@gotcos/glasses-server 6.36.9 → 6.36.11

package/CHANGELOG.md

@@ -1,5 +1,109 @@
 ## Unreleased
 
+## 6.36.11
+- **Fences now record WHY, so the population can be measured before anything resolves
+  automatically.** Two plans designed an automatic fence resolver and both were rejected —
+  the second because there has never been a single fence on the machine to look at. If
+  `timeout` dominates, the child had the full 21-minute budget to run tool calls before
+  SIGKILL and re-delivery would re-execute them, so no automatic clear is ever safe. That
+  question was unanswerable and now is not.
+- **`reaped` is reported by the adapter, never derived.** A signal-killed child reports
+  `code === null` and the handlers only assign `exitCode` for a numeric code — so deriving
+  "was it reaped" from `exitCode` reports NEVER REAPED for the dominant timeout shape
+  (SIGTERM, then SIGKILL), which is exactly backwards for the decision this data informs.
+  The first cut of this change did derive it. `AttachedTurnFailureResult` now carries
+  `reaped`, true from every settle reached via `close`/`error` and false only from the
+  force-settle that fires when `close` never arrived.
+- **An unreadable adapter result records nothing, not zeroes.** Reading `{}` and writing
+  `exitCode: null, childReaped: false` states two facts about a child nothing is known
+  about, indistinguishable on disk from a confirmed-unreaped timeout — corrupting the one
+  discriminator this evidence exists to establish.
+- **`fenceSite` says which site fired.** `adapterReason` cannot substitute: the catch site
+  inherits whatever the adapter last reported, so a route crash AFTER a clean delivery
+  records `ok` — the strongest possible reason NOT to re-deliver, which would otherwise
+  read as "nothing went wrong".
+- **One resolved reason for the record and the log.** The record said `unreadable_result`
+  while the breadcrumb said `unknown`, so an operator grepping for the sentinel found
+  nothing. This is the same contradiction 6.36.10 fixed at the other fence site,
+  re-committed one release later in the same handler; both now read one value.
+- **A release no longer destroys the evidence.** `releaseFence` deletes the row, and the
+  realistic first-fence sequence is: fence lands, Control's card appears, it is released,
+  the distribution is gone. The release breadcrumb now carries the whole record.
+- **Spawn identity as `{pid, startMs}` PAIRS.** `recordedPids` held bare pids and the
+  measured start was discarded; a pid alone cannot be told apart from a recycled one.
+  The evidence type is narrowed to the six adapter fields so the spread at the fence
+  sites cannot clobber the fence's own identity — `Partial<FenceEvidence>` permitted
+  `provider`, and the adapter result carries one that would write null and fail
+  `isFenceRecord` on the next read, silently un-enforcing the fence.
+- **All fields OPTIONAL and NOT in `isFenceRecord`.** That predicate is cast-based, so a
+  required field would type as present while being undefined at runtime; extending it
+  would reclassify existing rows as unrecognised and silently un-enforce them.
+- **DISK ONLY.** Nothing reaches the wire; asserted against the real `/fences` body and
+  the release preview.
+- **No behaviour change.** Nothing clears, nothing refuses differently. Upgrade, downgrade
+  and the COS Control card were all verified unaffected.
+- **Coverage, stated honestly.** The adapter's `reaped` contract is tested at the adapter,
+  driving a real `close(null)` — a route test could not cover it, because the route
+  fixtures supply `reaped` themselves and would pass with the adapter gutted. Mutation
+  results: caught — derived-from-exitCode, adapter stops reporting reaped, startMs zeroed,
+  empty spawn list, missing fenceSite, unrecorded adapterReason, wire leak. **Survived, and
+  therefore unverified: `stderrClass` is written but asserted nowhere, and `fail()`'s
+  `reaped: false` default on the `not_attempted` paths (which never fence).** The
+  `route_error` fence site remains reachable by no test.
+- **Known gaps, not fixed here:** `stderrClass` appears in the breadcrumbs but on a
+  default install (`COS_THREAD_FENCE_DURABLE` unset) nothing is written to disk at all;
+  and reading the distribution means reading `thread-fences.json` or the server log —
+  there is no UI for it.
+
+## 6.36.10
+- **A fenced thread had no exit and left no trace.** An ambiguous delivery fences the
+  target so a prompt cannot be double-delivered into a real conversation — that is
+  correct and stays. Everything around it was wrong: the fence lived in a process-local
+  Map, wrote no log line at either site, had no list, and had no release. It was
+  discoverable only by being refused, and the only thing that cleared it was a restart.
+- **Now listable and releasable without a restart.** `GET /api/agent-sessions/fences`
+  lists them; `POST /api/agent-sessions/fences/release` clears one. Addressed by DIGEST,
+  never by raw target key — the key embeds the private native thread id. Fails closed:
+  without `confirm: true` it returns 400 with a preview of what would be reopened. That
+  confirmation is a deliberate second call, NOT proof a human looked — the API token is
+  shared by the phone, the lens and every COS agent session, so nothing is structurally
+  prevented from asserting it. The comment says so rather than overclaiming.
+- **Durable storage ships INERT, behind `COS_THREAD_FENCE_DURABLE=1`, default off.**
+  Persisting the fence is the right direction, but durability without a reachable
+  release is a regression, not a fix: today "Restart Server" in COS Control clears a
+  fence, and making it survive restarts with no operator surface in Control would turn
+  an 8-second annoyance into a permanently dead thread needing a terminal. The flag
+  flips on when COS Control has a Fences card. The routes above already remove the
+  restart from the recovery path.
+- **A write can never erase what it could not read.** `TargetGuard` hydrates only the
+  rows it understood and saves its map wholesale, so a single unrecognised row — a
+  newer schema, a partial write, one bad field — would otherwise be erased by the next
+  fence on an unrelated thread, silently reopening every other fenced thread. Writes now
+  merge unrecognised rows back through. A corrupt file is quarantined to
+  `.corrupt-<ts>` rather than dropped, and uses `durableAtomicWriteFileSync` (fsync of
+  bytes, metadata and directory; randomized exclusive temp name) rather than the
+  lightweight cache writer.
+- **A release is persisted before it is reported.** Mutating memory first and reporting
+  success meant an operator could be told a thread was open, write to it, and find it
+  fenced again after the next restart with no record of why. A failed write now returns
+  500 `persist_failed` and the fence holds. `GET /fences` reports `degraded` when the
+  last write failed — a memory-only fence set is otherwise indistinguishable from a
+  durable one until the process restarts.
+- **Visible.** Breadcrumbs at both fence-set sites (tagged `ambiguous` vs `route_error`)
+  and both fence-hit routes (turn, attach). No raw target key in any of them. The
+  route_error line reports the hoisted pre-turn head rather than hardcoding
+  `unavailable`, which contradicted the record it had just written.
+- **Known, not fixed here:** releasing a fence does not by itself make the thread
+  attachable — the turn that fenced it left a binding holding the target for its
+  30-minute TTL, so the next attach refuses `native_target_busy`. There is no detach
+  route. `/attachability` still does not consult the fence, so the lens menu renders
+  Continue enabled on a fenced thread; the refusal is honest, the menu is not yet.
+- **Coverage.** 11 mutations against the new guards fail the suite, including all four
+  that survived the first QA pass (the release handle check, and the provider/reason/
+  fencedAt row validators). One documented survivor remains: removing the write-once
+  guard on `fence()`, which no route can reach because both fence sites sit inside the
+  `tryClaim` section. The code says so at the call site.
+
 ## 6.36.9
 - **The queue gate can now see COS's own bindings, which closes a 30-minute lockout.**
   `threadOccupancy` sees FOREIGN holders only — `OccupancyReason` has no

