diff --git a/lib/messages/inject/utils.ts b/lib/messages/inject/utils.ts
index 6d35e4c5..0c8b3c37 100644
--- a/lib/messages/inject/utils.ts
+++ b/lib/messages/inject/utils.ts
@@ -84,13 +84,19 @@ export function getModelInfo(messages: WithParts[]): LastUserModelContext {
     }
 }
 
+interface ResolvedContextLimit {
+    /** The user configured a value, even if it could not be turned into a token count. */
+    configured: boolean
+    tokens: number | undefined
+}
+
 function resolveContextTokenLimit(
     config: PluginConfig,
     state: SessionState,
     providerId: string | undefined,
     modelId: string | undefined,
     threshold: "max" | "min",
-): number | undefined {
+): ResolvedContextLimit {
     const parseLimitValue = (limit: number | `${number}%` | undefined): number | undefined => {
         if (limit === undefined) {
             return undefined
@@ -120,13 +126,13 @@ function resolveContextTokenLimit(
         const providerModelId = `${providerId}/${modelId}`
         const modelLimit = modelLimits[providerModelId]
         if (modelLimit !== undefined) {
-            return parseLimitValue(modelLimit)
+            return { configured: true, tokens: parseLimitValue(modelLimit) }
         }
     }
 
     const globalLimit =
         threshold === "max" ? config.compress.maxContextLimit : config.compress.minContextLimit
-    return parseLimitValue(globalLimit)
+    return { configured: globalLimit !== undefined, tokens: parseLimitValue(globalLimit) }
 }
 
 export function isContextOverLimits(
@@ -147,14 +153,28 @@ export function isContextOverLimits(
         "max",
     )
     const maxContextLimit =
-        resolvedMaxContextLimit === undefined
+        resolvedMaxContextLimit.tokens === undefined
             ? undefined
-            : resolvedMaxContextLimit + summaryTokenExtension
-    const minContextLimit = resolveContextTokenLimit(config, state, providerId, modelId, "min")
+            : resolvedMaxContextLimit.tokens + summaryTokenExtension
+    const resolvedMinContextLimit = resolveContextTokenLimit(
+        config,
+        state,
+        providerId,
+        modelId,
+        "min",
+    )
     const currentTokens = getCurrentTokenUsage(state, messages)
 
     const overMaxLimit = maxContextLimit === undefined ? false : currentTokens > maxContextLimit
-    const overMinLimit = minContextLimit === undefined ? true : currentTokens >= minContextLimit
+
+    // A threshold that was configured but cannot be evaluated yet (a percentage
+    // before the host reports the model context window, or an unparseable value)
+    // must not behave as if no threshold exists. Treating it as "over the lower
+    // limit" started compression far below the value the user asked for (#548).
+    const overMinLimit = resolvedMinContextLimit.configured
+        ? resolvedMinContextLimit.tokens !== undefined &&
+          currentTokens >= resolvedMinContextLimit.tokens
+        : true
 
