flint_lean/sentinel.rs
1//! The boundary verbs: `.flint/publish` and `.flint/sync` (plan §2.1,
2//! §2.2 — D1, D2, D3, D3.1, D12).
3//!
4//! The agent touches a file; the syncer consumes it, honors it with a
5//! real barrier (or a real sync), and answers with an ack. The
6//! discipline is `consume_inbox`'s, mirrored — exactly-once via
7//! integrate → persist → drop, idempotent between — with rename in
8//! place of CAS:
9//!
10//! 1. **Consume** — the sentinel is renamed out of the agent's reach
11//! into the state dir. Type-checked first (`S_ISREG` only: a FIFO
12//! would block the body read forever), body bounded at 64 KiB.
13//! **Settle-before-consume:** the poll never consumes while a
14//! pending record stands — a surviving pending must be honored,
15//! acked and retired first, or the consume would clobber it and
16//! orphan its nonces forever. Coalescing happens INSIDE the pending
17//! record: touches arriving during the min-interval wait append
18//! their nonces to the standing record.
19//! 2. **Honor** — run one full fused barrier (which claims the publish
20//! fence for its commit section only) or one sync.
21//! 3. **Ack** — written atomically AFTER the barrier's manifest CAS and
22//! baseline rewrite, carrying the FULL covered-nonce set: under
23//! coalescing an agent whose nonce rode behind a later touch would
24//! otherwise never see it and would re-touch in a loop, feeding the
25//! storm the rate limit exists to prevent.
26//! 4. **Retire** — the pending record is removed after the ack rename.
27//!
28//! **The uniform crash rule (D2, replacing the draft's per-crash-point
29//! matrix).** Pending-present-and-no-matching-ack is the SAME
30//! observable state for "crashed before the manifest CAS" and "crashed
31//! after step 7" — the baseline is rewritten only at step 7 — so acking
32//! from persisted state would assert publication of writes that never
33//! uploaded. On restart: pending + no matching ack ⇒ ALWAYS run a full
34//! barrier (idempotent; AdoptOwn/recent-uuids covers a half-published
35//! set), then ack with THAT barrier's installed seq. Pending + matching
36//! ack ⇒ retire only; the ack already names a real install.
37//!
38//! **A fence is a retry (design 2026-09-13 §4).** The lease is held per
39//! barrier, so a commit section deposed mid-way abandons THAT barrier:
40//! nothing was installed, the pending record stands, and the next tick
41//! honors it with a fresh claim. There is no refused-fenced ack and no
42//! fenced marker any more — the life-long lease they answered for is
43//! gone.
44
45use std::path::{Path, PathBuf};
46
47use serde::{Deserialize, Serialize};
48
49use super::control::{self, write_atomic};
50use super::state::ConflictRecord;
51use super::{now_unix, LeanError, LeanResult, Syncer};
52
53/// Bound on a sentinel body read. A larger file is truncated and the
54/// remainder ignored — never a wedge.
55const MAX_BODY: u64 = 64 * 1024;
56
57/// Bound on the covered-nonce set carried in one ack (oldest dropped).
58const MAX_NONCES: usize = 32;
59
60/// Bound on a scoped sync's entry list (§2.2 write containment).
61pub const MAX_SCOPE_ENTRIES: usize = 64;
62pub const MAX_SCOPE_ENTRY_LEN: usize = 1024;
63
64const BUDGET: &str = "sentinel-budget.json";
65
66#[derive(Debug, Clone, Copy, PartialEq, Eq)]
67pub enum Verb {
68 Publish,
69 Sync,
70}
71
72impl Verb {
73 pub fn as_str(&self) -> &'static str {
74 match self {
75 Verb::Publish => "publish",
76 Verb::Sync => "sync",
77 }
78 }
79 pub fn sentinel_name(&self) -> &'static str {
80 match self {
81 Verb::Publish => control::PUBLISH,
82 Verb::Sync => control::SYNC,
83 }
84 }
85 fn ack_name(&self) -> &'static str {
86 match self {
87 Verb::Publish => control::PUBLISH_ACK,
88 Verb::Sync => control::SYNC_ACK,
89 }
90 }
91 /// The standing pending record. `pub` because `flint-sync status`
92 /// reports on it and MUST NOT spell it a second time: it did, it
93 /// spelled it wrong ("pending-publish.json"), and the field that
94 /// answers "is my agent blocked on an ack?" therefore answered no
95 /// forever.
96 pub fn pending_name(&self) -> &'static str {
97 match self {
98 Verb::Publish => "publish.pending.json",
99 Verb::Sync => "sync.pending.json",
100 }
101 }
102 /// The consume staging name. A crash between the rename (which
103 /// removes the sentinel) and the pending write would otherwise lose
104 /// the touch: on restart the staging file is recovered into the
105 /// pending record instead.
106 fn staging_name(&self) -> &'static str {
107 match self {
108 Verb::Publish => "publish.consumed",
109 Verb::Sync => "sync.consumed",
110 }
111 }
112}
113
114/// The agent's optional JSON body.
115#[derive(Debug, Clone, Default, Deserialize)]
116struct SentinelBody {
117 #[serde(default)]
118 nonce: Option<String>,
119 #[serde(default)]
120 note: Option<String>,
121 /// `sync` only: path prefixes and exact paths (D4).
122 #[serde(default)]
123 scope: Option<Vec<String>>,
124}
125
126/// A consumed sentinel awaiting honor (`.flint-sync/<verb>.pending.json`).
127#[derive(Debug, Clone, Serialize, Deserialize)]
128pub struct PendingSentinel {
129 pub verb: String,
130 /// The latest covered touch, in nanoseconds — bare-touch agents
131 /// match their boundary on this.