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gotcos/glasses-server",
-  "version": "6.36.9",
+  "version": "6.36.11",
   "description": "COS Glasses \u2014 self-hosted AI heads-up-display server for Even G2 smart glasses, powered by Claude Code, Codex, or Cursor Agent CLI",
   "type": "module",
   "bin": {

package/server/index.ts

@@ -109,6 +109,7 @@ import { createThreadTurnQueueRouter, drainAllThreads } from './routes/thread-tu
 import { transcriptTurnEnded } from './lib/thread-turn-queue-store.js'
 import { transcriptPathFor } from './lib/native-head.js'
 import { deliverQueuedTurnOverLoopback } from './lib/thread-turn-queue-deliver.js'
+import { readFences, writeFences } from './lib/thread-fence-store.js'
 import type { QueuedThreadTurn } from './lib/thread-turn-queue.js'
 
 const PORT = parseInt(process.env.PORT ?? '3141', 10)
@@ -582,6 +583,23 @@ if (threadAttachEnabled()) {
 }
 
 app.use('/api', createAgentSessionBindingsRouter({
+  // Durable fences (6.36.10), OFF BY DEFAULT.
+  //
+  // The fence is the one piece of state whose loss writes twice into a real
+  // conversation, so persisting it is the right direction. But durability
+  // without a reachable release is a REGRESSION, not a fix: today "Restart
+  // Server" in COS Control clears a fence, and making it survive restarts with
+  // no operator surface in Control turns an 8-second annoyance into a
+  // permanently dead thread that needs a terminal to clear. Miles, 2026-08-12:
+  // "we couldn't do anything without bash, that shouldn't be the case."
+  //
+  // So the storage ships inert. GET /agent-sessions/fences and
+  // POST /agent-sessions/fences/release work either way — which already removes
+  // the restart from the recovery path — and this flag flips on once COS Control
+  // has a Fences card.
+  fencePersistence: process.env.COS_THREAD_FENCE_DURABLE === '1'
+    ? { load: readFences, save: writeFences }
+    : undefined,
   probes: occupancyProbes,
   dirs: occupancyDirs,
   now: () => Date.now(),

package/server/lib/agent-session-store.ts

@@ -1063,9 +1063,29 @@ async function enrichLiveClaude(row: AgentSessionRow, roots: AgentSessionRoots):
   const firstPrompt = await firstClaudeUserTitle(found)
   const title = peek.customTitle ?? firstPrompt ?? row.display_label
   const fullId = found.split('/').pop()?.replace(/\.jsonl$/i, '') || row.session_id
+  // THE LIVE ROW'S `modified` IS A HEARTBEAT, NOT AN MTIME, AND IT GOES STALE.
+  //
+  // `liveClaudeRows` builds it from the peer registry's `lastActiveAt` (routes/
+  // agent-sessions.ts:257). That value tracks the registry record, not the
+  // transcript, so a session that is actively writing keeps reporting whenever the
+  // registry last moved. Measured 2026-08-18 on three live sessions: the wire said
+  // 55.3m / 407.7m / 435.0m while the transcript had been written 0.1m / 0.2m /
+  // 5.1m earlier — under-reporting a live session by up to 7.2 hours.
+  //
+  // We are already holding the resolved transcript path and have already read it
+  // twice, so the true mtime costs one stat. A failed stat keeps the heartbeat:
+  // worse, but no worse than today.
+  //
+  // THAT FALLBACK IS DEFENSIVE AND UNREACHED BY THE SUITE. `if (!found) return row`
+  // above already catches the missing-file case, so reaching `: row.modified` needs a
+  // stat to fail on a file that was just resolved. A mutation of it survives; that is
+  // recorded rather than papered over.
+  const st = await fileStat(found)
+  const modified = st?.isFile ? isoFromMtime(st.mtimeMs) : row.modified
   return {
     ...row,
     session_id: fullId,
+    modified,
     display_label: title || row.display_label || 'Claude session',
     ...discussionFields(title || row.display_label, firstPrompt || '', peek.latestAssistant),
   }

