diff --git a/CLAUDE.md b/CLAUDE.md
index 40c0b57..16c815c 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -200,6 +200,23 @@ read them à la carte.
 In-flight or recent context that hasn't earned a section yet.
 Prune freely.
 
+- **Just landed:** proactive context-size compaction. `turn::
+  drive_turn` now has two compaction paths. The old *reactive*
+  one (claude prints `Prompt is too long` → `/compact` + retry)
+  is unchanged — by then the session is already past the window
+  and no turn can run on it. The new *proactive* one fires after
+  a clean turn whose last-inference context
+  (`Bus::last_ctx_usage().context_tokens()`) crossed a watermark:
+  `drive_turn` injects one synthetic notes-checkpoint turn
+  (`CHECKPOINT_PROMPT` — "flush durable state into /state now")
+  then runs `/compact`, so the agent can persist in-flight state
+  before the detail collapses into a summary. Watermark is
+  `HIVE_COMPACT_WATERMARK_TOKENS` (default 150k, ~75% of a 200k
+  window; `0` disables). New `maybe_checkpoint_and_compact`
+  helper, all best-effort (a failed checkpoint/compact is a Note,
+  never fails the turn). Both ag3nt + m1nd loops get it for free
+  via the shared `drive_turn`. `docs/turn-loop.md` gained a
+  Compaction section.
 - **Just landed:** dashboard side panel + forge-linked
   config/approvals + per-bucket model stats. (a) Long content
   (file previews, approval diffs, journald logs) opens in a
@@ -608,8 +625,8 @@ Prune freely.
   agent_web_port + collision banner + spawn/rebuild preflight,
   browser notifications, focus-preserving refresh, generalised
   <details data-restore-key> survival, prompt-on-submit pattern.
-- **Open threads:** two-step spawn, notes compaction,
-  unprivileged containers, Bash allow-list, xterm.js. The
+- **Open threads:** two-step spawn, unprivileged
+  containers, Bash allow-list, xterm.js. The
   deployment / gateway / privsep cluster is tracked as
   `area:ops` forge issues. (Landed since this note was first written:
   extra per-agent MCP servers, per-agent send allow-list,
diff --git a/docs/turn-loop.md b/docs/turn-loop.md
index 718bcc0..bfcfdbc 100644
--- a/docs/turn-loop.md
+++ b/docs/turn-loop.md
@@ -18,8 +18,9 @@ Each agent harness (`hive-ag3nt serve` or `hive-m1nd serve`) runs:
    over stdin.
 5. Stream stdout (JSON lines) into the bus as
    `LiveEvent::Stream(value)`. Pump stderr as `Note`.
-6. Wait for claude to exit. On `Prompt is too long`, run `/compact`
-   on the session once and retry the turn.
+6. Wait for claude to exit. Compaction is two-pronged — *reactive*
+   on `Prompt is too long` and *proactive* on a context watermark
+   (see [Compaction](#compaction) below).
 7. Emit `LiveEvent::TurnEnd { ok, note }`. Sleep `poll_ms` to avoid
    tight loops on transient failures.
 
@@ -43,14 +44,40 @@ override path: `HYPERHIVE_MODEL_FILE` env var for tests.
 `--continue` keeps a persistent session per agent (claude stores
 sessions in `~/.claude/projects/`, which is bind-mounted
 persistently). Auto-compact and auto-memory are disabled via
-`--settings` because hyperhive owns compaction (`/compact` on
-overflow, retry once; operator can also force one via `/api/compact`).
+`--settings` because hyperhive owns compaction — see
+[Compaction](#compaction) below.
 A one-shot `--continue` suppression is available via
 `POST /api/new-session` (or `/new-session` slash command in the
 per-agent terminal) — `Bus::take_skip_continue()` flips an
 `AtomicBool` once per turn, the next claude invocation drops
 `--continue`, every subsequent turn resumes normal behaviour.
 