132 pub consumed_mtime_unix_ns: u128,
133 pub consumed_at: u64,
134 /// EVERY coalesced nonce, oldest dropped past `MAX_NONCES`.
135 pub nonces: Vec<String>,
136 #[serde(default)]
137 pub note: Option<String>,
138 #[serde(default)]
139 pub scope: Option<Vec<String>>,
140 /// An unparsable or oversize body was honored as a bare touch — a
141 /// warning conflict record names it. Never a wedge, never a silent
142 /// drop.
143 #[serde(default)]
144 pub torn: bool,
145 /// A coalesced touch asked for the WHOLE tree (no `scope`); later
146 /// scoped touches must not narrow it (review 2026-09-12, inbox-7).
147 #[serde(default)]
148 pub whole_tree: bool,
149}
150
151/// The ack document (`.flint/<verb>.ack`).
152#[derive(Debug, Clone, Serialize, Deserialize)]
153pub struct Ack {
154 /// "ok" | "partial" | "refused-scope".
155 ///
156 /// `partial` is the honest answer when the boundary installed but a
157 /// path the agent declared is not in it (D1) —
158 /// `report.dropped` names them. An agent that treats it as failure
159 /// and re-touches is behaving correctly.
160 pub status: String,
161 pub nonces: Vec<String>,
162 pub sentinel_mtime_unix_ns: u128,
163 #[serde(skip_serializing_if = "Option::is_none")]
164 pub seq: Option<u64>,
165 #[serde(skip_serializing_if = "Option::is_none")]
166 pub manifest_etag: Option<String>,
167 /// "sentinel" | "sentinel-deferred" | "drain" | "recovered".
168 pub boundary: String,
169 pub completed_unix: u64,
170 /// Set on a refusal that is the AGENT's to fix (`refused-scope`):
171 /// what was wrong with the request, in words.
172 #[serde(default, skip_serializing_if = "Option::is_none")]
173 pub reason: Option<String>,
174 pub report: AckReport,
175}
176
177#[derive(Debug, Clone, Default, Serialize, Deserialize)]
178pub struct AckReport {
179 pub uploaded: usize,
180 pub deleted: usize,
181 pub parked: usize,
182 pub consumed: usize,
183 pub no_change: bool,
184 /// `sync` only: the applied/deleted/conflict transport (§2.2 — the
185 /// conflict report rides the ack in FULL: "never a silent winner"
186 /// must survive the file transport).
187 #[serde(default, skip_serializing_if = "Vec::is_empty")]
188 pub applied: Vec<String>,
189 #[serde(default, skip_serializing_if = "Vec::is_empty")]
190 pub conflicts: Vec<ConflictRecord>,
191 #[serde(default, skip_serializing_if = "Option::is_none")]
192 pub scope: Option<Vec<String>>,
193 /// Foreign changes seen but deferred to the inbox flow (D4).
194 #[serde(default)]
195 pub out_of_scope_foreign: usize,
196 /// Declared paths the boundary does NOT carry (a standing park).
197 /// Non-empty ⇒ `status: "partial"`: the §2.2 rule — the
198 /// conflict report rides the ack in full, never a silent loser —
199 /// generalized from `sync` to the publish verb.
200 #[serde(default, skip_serializing_if = "Vec::is_empty")]
201 pub dropped: Vec<String>,
202}
203
204/// The work meter (D3.1 — the hot-loops no-regression rule).
205///
206/// A counted budget charges a hot 2 GiB checkpoint the same one unit as
207/// a 4 KiB file, so a storming agent could drive `dirty_bytes × 60/hour`
208/// of extra upload while staying inside a green budget — a regression on
209/// the one performance sub-axis lean currently maxes (a tight local
210/// loop's republish is coalesced by cadence; sentinels un-coalesce it on
211/// demand). Metering work instead bounds sentinel-driven published bytes
212/// at `budget × whole_put_max` per hour independent of dirty-set size,
213/// and a workspace at the cap degrades to exactly today's cadence
214/// behavior — the definition of no regression.
215#[derive(Debug, Clone, Default, Serialize, Deserialize)]
216pub struct SentinelBudget {
217 /// (unix, units) charges inside the rolling hour.
218 pub charges: Vec<(u64, u64)>,
219 /// The last sentinel-honoring barrier, for the min-interval.
220 pub last_honor_unix: u64,
221}
222
223impl SentinelBudget {
224 fn prune(&mut self, now: u64) {
225 self.charges.retain(|(at, _)| now.saturating_sub(*at) < 3600);
226 }
227 pub fn spent(&self, now: u64) -> u64 {
228 self.charges
229 .iter()
230 .filter(|(at, _)| now.saturating_sub(*at) < 3600)
231 .map(|(_, u)| *u)
232 .sum()
233 }
234 pub fn remaining(&self, now: u64, budget: u64) -> u64 {
235 budget.saturating_sub(self.spent(now))
236 }
237}
238
239/// Why a standing pending sentinel is not being honored right now.
240/// The records a sync produced, given how many the log held before it.
241///
242/// `skip(before)` alone is only sound while the log GREW. A rotation
243/// during the sync can SHORTEN it, and skipping past a shorter list
244/// reports NO conflicts for a sync that produced some — the one answer
245/// this ack must never give. Over-reporting the survivors is the safe
246/// direction: an ack naming a conflict that predates the sync costs a
247/// reader a second look, an ack naming none costs them the conflict.
248pub(crate) fn conflicts_since(
249 before: usize,
250 after: Vec<ConflictRecord>,
251) -> Vec<ConflictRecord> {
252 if after.len() >= before {
253 after[before..].to_vec()
254 } else {
255 after
256 }
257}
258
259#[derive(Debug, Clone, Copy, PartialEq, Eq)]
260pub enum Due {
261 /// Honor it on this tick.