package/server/lib/attached-provider-adapter.ts

@@ -194,6 +194,18 @@ export interface AttachedTurnFailureResult {
    */
   detail: string | null
   exitCode: number | null
+  /**
+   * Did `close`/`exit` actually fire, i.e. did the kernel reap the child?
+   *
+   * NOT DERIVABLE FROM `exitCode`. A child killed by a signal reports
+   * `code === null`, and the handlers below only assign `exitCode` for a numeric
+   * code — so the dominant timeout shape (SIGTERM, then SIGKILL) is reaped while
+   * leaving `exitCode` null. Deriving reaping from the code therefore reports
+   * "never reaped" for exactly the case a reader most needs to identify.
+   * Only the force-settle path, which fires when `close` never arrived, is
+   * genuinely unreaped.
+   */
+  reaped: boolean
   stderrClass: AttachedStderrClass
   durationMs: number
 }
@@ -745,6 +757,10 @@ function fail(
     nativeThreadId: null,
     returnedNativeId: null,
     delivery,
+    // Conservative default: NOT OBSERVED reaped. Most `fail()` callers are
+    // `not_attempted` paths where no child exists, and the ones that do have a
+    // child override this from `settleFailure`.
+    reaped: false,
     reason,
     detail: null,
     exitCode: null,
@@ -1051,6 +1067,10 @@ function driveChild(input: DriveInput): Promise<AttachedTurnResult> {
         nativeThreadId,
         returnedNativeId: observedIds.length === 1 ? observedIds[0]! : null,
         exitCode,
+        // Every settle but the force-settle below is reached from `finishTerminal`,
+        // which only runs from the `close`/`error` handlers — so the child was
+        // reaped. The one exception overrides this explicitly.
+        reaped: true,
         stderrClass: classifyStderr(stderrSample),
         durationMs: duration(),
         ...over,
@@ -1193,7 +1213,7 @@ function driveChild(input: DriveInput): Promise<AttachedTurnResult> {
           // A child that survived SIGKILL cannot be reached from here, and
           // blocking forever would wedge the coordinator and any Control drain
           // behind it.
-          settleFailure('timeout', { detail: 'unreaped' })
+          settleFailure('timeout', { detail: 'unreaped', reaped: false })
         }, FORCE_SETTLE_MS)
       }, KILL_GRACE_MS)
     }, timeoutMs)

