@chroxify/eunoe 1.2.2 → 1.3.1

package/src/core/state.ts

@@ -0,0 +1,50 @@
+import type { Message } from "../context/content"
+import { isPrompt } from "../context/fold/turns"
+import { messageFingerprint } from "./fingerprint"
+
+export interface Handoff {
+  label: string
+  source: string
+  text: string
+}
+
+export interface ThreadState {
+  handoffs?: Record<string, Handoff | null>
+  cut?: string
+  keep?: number
+  cleared?: number
+  stuck?: number
+  host?: boolean
+  pins?: string[]
+  ratio?: number
+  limits?: Record<string, number>
+  recalls?: Record<string, string>
+}
+
+export interface ThreadStore {
+  read: (key: string) => ThreadState
+  save: (key: string, state: ThreadState) => void
+}
+
+export function memoryStore(initial: Record<string, ThreadState> = {}): ThreadStore {
+  const threads = new Map(Object.entries(initial).map(([key, state]) => [key, structuredClone(state)]))
+  return {
+    read: (key) => structuredClone(threads.get(key) ?? {}),
+    save: (key, state) => {
+      threads.set(key, structuredClone(state))
+    },
+  }
+}
+
+export function retireRecalls(state: ThreadState, folded: Message[][]) {
+  if (!state.recalls) return
+  for (const message of folded.flat()) if (isPrompt(message)) delete state.recalls[messageFingerprint(message)]
+}
+
+export function resetCut(state: ThreadState) {
+  delete state.cut
+  delete state.keep
+  delete state.cleared
+  delete state.stuck
+  delete state.host
+}

package/src/core/subagents.ts

@@ -0,0 +1,28 @@
+import { blocks, type Message } from "../context/content"
+import { splitTurns } from "../context/fold/turns"
+import { type RecallChunk, recallChunks } from "../context/recall"
+
+export interface Subagent {
+  description: string
+  messages: Message[]
+  toolUseId?: string
+}
+
+const closingTurn: Message = { role: "user", content: "" }
+
+export function launchTurnOf(parentMessages: Message[], toolUseId: string | undefined): number {
+  if (!toolUseId) return 0
+  const index = splitTurns(parentMessages).turns.findIndex((turn) => turn.some((message) => message.role === "assistant" && blocks(message).some((b) => b.type === "tool_use" && b.id === toolUseId)))
+  return index + 1
+}
+
+const subagentLabel = (description: string) => `subagent "${description}"`
+
+export function subagentRecall(messages: Message[], description: string, parentTurn: number): RecallChunk[] {
+  const from = subagentLabel(description)
+  return recallChunks([...messages, closingTurn], new Map()).map((chunk) => ({ ...chunk, from, turn: parentTurn }))
+}
+
+export function listSubagentRecall(parentMessages: Message[], subagents: Subagent[]): RecallChunk[] {
+  return subagents.flatMap((subagent) => subagentRecall(subagent.messages, subagent.description, launchTurnOf(parentMessages, subagent.toolUseId)))
+}