262 Ready,
263 /// Inside `sentinel_min_interval_secs` of the last honor: wait, and
264 /// let further touches coalesce.
265 MinInterval,
266 /// The hourly budget is exhausted: the boundary is still honored,
267 /// by the next FLOOR tick, and its ack is stamped
268 /// `sentinel-deferred`. Contents are never thinned (D1's corollary)
269 /// — cost is bounded by deferring the barrier, never by publishing
270 /// less than the boundary covers.
271 BudgetDeferred,
272}
273
274fn mtime_ns(m: &std::fs::Metadata) -> u128 {
275 m.modified()
276 .ok()
277 .and_then(|t| t.duration_since(std::time::UNIX_EPOCH).ok())
278 .map(|d| d.as_nanos())
279 .unwrap_or(0)
280}
281
282/// Read the sentinel body, bounded, and WITHOUT ever blocking on the
283/// open.
284///
285/// The type check in `consume_sentinel` is an `lstat`, and this is a
286/// second path resolution — so on its own the check is a TOCTOU: swap a
287/// FIFO in between and a plain `File::open` blocks forever waiting for a
288/// writer, wedging the poll arm and, behind it, every boundary this
289/// syncer owes (review: U23).
290///
291/// `O_NONBLOCK` closes it at the syscall rather than by winning a race:
292/// opening a writer-less FIFO returns immediately instead of blocking,
293/// and `O_NOFOLLOW` refuses a symlink swapped in for the same purpose.
294/// Both are no-ops on the regular file this is supposed to be reading.
295pub(crate) fn read_bounded(path: &Path) -> std::io::Result<(Vec<u8>, bool)> {
296 use std::io::Read;
297 use std::os::unix::fs::OpenOptionsExt;
298 let f = std::fs::File::options()
299 .read(true)
300 .custom_flags(libc::O_NONBLOCK | libc::O_NOFOLLOW)
301 .open(path)?;
302 // A FIFO that a writer DOES hold open reads EAGAIN rather than
303 // blocking; treat it as an empty body, which the torn-body rule
304 // already handles as a bare touch.
305 let meta = f.metadata()?;
306 if !meta.is_file() {
307 return Err(std::io::Error::new(
308 std::io::ErrorKind::InvalidInput,
309 "sentinel is not a regular file",
310 ));
311 }
312 let mut buf = Vec::new();
313 let mut take = f.take(MAX_BODY + 1);
314 take.read_to_end(&mut buf)?;
315 let oversize = buf.len() as u64 > MAX_BODY;
316 buf.truncate(MAX_BODY as usize);
317 Ok((buf, oversize))
318}
319
320impl Syncer {
321 fn pending_path(&self, verb: Verb) -> PathBuf {
322 self.cfg.state_dir().join(verb.pending_name())
323 }
324 fn staging_path(&self, verb: Verb) -> PathBuf {
325 self.cfg.state_dir().join(verb.staging_name())
326 }
327 fn budget_path(&self) -> PathBuf {
328 self.cfg.state_dir().join(BUDGET)
329 }
330
331 pub fn load_pending(&self, verb: Verb) -> LeanResult<Option<PendingSentinel>> {
332 let p = self.pending_path(verb);
333 if !p.exists() {
334 return Ok(None);
335 }
336 let bytes = std::fs::read(&p)?;
337 match serde_json::from_slice(&bytes) {
338 Ok(v) => Ok(Some(v)),
339 // Our own record is torn (a crash mid-rewrite): honor it as
340 // a bare touch rather than wedging on it.
341 Err(_) => Ok(Some(PendingSentinel {
342 verb: verb.as_str().into(),
343 consumed_mtime_unix_ns: 0,
344 consumed_at: now_unix(),
345 nonces: vec![],
346 note: None,
347 scope: None,
348 torn: true,
349 whole_tree: false,
350 })),
351 }
352 }
353
354 fn save_pending(&self, verb: Verb, p: &PendingSentinel) -> LeanResult<()> {
355 let bytes = serde_json::to_vec_pretty(p)
356 .map_err(|e| LeanError::State(format!("pending: {e}")))?;
357 write_atomic(&self.pending_path(verb), &bytes)
358 }
359
360 fn retire_pending(&self, verb: Verb) -> LeanResult<()> {
361 let p = self.pending_path(verb);
362 if p.exists() {
363 std::fs::remove_file(&p)?;
364 }
365 Ok(())
366 }
367
368 pub fn load_budget(&self) -> LeanResult<SentinelBudget> {
369 let p = self.budget_path();
370 if !p.exists() {
371 return Ok(SentinelBudget::default());
372 }
373 let bytes = std::fs::read(&p)?;
374 Ok(serde_json::from_slice(&bytes).unwrap_or_default())
375 }
376
377 fn save_budget(&self, b: &SentinelBudget) -> LeanResult<()> {
378 let bytes =
379 serde_json::to_vec_pretty(b).map_err(|e| LeanError::State(format!("budget: {e}")))?;
380 write_atomic(&self.budget_path(), &bytes)
381 }
382
383 /// D3.1: charge the meter for a completed sentinel honor.
384 /// `published_bytes == 0` (a no-diff honor) costs NOTHING — the
385 /// budget exists to bound work and a no-diff honor does none; the
386 /// min-interval remains its only bound.
387 pub fn charge_budget(&self, published_bytes: u64) -> LeanResult<u64> {
388 let now = now_unix();
389 let mut b = self.load_budget()?;
390 b.prune(now);
391 let units = if published_bytes == 0 {
392 0
393 } else {
394 published_bytes.div_ceil(self.cfg.whole_put_max.max(1)).max(1)
395 };
396 if units > 0 {
397 b.charges.push((now, units));
398 }
399 b.last_honor_unix = now;
400 self.save_budget(&b)?;
401 Ok(units)
402 }
403
404 /// Whether a standing pending sentinel may be honored on this tick.