package/server/lib/thread-fence-store.ts

@@ -0,0 +1,165 @@
+// Durable storage for target fences.
+//
+// A fence shuts a native thread that may already hold an undelivered COS turn. It
+// is the one piece of state whose LOSS writes twice into a real human conversation
+// — agent-session-binding-registry.ts records the incident verbatim: "the
+// process-local fence re-opened on restart and delivered a second copy."
+//
+// SEPARATE FROM THE DECISIONS, matching thread-turn-queue-store.ts: TargetGuard
+// holds the rules and takes load/save as injected callbacks, so its behaviour stays
+// testable in memory and only production touches the disk.
+//
+// UNDER THE DATA HOME, never the generation directory, which Update Server replaces
+// wholesale. Same lesson as the stranded voice profiles.
+//
+// NEVER TTL'd AND NEVER EVICTED. `heads` and fork refs both bound their maps, and
+// both evict in the SAFE direction — losing a head asks the user to acknowledge,
+// losing a fork ref makes them find the thread by hand. Losing a fence silently
+// reopens a thread that may hold an undelivered turn, so the only way an entry
+// leaves this file is an explicit operator release.
+
+import { durableAtomicWriteFileSync, loadJsonOrQuarantine } from './atomic-fs.js'
+import { dataPath } from './data-dir.js'
+
+export interface FenceRecord {
+  /** The raw target key. On disk only — it embeds the private native thread id and
+   *  is never emitted by any route. Callers outside this module use the digest. */
+  targetKey: string
+  provider: string
+  reason: string
+  /** The head digest as it stood BEFORE the ambiguous turn. Null ONLY when the
+   *  failure happened before the head was read — `head` is scoped to the try, so
+   *  the route-error site reads a hoisted copy rather than nothing. */
+  headBefore: string | null
+  turnId: string
+  bindingId: string | null
+  fencedAt: number
+
+  // ── EVIDENCE (all optional, all DISK-ONLY) ────────────────
+  //
+  // WHY THIS EXISTS. Two plans designed an automatic fence resolver and both were
+  // rejected, the second because we have never observed a single fence. The
+  // distribution decides whether any automatic path is safe: if `timeout`
+  // dominates, the child had the full 21-minute budget to run tool calls before
+  // SIGKILL and re-delivery re-executes them, so no automatic clear is ever safe.
+  // If the child exited on its own with a known code, the footing is much better.
+  //
+  // EVERY FIELD IS OPTIONAL, and deliberately NOT added to `isFenceRecord`. That
+  // predicate is cast-based (`r is FenceRecord`), so a REQUIRED field would type as
+  // present while being undefined at runtime and tsc could not catch it; and
+  // extending the validator would reclassify existing rows as unrecognised, routing
+  // them into the preserved-but-inert pile and silently un-enforcing real fences.
+  //
+  // NONE OF THIS REACHES THE WIRE. The router contract carries no pid and no native
+  // thread id; `listFences()` already omits `bindingId` as precedent.
+
+  /**
+   * The adapter's own verdict.
+   *
+   * REACHABILITY IS DECIDED BY `delivery === 'ambiguous'`, i.e. the stdin-write
+   * boundary — only a turn whose prompt bytes reached the child can fence. The
+   * reachable set is therefore NOT the full `AttachedTurnFailure` enum:
+   *   timeout, provider_exit_nonzero, no_native_id_returned, native_id_mismatch,
+   *   child_stdio_unavailable (detail `write_failed`), spawn_failed (ONLY via
+   *   detail `child_error`; its other forms are `not_attempted` and never fence),
+   *   plus two this route synthesises — `ok` (the adapter reported success with a
+   *   non-delivered state, a contradiction) and `unreadable_result`.
+   * `adapter_internal_error` can NEVER fence: every emission site is
+   * `not_attempted` or guarded. `route_error` comes from the catch site.
+   */
+  adapterReason?: string
+  /**
+   * WHICH fence site fired.
+   *
+   * `adapterReason` cannot substitute. The catch site inherits whatever the
+   * adapter last reported, so a route crash AFTER a clean delivery records
+   * `ok` — the single strongest reason NOT to re-deliver, which would read as
+   * "nothing went wrong". Only this field separates them.
+   */
+  fenceSite?: 'ambiguous' | 'route_error'
+  /** Bounded self-authored discriminator, e.g. `unreaped`. Never provider output. */
+  adapterDetail?: string | null
+  exitCode?: number | null
+  /**
+   * Reported BY THE ADAPTER, never derived from `exitCode`.
+   *
+   * A signal-killed child reports `code === null`, and the adapter only assigns
+   * `exitCode` for a numeric code — so deriving this would record "never reaped"
+   * for the dominant timeout shape (SIGTERM then SIGKILL), which is the exact
+   * case this evidence exists to identify. Undefined when the adapter result was
+   * unreadable: absent, never a fabricated `false`.
+   */
+  childReaped?: boolean
+  stderrClass?: string
+  /** How long the turn ran before it failed. A timeout is ~21 minutes of tool calls. */
+  durationMs?: number
+  /** Every child COS spawned for this turn, with its MEASURED start. Both are needed:
+   *  a pid alone cannot be distinguished from a recycled one. An EMPTY list means no
+   *  child was ever spawned, which no resolver may ever read as "nothing landed". */
+  spawns?: Array<{ pid: number; startMs: number }>
+}
+
+export function fencePath(): string {
+  return dataPath('thread-fences.json')
+}
+
+function isFenceRecord(r: unknown): r is FenceRecord {
+  return !!r && typeof r === 'object'
+    && typeof (r as FenceRecord).targetKey === 'string' && (r as FenceRecord).targetKey.length > 0
+    && typeof (r as FenceRecord).provider === 'string'
+    && typeof (r as FenceRecord).reason === 'string'
+    && typeof (r as FenceRecord).fencedAt === 'number'
+}
+
+/** The raw array on disk, or [] when the file is missing or was quarantined. */
+function rawRows(): unknown[] {
+  const loaded = loadJsonOrQuarantine<unknown>(fencePath())
+  if (loaded.status === 'corrupt') {
+    // Quarantined to `<path>.corrupt-<ts>` rather than discarded: the bytes are
+    // the only record of which threads were fenced, and this is the one state
+    // whose silent loss double-writes a real conversation.
+    console.warn(`[thread-fence-store] fence file was corrupt, quarantined as ${loaded.quarantinedAs}`)
+    return []
+  }
+  if (loaded.status !== 'ok') return []
+  if (!Array.isArray(loaded.data)) {
+    console.warn('[thread-fence-store] fence file is not an array — treating as empty')
+    return []
+  }
+  return loaded.data
+}
+
+/**
+ * Every stored fence this build can understand.
+ *
+ * A missing or corrupt file reads as empty. That fails OPEN, deliberately:
+ * failing closed would refuse every thread on the machine with no way back,
+ * while failing open is exactly the pre-6.36.10 behaviour (the fence was
+ * process-local and died on restart), so it cannot be a regression. A corrupt
+ * file is quarantined rather than dropped, so the evidence survives.
+ */
+export function readFences(): FenceRecord[] {
+  return rawRows().filter(isFenceRecord)
+}
+
+/**
+ * Replace the stored set, PRESERVING rows this build could not validate.
+ *
+ * THE MERGE IS THE WHOLE POINT. `TargetGuard` holds only the rows `readFences`
+ * understood and saves its map wholesale, so without this a single unrecognised
+ * row — a newer schema, a partial write, one bad field — would be erased by the
+ * next fence on an unrelated thread, silently reopening every other fenced
+ * thread. Preserved rows are inert (nothing enforces a fence that is not in the
+ * map) but they are never destroyed by a write that did not understand them.
+ *
+ * Uses the DURABLE writer, not the lightweight one: fsync of bytes, metadata and
+ * directory, plus a randomized exclusive temp name so two independent writers
+ * cannot share `<path>.tmp`.
+ */
+export function writeFences(rows: FenceRecord[]): void {
+  const preserved = rawRows().filter(r => !isFenceRecord(r))
+  if (preserved.length > 0) {
+    console.warn(`[thread-fence-store] preserving ${preserved.length} unrecognised fence row(s) through this write`)
+  }
+  durableAtomicWriteFileSync(fencePath(), `${JSON.stringify([...rows, ...preserved], null, 2)}\n`)
+}

package/server/routes/agent-session-bindings.ts

@@ -94,6 +94,7 @@
 // does not — a future remount above the parser — the POST routes see a non-object
 // and answer 400. They never treat an unparsed body as an empty one.
 
+import type { FenceRecord } from '../lib/thread-fence-store.js'
 import { Router, type Request, type Response } from 'express'
 import { createHash, randomUUID } from 'node:crypto'
 import {
@@ -268,6 +269,12 @@ export type AttachedTurnResult =
   | { ok: boolean; delivery: 'not_attempted' | 'aborted' | 'ambiguous' | 'delivered' }
 
 export interface AgentSessionBindingsDeps {
+  /**
+   * Durable fence storage. OPTIONAL, and omitting it is what keeps the existing
+   * suite in memory: a test that silently began writing the real data home would
+   * leak fences between cases and into the running server. Production wires it.
+   */
+  fencePersistence?: FencePersistence
   probes: OccupancyProbes
   dirs: OccupancyDirs
   /** Epoch ms. Injected so lease expiry is decidable in a test without waiting. */
@@ -829,24 +836,97 @@ export const MAX_TRACKED_HEADS = 512
 /** A COS session id may contain ':' and '/', which is exactly why it is never projected. */
 export const COS_SESSION_ID_RE = /^[A-Za-z0-9][A-Za-z0-9._:@/-]{0,127}$/
 