package/src/core/transcript.ts

@@ -0,0 +1,44 @@
+import { blocks, type Message } from "../context/content"
+import { isPrompt } from "../context/fold/turns"
+import { searchGuide } from "../context/prompts"
+import { renderIndex, renderMarkdown, renderSubagentMarkdown, renderSubagentXml, renderXml, type SubagentRecord, subagentFile, turnFile } from "../transcript/format"
+import { parseTurns, type SessionEntry, type TranscriptFormat, type Turn } from "../transcript/parse"
+import { launchTurnOf, type Subagent } from "./subagents"
+
+export interface TranscriptOptions {
+  format?: TranscriptFormat
+  subagents?: Subagent[]
+}
+
+const isNote = (message: Message) => message.role === "user" && !isPrompt(message) && !blocks(message).some((b) => b.type === "tool_result")
+
+const turnsOf = (messages: Message[], sidechain: boolean | "any" = false): Turn[] =>
+  parseTurns(
+    messages.map((message): SessionEntry => ({ type: message.role, isMeta: isNote(message), message })),
+    sidechain,
+  )
+
+function subagentRecords(messages: Message[], turns: Turn[], subagents: Subagent[], format: TranscriptFormat): SubagentRecord[] {
+  const perTurn = new Map<number, number>()
+  return subagents.map((subagent) => {
+    const parentTurn = launchTurnOf(messages, subagent.toolUseId)
+    const k = (perTurn.get(parentTurn) ?? 0) + 1
+    perTurn.set(parentTurn, k)
+    const file = subagentFile(parentTurn, k, format)
+    const launcher = turns[parentTurn - 1]?.parts.find((part) => part.kind === "tool" && part.id === subagent.toolUseId)
+    if (launcher?.kind === "tool") launcher.record = file
+    return { run: { description: subagent.description }, turns: turnsOf(subagent.messages, "any"), parentTurn, file }
+  })
+}
+
+export function renderTranscriptFiles(messages: Message[], { format = "markdown", subagents = [] }: TranscriptOptions = {}): Record<string, string> {
+  const turns = turnsOf(messages)
+  const records = subagentRecords(messages, turns, subagents, format)
+  const files: Record<string, string> = {}
+  for (const turn of turns) files[turnFile(turn, format)] = format === "xml" ? renderXml(turn) : renderMarkdown(turn)
+  for (const record of records) files[record.file] = format === "xml" ? renderSubagentXml(record) : renderSubagentMarkdown(record)
+  files["index.md"] = renderIndex(turns, format, records)
+  return files
+}
+
+export const transcriptGuide = (dir: string, format: TranscriptFormat = "markdown") => searchGuide({ mode: format, dir, jsonl: "" })

package/src/{proxy → core}/usage.ts

@@ -1,4 +1,8 @@
-type Usage = Record<string, number>
+import type { Body } from "../context/content"
+import { sizeOf } from "./estimate"
+import type { ThreadStore } from "./state"
+
+export type Usage = Record<string, number>
 
 function usageIn(line: string): Usage | null {
   if (!line.startsWith("data:") || !line.includes("message_start")) return null
@@ -58,3 +62,41 @@ export function watchUsage(response: Response, onUsage: (usage: Usage) => void):
 export function inputTokens(usage: Usage) {
   return (usage.input_tokens ?? 0) + (usage.cache_read_input_tokens ?? 0) + (usage.cache_creation_input_tokens ?? 0)
 }
+
+const OVERFLOW_MARGIN = 1.1
+
+const charsOf = (sent: Body | number) => (typeof sent === "number" ? sent : sizeOf(sent))
+
+function overflowOf(errorText: string): { tokens: number; maximum: number | null } | null {
+  const match = errorText.match(/prompt is too long[^0-9]*(\d+)\s*tokens?\s*>\s*(\d+)/i)
+  const tokens = Number(match?.[1])
+  const maximum = Number(match?.[2])
+  if (!Number.isFinite(tokens) || tokens <= 0) return null
+  return { tokens, maximum: Number.isFinite(maximum) && maximum > 0 ? maximum : null }
+}
+
+export function recordUsage(store: ThreadStore, threadKey: string, sent: Body | number, usage: Usage): boolean {
+  const input = inputTokens(usage)
+  const chars = charsOf(sent)
+  if (input <= 0 || chars <= 0) return false
+  const state = store.read(threadKey)
+  state.ratio = input / chars
+  store.save(threadKey, state)
+  return true
+}
+
+export const limitKey = (model: unknown) => String(model ?? "")
+
+export function recordOverflow(store: ThreadStore, threadKey: string, sent: Body | number, errorText: string, model: unknown = typeof sent === "number" ? undefined : sent.model): number | null {
+  const overflow = overflowOf(errorText)
+  const chars = charsOf(sent)
+  if (!overflow || chars <= 0) return null
+  const state = store.read(threadKey)
+  state.ratio = Math.max(state.ratio ?? 0, overflow.tokens / chars) * OVERFLOW_MARGIN
+  if (overflow.maximum) {
+    const key = limitKey(model)
+    state.limits = { ...state.limits, [key]: Math.min(state.limits?.[key] ?? Number.POSITIVE_INFINITY, overflow.maximum) }
+  }
+  store.save(threadKey, state)
+  return overflow.tokens
+}