     return {
         overMaxLimit,
diff --git a/tests/nudge-thresholds.test.ts b/tests/nudge-thresholds.test.ts
new file mode 100644
index 00000000..21613a31
--- /dev/null
+++ b/tests/nudge-thresholds.test.ts
@@ -0,0 +1,131 @@
+import assert from "node:assert/strict"
+import test from "node:test"
+import { mkdtempSync, rmSync } from "node:fs"
+import { join } from "node:path"
+import { tmpdir } from "node:os"
+import type { WithParts } from "../lib/state"
+import type { PluginConfig } from "../lib/config"
+
+const root = mkdtempSync(join(tmpdir(), "dcp-nudges-"))
+process.env.XDG_DATA_HOME = join(root, "data")
+process.env.XDG_CONFIG_HOME = join(root, "config")
+process.env.OPENCODE_CONFIG_DIR = join(root, "config", "opencode")
+test.after(() => rmSync(root, { recursive: true, force: true }))
+
+const { createSessionState } = await import("../lib/state")
+const { isContextOverLimits } = await import("../lib/messages/inject/utils")
+
+function config(overrides: Record<string, unknown> = {}): PluginConfig {
+    return {
+        compress: {
+            mode: "range",
+            permission: "allow",
+            protectedTools: [],
+            protectUserMessages: false,
+            nudgeFrequency: 1,
+            ...overrides,
+        },
+        manualMode: { enabled: false, automaticStrategies: true },
+        experimental: { allowSubAgents: false },
+        turnProtection: { enabled: false, turns: 4 },
+        strategies: { deduplication: { enabled: false }, purgeErrors: { enabled: false } },
+        pruneNotification: "off",
+        protectedFilePatterns: [],
+    } as unknown as PluginConfig
+}
+
+function message(
+    sessionID: string,
+    id: string,
+    role: "user" | "assistant",
+    created: number,
+    text: string,
+): WithParts {
+    return {
+        info: {
+            id,
+            role,
+            sessionID,
+            agent: "build",
+            model: { providerID: "lab", modelID: "test" },
+            time: { created },
+            tokens:
+                role === "assistant"
+                    ? { input: 1_000, output: 100, reasoning: 0, cache: { read: 0, write: 0 } }
+                    : undefined,
+        },
+        parts: [{ id: `prt_${id}`, messageID: id, sessionID, type: "text", text }],
+    } as unknown as WithParts
+}
+
+function conversation(sessionID: string): WithParts[] {
+    return [
+        message(sessionID, "msg_1", "user", 1, "first question"),
+        message(sessionID, "msg_2", "assistant", 2, "first answer"),
+        message(sessionID, "msg_3", "user", 3, "second question"),
+        message(sessionID, "msg_4", "assistant", 4, "second answer"),
+    ]
+}
+
+// A percentage limit cannot be resolved until the host has reported the model
+// context window. While it is unknown the old code fell back to "over the lower
+// threshold", which is the opposite of what a configured threshold asks for and
+// is what made compression start at 4% of the window (#548).
+test("an unresolvable percentage threshold does not enable nudges", () => {
+    const sessionID = `ses_thresh_${Date.now()}`
+    const state = createSessionState()
+    const messages = conversation(sessionID)
+    const settings = config({ minContextLimit: "15%", maxContextLimit: "80%" })
+
+    // The model context window has not been reported yet.
+    state.modelContextLimit = undefined
+
+    const limits = isContextOverLimits(settings, state, "lab", "test", messages)
+
+    assert.equal(limits.overMinLimit, false, "percentage threshold silently became permissive")
+    assert.equal(limits.overMaxLimit, false)
+})
+
+test("an unparseable threshold does not enable nudges", () => {
+    const sessionID = `ses_thresh_bad_${Date.now()}`
+    const state = createSessionState()
+    const messages = conversation(sessionID)
+    const settings = config({ minContextLimit: "not-a-number", maxContextLimit: "80%" })
+    state.modelContextLimit = 1_000_000
+
+    const limits = isContextOverLimits(settings, state, "lab", "test", messages)
+
+    assert.equal(limits.overMinLimit, false)
+})
+
+test("a resolved threshold is still honoured in both directions", () => {
+    const sessionID = `ses_thresh_ok_${Date.now()}`
+    const messages = conversation(sessionID)
+
+    const below = isContextOverLimits(
+        config({ minContextLimit: "50%", maxContextLimit: "80%" }),
+        stateForThreshold(sessionID, "50%"),
+        "lab",
+        "test",
+        messages,
+    )
+    assert.equal(below.overMinLimit, false, "threshold fired below the configured value")
+    assert.equal(below.overMaxLimit, false)
+
+    const above = isContextOverLimits(
+        config({ minContextLimit: 1, maxContextLimit: "80%" }),
+        stateForThreshold(sessionID, "50%"),
+        "lab",
+        "test",
+        messages,
+    )
+    assert.equal(above.overMinLimit, true)
+    assert.equal(above.overMaxLimit, false)
+})
+
+function stateForThreshold(sessionID: string, _label: string) {
+    const state = createSessionState()
+    state.sessionId = sessionID
+    state.modelContextLimit = 1_000_000
+    return state
+}