Relay 2.2.2: fix job delivery, backend shutdown and review limits

.claude-plugin/marketplace.json

@@ -5,13 +5,13 @@
   },
   "metadata": {
     "description": "Run jobs on Codex, Grok Build, ZCode and Kimi Code from Claude Code like Claude agents.",
-    "version": "2.2.1"
+    "version": "2.2.2"
   },
   "plugins": [
     {
       "name": "relay",
       "description": "Run jobs on Codex, Grok Build, ZCode and Kimi Code from Claude Code like Claude agents.",
-      "version": "2.2.1",
+      "version": "2.2.2",
       "author": {
         "name": "Libar"
       },

AGENTS.md

@@ -22,3 +22,8 @@ tests, fixtures, docs, skills, CHANGELOG entries, commit messages and pull reque
 - Every version field moves together with `node scripts/bump-version.mjs <version>`.
   `npm run check-version` checks them.
 - The CHANGELOG groups entries under the release heading. Don't add a heading per development version.
+
+## Release maintenance
+
+2.x is the final Relay line and takes bug fixes only: no new features. Validate its integration in
+Claude Code after harness upgrades.

README.md

@@ -14,6 +14,8 @@ with a state block that says what happened.
 1.0.6. [NOTICE](NOTICE) keeps the attribution and license terms. [docs/design.md](docs/design.md)
 explains the design.
 
+2.x is the final Relay line and takes bug fixes only.
+
 ## Backends
 
 A backend is named by its harness. The model family goes in `model:`.
@@ -35,8 +37,8 @@ Requirements:
 - Claude Code 2.1.271 or later, for hooks on `SubagentHandback`.
 - Node.js 22 or later.
 - macOS or Linux. The process sweep reads the process table with `ps`.
-- At least one backend CLI, signed in: the Codex CLI, Grok Build, ZCode or Kimi Code. Release 2.1.2
-  was checked live with Codex 0.163.0, Grok Build 1.0.60, ZCode 3.15.3 and Kimi Code 2.2.0.
+- At least one backend CLI, signed in: the Codex CLI, Grok Build, ZCode or Kimi Code. Release 2.2.2
+  was checked live with Codex 0.166.0, Grok Build 1.0.70, ZCode 3.16.2 and Kimi Code 2.4.1.
 
 ```bash
 claude plugin marketplace add libar-dev/relay-plugin-cc
@@ -218,8 +220,8 @@ is recorded as a job, so `/relay:status` lists it.
 | `--wait`, `--background` | run in the foreground, or as a Claude background task; without either, a review you type asks, and a review Claude starts runs in the foreground |
 | `[focus]` | text after the flags; `/relay:adversarial-review` only |
 
-Grok, ZCode and Kimi read only the diff inlined in their prompt, so they refuse a change above two files
-or 256 KiB before anything starts. An adversarial answer that does not match the review schema
+Grok, ZCode and Kimi read only the diff inlined in their prompt, so they refuse a change above 256 KiB
+before anything starts. An adversarial answer that does not match the review schema
 fails the review with the raw answer in the report.
 
 ### The stop-time review gate
@@ -319,7 +321,7 @@ runs on command hooks and the transcript.
 
 ## Development
 
-`package.json` names the package `@libar-dev/relay-plugin-cc`, version 2.2.0. It is private and
+`package.json` names the package `@libar-dev/relay-plugin-cc`, version 2.2.2. It is private and
 not published to npm. Clone the repository and run:
 
 ```bash
@@ -347,6 +349,8 @@ them); the version named above in this section is set by hand.
 
 Read [docs/architecture.md](docs/architecture.md) for the code layout and module rules,
 [docs/backends.md](docs/backends.md) for the backend contract,
-[docs/design.md](docs/design.md) for the behavior and the reasons behind it, and
-[docs/kimi.md](docs/kimi.md) for what the Kimi backend rests on and how to check a new Kimi version, and
-[docs/thread-surveys.md](docs/thread-surveys.md) for the surveys of real use that fixes came from.
+[docs/design.md](docs/design.md) for the behavior and the reasons behind it,
+[docs/kimi.md](docs/kimi.md) for what the Kimi backend rests on and how to check a new Kimi version,
+[docs/thread-surveys.md](docs/thread-surveys.md) for the surveys of real use that fixes came from, and
+[docs/relay-2.2.2-validation.md](docs/relay-2.2.2-validation.md) for this release's review findings,
+regressions and live-validation evidence.