package/src/core/window.ts

@@ -0,0 +1,69 @@
+export type WindowLookup = (model: unknown, requestHeaders: Headers, upstream?: string) => Promise<number | null>
+
+export interface WindowLookupOptions {
+  known?: Map<string, number>
+  onLearn?: (windows: Map<string, number>) => void
+  timeoutMs?: number
+  retryAfterMs?: number
+}
+
+const LOOKUP_TIMEOUT_MS = 1_500
+
+const RETRY_AFTER_MS = 10 * 60_000
+
+const LONG_SUFFIX = /\[1m\]$/i
+
+const SUFFIX = /\[[^\]]*\]$/
+
+function lookupHeaders(requestHeaders: Headers): Headers {
+  const headers = new Headers()
+  for (const name of ["authorization", "x-api-key", "anthropic-version", "anthropic-beta"]) {
+    const value = requestHeaders.get(name)
+    if (value) headers.set(name, value)
+  }
+  if (!headers.has("anthropic-version")) headers.set("anthropic-version", "2023-06-01")
+  return headers
+}
+
+function familyWindow(model: string): number | null {
+  if (LONG_SUFFIX.test(model)) return 1_000_000
+  const id = model.replace(SUFFIX, "")
+  if (/^claude-(?:(?:opus|sonnet|haiku|fable)-5|opus-4-8)/.test(id)) return 1_000_000
+  if (id.startsWith("claude-")) return 200_000
+  return null
+}
+
+export function createWindowLookup({ known = new Map<string, number>(), onLearn, timeoutMs = LOOKUP_TIMEOUT_MS, retryAfterMs = RETRY_AFTER_MS }: WindowLookupOptions = {}): WindowLookup {
+  const lookups = new Map<string, Promise<number | null>>()
+  const failedAt = new Map<string, number>()
+  const lookup = async (id: string, requestHeaders: Headers, upstream: string): Promise<number | null> => {
+    const response = await fetch(`${upstream}/v1/models/${encodeURIComponent(id)}`, { headers: lookupHeaders(requestHeaders), signal: AbortSignal.timeout(timeoutMs) })
+    if (!response.ok) return null
+    const info = (await response.json()) as { max_input_tokens?: number }
+    if (typeof info.max_input_tokens !== "number" || info.max_input_tokens <= 0) return null
+    known.set(id, info.max_input_tokens)
+    onLearn?.(known)
+    return info.max_input_tokens
+  }
+  return (model, requestHeaders, upstream = "https://api.anthropic.com") => {
+    const given = String(model ?? "")
+    const id = given.replace(SUFFIX, "")
+    if (!id) return Promise.resolve(null)
+    const window = known.get(id)
+    if (window) return Promise.resolve(window)
+    const fallback = familyWindow(given)
+    const failed = failedAt.get(id)
+    if (failed !== undefined && Date.now() - failed < retryAfterMs) return Promise.resolve(fallback)
+    const pending =
+      lookups.get(id) ??
+      lookup(id, requestHeaders, upstream)
+        .catch(() => null)
+        .then((found) => {
+          if (found === null) failedAt.set(id, Date.now())
+          return found
+        })
+        .finally(() => lookups.delete(id))
+    lookups.set(id, pending)
+    return pending.then((found) => found ?? fallback)
+  }
+}

package/src/proxy/rewrite.ts

@@ -1,20 +1,15 @@
+import { existsSync, readFileSync } from "node:fs"
 import type { Config } from "../config"
-import type { Body, Message } from "../context/content"
-import { appendSystem, tailMessages, trimMessages } from "../context/fold/strategies"
-import { splitTurns, turnFingerprint } from "../context/fold/turns"
-import { SYSTEM_PROMPT, searchGuide, searchOnlyInstructions, trimInstructions } from "../context/prompts"
-import { identityReranker, type Reranker, withRecall } from "../context/recall"
-import { jevReranker, rerankKey } from "../context/recall/rerank"
+import { paths } from "../config/paths"
+import type { Reranker } from "../context/recall"
+import { jevReranker } from "../context/recall/rerank"
+import { type RewriteInput as CoreInput, prefetchRecall as prefetchCore, type Rewrite, rewrite as rewriteCore, rewriteAsync as rewriteCoreAsync } from "../core/rewrite"
 import type { TranscriptRef } from "../transcript/search"
-import { compact } from "./compaction"
-import { DEFAULT_RATIO, estimator, sizeOf, targetSize } from "./estimate"
-import { messageFingerprint } from "./fingerprint"
-import { activeHandoffs, attachHandoffs, withHandoffs } from "./handoff"
-import { activeRecalls, applyRecall, prefetchFor, wantsRecall } from "./recall"
-import { readState, saveState } from "./state"
+import { sessionSources } from "./sources"
+import { fileStore } from "./state"
 
