@staix/agent-hub 0.30.0 → 0.32.0

package/docs/verification/2026-11-06-rooms-presence-cost.md

@@ -0,0 +1,152 @@
+# Rooms presence producer cost
+
+Refs #1204, CI task 742. Historical macOS comparisons and Linux candidate
+measurements and targeted hosted macOS candidate acceptance are recorded below.
+
+## Failed jobs and revision attribution
+
+- [PR 1217 macOS job](https://github.com/STAIxBWLB/agent-hub/actions/runs/42098376188/job/127501666095)
+  checked out `92486ab5c48b4c0281643750590fa6693aaaa7e6` and timed out the
+  unchanged 1,000-flip test at 20,025.68 ms.
+- [PR 1215 macOS job](https://github.com/STAIxBWLB/agent-hub/actions/runs/42074839676/job/127420743582)
+  checked out `dc9e482c659a29fa3f6b4597bc04219f5b5f3fbd` and timed out at
+  20,034.80 ms.
+- The workflow checks out the PR head. Neither failed head contains T2 commit
+  `3185dd4c57e36213a8ca398dd304f3284daa03ed`. Their rooms files, Bus,
+  delivery journal and delivery test are byte-identical to
+  `b7aec7d5621e25e0ea1df1e5b8b7366b92512bfe` in those paths.
+- The requested b7..3185 range has only two intervening commits: handoff T2
+  `b5c2525495b42a986207308af815e530f78c0c61`, then rooms T2. The old
+  producer path reproduces at b7 itself, so the observed failures do not establish
+  T2 as a first bad commit.
+
+## Measured producer paths
+
+The repo-only preload `scripts/diagnose-presence-cost-742.ts` counts fixed
+method names and inclusive wall time, never argument text or ids. The original
+test and its 20-second guard are unchanged.
+
+The natural run is separate from a controlled scheduling probe. In the probe,
+the first hot publication waits for the existing serial seal before calling the
+original producer. No state flag is overwritten. This exercises an idle producer
+instead of relying on a boot-time publication still being in flight.
+
+Pinned Bun 1.4.2, Linux x64, frozen installs on dgx02:
+
+| Runtime and mode | Flip phase | Sequence reservations in phase | Store transactions in phase | Bus persists, whole case |
+| --- | ---: | ---: | ---: | ---: |
+| Pre-fix natural, `0a32d86a` | 2 ms | 0 | 0 | 4 |
+| b7, controlled idle entry | 10,645 ms | 1,000 | 3,173 | 4 |
+| b5c, controlled idle entry | 11,324 ms | 1,000 | 3,161 | 4 |
+| Pre-fix `ab751794`, controlled idle entry | 11,868 ms | 1,000 | 4,145 | 4 |
+| Fixed FIFO candidate `2452b905`, controlled idle entry | 109 ms | 7 | 33 | 4 |
+
+The pre-fix ab751 phase spent 2,401 ms inside sequence reservation and 5,205 ms
+inside Store transactions. These measurements overlap and must not be added.
+Transaction timing includes SQL and durable I/O; this is not a separate fsync
+syscall trace. The Store uses DELETE journaling and synchronous EXTRA. The data
+does not attribute all remaining time to either crypto or the filesystem.
+
+b7 and b5c were separate historical checkouts with the same diagnostic preload.
+0a32/ab751 include only close guards and the diagnostic on top of 3185; their room
+producer was still unchanged. The b7/b5c comparison and named-red checks ran from
+the diagnostic harness at `11e3121f51eff2a9cd60eb95400e568051163072`.
+
+The code defect is cancellation of scheduled coalesced work: ordinary calls
+clear the pending republish before starting another seal. If the producer is
+already busy, calls collapse; if it starts idle, an awaited loop can repeatedly
+reserve durable counters. Task 740's persist fix remains effective: all measured
+cases retain four Bus persists.
+
+## Fix and regressions
+
+- Ordinary calls reuse pending republish; response/forced-request priority stays.
+- Chat observations precede the chat, and the chat slot is reserved synchronously
+  before awaiting the earlier presence. This avoids overtaking by a concurrent
+  chat. Tests check ciphertext and real Store insertion order.
+- Receipt durability, ACK ordering, caps and actual-send exposure/current fences
+  remain covered by unchanged backpressure tests and new focused regressions.
+- Closed public and deferred publication return before touching the journal or
+  observers. Live storage failures still propagate.
+- A failed presence barrier remains an error of its reserved chat, rather than
+  a successful empty send or an unhandled rejection.
+
+## Hosted macOS reference comparison
+
+The [reference job](https://github.com/STAIxBWLB/agent-hub/actions/runs/42106515415/job/127522644864)