405 pub fn sentinel_due(&self) -> LeanResult<Due> {
406 let now = now_unix();
407 let b = self.load_budget()?;
408 if b.last_honor_unix > 0
409 && now.saturating_sub(b.last_honor_unix) < self.cfg.sentinel_min_interval_secs
410 {
411 return Ok(Due::MinInterval);
412 }
413 if b.remaining(now, self.cfg.sentinel_hourly_budget) == 0 {
414 return Ok(Due::BudgetDeferred);
415 }
416 Ok(Due::Ready)
417 }
418
419 /// One poll tick: `lstat` exactly two fixed paths — no inotify
420 /// dependency, no directory scan, ~2 lstats/s when idle.
421 ///
422 /// Returns the verbs whose sentinel was consumed (or coalesced) on
423 /// this tick.
424 pub fn poll_sentinels(&mut self) -> LeanResult<Vec<Verb>> {
425 let mut consumed = vec![];
426 for verb in [Verb::Publish, Verb::Sync] {
427 if self.consume_sentinel(verb)? {
428 consumed.push(verb);
429 }
430 }
431 Ok(consumed)
432 }
433
434 /// Recover a consume that crashed between the rename and the
435 /// pending write. Called at startup, before the first poll.
436 pub fn recover_consume_staging(&mut self) -> LeanResult<()> {
437 for verb in [Verb::Publish, Verb::Sync] {
438 let staging = self.staging_path(verb);
439 if staging.exists() {
440 let meta = std::fs::metadata(&staging)?;
441 let (body, oversize) = read_bounded(&staging)?;
442 self.fold_into_pending(verb, mtime_ns(&meta), &body, oversize)?;
443 std::fs::remove_file(&staging)?;
444 }
445 }
446 Ok(())
447 }
448
449 pub(crate) fn consume_sentinel(&mut self, verb: Verb) -> LeanResult<bool> {
450 let path = self.control_path(verb.sentinel_name());
451 let meta = match std::fs::symlink_metadata(&path) {
452 Ok(m) => m,
453 Err(_) => return Ok(false),
454 };
455 // Type check FIRST: a FIFO would block the body read forever;
456 // a directory, socket or symlink at the sentinel path is not a
457 // touch this syncer will act on.
458 if !meta.is_file() {
459 // ONCE per process, not once per poll tick. A FIFO or dir
460 // parked at the sentinel path is a standing condition, and
461 // appending a record every 10 s grows a file that
462 // `load_conflicts` parses WHOLE — twice per sync honor, and
463 // again per status read and per scrape (review: U23). The
464 // condition is worth recording; the repetition is not.
465 let slot = format!("{}/{}", super::CONTROL_DIR, verb.sentinel_name());
466 if self.noted_not_regular.insert(slot.clone()) {
467 self.state.append_conflict(&ConflictRecord {
468 path: slot,
469 foreign_etag: String::new(),
470 preserved_key: None,
471 kind: "sentinel-not-regular-file".into(),
472 at_unix: now_unix(),
473 })?;
474 }
475 return Ok(false);
476 }
477 let ns = mtime_ns(&meta);
478 let (body, oversize) = match read_bounded(&path) {
479 Ok(v) => v,
480 Err(_) => (Vec::new(), false),
481 };
482 // The consume act: rename the sentinel out of the agent's reach.
483 // Staging first, pending second — a crash between is recovered
484 // at startup from the staging file, so a touch is never lost
485 // ACROSS A RESTART.
486 //
487 // Scope worth naming (review: U37): recovery runs ONLY at
488 // startup (`settle_pending_at_startup`, called once before the
489 // run loop). A *transient* fold failure mid-run therefore leaves
490 // a consumed touch stranded in the staging file until the
491 // process restarts, and the next consume's rename clobbers it,
492 // orphaning its nonce. Recovering in the poll arm as well would
493 // close it; nothing does today.
494 self.noted_not_regular.remove(&format!("{}/{}", super::CONTROL_DIR, verb.sentinel_name()));
495 let staging = self.staging_path(verb);
496 std::fs::rename(&path, &staging)?;
497 self.fold_into_pending(verb, ns, &body, oversize)?;
498 let _ = std::fs::remove_file(&staging);
499 Ok(true)
500 }
501
502 /// Build or coalesce into the standing pending record.
503 ///
504 /// Settle-before-consume in its positive form: an existing pending
505 /// record is never overwritten. Its nonce set grows, its covered
506 /// mtime advances to the latest touch, and its scope (sync) unions —
507 /// so the honor that eventually runs covers every coalesced touch,
508 /// and every one of their nonces appears in the ack.
