diff --git a/src/agent.zig b/src/agent.zig
index 60edc4b1..6eec14be 100644
--- a/src/agent.zig
+++ b/src/agent.zig
@@ -225,6 +225,7 @@ pub const Agent = struct {
     compact_summary_failures: u8 = 0, // #379: consecutive complete-but-unusable (empty/truncated) summaries
     compact_pin_degraded: bool = false, // #581: pinned suffix over budget — skip+say once per unresolved turn
     compact_stall: @import("compact_cut.zig").Stall = .{}, // #706: same-boundary no-progress cuts
+    compact_backoff: @import("compact_backoff.zig").Backoff = .{}, // #1686: compactions that bought no room
     empty_completion_retries: u8 = 0, // degenerate empty completions re-asked this turn (agent_empty_completion.zig)
     length_cutoffs: u8 = 0, // cut-off or runaway replies rewound this turn (agent_length_cutoff.zig)
     named_work_nudges: u8 = 0, // named-file / zero-tool nudge this turn (named_work.zig)
diff --git a/src/agent_server_compact.zig b/src/agent_server_compact.zig
index 58a4840d..be416349 100644
--- a/src/agent_server_compact.zig
+++ b/src/agent_server_compact.zig
@@ -107,7 +107,8 @@ pub fn directive(self: anytype) ?u64 {
     if (self.provider.serverCompactUrl() != null) return null;
     if (self.server_compaction_request and manualRoute(self.provider) == .in_stream) return 1_000;
     if (!enabled(self.provider)) return null;
-    const threshold = self.provider.compactAt();
+    // #1686: while a compaction that bought no room holds the next one.
+    const threshold = @import("compact_backoff.zig").threshold(self.compact_backoff, self.provider.compactAt(), self.provider.context);
     // A request carrying the directive answers every GPT-6 `response.steer`
     // with steering_not_supported (probed on Codex). It is inert below its
     // threshold, so a root turn that can steer leaves it off until the
diff --git a/src/agent_turn.zig b/src/agent_turn.zig
index 09666286..97f4fcc9 100644
--- a/src/agent_turn.zig
+++ b/src/agent_turn.zig
@@ -2,6 +2,7 @@
 
 const std = @import("std");
 const Agent = @import("agent.zig").Agent;
+const main_mod = @import("main.zig");
 const empty_completion = @import("agent_empty_completion.zig");
 const turn_inbox = @import("turn_inbox.zig");
 
@@ -46,7 +47,9 @@ pub fn run(self: *Agent, root_turn_prepared: *std.atomic.Value(bool)) anyerror![
         // meter (last_context_tokens, server-reported) catches up. Estimate the
         // full input locally and compact BEFORE sending so we never ship an
         // over-cap request (codex run_pre_sampling_compact / opencode isOverflow).
-        if (self.inputOverCompactThreshold() and !readBatchFirst(self)) {
+        const over_compact = self.inputOverCompactThreshold();
+        @import("compact_backoff.zig").step(&self.compact_backoff, self.history_rewrites, over_compact);
+        if (over_compact and !readBatchFirst(self) and !compactionHeld(self)) {
             // Bracket the compaction with codex-WS resets, mirroring how every
             // other compactOrRecover call site is bookended by runTurn's
             // closeCodexWs (299 + defer 300). Mid-turn the WS can be live with a
@@ -103,6 +106,22 @@ pub fn run(self: *Agent, root_turn_prepared: *std.atomic.Value(bool)) anyerror![
     }
 }
 