 export interface RewriteInput {
-  body: Body
+  body: CoreInput["body"]
   sessionId: string | null
   config: Config
   transcript: TranscriptRef | null
@@ -23,146 +18,34 @@ export interface RewriteInput {
   omittedBefore?: number
 }
 
-export interface Rewrite {
-  body: Body
-  threadKey: string | null
-  note: Record<string, unknown>
+export function rerankKey(): string | null {
+  const env = process.env.TYPESAFE_API_KEY?.trim()
+  if (env) return env
+  return existsSync(paths.typesafeKey) ? readFileSync(paths.typesafeKey, "utf8").trim() : null
 }
 
-interface Prefetched {
-  reranked: boolean
-  rerankError: string | null
-  recalled: number
-}
-
-const pass = (body: Body): Rewrite => ({ body, threadKey: null, note: { action: "pass" } })
-
-const isAgentThread = (body: Body) => Array.isArray(body.tools) && body.tools.length > 0
-
-const withSystemPrompt = (system: Body["system"], extra?: string | null) => appendSystem(extra ? appendSystem(system, extra) : system, SYSTEM_PROMPT)
-
-const ratioOf = (threadKey: string) => readState(threadKey).ratio ?? DEFAULT_RATIO
-
 function rerankModel(config: Config): string | null {
   const model = config.eval?.rerank?.model ?? config.rerank
   return model && model !== "off" ? model : null
 }
 
