": { "deepseek": { "entries": 32, "chars": 98412 } },
    "byModel": { "deepseek-reasoner": { "entries": 32, "chars": 98412 } },
    "oldestEntry": "2026-05-13T10:00:00.000Z",
    "newestEntry": "2026-05-13T11:42:11.000Z"
  },
  "entries": [
    {
      "toolCallId": "call_abc",
      "provider": "deepseek",
      "model": "deepseek-reasoner",
      "reasoning": "...",
      "charCount": 3128,
      "createdAt": "...",
      "expiresAt": "..."
    }
  ]
}

Operational Notes

• Cleanup: cleanupReasoningCache() purges expired memory entries and runs DELETE FROM reasoning_cache WHERE expires_at <= unixepoch('now'). Health-check workers call this periodically.
• Crash recovery: After a restart, memory is empty but the DB still holds unexpired entries. The first lookup for a given tool_call_id is a DB hit; subsequent lookups are memory hits.
• No reasoning, no cache: cacheReasoningFromAssistantMessage returns 0 when the assistant message has no reasoning_content / reasoning field, so non-thinking responses cost nothing.
• Write is gated too: both call sites in chatCore.ts (non-streaming and streaming) only call cacheReasoningFromAssistantMessage() when requiresReasoningReplay(provider, model) is true — the same predicate the read side checks. Installs that never touch a replay provider stop paying for the write, the index update, and the try/catch on every reasoning-bearing response.
• Non-strict providers: When requiresReasoningReplay is false and the target format is OpenAI, the translator strips any reasoning_content field from outgoing messages — OpenAI Chat Completions does not accept it.

See Also

• RESILIENCE_GUIDE.md — circuit breakers, cooldowns, model lockouts
• TROUBLESHOOTING.md — diagnosing upstream 400s
• Source: src/lib/db/reasoningCache.ts, open-sse/services/reasoningCache.ts, open-sse/translator/index.ts
• Migration: src/lib/db/migrations/033_create_reasoning_cache.sql
• API route: src/app/api/cache/reasoning/route.ts
• Original issue: #1628