+ran on Darwin arm64 with Bun 1.4.2. Each result below is the unchanged test with
+the same diagnostic preload; controlled scheduling is labelled separately.
+
+| Runtime and mode | Whole test | Flip phase | Reservations | Store transaction time in phase |
+| --- | ---: | ---: | ---: | ---: |
+| b7, natural | 17,806.52 ms | 17,110 ms | 1,000 | 1,194 ms |
+| b7, controlled idle entry | 18,798.21 ms | 16,656 ms | 1,000 | 1,438 ms |
+| b5c, natural | 698.92 ms | 4 ms | 0 | 0 ms |
+| b5c, controlled idle entry | 17,084.92 ms | 16,440 ms | 1,000 | 1,065 ms |
+| 3185, natural | 710.99 ms | 2 ms | 0 | 0 ms |
+| 3185, controlled idle entry | 17,604.82 ms | 16,902 ms | 1,000 | 1,139 ms |
+
+The slow producer therefore costs about 16.4-17.1 ms per flip on this runner.
+For 3185, sequence reservation accounts for 511 ms inside the overlapping
+1,139 ms Store transaction time; durable I/O is not the dominant measured time.
+The remaining seal/async producer work was not split into separate CPU/syscall
+traces. All cases retain four Bus persists, confirming the task-740 fix remains.
+
+The range's lower bound is not unconditionally fast: b7 naturally selects the
+slow path in this job. A monotonic good/bad bisect cannot assign the defect to a
+newer commit; exhaustive comparison of the two intervening commits shows the
+pre-existing scheduling path. This is not a claim that the failed jobs merely
+needed their timeout increased.
+
+The job completed all six cases above, then failed because abbreviated reference
+924 was absent from that checkout. That final workflow error does not invalidate
+the completed cases, but supplies no 924 measurement. The follow-up explicitly
+fetches its full SHA and checks the candidate separately.
+
+Controlled red evidence uses unchanged historical source with only the new
+regression file copied into an isolated checkout:
+
+- At 3185, the count test fails its named assertion: expected one reserve,
+  received 1,001.
+- At pre-FIFO-fix 7b99, the concurrent-chat test fails its named assertion with
+  B before A.
+- The existing seed classifier accepted both as assertion failures, not timeouts
+  or runtime diagnostics.
+
+## Hosted macOS candidate acceptance
+
+The [candidate job](https://github.com/STAIxBWLB/agent-hub/actions/runs/42110940560/job/127537997562)
+passed on Darwin arm64, Bun 1.4.2 at
+`e269d5c3be4a81fbd95220d760fb1f104fae3e3f`:
+
+- The unchanged, unprofiled AC1 test passed in 787.72 ms under its original
+  20-second guard.
+- All 19 focused close, coalescing and unchanged backpressure tests passed,
+  with 210 expectations and no failures.
+- Natural candidate profile: whole test 692.06 ms, flip phase 2 ms, no sequence
+  reservations or Store transactions in that phase.
+- Controlled idle-entry candidate profile: whole test 783.51 ms, flip phase
+  114 ms, six reservations and 26 Store transactions in that phase.
+- The explicitly fetched historical 924 head passed naturally in 669.40 ms,
+  but controlled idle entry took 18,277.53 ms (17,679 ms in the flip phase,
+  1,000 reservations, 2,883 Store transactions). Its transaction time was
+  2,082 ms, overlapping the 1,078 ms sequence-reservation time.
+
+This confirms the scheduling-dependent cause and candidate improvement on
+macOS without changing an assertion or timeout. The job is targeted evidence,
+not a full PR gate. Production source is unchanged since 11e312; subsequent
+changes add evidence and remove the temporary diagnostic workflows.
+
+## Evidence limits
+
+- Latest focused source check at 11e312 passed 12 tests, 0 failed, 94 expectations.
+  Earlier 2452 passed 18 focused/backpressure tests and the controlled AC1 case.
+- The first reference run measured b7, b5c and 3185; the successful candidate
+  job additionally measured the full 924 SHA. Neither temporary workflow is
+  retained in the final change.
+- The original failed macOS jobs did not record per-flip counters. Attribution
+  of those exact executions to this reproduced path remains an inference;
+  hosted reference measurements establish the slow path; the separate candidate
+  job supplies targeted macOS acceptance. Linux timing alone is not macOS acceptance.
+- No original test assertion or timeout was changed. No Mac-local gate, live
+  room enrollment, native-account acceptance or release is claimed.