509 fn fold_into_pending(
510 &mut self,
511 verb: Verb,
512 ns: u128,
513 body: &[u8],
514 oversize: bool,
515 ) -> LeanResult<()> {
516 let trimmed = body.iter().all(|b| b.is_ascii_whitespace());
517 let (parsed, torn) = if trimmed {
518 (SentinelBody::default(), false) // bare touch
519 } else if oversize {
520 (SentinelBody::default(), true)
521 } else {
522 match serde_json::from_slice::<SentinelBody>(body) {
523 Ok(v) => (v, false),
524 Err(_) => (SentinelBody::default(), true),
525 }
526 };
527 if torn {
528 self.state.append_conflict(&ConflictRecord {
529 path: format!("{}/{}", super::CONTROL_DIR, verb.sentinel_name()),
530 foreign_etag: String::new(),
531 preserved_key: None,
532 kind: "sentinel-torn-body".into(),
533 at_unix: now_unix(),
534 })?;
535 }
536 let mut pending = self.load_pending(verb)?.unwrap_or(PendingSentinel {
537 verb: verb.as_str().into(),
538 consumed_mtime_unix_ns: 0,
539 consumed_at: now_unix(),
540 nonces: vec![],
541 note: None,
542 scope: None,
543 torn: false,
544 whole_tree: false,
545 });
546 pending.consumed_mtime_unix_ns = pending.consumed_mtime_unix_ns.max(ns);
547 pending.consumed_at = now_unix();
548 pending.torn |= torn;
549 if let Some(n) = parsed.nonce {
550 let n: String = n.chars().take(128).collect();
551 if !pending.nonces.contains(&n) {
552 pending.nonces.push(n);
553 }
554 let excess = pending.nonces.len().saturating_sub(MAX_NONCES);
555 if excess > 0 {
556 pending.nonces.drain(..excess);
557 }
558 }
559 if let Some(note) = parsed.note {
560 pending.note = Some(note.chars().take(4096).collect());
561 }
562 // Scope folding (review 2026-09-12, inbox-7 / inbox-10): an
563 // UNSCOPED sync touch means the whole tree and outranks every
564 // scoped one it coalesces with, in either order — the whole-tree
565 // agent matches its nonce on this ack and must get what it asked
566 // for. Scoped touches union, and the union is NOT truncated
567 // here: an oversize or all-invalid scope is answered at the
568 // honor as `refused-scope`, never silently narrowed.
569 match parsed.scope {
570 None if matches!(verb, Verb::Sync) => {
571 pending.scope = None;
572 pending.whole_tree = true;
573 }
574 Some(scope) if !pending.whole_tree => {
575 let mut merged = pending.scope.take().unwrap_or_default();
576 for e in scope {
577 if e.len() > MAX_SCOPE_ENTRY_LEN {
578 continue;
579 }
580 if !merged.contains(&e) {
581 merged.push(e);
582 }
583 }
584 pending.scope = Some(merged);
585 }
586 _ => {}
587 }
588 self.save_pending(verb, &pending)?;
589 Ok(())
590 }
591
592 /// A boundary asked for through a door that is NOT the file
593 /// protocol — the gateway's inbox field (§2.5) or the UDS socket.
594 ///
595 /// It folds into the same pending record a `.flint/publish` touch
596 /// would, which is the whole design of the layered doors: the extra
597 /// doors are sugar over one consume path, so min-interval,
598 /// coalescing, the work-metered budget, the covered-nonce ack and
599 /// the crash rules apply to them without a second implementation
600 /// that could disagree with the first.
601 pub fn request_boundary(&mut self, nonce: &str, note: Option<String>) -> LeanResult<()> {
602 let body = serde_json::json!({ "nonce": nonce, "note": note });
603 let ns = std::time::SystemTime::now()
604 .duration_since(std::time::UNIX_EPOCH)
605 .map(|d| d.as_nanos())
606 .unwrap_or(0);
607 self.fold_into_pending(Verb::Publish, ns, body.to_string().as_bytes(), false)
608 }
609
610 fn ack_path(&self, verb: Verb) -> PathBuf {
611 self.control_path(verb.ack_name())
612 }
613
614 pub fn read_ack(&self, verb: Verb) -> Option<Ack> {
615 let bytes = std::fs::read(self.ack_path(verb)).ok()?;
616 serde_json::from_slice(&bytes).ok()
617 }
618
619 fn write_ack(&self, verb: Verb, ack: &Ack) -> LeanResult<()> {
620 let bytes =
621 serde_json::to_vec_pretty(ack).map_err(|e| LeanError::State(format!("ack: {e}")))?;
622 write_atomic(&self.ack_path(verb), &bytes)?;
623 self.trace("ack", serde_json::json!({"verb": format!("{verb:?}"), "nonces": ack.nonces, "status": ack.status,
624 "seq": ack.seq, "dropped": ack.report.dropped, "boundary": ack.boundary}));
625 Ok(())
626 }
627
628 /// Does a standing ack already answer this pending record? Used by
629 /// the restart rule: matching ⇒ retire only (the ack names a real
630 /// install); not matching ⇒ run a full barrier and ack from THAT.
631 pub(crate) fn ack_matches(&self, verb: Verb, pending: &PendingSentinel) -> bool {
632 let Some(ack) = self.read_ack(verb) else { return false };
633 // Compare values the SYNCER minted (review: U22). The ack is an
634 // ordinary file in `.flint/`, writable by every process sharing
635 // the mount, and `sentinel_mtime_unix_ns` is the agent's own
636 // file mtime — which is not monotone even without an adversary:
637 // `touch -t`, a clock step, a restored file and a tar extract
638 // all move it backwards. Deciding "has this boundary already
639 // run?" from it lets a STALE ack retire a FRESH request, and
640 // the request is then retired having never run.
641 //
642 // `consumed_at` and `completed_unix` are both minted here, in
643 // that order, so their ordering answers the question the mtime
644 // was being asked.
645 if ack.completed_unix < pending.consumed_at {
646 return false;
647 }
648 if ack.sentinel_mtime_unix_ns < pending.consumed_mtime_unix_ns {
649 return false;
650 }
651 // A bare touch carries no nonce, so `all()` over an empty set is
652 // vacuously TRUE and the timestamps become the entire test.
653 // Require the ack to name this exact touch instead.
654 if pending.nonces.is_empty() {
655 return ack.sentinel_mtime_unix_ns == pending.consumed_mtime_unix_ns;
656 }
657 pending.nonces.iter().all(|n| ack.nonces.contains(n))
658 }
659
660 /// Honor a standing pending sentinel, if one stands and is due.
661 ///
662 /// `forced` = this is a floor tick picking up a budget-deferred
663 /// boundary; the ack is stamped `sentinel-deferred`. `Ok(None)`
664 /// means nothing was owed (or it is not due yet).