+/** What a fence records about the turn that set it. */
+export interface FenceEvidence {
+  provider: string
+  /** The adapter's verdict, or 'route_error' at the catch site. DISK-ONLY. */
+  adapterReason?: string
+  /** WHICH site fenced. `adapterReason` cannot substitute: the catch site inherits
+   *  whatever the adapter reported, so a fence can read 'ok' there. DISK-ONLY. */
+  fenceSite?: 'ambiguous' | 'route_error'
+  adapterDetail?: string | null
+  exitCode?: number | null
+  childReaped?: boolean
+  stderrClass?: string
+  durationMs?: number
+  spawns?: Array<{ pid: number; startMs: number }>
+  /** The head BEFORE the ambiguous turn. Null ONLY when the failure happened
+   *  before the head was read. */
+  headBefore: string | null
+  turnId: string
+  bindingId: string | null
+  now: number
+}
+
+export type ReleaseOutcome =
+  | { ok: true; row: FenceRecord }
+  | { ok: false; reason: 'unknown_fence' | 'persist_failed' }
+
+/** Injected so TargetGuard stays testable in memory; production wires the store. */
+export interface FencePersistence {
+  load: () => FenceRecord[]
+  save: (rows: FenceRecord[]) => void
+}
+
 /**
- * The in-process half of plan 4.5 and 4.6: one COS turn per native target, and a
- * target that may already hold an undelivered turn stays shut.
+ * One COS turn per native target, and a target that may already hold an
+ * undelivered turn stays shut.
+ *
+ * CLAIMS ARE PROCESS-LOCAL; FENCES ARE DURABLE (6.36.10). The two states fail in
+ * opposite directions, which is why only one of them is persisted. Losing a claim
+ * on restart is safe — no turn is running after a restart anyway. Losing a FENCE
+ * reopens a thread that may already hold an undelivered turn, and the binding
+ * registry records what that cost: "the process-local fence re-opened on restart
+ * and delivered a second copy." So fences are written through to disk and
+ * rehydrated in the constructor, and the ONLY way one leaves the map is an
+ * explicit operator release (`releaseFence`).
  *
- * NOT DURABLE, AND THAT IS A STATED GAP. Plan 4.5 wants the reservation
- * persisted in the job journal and rehydrated on boot, and 4.6 item 3 wants the
- * fence to have its own lifecycle and an operator release path. Both belong to
- * Phase 2, which owns that journal. What lives here is the process-lifetime
- * version, which is enough to make the two properties true for a running server
- * and fails in the safe direction on restart: a claim is released (no turn is
- * running after a restart anyway) and a FENCE is lost, which is the one that
- * matters and is why it is called out rather than implied.
+ * Persistence is INJECTED rather than imported. A test that silently began writing
+ * the real data home would leak fences between cases and into the running server,
+ * so the suite runs with `null` and stays in memory.
  */
 class TargetGuard {
   /** targetKey -> turnId of the single COS turn allowed to be in flight. */
   private readonly claims = new Map<string, string>()
-  /** targetKey -> why no further turn may be delivered. */
-  private readonly fences = new Map<string, WriteRefusal>()
+  /** targetKey -> the fence record. DURABLE as of 6.36.10: persistence is injected
+   *  so tests stay in memory and only production touches the data home. */
+  private readonly fences = new Map<string, FenceRecord>()
+  private readonly persistence: FencePersistence | null
+  private persistDegraded = false
+
+  constructor(persistence: FencePersistence | null = null) {
+    this.persistence = persistence
+    if (persistence === null) return
+    // Rehydrate BEFORE the router serves. A fence that died on restart is exactly
+    // how a second copy of a turn reached a real transcript.
+    try {
+      for (const row of persistence.load()) this.fences.set(row.targetKey, row)
+    } catch (error) {
+      console.error(`[agent-session-bindings] fence rehydrate failed: ${error instanceof Error ? error.message : error}`)
+    }
+  }
+
+  /** True when the durable write succeeded (or there is nothing to persist to). */
+  private persistFences(rows: FenceRecord[]): boolean {
+    if (this.persistence === null) return true
+    try {
+      this.persistence.save(rows)
+      this.persistDegraded = false
+      return true
+    } catch (error) {
+      // The in-memory fence still holds for this process, so the thread stays shut
+      // NOW; what is lost is survival across a restart. Loud, and surfaced on
+      // GET /fences — a silent fallback is indistinguishable from working.
+      this.persistDegraded = true
+      console.error(`[agent-session-bindings] fence persist FAILED (fences hold in memory only): ${error instanceof Error ? error.message : error}`)
+      return false
+    }
+  }
+
+  /** Whether the last durable write failed. Reported, never inferred. */
+  degraded(): boolean {
+    return this.persistDegraded
+  }
   /** bindingId -> the head digest this binding is currently reconciled to. */
   private readonly heads = new Map<string, string>()
 
@@ -869,12 +949,80 @@ class TargetGuard {
     if (this.claims.get(targetKey) === turnId) this.claims.delete(targetKey)
   }
 