-function evalStrategy(input: RewriteInput, turns: Message[][], threadKey: string, guide: string | null): Rewrite | null {
-  const { body, config } = input
-  const strategy = config.eval?.strategy
-  if (strategy !== "guide" && strategy !== "trim" && strategy !== "tail") return null
-  if (!isAgentThread(body)) return pass(body)
-  const counts = { turns: turns.length, messagesIn: body.messages.length }
-  if (strategy === "guide") {
-    return { body: { ...body, system: withSystemPrompt(body.system, guide ? searchOnlyInstructions(guide) : null) }, threadKey, note: { action: "guide", ...counts } }
-  }
-  if (strategy === "trim") {
-    const messages = trimMessages(body.messages, { evidence: config.eval?.evidence ?? true })
-    return { body: { ...body, system: withSystemPrompt(body.system, trimInstructions(guide)), messages }, threadKey, note: { action: "trim", ...counts, messagesOut: messages.length } }
-  }
-  const based = { ...body, system: withSystemPrompt(body.system) }
-  const size = estimator(based, ratioOf(threadKey))
-  let from = 1
-  if (input.window) {
-    const target = targetSize(size, input.window)
-    from = turns.length - 1
-    while (from > 1 && size.of(tailMessages(body.messages, from - 1, guide)) <= target) from -= 1
-  }
-  const messages = tailMessages(body.messages, from, guide)
-  return { body: { ...based, messages }, threadKey, note: { action: "tail", keptTurns: turns.length - from, ...counts, messagesOut: messages.length, estimate: size.of(messages) } }
-}
-
-function strategize(input: RewriteInput): Rewrite {
-  const { body, sessionId, config, transcript } = input
-  if (config.mode === "off" || !sessionId || !Array.isArray(body.messages) || body.messages.length === 0) return pass(body)
-  const { turns } = splitTurns(body.messages)
-  if (turns.length === 0) return pass(body)
-  const threadKey = `${sessionId}-${turnFingerprint(turns[0])}`
-  const guide = transcript ? searchGuide(transcript) : null
-  const state = readState(threadKey)
-  if (attachHandoffs(body.messages, config, sessionId, state)) saveState(threadKey, state)
-
-  const forced = evalStrategy(input, turns, threadKey, guide)
-  if (forced) return forced
-
-  const based = isAgentThread(body) ? { ...body, system: withSystemPrompt(body.system) } : body
-  if (body.messages.length < 2) return { body: based, threadKey, note: { action: "full", turns: turns.length } }
-  const handoffs = activeHandoffs(state.handoffs)
-  const recalls = wantsRecall(input) ? activeRecalls(state.recalls) : {}
-  const result = compact({
-    body: based,
-    config,
-    state,
-    window: input.window,
-    guide,
-    transcriptDir: transcript && transcript.mode !== "jsonl" ? transcript.dir : null,
-    mayCompact: input.mayCompact,
-    omittedBefore: input.omittedBefore,
-    decorate: (messages) => withRecall(withHandoffs(messages, handoffs), messageFingerprint, recalls),
-  })
-  if (result.stateChanged) saveState(threadKey, state)
-  return { body: result.messages === body.messages ? based : { ...based, messages: result.messages }, threadKey, note: result.note }
-}
-
-function withAttachedHandoffs(result: Rewrite): Rewrite {
-  if (!result.threadKey) return result
-  const handoffs = activeHandoffs(readState(result.threadKey).handoffs)
-  const messages = withHandoffs(result.body.messages, handoffs)
-  if (messages === result.body.messages) return result
-  const attached = Object.values(handoffs)
-  const note = { ...result.note, handoff: attached.reduce((sum, h) => sum + h.text.length, 0), handoffFrom: attached.map((h) => h.label).join(", ") }
-  return { ...result, body: { ...result.body, messages }, note }
-}
-
-function once(input: RewriteInput): Rewrite {
-  const result = withAttachedHandoffs(strategize(input))
-  return wantsRecall(input) ? applyRecall(input, result) : result
-}
-
-function sentEstimate(result: Rewrite): number {
-  return Math.round(sizeOf(result.body) * (result.threadKey ? ratioOf(result.threadKey) : DEFAULT_RATIO))
-}
-
-function overLimit(input: RewriteInput, result: Rewrite): boolean {
-  if (!input.window || !result.threadKey || input.mayCompact === false || result.note.action === "pass") return false
-  return sentEstimate(result) > input.config.compactAt * input.window
-}
-
-function carried(from: Rewrite, to: Rewrite): Rewrite {
-  const actionsOf = ({ note }: Rewrite) => (note.actions as string[] | undefined) ?? []
-  const actions = [...actionsOf(from), ...actionsOf(to)]
-  if (!from.note.compacted && actions.length === 0) return to
-  return { ...to, note: { ...to.note, compacted: Boolean(to.note.compacted || from.note.compacted), actions } }
-}
+const toCore = ({ sessionId, transcript, ...input }: RewriteInput): CoreInput => ({ ...input, threadId: sessionId, store: fileStore, sources: sessionSources(transcript, input.config, sessionId) })
 
-export function rewrite(input: RewriteInput, before?: Rewrite): Rewrite {
-  const result = before ? carried(before, once(input)) : once(input)
-  if (!overLimit(input, result)) return result
-  const recut = carried(result, once(input))
-  if (!overLimit(input, recut)) return { ...recut, note: { ...recut.note, fit: "recut" } }
-  const bare = carried(recut, once({ ...input, config: { ...input.config, eval: { ...input.config.eval, recall: false } } }))
-  return { ...bare, note: { ...bare.note, fit: overLimit(input, bare) ? "over" : "bare", fitEstimate: sentEstimate(bare) } }
-}
+export const rewrite = (input: RewriteInput): Rewrite => rewriteCore(toCore(input))
 
-export async function prefetchRecall(input: RewriteInput, rerank: Reranker): Promise<Prefetched | null> {
-  if (!wantsRecall(input)) return null
-  const recalled = await prefetchFor(input, strategize(input), rerank)
-  return recalled === null ? null : { reranked: rerank !== identityReranker, rerankError: null, recalled }
-}
+export const prefetchRecall = (input: RewriteInput, rerank: Reranker) => prefetchCore(toCore(input), rerank)
 