+### Compaction
+
+claude's own in-session auto-compact is off (`--settings`); hyperhive
+owns it explicitly in `turn::drive_turn`. There are two triggers:
+
+- **Reactive** — claude-code prints `Prompt is too long` (the
+  `PROMPT_TOO_LONG_MARKER`). The session is *already* past the context
+  window, so no turn can run on it — `drive_turn` runs `/compact`
+  straight away and retries the same wake-up prompt once. No
+  notes-checkpoint turn is possible here: the detail is gone.
+- **Proactive** — a turn finishes cleanly but the last inference's
+  context size (`Bus::last_ctx_usage().context_tokens()`) is at or
+  above a watermark. While the session is still healthy, `drive_turn`
+  injects one synthetic *notes-checkpoint* turn (`CHECKPOINT_PROMPT`
+  — "context is filling up, flush durable state into `/state` now")
+  and *then* runs `/compact`. This gives the agent a chance to
+  persist in-flight task state, decisions, and file paths before the
+  conversation detail collapses into a summary.
+
+The watermark is `HIVE_COMPACT_WATERMARK_TOKENS` (default `150_000`,
+~75% of a 200k window); set it to `0` to disable proactive compaction
+entirely (the reactive path always applies). The proactive path is
+best-effort — a failed checkpoint turn or `/compact` is surfaced as a
+`Note` but never fails the turn that already succeeded. The operator
+can also force a compaction any time via `/api/compact`.
+
 The child runs with `cwd = /state` (when the bind exists; falls
 back to the parent's cwd in dev), so any relative path in a tool
 call (`Read foo.md`, `Bash ls`, `Write notes.md`) lands in the
diff --git a/hive-ag3nt/src/turn.rs b/hive-ag3nt/src/turn.rs
index 99ca80e..3cb28fc 100644
--- a/hive-ag3nt/src/turn.rs
+++ b/hive-ag3nt/src/turn.rs
@@ -34,6 +34,33 @@ const CLAUDE_SETTINGS: &str = include_str!("../prompts/claude-settings.json");
 /// claude exit with a useful error in the live view.
 const PROMPT_TOO_LONG_MARKER: &str = "Prompt is too long";
 
+/// Token watermark for *proactive* compaction. Once a turn finishes with
+/// the last inference's context size at or above this many tokens,
+/// `drive_turn` runs one dedicated notes-checkpoint turn (so the agent
+/// can flush durable state into `/state`) and then `/compact` — while the
+/// session is still healthy enough to run a turn at all. This is distinct
+/// from the reactive `PROMPT_TOO_LONG_MARKER` path, which only fires once
+/// the session is *already* past the window: at that point no turn can
+/// run on it, so the reactive path just compacts + retries with no
+/// checkpoint. Default is ~75% of a 200k-token window; override via
+/// `HIVE_COMPACT_WATERMARK_TOKENS`, or set that to `0` to disable
+/// proactive compaction entirely (the reactive path always applies).
+const DEFAULT_COMPACT_WATERMARK_TOKENS: u64 = 150_000;
+
+/// Synthetic wake prompt for the proactive notes-checkpoint turn. Not an
+/// inbox message — the harness injects it directly so the agent gets one
+/// turn to persist durable state before `/compact` collapses the
+/// turn-by-turn history into a summary.
+const CHECKPOINT_PROMPT: &str = "[system] Context checkpoint — no inbox message to handle.\n\n\
+Your conversation context has grown large and the harness is about to run `/compact`, \
+which collapses the detailed turn-by-turn history into a short summary. Anything you \
+do not persist now is effectively lost after the next turn.\n\n\
+Use THIS turn to flush anything worth keeping into your durable `/state` files: update \
+your notes / CLAUDE.md / TODO.md with in-flight task state, decisions made, important \
+file paths, and whatever you would need to resume cleanly with only a summary of this \
+conversation to go on. Do not start new work or reply to anyone — just write your notes \
+and end the turn.";
+
 /// The set of files claude reads on every invocation: the MCP server
 /// config (`--mcp-config`), static settings (`--settings`), and the
 /// pre-rendered role/tools system prompt (`--system-prompt-file`).
@@ -129,10 +156,32 @@ pub enum TurnOutcome {
     Failed(anyhow::Error),
 }
 
