@ssheleg/agent-sync 1.18.7 → 1.19.1

package/CHANGELOG.md

@@ -1,3 +1,88 @@
+## v1.19.1 — re-acquiring your own lease now refreshes it
+
+**`renew` was fixed for exactly this and `acquire` was not, in the same file.**
+`_refresh_lease`'s docstring records the bug it exists to have fixed: *"`renew`
+appended `op=renew` to the RECORD plane — which has not decided a lease since 1.0.0 —
+and touched a throttle file. The lock's own `ts` was written once, by `acquire`. So a
+run holding a lease lost it at TTL while still working."* The repair landed in `renew`.
+The own-lock branch of `acquire` kept doing precisely what that paragraph describes.
+
+**Two harms, and the second is the one that bites.** The lock stayed expired, so
+`classify_lock` read it as spent and another run's `acquire` took it over while this one
+believed it held — measured 2026-09-01: a foreign expired lock IS taken over, timestamp
+and host rewritten. And `_touch_renew` resets the marker `renew()` throttles on, so a
+re-acquire **actively suppressed the next real refresh** while refreshing nothing
+itself. That is the mechanism behind a lease expiring three times in one run against a
+450-step CI job.
+
+Measured before and after, on a lock aged past its own TTL:
+
+```
+old: won mykey (…)   ts after re-acquire: 2026-08-29T20:59:04Z   still expired
+new: won mykey (…)   ts after re-acquire: 2026-09-01T12:38:48Z
+```
+
+`check_acquire_refreshes_its_own_lease` is beside `check_renew_extends_the_lease`,
+watched failing against the previous commit with both of its assertions.
+
+**One finding retracted in the same run, and it belongs here.** `reap` refuses to clear
+a lock three days past a forty-five-minute TTL when this run's identity is the shared
+fallback, and that reads like a deadlock. It is not: `acquire` takes over an expired
+lock regardless of who wrote it — measured, timestamp and host rewritten — so the
+expired lock blocks nothing, and `reap`'s stricter bar for a *tidying* operation is
+correct as designed. The claim that expiry alone should license a delete was wrong
+about which operation matters.
+
+## v1.19.0 — the skill now promises the finish it has always performed
+
+Wave-4 close-out of the 2026-08-29 family audit (ASY-10), filed by this
+repository's own first eval run. One front-matter line changes; no behaviour
+does.
+
+- **ASY-10 — the submodule-finish request reached nothing, and now reaches this
+  skill on every sample.** `finish` exists for exactly one silent failure — a
+  submodule pushed while the parent still points at the commit before it — and
+  the description advertised none of its words, so "Finish a submodule change so
+  the parent points at the pushed commit" (eval q06) routed past the skill on
+  every probe of the v1.18.7 run. The *Use when* clause now names it, in the
+  words `finish` itself prints (`finishing work across a parent repo and its
+  submodules - clean, pushed, and pointed at`), and the trigger list carries the
+  symptom in both languages: `'the submodule is pushed, the parent points at the
+  old commit' / 'сабмодуль запушен, родитель на старом коммите'`. The
+  description is **951 characters — against the 970 house working limit, not the
+  1024 spec cap**: a first draft at 1017 passed `test/validate.py` and was
+  refused by the family's pinned auditor (`GAP DESC_HEADROOM`, exit 1, a CI
+  failure), so it was rewritten rather than shipped. `yaml.safe_load` reads the
+  front matter — the two things this family has shipped broken before.
+- **The fix is measured, not asserted, and both arms ran in this session.** Two
+  corpus files differing in exactly one line (`- agent-sync: …`, old vs new),
+  one enforced protocol, fresh context-free subagents: **q06 went 0/3 → 3/3 on
+  haiku and 0/3 → 3/3 on sonnet**, and every number in the results file comes
+  from the text that ships — the discarded 1017-char draft was re-measured from
+  scratch after the rewrite rather than having its numbers carried over. The
+  pilot probes are reported too, because
+  they changed the protocol: offered the corpus rather than required to read it,
+  all three sonnet probes answered with no tool call at all, from the machine's
+  own installed listing — so the read was made mandatory and provable, and both
+  arms re-run under it.
+- **Every other query was re-probed, because a widened trigger steals a
+  neighbour's.** All twelve queries × two models: **12/12 on both, train 6/6 and
+  validation 6/6**, up from 11/12. Nothing moved except q06 — including q07
+  (solo git branching) and q08 (the full delivery cycle), the two negatives with
+  the most to lose from submodule and finish vocabulary. Tables, the 68-probe
+  accounting across three arms and the limits are in `test/evals/RESULTS.md`;
+  the v1.18.7 rows are left standing beside the new ones rather than rewritten.
+- Ten routed triggers still advertised — the umbrella's own
+  `advertised_check.js` was run against this checkout (`ok: agent-sync
+  advertises all 10 routed trigger(s)`), not a copy of its table.
+- **Filed, not fixed: AS-10.** `npm test` accepts a description up to 1024
+  characters and CI's pinned auditor refuses one past 970, so a description in
+  that band goes green locally and red on the runner. That is AS-06's shape a
+  second time — the same auditor, the other half of the same budget — and the
+  fix belongs in `test/validate.py` with a planted self-test, not inside a
+  release whose subject is one description line. The retro entry records what
+  the divergence cost here.
+
 ## v1.18.7 — the evals stop being a claim, and the manifests say what shapes them
 