-export async function rewriteAsync(input: RewriteInput): Promise<Rewrite> {
-  const model = wantsRecall(input) ? rerankModel(input.config) : null
+export async function rewriteAsync(input: RewriteInput): Promise<Rewrite & { note: { reranked?: string; rerankError?: string } }> {
+  const core = toCore(input)
+  const model = core.config.mode !== "off" && (core.config.eval?.recall ?? true) ? rerankModel(input.config) : null
   const key = model ? rerankKey() : null
-  if (!model || !key) return rewrite(input)
+  if (!model || !key) return rewriteCore(core)
   const failure = { reason: null as string | null }
   const reranker = jevReranker(model, key, input.config.eval?.rerank?.endpoint, (reason) => {
     failure.reason = reason
   })
-  const first = strategize(input)
-  const recalled = await prefetchFor(input, first, reranker)
-  const final = rewrite(input, first)
-  if (recalled === null) return final
+  const result = await rewriteCoreAsync(core, reranker)
+  if (result.note.recalled === undefined) return result
   const outcome = failure.reason ? { rerankError: `${model}: ${failure.reason}` } : { reranked: model }
-  return { ...final, note: { ...final.note, recalled, ...outcome } }
+  return { ...result, note: { ...result.note, ...outcome } }
 }

package/src/proxy/server.ts

@@ -1,17 +1,16 @@
 import path from "node:path"
 import { type Config, claudeConfigDirs, readConfig, settingsFor } from "../config"
-import { type Body, wellFormed } from "../context/content"
+import type { Body } from "../context/content"
+import { contextFingerprint } from "../core/fingerprint"
+import { claudeCodeSessionId, createForwarder, type Prepared } from "../core/forward"
 import { findSessionFile } from "../transcript/locate"
 import { prepareTranscript } from "../transcript/search"
 import { VERSION } from "../version"
 import { applyCanonical } from "./canonical"
-import { sizeOf } from "./estimate"
-import { type ContextFingerprint, contextFingerprint } from "./fingerprint"
 import { logRequest } from "./log"
 import { rewriteAsync } from "./rewrite"
 import { recordSession } from "./sessions"
-import { readState, saveState } from "./state"
-import { inputTokens, watchUsage } from "./usage"
+import { fileStore } from "./state"
 import { resolveWindow } from "./window"
 
 export interface Health {
@@ -22,48 +21,14 @@ export interface Health {
   startedAt?: number
 }
 
-interface Outgoing {
-  note: Record<string, unknown>
-  threadKey: string | null
-  bodyText: string | undefined
-  sentSize: number
-  givenSize: number
-  fixedSize: number
-  sent: (ContextFingerprint & { session: string | null }) | null
-}
-
 export const HEALTH_PATH = "/_eunoe/health"
 
 const startedAt = Date.now()
 
 const trustedHeader = (request: Request, config: Config, name: string) => (config.eval?.trustHeaders === true ? request.headers.get(`x-eunoe-${name}`) : null)
 
-const clipped = (error: string) => error.slice(0, 500)
-
-function sessionIdOf(headers: Headers, body: Body): string | null {
-  const header = headers.get("x-claude-code-session-id")
-  if (header) return header
-  try {
-    return JSON.parse(body.metadata?.user_id ?? "{}").session_id ?? null
-  } catch {
-    return null
-  }
-}
-
-function forwardableHeaders(headers: Headers) {
-  const out = new Headers(headers)
-  out.delete("content-encoding")
-  out.delete("content-length")
-  return out
-}
-
-function overflowTokens(errorText: string): number | null {
-  const tokens = Number(errorText.match(/prompt is too long[^0-9]*(\d+)\s*tokens?\s*>\s*(\d+)/i)?.[1])
-  return Number.isFinite(tokens) && tokens > 0 ? tokens : null
-}
-
-async function prepare(request: Request, body: Body, base: Config, isCountTokens: boolean) {
-  const sessionId = sessionIdOf(request.headers, body)
+async function prepareSession(request: Request, body: Body, base: Config, isCountTokens: boolean) {
+  const sessionId = claudeCodeSessionId(request.headers, body)
   const config = settingsFor(base, sessionId)
   const trusted = config.eval?.trustHeaders === true
   const transcriptOverride = trustedHeader(request, config, "transcript")
@@ -87,132 +52,29 @@ async function prepare(request: Request, body: Body, base: Config, isCountTokens
   })
 }
 