665 pub async fn honor_pending(&mut self, verb: Verb, forced: bool) -> LeanResult<Option<Ack>> {
666 self.honor_pending_as(verb, forced, None).await
667 }
668
669 /// `honor_pending` under an explicit provenance stamp. The drain
670 /// needs it: it rewrites the ack to `drain`, and a manifest still
671 /// stamped `sentinel-deferred` would have the bucket and the ack
672 /// naming two different clocks for one boundary.
673 pub async fn honor_pending_as(
674 &mut self,
675 verb: Verb,
676 forced: bool,
677 source: Option<&str>,
678 ) -> LeanResult<Option<Ack>> {
679 let Some(pending) = self.load_pending(verb)? else { return Ok(None) };
680 if !forced {
681 match self.sentinel_due()? {
682 Due::Ready => {}
683 // Not honored on this tick; the floor tick will (D3).
684 Due::MinInterval | Due::BudgetDeferred => return Ok(None),
685 }
686 }
687 // The restart rule: a matching ack already names a real
688 // install — retire, never re-run.
689 if self.ack_matches(verb, &pending) {
690 self.retire_pending(verb)?;
691 return Ok(None);
692 }
693
694 let ack = match verb {
695 Verb::Publish => self.honor_publish(&pending, forced, source).await,
696 Verb::Sync => self.honor_sync(&pending, forced).await,
697 };
698 match ack {
699 Ok(ack) => {
700 self.write_ack(verb, &ack)?;
701 self.retire_pending(verb)?;
702 Ok(Some(ack))
703 }
704 Err(e) => Err(e),
705 }
706 }
707
708 async fn honor_publish(
709 &mut self,
710 pending: &PendingSentinel,
711 forced: bool,
712 source: Option<&str>,
713 ) -> LeanResult<Ack> {
714 // The DECLARED form (D1): a delete the agent made before the
715 // touch is part of the coherent point it declared, so this
716 // barrier confirms first-absence paths instead of acking a
717 // boundary that withholds them to the next floor tick.
718 // The ack and the manifest must agree on which clock published:
719 // a budget-deferred boundary reads `sentinel-deferred` in both.
720 let before = self.state.load_conflicts()?.len();
721 let report = self
722 .declared_barrier_as(
723 source.unwrap_or(if forced { "sentinel-deferred" } else { "sentinel" }),
724 )
725 .await?;
726 // Review 2026-09-12, ack-3: the boundary's own records (a
727 // `consume-dirty` the agent won, a refused removal) ride the
728 // publish ack as they ride the sync ack — AGENTS.md says so.
729 let conflicts = conflicts_since(before, self.state.load_conflicts()?);
730 let units = self.charge_budget(report.published_bytes)?;
731 let _ = units;
732 let baseline = self.state.load_baseline()?;
733 // Review 2026-09-12, inbox-1: a boundary with a standing park does
734 // NOT carry the agent's file — `ok` promised "the boundary is in
735 // the bucket". It is `partial`, and `report.dropped` names the
736 // paths.
737 Ok(Ack {
738 status: if report.parked.is_empty() { "ok".into() } else { "partial".into() },
739 nonces: pending.nonces.clone(),
740 sentinel_mtime_unix_ns: pending.consumed_mtime_unix_ns,
741 seq: report.seq,
742 manifest_etag: baseline.manifest_etag.clone(),
743 boundary: if forced { "sentinel-deferred".into() } else { "sentinel".into() },
744 completed_unix: now_unix(),
745 reason: None,
746 report: AckReport {
747 uploaded: report.uploaded.len(),
748 deleted: report.deleted.len(),
749 parked: report.parked.len(),
750 consumed: report.consumed,
751 no_change: report.no_change,
752 conflicts,
753 dropped: report.parked.clone(),
754 ..Default::default()
755 },
756 })
757 }
758
759 async fn honor_sync(&mut self, pending: &PendingSentinel, forced: bool) -> LeanResult<Ack> {
760 if let Some(reason) = Self::scope_refusal(&pending.scope) {
761 // Review 2026-09-12, ack-1 / inbox-3: the refusal used to be
762 // an ERROR the loop retried every tick — never acked, never
763 // retired, and returned ahead of the publish honor and the
764 // cadence barrier, so one `{"scope":[]}` stopped every
765 // boundary for the life of the workspace. An invalid request
766 // is the agent's to fix: answered, retired, and nothing else
767 // waits on it.
768 return Ok(Ack {
769 status: "refused-scope".into(),
770 nonces: pending.nonces.clone(),
771 sentinel_mtime_unix_ns: pending.consumed_mtime_unix_ns,
772 seq: None,
773 manifest_etag: None,
774 boundary: if forced { "sentinel-deferred".into() } else { "sentinel".into() },
775 completed_unix: now_unix(),
776 reason: Some(reason),
777 report: AckReport { scope: pending.scope.clone(), ..Default::default() },
778 });
779 }
780 let before = self.state.load_conflicts()?.len();
781 let report = self.sync_scoped(pending.scope.clone()).await?;
782 let conflicts = conflicts_since(before, self.state.load_conflicts()?);
783 // A sync publishes no bytes: it costs no budget units, only the
784 // min-interval (which it shares with publish).
785 self.charge_budget(0)?;
786 Ok(Ack {
787 status: "ok".into(),
788 nonces: pending.nonces.clone(),
789 sentinel_mtime_unix_ns: pending.consumed_mtime_unix_ns,
790 seq: Some(report.seq),
791 manifest_etag: None,
792 boundary: if forced { "sentinel-deferred".into() } else { "sentinel".into() },
793 completed_unix: now_unix(),
794 reason: None,
795 report: AckReport {
796 consumed: report.applied.len(),
797 deleted: report.deleted.len(),
798 applied: report.applied.clone(),
799 conflicts,
800 scope: pending.scope.clone(),
801 out_of_scope_foreign: report.out_of_scope_foreign,
802 ..Default::default()
803 },
804 })
805 }
806
807 /// Review 2026-09-12, ack-1: an invalid sync scope is the AGENT's
808 /// error and is answered as one (`refused-scope`), never retried.