package/docs/verification/claude-shell-wrapper-native.py

@@ -68,7 +68,7 @@ results=[]
 for raw_binary in args.claude:
  binary=pathlib.Path(raw_binary).resolve(); identity=ROOT/('identity-'+binary.name); identity.mkdir()
  version=subprocess.run([str(binary),'--version'],env={'HOME':str(identity),'CLAUDE_CONFIG_DIR':str(identity/'config'),'PATH':'/usr/bin:/bin'},capture_output=True,text=True,timeout=10)
- assert version.returncode==0 and version.stdout.strip().split()[0] in ('2.1.309','2.1.310','2.1.314','2.1.316'),(version.returncode,version.stdout)
+ assert version.returncode==0 and version.stdout.strip().split()[0] in ('2.1.309','2.1.310','2.1.314','2.1.316','2.1.318'),(version.returncode,version.stdout)
  label=version.stdout.strip().split()[0]; digest=sha(binary)
  for mode in ('context','normalized'):
   d=ROOT/(label+'-'+mode); d.mkdir(); home=d/'home';home.mkdir(); work=d/"work 'quoted'";work.mkdir();(work/'sub').mkdir()

package/docs/verified.json

@@ -2,7 +2,7 @@
   "schemaVersion": 1,
   "documents": {
     "README.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "2f4c372a6ce48dc01acb4a002f6a46e9557cbd17",
       "paths": [
         "package.json",
         "src/",
@@ -11,14 +11,14 @@
       ]
     },
     "docs/security.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "2f4c372a6ce48dc01acb4a002f6a46e9557cbd17",
       "paths": [
         "src/",
         "templates/"
       ]
     },
     "docs/operations.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "2f4c372a6ce48dc01acb4a002f6a46e9557cbd17",
       "paths": [
         "package.json",
         "src/",
@@ -35,7 +35,7 @@
       ]
     },
     "docs/quickstart.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "2f4c372a6ce48dc01acb4a002f6a46e9557cbd17",
       "paths": [
         "package.json",
         "src/",
@@ -43,7 +43,7 @@
       ]
     },
     "docs/agent-notes/adapters.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "2f4c372a6ce48dc01acb4a002f6a46e9557cbd17",
       "paths": [
         "src/adapters/",
         "src/pi/",
@@ -81,7 +81,7 @@
       ]
     },
     "docs/agent-notes/benchmarks.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "af37f8fd8b44a2c566aa47629ec5b02ee9f91d69",
       "paths": [
         "scripts/benchmarks/",
         "test/benchmarks/",
@@ -93,7 +93,7 @@
       ]
     },
     "docs/agent-notes/budget.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "2f4c372a6ce48dc01acb4a002f6a46e9557cbd17",
       "paths": [
         "src/hub/budget.ts",
         "src/cli/statusline-tee.ts",
@@ -117,7 +117,7 @@
       ]
     },
     "docs/agent-notes/bus.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "2f4c372a6ce48dc01acb4a002f6a46e9557cbd17",
       "paths": [
         "src/hub/bus.ts",
         "src/hub/envelope.ts",
@@ -132,7 +132,7 @@
       ]
     },
     "docs/agent-notes/daemon.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "2f4c372a6ce48dc01acb4a002f6a46e9557cbd17",
       "paths": [
         "src/hub/daemon.ts",
         "src/hub/control-client.ts",
@@ -223,7 +223,7 @@
       ]
     },
     "docs/agent-notes/local-worker.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "af37f8fd8b44a2c566aa47629ec5b02ee9f91d69",
       "paths": [
         "src/adapters/local-worker.ts",
         "src/local/",
@@ -243,7 +243,7 @@
       ]
     },
     "docs/agent-notes/tasks.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "2f4c372a6ce48dc01acb4a002f6a46e9557cbd17",
       "paths": [
         "src/hub/tasks.ts",
         "src/hub/board.ts",
@@ -291,7 +291,7 @@
       ]
     },
     "docs/agent-notes/tests.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "4ad77748c96383bdef8e55a11ff53808a1359500",
       "paths": [
         "test/",
         "scripts/check.sh",
@@ -320,7 +320,7 @@
       ]
     },
     "docs/agent-notes/README.md": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "2f4c372a6ce48dc01acb4a002f6a46e9557cbd17",
       "paths": [
         "src/cli/console-state.ts",
         "src/cli/attention.ts",
@@ -329,7 +329,7 @@
       ]
     },
     "src/cli/help.ts": {
-      "verifiedAgainst": "8612cfd740abd013689532d993ebc981548510bb",
+      "verifiedAgainst": "af37f8fd8b44a2c566aa47629ec5b02ee9f91d69",
       "paths": [
         "src/cli/main.ts",
         "src/cli/launch.ts",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@staix/agent-hub",
-  "version": "0.30.0",
+  "version": "0.32.0",
   "description": "Native multi-agent hub: Claude Code, Codex, Kimi Code, Pi and local inference as peers in one project",
   "license": "MIT",
   "type": "module",
@@ -17,18 +17,17 @@
     "bun": ">=1.4.2"
   },
   "dependencies": {
-    "@staix/agent-mem": "0.20.1"
+    "@staix/agent-mem": "0.21.0"
   },
   "devDependencies": {
     "@modelcontextprotocol/sdk": "^1.27.1",
+    "@staix/agent-console": "0.1.3",
     "@types/bun": "^1.3.0",
     "typescript": "^5.8.0"
   },
   "files": [
     ".claude-plugin",
     "src",
-    "src/pet/window.jxa.js",
-    "src/pet/window.html",
     "plugins",
     "templates",
     "scripts/check-docs.mjs",

package/plugins/agent-hub/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-hub",
-  "version": "0.30.0",
+  "version": "0.32.0",
   "description": "Channel between Claude Code and the agent-hub daemon: peer messages from Codex, Kimi and the local worker arrive as channel events; hub_send replies.",
   "author": {
     "name": "Young Joon Lee",