-/// Drive one turn end-to-end, transparently compacting + retrying once on
-/// `Prompt is too long`. Both the sub-agent and manager loops call this.
+/// Resolve the proactive-compaction watermark: `HIVE_COMPACT_WATERMARK_TOKENS`
+/// if set to a valid integer, else `DEFAULT_COMPACT_WATERMARK_TOKENS`. A
+/// value of `0` disables proactive compaction.
+fn compact_watermark_tokens() -> u64 {
+    std::env::var("HIVE_COMPACT_WATERMARK_TOKENS")
+        .ok()
+        .and_then(|s| s.trim().parse::<u64>().ok())
+        .unwrap_or(DEFAULT_COMPACT_WATERMARK_TOKENS)
+}
+
+/// Drive one turn end-to-end. Two compaction paths layer on top of the
+/// raw `run_turn`:
+///
+/// - **Reactive** — `run_turn` returns `PromptTooLong`: the session is
+///   already past the context window and *no* turn can run on it, so we
+///   compact immediately and retry the same wake-up prompt once. No
+///   notes-checkpoint turn is possible here — the detail is gone.
+/// - **Proactive** — the turn finished cleanly but its context size has
+///   crept past the watermark: while the session is still healthy we
+///   give the agent one dedicated turn to checkpoint its `/state` notes,
+///   then compact. This keeps a later turn from hitting the reactive
+///   path (where there is no chance to save anything first).
+///
+/// Both the sub-agent and manager loops call this.
 pub async fn drive_turn(prompt: &str, files: &TurnFiles, bus: &Bus) -> TurnOutcome {
-    match run_turn(prompt, files, bus).await {
+    let outcome = match run_turn(prompt, files, bus).await {
         TurnOutcome::PromptTooLong => {
             if let Err(e) = compact_session(files, bus).await {
                 tracing::warn!(error = %format!("{e:#}"), "compact failed");
@@ -141,6 +190,56 @@ pub async fn drive_turn(prompt: &str, files: &TurnFiles, bus: &Bus) -> TurnOutco
             run_turn(prompt, files, bus).await
         }
         other => other,
+    };
+    // Proactive: a turn just completed on a still-healthy session. If its
+    // context crossed the watermark, checkpoint + compact before a later
+    // turn overflows into the reactive path. Best-effort — never changes
+    // the outcome of the turn that already succeeded.
+    if matches!(outcome, TurnOutcome::Ok) {
+        maybe_checkpoint_and_compact(files, bus).await;
+    }
+    outcome
+}
+
+/// Proactive post-turn compaction. If the last inference's context size
+/// has crossed the watermark, run one notes-checkpoint turn so the agent
+/// can persist durable state, then `/compact`. Best-effort: a failed
+/// checkpoint or compaction is logged + surfaced as a Note but never
+/// fails the turn that already succeeded.
+async fn maybe_checkpoint_and_compact(files: &TurnFiles, bus: &Bus) {
+    let watermark = compact_watermark_tokens();
+    if watermark == 0 {
+        return; // proactive compaction disabled
+    }
+    let Some(used) = bus.last_ctx_usage().map(|u| u.context_tokens()) else {
+        return; // no usage reading yet — nothing to compare against
+    };
+    if used < watermark {
+        return;
+    }
+    bus.emit(LiveEvent::Note {
+        text: format!(
+            "context at {used} tokens (watermark {watermark}) — running a \
+             notes-checkpoint turn before /compact"
+        ),
+    });
+    // Give the agent one turn to flush durable state into /state. If the
+    // session is somehow already too far gone to run even this, fall
+    // through to compaction anyway — the checkpoint is best-effort.
+    match run_turn(CHECKPOINT_PROMPT, files, bus).await {
+        TurnOutcome::Ok => {}
+        TurnOutcome::PromptTooLong => bus.emit(LiveEvent::Note {
+            text: "checkpoint turn overflowed the window — compacting without it".into(),
+        }),
+        TurnOutcome::Failed(e) => bus.emit(LiveEvent::Note {
+            text: format!("checkpoint turn failed ({e:#}) — compacting anyway"),
+        }),
+    }
+    if let Err(e) = compact_session(files, bus).await {
+        tracing::warn!(error = %format!("{e:#}"), "post-checkpoint compact failed");
+        bus.emit(LiveEvent::Note {
+            text: format!("/compact after checkpoint failed: {e:#}"),
+        });
     }
 }