-  fence(targetKey: string, reason: WriteRefusal): void {
-    if (!this.fences.has(targetKey)) this.fences.set(targetKey, reason)
+  /**
+   * Write-once: the FIRST reason wins, so a later ambiguity cannot overwrite the
+   * evidence chain of an unresolved one.
+   *
+   * DEFENSIVE, AND UNVERIFIED BY EXECUTION. Both fence sites sit inside the
+   * `tryClaim` section, which serialises them per target, and a fenced target is
+   * refused at the check before it can reach either site again — so no route can
+   * currently fence the same key twice, and a mutation removing this guard passes
+   * the whole suite. It is kept because the one path that could reach it (the
+   * ambiguous site fences, then the response throws into the catch, which fences
+   * `claimedKey` again) would otherwise replace a record carrying `bindingId` with
+   * one carrying null. Do not read the passing suite as coverage of this line.
+   */
+  fence(targetKey: string, reason: WriteRefusal, evidence: FenceEvidence): void {
+    if (this.fences.has(targetKey)) return
+    this.fences.set(targetKey, {
+      targetKey,
+      provider: evidence.provider,
+      reason,
+      headBefore: evidence.headBefore,
+      turnId: evidence.turnId,
+      bindingId: evidence.bindingId,
+      fencedAt: evidence.now,
+      adapterReason: evidence.adapterReason,
+      fenceSite: evidence.fenceSite,
+      adapterDetail: evidence.adapterDetail,
+      exitCode: evidence.exitCode,
+      childReaped: evidence.childReaped,
+      stderrClass: evidence.stderrClass,
+      durationMs: evidence.durationMs,
+      spawns: evidence.spawns,
+    })
+    this.persistFences([...this.fences.values()])
   }
 
   fencedReason(targetKey: string): WriteRefusal | null {
-    return this.fences.get(targetKey) ?? null
+    const row = this.fences.get(targetKey)
+    return row === undefined ? null : (row.reason as WriteRefusal)
+  }
+
+  /** Every fence, REDACTED for the wire: the raw targetKey embeds the private
+   *  native thread id, so callers address a fence by its deterministic digest. */
+  listFences(): Array<{ target: string; provider: string; reason: string; headBefore: string | null; turnId: string; fencedAt: number }> {
+    return [...this.fences.values()].map(row => ({
+      target: opaqueRevision(row.targetKey),
+      provider: row.provider,
+      reason: row.reason,
+      headBefore: row.headBefore,
+      turnId: row.turnId,
+      fencedAt: row.fencedAt,
+    }))
+  }
+
+  /**
+   * The operator release. THE ONLY WAY A FENCE LEAVES THIS MAP.
+   *
+   * Addressed by digest, never by raw target key. Returns false when no fence
+   * matches, so a stale handle reports honestly instead of silently succeeding.
+   */
+  releaseFence(targetDigest: string): ReleaseOutcome {
+    for (const [key, row] of this.fences) {
+      // THE authority on which fence a handle names. The route also looks the row
+      // up for its preview, but the release decision is made here — a duplicate
+      // lookup upstream would leave this comparison enforced by nothing.
+      if (opaqueRevision(row.targetKey) !== targetDigest) continue
+      // PERSIST FIRST. Reporting a release that was not durably recorded is how an
+      // operator is told a thread is open, writes to it, and finds it fenced again
+      // after the next restart with no record of why.
+      const remaining = [...this.fences.values()].filter(r => r.targetKey !== key)
+      if (!this.persistFences(remaining)) return { ok: false, reason: 'persist_failed' }
+      this.fences.delete(key)
+      return { ok: true, row }
+    }
+    return { ok: false, reason: 'unknown_fence' }
   }
 
   /**
@@ -1135,7 +1283,7 @@ export function createAgentSessionBindingsRouter(deps: AgentSessionBindingsDeps)
   const canListBindings = bindingDepsUsable(deps)
   const canWriteBindings = bindingWriteDepsUsable(deps)
   const detect = deps?.occupancy ?? threadOccupancy
-  const guard = new TargetGuard()
+  const guard = new TargetGuard(deps.fencePersistence ?? null)
   const ownership = deps?.ownership ?? { record: recordCosSpawn, release: releaseCosSpawn }
   // One per router. Injectable so the follow-on (attach accepting a `forkRef`)
   // shares this instance rather than standing up a second, disconnected one.
@@ -1216,6 +1364,59 @@ export function createAgentSessionBindingsRouter(deps: AgentSessionBindingsDeps)
     }
   }
 
+  // ------------------------------------------------------------------ fences
+  //
+  // Mounted BEFORE the parameterised routes so `fences` can never be read as a
+  // provider. Both are operator surfaces: a fenced thread was previously
+  // discoverable only by trying to use it and being refused.
+  //
+  // A fence is addressed by DIGEST. The raw target key embeds the private native
+  // thread id, and the redaction contract at the top of this file is absolute.
+
+  router.get('/agent-sessions/fences', (_req, res) => {
+    // A fence list is a liveness answer; a cached one is worse than none.
+    res.set('Cache-Control', 'private, no-store')
+    // `degraded` is reported, never inferred: a memory-only fence set behaves
+    // identically to a durable one until the process restarts, so a silent
+    // fallback would be indistinguishable from working.
+    res.json({ fences: guard.listFences(), degraded: guard.degraded() })
+  })
+
+  router.post('/agent-sessions/fences/release', (req, res) => {
+    const body = (req.body ?? {}) as { target?: unknown; confirm?: unknown }
+    const target = body.target
+    if (typeof target !== 'string' || target.length === 0 || target.length > 256) {
+      res.status(400).json({ released: false, reason: 'invalid_request' })
+      return
+    }
+    res.set('Cache-Control', 'private, no-store')
+    if (body.confirm !== true) {
+      // FAILS CLOSED, like every other destructive COS call. NOTE WHAT THIS IS
+      // AND IS NOT: it is a deliberate second call, not proof a human looked.
+      // The API token is shared by the phone, the lens and every COS agent
+      // session, so nothing is structurally prevented from asserting `confirm`.
+      // It stops an accidental release, not an automated one.
+      const preview = guard.listFences().find(f => f.target === target) ?? null
+      res.status(400).json({ released: false, reason: 'confirmation_required', preview })
+      return
+    }
+    // The guard decides. It re-matches the handle itself rather than trusting a
+    // lookup performed up here, and it persists BEFORE it mutates.
+    const outcome = guard.releaseFence(target)
+    if (!outcome.ok) {
+      const status = outcome.reason === 'unknown_fence' ? 404 : 500
+      res.status(status).json({ released: false, reason: outcome.reason })
+      return
+    }
+    // THE EVIDENCE DIES WITH THE ROW. `releaseFence` deletes it, and the realistic
+    // sequence is: first fence ever lands -> Control's card appears -> it is
+    // released -> the distribution this evidence exists to collect is gone. So the
+    // release line carries the whole record, not just its identity.
+    const ev = outcome.row
+    console.warn(`[agent-session-bindings] fence RELEASED by operator target=${target} provider=${ev.provider} fencedAt=${ev.fencedAt} fenceSite=${ev.fenceSite ?? 'unknown'} adapterReason=${ev.adapterReason ?? 'unknown'} detail=${ev.adapterDetail ?? 'none'} exitCode=${ev.exitCode ?? 'null'} childReaped=${ev.childReaped ?? 'unknown'} stderrClass=${ev.stderrClass ?? 'none'} durationMs=${ev.durationMs ?? 'unknown'} spawnCount=${ev.spawns?.length ?? 0}`)
+    res.json({ released: true, target, provider: outcome.row.provider })
+  })
+
   router.get('/agent-sessions/:provider/:threadId/attachability', (req, res) => {
     // An occupancy verdict is a liveness answer with a lifetime of roughly now.
     // A cached `attachable: true` is indistinguishable from a stale one, which is
@@ -1375,7 +1576,10 @@ export function createAgentSessionBindingsRouter(deps: AgentSessionBindingsDeps)
       // again just because the binding that delivered it is gone. Checked here as
       // well as in the turn route, because a fresh attach is the obvious way around
       // a per-binding fence.
-      if (fenced !== null) return refuseAttach(res, fenced)
+      if (fenced !== null) {
+        console.warn(`[agent-session-bindings] fence hit route=attach provider=${providerParam} target=${opaqueRevision(key)}`)
+        return refuseAttach(res, fenced)
+      }
 