 Wave-3 close-out of the 2026-08-29 family audit (ASY-03, ASY-09). No behavior

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ssheleg/agent-sync",
-  "version": "1.18.7",
+  "version": "1.19.1",
   "description": "Let concurrent coding agents share one project without colliding — leases with TTL, race-free id reservation, a run journal and a generated board, over a pluggable knowledge cloud.",
   "bin": {
     "agent-sync": "bin/agent-sync.js"

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

@@ -2,7 +2,7 @@
   "$schema": "https://json.schemastore.org/claude-code-plugin-manifest.json",
   "name": "agent-sync",
   "displayName": "Agent Sync",
-  "version": "1.18.7",
+  "version": "1.19.1",
   "description": "Coordination layer for multi-agent repositories — leases with TTL, race-free ID reservation, a run journal, a cross-repo signal feed and a generated board, over a pluggable knowledge cloud.",
   "author": {
     "name": "ssheleg",

package/plugins/agent-sync/skills/agent-sync/SKILL.md

@@ -1,10 +1,10 @@
 ---
 name: agent-sync
-description: "Use when several coding agents work one repository at the same time and must not collide - claiming a task, reserving the next decision/question/ticket id, journaling a run, filing or answering a cross-repo dependency, or regenerating the shared board. Triggers - 'claim this task' / 'возьми задачу', 'who is working on X' / 'кто сейчас делает X', 'reserve an id' / 'зарезервируй id', 'sync the board' / 'обнови доску', 'set up agent coordination' / 'настрой координацию агентов', /agent-sync. Use it BEFORE editing any shared registry file (decisions, open questions, roadmap, workstreams, dependencies) in a project that has .claude/agent-sync.json, even when the user never mentions coordination - an unclaimed edit to those files is how two agents overwrite each other."
+description: "Use when several coding agents work one repository at the same time and must not collide - claiming a task, reserving the next decision/question/ticket id, journaling a run, moving a cross-repo dependency, regenerating the board, or finishing work across a parent repo and its submodules - clean, pushed, and pointed at. Triggers - 'claim this task' / 'возьми задачу', 'who is working on' / 'кто сейчас делает', 'reserve an id' / 'зарезервируй id', 'sync the board' / 'обнови доску', 'set up agent coordination' / 'настрой координацию агентов', 'the submodule is pushed, the parent points at the old commit' / 'сабмодуль запушен, родитель на старом коммите', /agent-sync. Use it BEFORE editing any shared registry file (decisions, open questions, roadmap, workstreams, dependencies) in a project that has .claude/agent-sync.json, even when the user never mentions coordination - an unclaimed edit to those files is how two agents overwrite each other."
 compatibility: "Requires the task-pipeline skill for its stages (npx sshlg-skills install). Needs python3 3.9+ (stdlib only, HTTP included - nothing to pip install) and bash for the hooks. The knowledge backend is configured per project; with none configured it degrades to git-file leases. Enforcement hooks are Claude Code only - on other agents the same checks run as a self-check."
 license: MIT
 metadata:
-  version: "1.18.7"
+  version: "1.19.1"
   author: ssheleg
 ---
 

package/plugins/agent-sync/skills/agent-sync/scripts/agent_sync.py

@@ -33,7 +33,7 @@ from datetime import datetime, timezone
 from pathlib import Path
 from typing import Any
 
-VERSION = "1.18.7"
+VERSION = "1.19.1"
 
 CONFIG_PATH = Path(".claude/agent-sync.json")
 ENV_FILE = Path(".env.agent-sync")
@@ -1646,6 +1646,21 @@ class Sync:
             except (json.JSONDecodeError, OSError):
                 held = {}
             if held.get("run") == self.rid:
+                # MOVE THE LOCK'S OWN `ts`, not just the throttle marker. This branch did
+                # exactly what `_refresh_lease`'s docstring describes as the bug it exists
+                # to have fixed — touch the throttle file and leave the timestamp the lease
+                # is arbitrated by — and the repair landed in `renew` and not here, in the
+                # same file, four minor versions ago.
+                #
+                # Two harms, and the second is the one that bites. The lock stayed expired,
+                # so `classify_lock` read it as spent and another run's `acquire` took it
+                # over while this one believed it held (measured 2026-09-01: a foreign
+                # expired lock IS taken over, timestamp and host rewritten). And
+                # `_touch_renew` resets the marker `renew()` throttles on — so a re-acquire
+                # actively SUPPRESSED the next real refresh, refreshing nothing itself.
+                # That is the mechanism behind a lease expiring three times in one run
+                # against a 450-step CI job on 2026-09-01.
+                self._refresh_lease(key)
                 self._touch_renew()
                 return True, self.rid
             if time.time() <= parse_iso(held.get("ts", "")) + int(held.get("ttl", self.ttl)):