-function saveRatio(threadKey: string, ratio: (previous: number | undefined) => number) {
-  const state = readState(threadKey)
-  state.ratio = ratio(state.ratio)
-  saveState(threadKey, state)
+async function prepare(request: Request, given: Body, config: Config, mayCompact: boolean): Promise<Prepared> {
+  const canonicalFile = trustedHeader(request, config, "canonical")
+  const canonical = canonicalFile ? applyCanonical(given, canonicalFile) : null
+  const result = await prepareSession(request, canonical?.body ?? given, config, !mayCompact)
+  const context = config.eval ? { context: { session: claudeCodeSessionId(request.headers, given), ...contextFingerprint(result.body), given: contextFingerprint(given) } } : {}
+  return { ...result, note: { ...result.note, ...(canonical && { canonical: canonical.status }), ...context } }
 }
 
 interface Timing {
   countTokensTimeoutMs: number
 }
 
-const COUNT_TOKENS_ATTEMPTS = 3
-
-export async function handle(request: Request, timing: Timing = { countTokensTimeoutMs: 10_000 }): Promise<Response> {
-  const began = Date.now()
-  try {
-    return await route(request, began, timing)
-  } catch (error) {
-    logRequest({ path: new URL(request.url).pathname, status: 502, action: "upstream-unreachable", error: clipped(String(error)), ms: Date.now() - began })
-    return Response.json({ type: "error", error: { type: "api_error", message: `eunoe could not reach the API: ${String(error)}` } }, { status: 502 })
-  }
-}
-
-async function route(request: Request, began: number, timing: Timing): Promise<Response> {
-  let stalled = 0
-  let prep = 0
-  let upload = 0
-  const log = (record: Record<string, unknown>) => logRequest({ ...record, ms: Date.now() - began, prep, upload, ...(stalled > 0 && { stalled }) })
+export function handle(request: Request, timing: Timing = { countTokensTimeoutMs: 10_000 }): Promise<Response> {
   const config = readConfig()
-  const url = new URL(request.url)
-  if (url.pathname === HEALTH_PATH) return Response.json({ ok: true, version: VERSION, mode: config.mode, pid: process.pid, startedAt } satisfies Health)
-  const headers = new Headers(request.headers)
-  for (const name of ["host", "content-length", "accept-encoding"]) headers.delete(name)
-
-  let givenText: string | undefined
-  if (request.method !== "GET" && request.method !== "HEAD") givenText = await request.text()
-  upload = Date.now() - began
-  const isMessages = request.method === "POST" && url.pathname.startsWith("/v1/messages")
-  const isCountTokens = url.pathname.includes("count_tokens")
-
-  const build = async (): Promise<Outgoing> => {
-    const passthrough: Outgoing = { note: { action: "pass" }, threadKey: null, bodyText: givenText, sentSize: 0, givenSize: 0, fixedSize: 0, sent: null }
-    if (!givenText || !isMessages) return passthrough
-    try {
-      const given = JSON.parse(givenText) as Body
-      const canonicalFile = trustedHeader(request, config, "canonical")
-      const canonical = canonicalFile ? applyCanonical(given, canonicalFile) : null
-      const result = await prepare(request, canonical?.body ?? given, config, isCountTokens)
-      const sentSize = sizeOf(result.body)
-      return {
-        note: canonical ? { ...result.note, canonical: canonical.status } : result.note,
-        threadKey: result.threadKey,
-        bodyText: result.body !== given ? JSON.stringify(result.body, wellFormed) : givenText,
-        sentSize,
-        givenSize: result.body === given ? sentSize : sizeOf(given),
-        fixedSize: sizeOf({ ...result.body, messages: [] }),
-        sent: config.eval ? { session: sessionIdOf(request.headers, given), ...contextFingerprint(result.body), given: contextFingerprint(given) } : null,
-      }
-    } catch (error) {
-      return { ...passthrough, note: { action: "error-passthrough", error: String(error) } }