809 /// `None` = the scope is acceptable (or absent: the whole tree).
810 fn scope_refusal(scope: &Option<Vec<String>>) -> Option<String> {
811 let raw = scope.as_ref()?;
812 if raw.len() > MAX_SCOPE_ENTRIES {
813 return Some(format!(
814 "scope names {} entries; at most {MAX_SCOPE_ENTRIES} are allowed per sync",
815 raw.len()
816 ));
817 }
818 if super::sync::Scope::new(raw).is_empty() {
819 return Some(format!(
820 "scope names {} entr{} and none is valid: each must be a relative path inside the \
821 workspace with no `.` or `..` component (an empty scope is refused, not widened \
822 to the whole tree)",
823 raw.len(),
824 if raw.len() == 1 { "y" } else { "ies" }
825 ));
826 }
827 None
828 }
829
830 /// The startup settle (the uniform crash rule). Runs before the
831 /// first poll arms: a surviving pending must be honored, acked and
832 /// retired before any fresh sentinel may be consumed.
833 pub async fn settle_pending_at_startup(&mut self) -> LeanResult<()> {
834 self.recover_consume_staging()?;
835 for verb in [Verb::Publish, Verb::Sync] {
836 let Some(pending) = self.load_pending(verb)? else { continue };
837 if self.ack_matches(verb, &pending) {
838 // Crashed after the ack, before the retire.
839 self.retire_pending(verb)?;
840 continue;
841 }
842 // Pending + no matching ack is the same observable state
843 // for crash-before-CAS and crash-after-step-7: run a full
844 // barrier and ack from ITS install, never from persisted
845 // state.
846 self.honor_pending(verb, true).await?;
847 }
848 Ok(())
849 }
850}
851
852/// What a floor tick did.
853#[derive(Debug, Default)]
854pub struct FloorOutcome {
855 pub seq: Option<u64>,
856 pub no_change: bool,
857 /// Acks written because a budget-deferred (or min-interval-held)
858 /// pending sentinel was picked up by this floor tick.
859 pub acks: Vec<Ack>,
860 pub uploaded: usize,
861 pub deleted: usize,
862 pub consumed: usize,
863 /// Why visibility is withheld right now, straight off the gauges —
864 /// so the per-tick stderr line is greppable and structured, which
865 /// until Phase 6 is the ONLY signal surface an operator has.
866 pub withheld_reason: Option<String>,
867 /// A foreign 412 parked at least one path on this tick.
868 pub parked: usize,
869 /// Carried out of the barrier only to feed the news ticker.
870 observed_etag: Option<String>,
871}
872
873impl Syncer {
874 /// Move the news ticker from what the barrier already learned (D5).
875 /// Never issues a request of its own.
876 fn ticker_from(&self, seq: Option<u64>, etag: Option<String>) -> LeanResult<()> {
877 let baseline_seq = self.state.load_baseline()?.seq;
878 // A merge that carried another writer's change into the manifest
879 // moves the baseline to that document while the change waits in
880 // this writer's queue (or, after a restart, its intent journal) for
881 // the next consume. Until then the tree is integrated only through
882 // what the ticker said while nothing waited.
883 let waiting = !self.state.load_foreign_queue()?.is_empty()
884 || !self.state.load_intent()?.installed_foreign.is_empty();
885 let integrated = if waiting {
886 baseline_seq.min(self.load_remote_seq().integrated_seq)
887 } else {
888 baseline_seq
889 };
890 self.touch_remote_seq(seq, etag, integrated)
891 }
892
893 /// One poll tick of the sentinel arm: consume what is there, then
894 /// honor if the min-interval and the work budget allow.
895 ///
896 /// In `cadence` mode the arm still consumes and honors — a sentinel
897 /// triggers a fused barrier there too; the mode's meaning is that
898 /// there is no separate citation lane, not that the verbs are dead.
899 pub async fn sentinel_tick(&mut self) -> LeanResult<Vec<Ack>> {
900 let posture = self.load_posture()?;
901 if posture.as_ref().map(|p| !p.enabled).unwrap_or(false) {
902 return Ok(vec![]);
903 }
904 self.poll_sentinels()?;
905 let mut acks = vec![];
906 // Sync before publish: a coalesced pair means "pull, then
907 // publish my coherent point" — the other order would publish
908 // against a tree the agent expected to have refreshed.
909 for verb in [Verb::Sync, Verb::Publish] {
910 match self.honor_pending(verb, false).await {
911 Ok(Some(a)) => acks.push(a),
912 Ok(None) => {}
913 Err(e) => {
914 // Review 2026-09-12, ack-1: a non-fence honor error is
915 // THIS verb's problem. Returning here left the other
916 // verb unhonoured for as long as the error recurred —
917 // for a deterministic one, forever. The pending stands
918 // and is retried; the tick goes on.
919 eprintln!("flint-sync: {verb:?} honor failed (pending kept, retrying): {e}");
920 }
921 }
922 }
923 if !acks.is_empty() {
924 let last = acks.last().and_then(|a| a.seq);
925 self.ticker_from(last, None)?;
926 }
927 Ok(acks)
928 }
929
930 /// One floor tick: either honor a standing
931 /// pending sentinel that the budget or min-interval held back — the
932 /// boundary is honored by a REAL barrier, its ack stamped
933 /// `sentinel-deferred` — or run the ordinary cadence barrier.
934 ///
935 /// This is the arm B16 regresses on: the interval rewrite must not
936 /// let the poll arm starve cadence.
