fix: persist modelContextLimit and skip min-threshold checks when unknown

Root cause of 'restart then immediately asked to compress': opencode runs
the chat.message transform before the system.prompt hook, so on the first
message after a restart state.modelContextLimit was not yet cached and
not persisted. resolveContextTokenLimit then failed for percentage
limits and overMinLimit fell back to unconditional true, injecting
compression nudges on every turn for a normal ~300K context on 1M models.

Fix:
- persist modelContextLimit in session state (save + load) so restarts
  restore it before the first chat.message transform
- overMinLimit now returns false when the limit cannot be resolved
  (skip nudge instead of unconditionally triggering); the real model
  limit is still enforced by the API layer

Tests: 109/109 pass (added min-threshold skip, 300K no-false-alarm,
persistence round-trip and sync-block regressions).

Author: Neko-Yukari

lib/messages/inject/utils.ts

@@ -154,7 +154,11 @@ export function isContextOverLimits(
     const currentTokens = getCurrentTokenUsage(state, messages)
 
     const overMaxLimit = maxContextLimit === undefined ? false : currentTokens > maxContextLimit
-    const overMinLimit = minContextLimit === undefined ? true : currentTokens >= minContextLimit
+    // minContextLimit 无法解析（如重启后第一轮 modelContextLimit 尚未缓存/持久化）时
+    // 不能无条件触发：在 1M 模型上 fallback 会把 300K 的正常上下文误判为超限，
+    // 每轮注入压缩提醒。此时跳过 nudge，等阈值可用后再正常工作。
+    const overMinLimit =
+        minContextLimit === undefined ? false : currentTokens >= minContextLimit
 
     return {
         overMaxLimit,

lib/state/persistence.ts

@@ -36,6 +36,7 @@ export interface PersistedNudges {
 export interface PersistedSessionState {
     sessionName?: string
     manualMode?: boolean
+    modelContextLimit?: number
     prune: PersistedPrune
     nudges: PersistedNudges
     stats: SessionStats
@@ -90,6 +91,10 @@ export async function saveSessionState(
         const state: PersistedSessionState = {
             sessionName: sessionName,
             manualMode: !!sessionState.manualMode,
+            // modelContextLimit 必须持久化：opencode 的 system.prompt hook 在
+            // chat.message hook 之后才缓存该值，重启后第一轮注入 nudge 时内存中
+            // 尚未缓存 → resolveContextTokenLimit 失败 → 无法按百分比解析阈值。
+            modelContextLimit: sessionState.modelContextLimit ?? undefined,
             prune: {
                 tools: Object.fromEntries(sessionState.prune.tools),
                 messages: serializePruneMessagesState(sessionState.prune.messages),

lib/state/state.ts

@@ -171,6 +171,13 @@ export async function ensureSessionInitialized(
         state.manualMode = persisted.manualMode ? "active" : false
     }
 
+    // 重启后第一轮 chat.message hook 先于 system.prompt hook 运行，
+    // 若不恢复持久化的 modelContextLimit，阈值无法按百分比解析，
+    // 导致 1M 模型上 300K 的正常上下文在重启后立即触发压缩提醒。
+    if (typeof persisted.modelContextLimit === "number") {
+        state.modelContextLimit = persisted.modelContextLimit
+    }
+
     state.prune.tools = loadPruneMap(persisted.prune.tools)
     state.prune.messages = loadPruneMessagesState(persisted.prune.messages)
     state.nudges.contextLimitAnchors = new Set<string>(persisted.nudges.contextLimitAnchors || [])

package-lock.json

@@ -7,7 +7,10 @@ import { createSessionState, type WithParts } from "../lib/state"
 import type { CompressionBlock } from "../lib/state"
 import { getCurrentTokenUsage } from "../lib/token-utils"
 
-function buildConfig(maxContextLimit: number, minContextLimit = 1): PluginConfig {
+function buildConfig(
+    maxContextLimit: number | `${number}%`,
+    minContextLimit: number | `${number}%` = 1,
+): PluginConfig {
     return {
         enabled: true,
         debug: false,
@@ -298,3 +301,39 @@ test("isContextOverLimits does not extend the max threshold when summaryBuffer i
 
     assert.equal(overLimit.overMaxLimit, true)
 })
+
+
+test("isContextOverLimits skips min threshold when modelContextLimit is unknown", () => {
+    // 回归：modelContextLimit 未缓存/未持久化（如重启后第一轮）时，
+    // 修复前 overMinLimit 无条件 true（每轮注入压缩提醒）；
+    // 修复后应跳过判定，避免 1M 模型上 300K 正常上下文被误判。
+    const messages = buildCompactedMessages()
+    messages.push(buildPostCompactionAssistantMessage())
+    const state = createSessionState() // modelContextLimit = undefined
+
+    const pctConfig = buildConfig("85%", "60%")
+    const result = isContextOverLimits(pctConfig, state, undefined, undefined, messages)
+    assert.equal(result.overMinLimit, false)
+    assert.equal(result.overMaxLimit, false)
+})
+
+test("isContextOverLimits does not force compression for large-but-normal context when limit is unknown", () => {
+    // 关键回归：1M 模型上 300K 上下文（30%）在 modelContextLimit 未知时
+    // 绝不能触发强制压缩警告（修复前 min 侧 fallback 误判导致误压缩）。
+    const messages = buildCompactedMessages()
+    messages.push(buildPostCompactionAssistantMessage())
+    const state = createSessionState()
+
+    const lastMsg = messages[messages.length - 1]
+    ;(lastMsg.info as any).tokens = {
+        input: 300000,
+        output: 500,
+        reasoning: 0,
+        cache: { read: 100, write: 0 },
+    }
+
+    const pctConfig = buildConfig("85%", "60%")
+    const result = isContextOverLimits(pctConfig, state, undefined, undefined, messages)
+    assert.equal(result.overMaxLimit, false)
+    assert.equal(result.overMinLimit, false)
+})