-    }
-  }
-  const target = config.upstream + url.pathname + url.search
-  const send = async (body: string | undefined): Promise<Response> => {
-    if (!isCountTokens) return fetch(target, { method: request.method, headers, body })
-    for (;;) {
-      try {
-        return await fetch(target, { method: request.method, headers, body, signal: AbortSignal.timeout(timing.countTokensTimeoutMs) })
-      } catch (error) {
-        const timedOut = error instanceof Error && error.name === "TimeoutError"
-        if (!timedOut || stalled + 1 >= COUNT_TOKENS_ATTEMPTS) throw error
-        stalled += 1
-      }
-    }
-  }
-
-  let out = await build()
-  prep = Date.now() - began - upload
-  let upstream = await send(out.bodyText)
-  const errorResponse = (text: string) => new Response(text, { status: upstream.status, statusText: upstream.statusText, headers: forwardableHeaders(upstream.headers) })
-
-  if (upstream.status >= 400 && out.note.action !== "pass") {
-    const errorText = await upstream.text()
-    const overflow = overflowTokens(errorText)
-    const logged = { path: url.pathname, status: upstream.status, thread: out.threadKey, ...out.note }
-    if (!overflow || !out.threadKey || out.sentSize === 0 || isCountTokens) {
-      log({ ...logged, error: clipped(errorText) })
-      return errorResponse(errorText)
-    }
-    const sentSize = out.sentSize
-    saveRatio(out.threadKey, (previous) => Math.max(previous ?? 0, overflow / sentSize) * 1.1)
-    log({ ...logged, overflow, error: clipped(errorText) })
-    out = await build()
-    out = { ...out, note: { ...out.note, healed: overflow } }
-    upstream = await send(out.bodyText)
-    if (upstream.status >= 400) {
-      const retryError = await upstream.text()
-      log({ path: url.pathname, status: upstream.status, thread: out.threadKey, ...out.note, error: clipped(retryError) })
-      return errorResponse(retryError)
-    }
-  }
-
-  const { note, threadKey, sentSize, givenSize, fixedSize, sent } = out
-  const response = new Response(upstream.body, { status: upstream.status, statusText: upstream.statusText, headers: forwardableHeaders(upstream.headers) })
-  if (note.action === "pass" || !isMessages || isCountTokens) {
-    if (note.action !== "pass" || sent) log({ path: url.pathname, status: upstream.status, ...note, ...(sent && { context: sent }) })
-    return response
-  }
-
-  return watchUsage(response, (usage) => {
-    const input = inputTokens(usage)
-    log({
-      path: url.pathname,
-      status: upstream.status,
-      thread: threadKey,
-      ...note,
-      input,
-      cacheRead: usage.cache_read_input_tokens ?? 0,
-      cacheWrite: usage.cache_creation_input_tokens ?? 0,
-      uncached: usage.input_tokens ?? 0,
-      charsIn: givenSize,
-      charsOut: sentSize,
-      charsFixed: fixedSize,
-      ...(sent && { context: sent }),
-    })
-    if (threadKey && input > 0 && sentSize > 0) saveRatio(threadKey, () => input / sentSize)
+  if (new URL(request.url).pathname === HEALTH_PATH) return Promise.resolve(Response.json({ ok: true, version: VERSION, mode: config.mode, pid: process.pid, startedAt } satisfies Health))
+  const forward = createForwarder({
+    store: fileStore,
+    upstream: config.upstream,
+    prepare: (incoming, body, mayCompact) => prepare(incoming, body, config, mayCompact),
+    record: logRequest,
+    countTokensTimeoutMs: timing.countTokensTimeoutMs,
   })
+  return forward(request)
 }
 
 export function serve(port: number) {

package/src/proxy/sessions.ts

@@ -1,6 +1,6 @@
 import { closeSync, mkdirSync, openSync, readFileSync, readSync, writeFileSync } from "node:fs"
-import type { Mode } from "../config"
 import { EUNOE_DIR, paths } from "../config/paths"
+import type { Mode } from "../core/config"
 
 interface SeenSession {
   lastSeen: string