docs/design.md

@@ -238,8 +238,10 @@ server that dies during initialization leaves an identity, and records its end:
 signal, an unfinished last line, the last 4096 bytes of stderr. `server.stderr.log` keeps at most
 1 MiB. JSONL lines have no size cap, so multi-megabyte results reach `events.jsonl` intact, and an
 unfinished last line is never parsed. A server that relay closes gets its stdin ended, then
-SIGTERM 50 ms later if it still runs, so a closed server's end record can show SIGTERM; that is
-relay's own close, not a crash, and the job's state was settled before it. When the server exits
+SIGTERM after 50 ms and SIGKILL after another 4,950 ms if it still runs. A process that survives
+SIGKILL for another five seconds is abandoned: its pipes close, its process is unreferenced, and its
+server record stays open for the next worker's sweep. Its late exit does not write to a job another
+worker may own. A close timeout is reported as such, not as an observed exit. When the server exits
 while a child of it holds a pipe, the client reads for 200 ms more, then closes the pipes, so the
 end is recorded and every pending request settles; a request to a server that has exited fails at
 once.
@@ -334,9 +336,11 @@ Before every relay tool call and at SubagentStop, the hook reads the transcript
 received message not yet delivered, in order: steered while the job runs, continued once it ended.
 
 - A message is recognised by its entry type and origin, never by its text: a `queued_command`
-  attachment from the coordinator (a `SendMessage` while the relay works), a `user` entry from the
-  coordinator (a message after the hand-back, taken out of the harness's fixed wrapper), or a plain
-  `user` turn after the prompt that is neither metadata (`isMeta`) nor a tool result.
+  attachment or `user` entry from the coordinator, a peer or a person, or a plain `user` turn after
+  the prompt that is neither metadata nor a tool result. Peer text comes from `origin.body`; the
+  coordinator's and person's fixed wrappers are removed. Text from any of those origins survives a
+  shared text/tool-result entry, while a tool-result-only entry stays out. A task notification stays
+  out too, and an entry of any other origin is reported as a violation rather than delivered.
 - A message that arrives before the job exists is delivered by `start` once the job exists.
 - Each delivery is recorded in `inbox/relay/<transcript entry uuid>.json` under the relay lock, with
   the inbox message id chosen before posting. A hook cut off half way posts again under the same
@@ -704,12 +708,13 @@ or `adversarial-review`, so `status` lists it. A Codex review turns off the job
   ZCode and Kimi, which have no schema parameter, get it in the prompt. A Grok review always runs
   under Grok's `read-only` profile, and a Kimi review refuses the file writes and terminal
   commands Kimi sends through the client, which is not an OS sandbox (Known limits, Kimi).
-- **Inline limits.** A Grok, ZCode or Kimi review reads only the inlined change (read-only ZCode and
-  Kimi have no shell to run git), so it sees the whole change or does not start. Untracked files count toward
-  the limits and are inlined whole, and a binary file is named and counts as a file. Above two files
-  or 256 KiB, the review is refused before any job or process exists, with a message naming
-  `--base`, `--scope` and Codex. A Codex adversarial review above the limits gets a summary and
-  reads the diff itself with read-only git commands; untracked files there are cut at 24 KiB each.
+- **Inline limit.** A Grok, ZCode or Kimi review reads only the inlined change (read-only ZCode and
+  Kimi have no shell to run git), so it sees the whole change or does not start. The actual serialized
+  repository context, including Git's rendered names and decoded file contents, must fit 256 KiB.
+  Above that, the review is refused before any job or process exists, with a message naming `--base`,
+  `--scope` and Codex. A file already known to be oversized is named without being read. A Codex
+  adversarial review above two files or 256 KiB gets a summary and reads the diff itself with
+  read-only git commands; untracked files there are cut at 24 KiB each.
 - **The schema.** An adversarial answer counts only if it matches
   `schemas/review-output.schema.json`: types, enums, required fields, ranges, array items and no
   extra properties. One JSON code fence around the whole answer is unwrapped. Anything else fails