937 pub async fn floor_tick(&mut self) -> LeanResult<FloorOutcome> {
938 let mut out = FloorOutcome::default();
939 // A held-back sync pending is settled first, for the same
940 // reason as in the poll arm.
941 //
942 // Only a PUBLISH honor discharges the floor's own barrier: a
943 // sync ack carries a seq (the manifest it synced against) but
944 // publishes nothing, so treating it as "the floor ran" would
945 // skip cadence for a tick — a silent RPO regression.
946 let mut published = false;
947 for verb in [Verb::Sync, Verb::Publish] {
948 if self.load_pending(verb)?.is_some() {
949 match self.honor_pending(verb, true).await {
950 Ok(Some(a)) => {
951 out.seq = a.seq.or(out.seq);
952 if verb == Verb::Publish && a.status == "ok" {
953 published = true;
954 out.uploaded = a.report.uploaded;
955 out.deleted = a.report.deleted;
956 out.consumed = a.report.consumed;
957 out.no_change = a.report.no_change;
958 }
959 out.acks.push(a);
960 }
961 Ok(None) => {}
962 Err(e) => {
963 // Review 2026-09-12, ack-1: same rule as the poll
964 // arm, and here the stake is the cadence barrier
965 // below, which used to go unrun behind the error.
966 eprintln!("flint-sync: {verb:?} honor failed on the floor (pending kept): {e}");
967 }
968 }
969 }
970 }
971 // A publish sentinel honored on this tick already ran the fused
972 // barrier the floor owed; running a second one would be pure
973 // churn.
974 if !published {
975 let ran = self.cadence_barrier().await.map(|r| {
976 out.seq = r.seq;
977 out.no_change = r.no_change;
978 out.uploaded = r.uploaded.len();
979 out.deleted = r.deleted.len();
980 out.consumed = r.consumed;
981 out.parked = r.parked.len();
982 out.observed_etag = r.observed_etag.clone();
983 r.observed_seq
984 });
985 match ran {
986 Ok(observed) => {
987 // The floor tick is the one scheduled request every
988 // writer makes, idle or not, so it is where a
989 // credential pause is seen to END — and, below, to
990 // begin. A claim sees both too, but an idle writer
991 // never claims. Best-effort: a gauge that failed to
992 // write must not fail a tick that succeeded.
993 let _ = self.clear_auth_pause();
994 let etag = out.observed_etag.take();
995 self.ticker_from(observed, etag)?;
996 // A foreign 412 outranks "waiting for a boundary":
997 // those paths are not ours to publish at all, and
998 // saying "quiesce-pending" would send an operator
999 // looking for a clock instead of a conflict record.
1000 let forced = (out.parked > 0).then_some(super::gauges::Withheld::Parked412);
1001 out.withheld_reason = self.write_gauges(forced)?.withheld_reason;
1002 return Ok(out);
1003 }
1004 Err(e) => {
1005 if e.is_auth() {
1006 let _ = self.note_auth_pause();
1007 }
1008 return Err(e);
1009 }
1010 }
1011 }
1012 self.ticker_from(out.seq, None)?;
1013 out.withheld_reason = self.write_gauges(None)?.withheld_reason;
1014 Ok(out)
1015 }
1016
1017 /// The preStop drain (D10 rule 1): settle every owed ack BEFORE the
1018 /// lease release. A pending sentinel at SIGTERM is answered by the
1019 /// drain itself — the container-restart case, where the emptyDir and
1020 /// the agent both survive, must not strand a waiting agent.
1021 pub async fn drain(&mut self) -> LeanResult<Vec<Ack>> {
1022 let mut acks = vec![];
1023 // Only an ok PUBLISH ack means a boundary already ran. A sync
1024 // ack carries a seq too — the manifest it synced AGAINST — while
1025 // publishing nothing, so asking "did any ack carry a seq?" lets
1026 // a pending `.flint/sync` at SIGTERM cancel the drain's own
1027 // cite-everything pass and forfeit every byte since the last
1028 // boundary. The floor arm's `sync_ack_is_not_a_floor` comment
1029 // names this exact trap; the drain repeated it.
1030 let mut published = false;
1031 for verb in [Verb::Sync, Verb::Publish] {
1032 let is_publish = matches!(verb, Verb::Publish);
1033 if self.load_pending(verb)?.is_some() {
1034 match self.honor_pending_as(verb, true, Some("drain")).await {
1035 Ok(Some(mut a)) => {
1036 a.boundary = "drain".into();
1037 // Re-write with the drain stamp so the agent can
1038 // tell a drained boundary from a live one.
1039 let bytes = serde_json::to_vec_pretty(&a)
1040 .map_err(|e| LeanError::State(format!("ack: {e}")))?;
1041 write_atomic(&self.ack_path(verb), &bytes)?;
1042 published |= is_publish && a.status == "ok";
1043 acks.push(a);
1044 }
1045 Ok(None) => {}
1046 Err(e) => return Err(e),
1047 }
1048 }
1049 }
1050 if !published {
1051 // D10: the drain publishes EVERYTHING it can, as one declared barrier.
1052 let r = self.declared_barrier_as("drain").await?;
1053 self.ticker_from(r.observed_seq, r.observed_etag.clone())?;
1054 // Review 2026-09-12, inbox-1: a drain that leaves a path
1055 // parked has not published the tree; attesting it let the
1056 // node remove the agent's only copy.
1057 if !r.parked.is_empty() {
1058 return Err(LeanError::State(format!(
1059 "drain: {} path(s) could not be published (parked on a foreign version whose \
1060 preserve failed): {:?} — not attesting",
1061 r.parked.len(),
1062 r.parked
1063 )));
1064 }
1065 }
1066 Ok(acks)
1067 }
1068}