       const resolve = deps.resolveTarget
       if (typeof resolve !== 'function') return refuseAttach(res, 'target_unresolvable')
@@ -1623,8 +1827,24 @@ export function createAgentSessionBindingsRouter(deps: AgentSessionBindingsDeps)
     const turnId = mintId()
     /** The target we hold a claim on, released in the finally. */
     let claimedKey: string | null = null
+    // Hoisted so the CATCH site can fence with evidence. `binding` and `head` are
+    // both declared inside the try, so neither is in scope where the route-error
+    // fence is set — without these it would store a fence it can say nothing about.
+    let fenceProvider = ''
+    let preTurnHeadDigest: string | null = null
     /** Children the adapter reported, released in the finally. */
-    const recordedPids: number[] = []
+    // PAIRS, not bare pids. A pid alone cannot be told apart from a recycled one,
+    // and `startMs` is measured in the onSpawn closure and was previously discarded.
+    const recordedPids: Array<{ pid: number; startMs: number }> = []
+    // Hoisted so BOTH fence sites can record what the adapter actually reported.
+    // `result` is scoped inside the try; the ambiguous site sits after the catch.
+    // NARROW ON PURPOSE. `Partial<FenceEvidence>` would permit provider/headBefore/
+    // turnId/bindingId/now, and the spread sits AFTER them at both fence sites — so
+    // a future field could silently overwrite the fence's identity. The adapter
+    // result literally carries a `provider`, which would write null and fail
+    // `isFenceRecord` on the next read, silently un-enforcing the fence.
+    let adapterEvidence: Partial<Pick<FenceEvidence,
+      'adapterReason' | 'adapterDetail' | 'exitCode' | 'childReaped' | 'stderrClass' | 'durationMs'>> = {}
     let pinnedBindingId: string | null = null
     let requestNow = 0
     /**
@@ -1789,6 +2009,7 @@ export function createAgentSessionBindingsRouter(deps: AgentSessionBindingsDeps)
       if (gate?.ok !== true) return refuseTurn(registryRefusal(gate?.reason))
 
       const binding = deps.bindings.get!(bindingId)
+      if (binding) fenceProvider = binding.provider
       // Only `active` runs work. `staging` is the pre-commit state of the journaled
       // Chat handoff and must never execute against a Chat that can still roll back.
       const usable = assertUsable(binding ?? null, now)
@@ -1805,6 +2026,7 @@ export function createAgentSessionBindingsRouter(deps: AgentSessionBindingsDeps)
       const key = binding.targetKey
       const fenced = guard.fencedReason(key)
       if (fenced !== null) {
+        console.warn(`[agent-session-bindings] fence hit route=turn provider=${binding.provider} target=${opaqueRevision(key)} turnId=${turnId}`)
         return refuseTurn(fenced, { retryable: false, deliveryState: 'unknown' })
       }
 
@@ -1823,6 +2045,7 @@ export function createAgentSessionBindingsRouter(deps: AgentSessionBindingsDeps)
 
       const head = await readHead(binding.provider, binding.nativeThreadId)
       if (head === null) return refuseTurn('native_head_unavailable')
+      preTurnHeadDigest = head.digest
 