+/// #1686: the previous compaction left the context over the line; wait for
+/// real growth (or the 95% wall) instead of compacting on every step.
+fn compactionHeld(self: *Agent) bool {
+    const context = self.effectiveContextTokens();
+    const backoff = @import("compact_backoff.zig");
+    switch (backoff.gate(&self.compact_backoff, context, self.provider.context, self.provider.nearContextLimit(context))) {
+        .compact => return false,
+        .hold => return true,
+        .hold_started => {
+            if (self.tracer) |tr| tr.note("compact", "held: the last compaction left the context over the line");
+            if (!main_mod.json_mode) self.say("[compaction freed too little room; compacting again once the context grows]\n", .{}) catch {};
+            return true;
+        },
+    }
+}
+
 /// Over the compaction line, but the history ends on a tool batch the model
 /// has not read and almost everything else is that batch. A compaction keeps
 /// the batch verbatim (compact_cut.unreadToolBatch), so it could only
diff --git a/src/compact_backoff.zig b/src/compact_backoff.zig
new file mode 100644
index 00000000..2739e07e
--- /dev/null
+++ b/src/compact_backoff.zig
@@ -0,0 +1,95 @@
+//! #1686: an automatic compaction that leaves the context over the compaction
+//! point fires again on the very next step. When the provider's compacted
+//! state, or a summary that must keep a long live turn verbatim, cannot get
+//! under the line, every step paid a compaction and the model reread what it
+//! had just lost. A compaction that bought no room holds the next one until
+//! the context grows by a twentieth of the window, or reaches the 95% wall
+//! where recovery may trim. The in-stream server directive follows the hold.
+
+const std = @import("std");
+
+pub const Backoff = struct {
+    rewrites: u32 = 0, // history_rewrites at the previous step
+    steps: u32 = std.math.maxInt(u32), // model steps since history was last rewritten
+    resume_at: u64 = 0, // context that ends the hold; 0 when not holding
+};
+
+pub const Gate = enum { compact, hold, hold_started };
+
+/// Once per model step, before the pre-send gate. `over`: the context is at or
+/// past the compaction point.
+pub fn step(b: *Backoff, rewrites: u32, over: bool) void {
+    if (rewrites != b.rewrites) {
+        b.rewrites = rewrites;
+        b.steps = 0;
+    } else b.steps +|= 1;
+    if (!over) b.resume_at = 0; // real room: compaction worked
+}
+
+/// The pre-send gate found `context` over the compaction point.
+pub fn gate(b: *Backoff, context: u64, window: u64, near_limit: bool) Gate {
+    if (near_limit) {
+        b.resume_at = 0;
+        return .compact;
+    }
+    if (b.resume_at != 0) {
+        if (context < b.resume_at) return .hold;
+        b.resume_at = 0;
+        return .compact;
+    }
+    // History was rewritten this step or the one before and is still over.
+    if (b.steps > 1) return .compact;
+    b.resume_at = context +| @max(window / 20, 1);
+    return .hold_started;
+}
+
+/// The in-stream compaction threshold while holding: the server must not
+/// compact on every request either. Never past the 95% wall.
+pub fn threshold(b: Backoff, base: u64, window: u64) u64 {
+    if (b.resume_at <= base) return base;
+    return @max(base, @min(b.resume_at, window - window / 20));
+}
+
+test "a compaction that leaves the context over the line holds the next one" {
+    var b: Backoff = .{};
+    const window: u64 = 800_000;
+    // Long-running session: no recent rewrite, so the first trip compacts.
+    b.steps = 10;
+    step(&b, 0, true);
+    try std.testing.expectEqual(Gate.compact, gate(&b, 645_000, window, false));
+    // The compaction rewrote history, and the next step is still over.
+    step(&b, 1, true);
+    try std.testing.expectEqual(Gate.hold_started, gate(&b, 641_000, window, false));
+    try std.testing.expectEqual(@as(u64, 681_000), b.resume_at);
+    try std.testing.expectEqual(@as(u64, 681_000), threshold(b, 640_000, window));
+    step(&b, 1, true);
+    try std.testing.expectEqual(Gate.hold, gate(&b, 660_000, window, false));
+    // Grown by a twentieth of the window: compact once more.
+    step(&b, 1, true);
+    try std.testing.expectEqual(Gate.compact, gate(&b, 682_000, window, false));
+    try std.testing.expectEqual(@as(u64, 640_000), threshold(b, 640_000, window));
+}
+
+test "the wall and real relief both end a hold" {
+    var b: Backoff = .{};
+    const window: u64 = 800_000;
+    step(&b, 1, true);
+    try std.testing.expectEqual(Gate.hold_started, gate(&b, 750_000, window, false));
+    try std.testing.expectEqual(@as(u64, 760_000), threshold(b, 640_000, window)); // capped at 95%
+    step(&b, 1, true);
+    try std.testing.expectEqual(Gate.compact, gate(&b, 765_000, window, true));
+
+    step(&b, 2, true);
+    try std.testing.expectEqual(Gate.hold_started, gate(&b, 650_000, window, false));
+    step(&b, 2, false);
+    try std.testing.expectEqual(@as(u64, 0), b.resume_at);
+    try std.testing.expectEqual(@as(u64, 640_000), threshold(b, 640_000, window));
+}
+
+test "a compaction that bought several steps is not a stall" {
+    var b: Backoff = .{};
+    step(&b, 1, false);
+    step(&b, 1, false);
+    step(&b, 1, true);
+    try std.testing.expectEqual(Gate.compact, gate(&b, 641_000, 800_000, false));
+}
diff --git a/src/main.zig b/src/main.zig
index c77ae41a..9779da97 100644
--- a/src/main.zig
+++ b/src/main.zig
@@ -570,7 +570,7 @@ test { // ── Unit tests (`zig build test`): pull in tests from imported modu
     _ = @import("agent_gateway_retry.zig"); // #1019: flake vs overflow classification must stay reachable
     _ = @import("agent_responses.zig"); // #1019: type:error frame message extraction
     _ = @import("agent_server_compact.zig"); // server-side autocompact (codex Responses)
-    _ = .{ @import("compact_status.zig"), @import("compact_mix.zig") }; // compact status + GRAFF_COMPACT_MIX mixture arm
+    _ = .{ @import("compact_status.zig"), @import("compact_mix.zig"), @import("compact_backoff.zig") }; // compact status + GRAFF_COMPACT_MIX mixture arm + #1686 backoff
     _ = .{ @import("agent_clef_compact.zig"), @import("agent_clef_compact_tests.zig"), @import("agent_clef_archive.zig") }; // gateway /v1/compact client (Clef pruning) + archive mode
     _ = @import("agent_compact_recover.zig"); // compaction last-resort recovery (split for the ceiling)
     _ = @import("agent_request_search_tests.zig");