       // The attach baseline, advanced by each completed turn and by each explicit
       // Continue Anyway. Without the advance the SECOND turn on a binding always
@@ -1907,10 +2130,35 @@ export function createAgentSessionBindingsRouter(deps: AgentSessionBindingsDeps)
             // before the prompt rather than deliver a turn that poisons the next
             // occupancy check.
             if (outcome !== 'recorded') return false
-            recordedPids.push(pid)
+            recordedPids.push({ pid, startMs })
             return true
           },
         })
+        // Read the adapter's OWN account before it goes out of scope. This is the
+        // whole point of the evidence work: `classifyDelivery` collapses six distinct
+        // failures into one word, and the difference between them is what decides
+        // whether an automatic resolver could ever be safe. A 21-minute `timeout`
+        // means the child ran tool calls; a `provider_exit_nonzero` with a code means
+        // it exited on its own. Defensive reads — the adapter is injected in tests.
+        // AN UNREADABLE RESULT RECORDS NOTHING, not zeroes. Reading `{}` and
+        // deriving `exitCode: null, childReaped: false` states two facts about a
+        // child nothing is known about, and writes them indistinguishably from a
+        // confirmed-unreaped timeout — corrupting the one discriminator this
+        // evidence exists to establish.
+        const r = (result !== null && typeof result === 'object')
+          ? result as Record<string, unknown>
+          : null
+        adapterEvidence = r === null ? {} : {
+          adapterReason: typeof r.reason === 'string' ? r.reason : (r.ok === true ? 'ok' : undefined),
+          adapterDetail: typeof r.detail === 'string' ? r.detail : null,
+          exitCode: typeof r.exitCode === 'number' ? r.exitCode : null,
+          // FROM THE ADAPTER, never derived from exitCode. A signal-killed child
+          // reports `code === null`, so deriving it would report "never reaped" for
+          // the dominant timeout shape — backwards for the decision this informs.
+          childReaped: typeof r.reaped === 'boolean' ? r.reaped : undefined,
+          stderrClass: typeof r.stderrClass === 'string' ? r.stderrClass : undefined,
+          durationMs: typeof r.durationMs === 'number' ? r.durationMs : undefined,
+        }
         delivery = classifyDelivery(result)
       } catch (error) {
         console.error(`[agent-session-bindings] adapter threw: ${error instanceof Error ? error.message : error}`)
@@ -1928,7 +2176,30 @@ export function createAgentSessionBindingsRouter(deps: AgentSessionBindingsDeps)
         // Fenced under its own reason, not this turn's: `delivery_ambiguous`
         // describes what happened to THIS request, while a later caller needs to
         // be told the thread is shut and why it must be inspected first.
-        guard.fence(key, 'native_target_fenced')
+        // ONE resolved value for the record AND the log. The previous release
+        // fixed exactly this contradiction at the other fence site and this one
+        // re-committed it: the record said `unreadable_result` while the log said
+        // `unknown`, so an operator grepping for the sentinel found nothing.
+        const recordedReason = adapterEvidence.adapterReason ?? 'unreadable_result'
+        guard.fence(key, 'native_target_fenced', {
+          provider: binding.provider,
+          headBefore: head.digest,
+          turnId,
+          bindingId,
+          now: Date.now(),
+          ...adapterEvidence,
+          // NEVER blank. An unreadable adapter result and "nobody recorded it" are
+          // different facts and a missing field cannot tell them apart -- which is
+          // the whole reason this evidence exists.
+          adapterReason: recordedReason,
+          fenceSite: 'ambiguous',
+          spawns: [...recordedPids],
+        })
+        // A fence shuts a thread until a human acts, and until now it wrote NO log
+        // line at either site — so a fenced thread was discoverable only by trying
+        // to use it (Miles, 2026-08-18). Never log `key`: it embeds the private
+        // native thread id, which this router does not emit anywhere.
+        console.warn(`[agent-session-bindings] fence set site=ambiguous provider=${binding.provider} target=${opaqueRevision(key)} turnId=${turnId} bindingId=${bindingId} headBefore=${head.digest} adapterReason=${recordedReason} detail=${adapterEvidence.adapterDetail ?? 'none'} exitCode=${adapterEvidence.exitCode ?? 'null'} childReaped=${adapterEvidence.childReaped ?? 'unknown'} stderrClass=${adapterEvidence.stderrClass ?? 'none'} durationMs=${adapterEvidence.durationMs ?? 'unknown'} spawnCount=${recordedPids.length}`)
         return reportAmbiguous()
       }
 
@@ -1958,13 +2229,35 @@ export function createAgentSessionBindingsRouter(deps: AgentSessionBindingsDeps)
       if (deliveryAttempted) {
         // A bug in this route that happened AROUND a delivery is indistinguishable
         // from a delivery.
-        if (claimedKey !== null) guard.fence(claimedKey, 'native_target_fenced')
+        if (claimedKey !== null) {
+          guard.fence(claimedKey, 'native_target_fenced', {
+            provider: fenceProvider,
+            headBefore: preTurnHeadDigest,
+            turnId,
+            bindingId: null,
+            now: Date.now(),
+            // A bug in THIS route, not a provider outcome. Named rather than left
+            // blank so a later reader can tell "the adapter reported nothing" apart
+            // from "nobody recorded it". Any adapter evidence captured before the
+            // throw is still carried.
+            ...adapterEvidence,
+            adapterReason: adapterEvidence.adapterReason ?? 'route_error',
+            fenceSite: 'route_error',
+            spawns: [...recordedPids],
+          })
+          // `head` is scoped to the try, so `preTurnHeadDigest` is hoisted to the
+          // handler specifically to reach this site. It is null ONLY when the throw
+          // happened before the head was read. An earlier version of this line
+          // hardcoded `unavailable` and so contradicted the record it had just
+          // written — an operator would read "no baseline" off a fence that has one.
+          console.warn(`[agent-session-bindings] fence set site=route_error target=${opaqueRevision(claimedKey)} turnId=${turnId} headBefore=${preTurnHeadDigest ?? 'unavailable'} adapterReason=${adapterEvidence.adapterReason ?? 'route_error'} detail=${adapterEvidence.adapterDetail ?? 'none'} exitCode=${adapterEvidence.exitCode ?? 'null'} childReaped=${adapterEvidence.childReaped ?? 'unknown'} stderrClass=${adapterEvidence.stderrClass ?? 'none'} durationMs=${adapterEvidence.durationMs ?? 'unknown'} spawnCount=${recordedPids.length}`)
+        }
         reportAmbiguous()
       } else {
         refuseTurn('turn_failed')
       }
     } finally {
-      for (const pid of recordedPids) {
+      for (const { pid } of recordedPids) {
         try {
           ownership.release(pid)
         } catch (error) {