PCIRCLE-AI · kevintseng · Sep 4, 2026 · Sep 4, 2026 · Sep 4, 2026 · Sep 4, 2026
diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
@@ -8,7 +8,7 @@
       "name": "memesh",
       "source": "./",
       "description": "MeMesh \u2014 agentic memory for coding agents. Captured from the agent's real work via hooks, recalled when it acts. One SQLite file, zero cloud required.",
-      "version": "4.8.3",
+      "version": "4.8.4",
       "author": {
         "name": "PCIRCLE AI"
       },

diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
@@ -4,7 +4,7 @@
   "author": {
     "name": "PCIRCLE AI"
   },
-  "version": "4.8.3",
+  "version": "4.8.4",
   "mcpServers": "./.claude-plugin/mcp.json",
   "homepage": "https://github.com/PCIRCLE-AI/memesh",
   "repository": "https://github.com/PCIRCLE-AI/memesh",

diff --git a/.codex-plugin/plugin.json b/.codex-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "memesh",
   "description": "MeMesh — agentic memory for coding agents.",
-  "version": "4.8.3",
+  "version": "4.8.4",
   "mcpServers": "./.codex-plugin/mcp.json"
 }
diff --git a/.gitignore b/.gitignore
@@ -123,3 +123,6 @@ docs/guides/
 PRE_RELEASE_CHECKLIST.md
 docs/notes/
 .smoke-test*.db-*
+
+# === qa:live-journey receipts (owner-machine evidence, never shipped) ===
+.qa/
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -4,15 +4,34 @@ All notable changes to MeMesh are documented here.
 
 ## [Unreleased]
 
+## [4.8.4] — 2026-09-04
+
 ### Added
 
 - **`npm run qa:pre-release` — one door before a release, that says what it did
   not check.** It runs `npm run build`, `verify:artifact` (lint, typecheck,
-  version coherence, doc claims, the isolated suite, the packed artifact and
-  every derived upgrade path — the same sequence `prepublishOnly` runs, now
-  named once instead of copied) and `audit:memory`, reports each step's real
-  exit code, and prints the checks it cannot run: the interactive live journey,
-  the post-release check, and the entry-point start gate.
+  version coherence, doc claims, the isolated suite, the packed artifact,
+  every derived upgrade path, and now the entry-point start gate below — the
+  same sequence `prepublishOnly` runs, now named once instead of copied) and
+  `audit:memory`, reports each step's real exit code, and prints the checks it
+  still cannot run here: the interactive live journey (gated separately by
+  `release:finish`, below) and the post-release check.
+- **A release gate that actually starts every shipped entry point.** Eight
+  binaries and nine hooks ship; every existing check only asked whether a
+  file existed, whether JSON parsed, or whether a referenced path resolved —
+  none of them ran the code. `scripts/check-entry-points-start.mjs`
+  (`npm run check:entry-points-start`, folded into `verify:release`) spawns
+  each of the 17 for real against a throwaway `MEMESH_DIR`: a CLI must accept
+  `--version`, an MCP server must exit 0 on stdin EOF, a long-lived daemon
+  must reach its "running" signal (a log line, or its socket file appearing),
+  and a host runtime with no config must fail closed with a named message,
+  never a raw stack trace. It also fails on any unresolved `${...}` left in
+  `.mcp.json` or `hooks/hooks.json`, evaluated against each manifest's own
+  real substitution environment — `CLAUDE_PLUGIN_ROOT` is always defined for
+  the plugin-loader-only `hooks/hooks.json`, but not for `.mcp.json`, which
+  Claude Code also auto-discovers with no plugin loader involved at all. The
+  Windows skip list is pinned to exactly one entry (`memesh-router`, no
+  `AF_UNIX` there) so it cannot grow silently.
 - **`npm run qa:post-release` — the check that runs on the machine, not on a
   fresh clone.** Every gate here runs on a fresh checkout or a fresh install,
   and all three release incidents lived in state that already existed: v4.7.0
@@ -26,9 +45,55 @@ All notable changes to MeMesh are documented here.
   resolve, not the first one: two installs on one PATH, four releases apart,
   is the shape of the incident, and a check that stops at the first hit cannot
   see it.
+- **Repeatable owner-run live delivery checks.** `npm run qa:live-journey -- --host codex|claude`
+  (`scripts/qa/live-journey.mjs`) starts this checkout's router in a throwaway
+  `MEMESH_DIR`, registers one real host session, sends one exact-session
+  envelope, and passes only on model-visible proof: the Codex path requires the
+  model to quote the envelope's `message_id` and `delivery_id` back on its next
+  turn; the Claude path requires an `intake` receipt written by the interactive
+  session the operator launched with the printed command. Both then stop the
+  session and require `recipient_unavailable` while the durable row stays
+  fetchable and the router still answers `discover`. It refuses to run against
+  `$HOME/.memesh`, reads no auth files, never runs in CI, and records its
+  limitations in the JSON report — including that the Codex registration is
+  harness-driven because `codex exec --ignore-user-config` bypasses the plugin
+  `SessionStart` hook, and that print-mode Claude is unsupported (#275).
+  Closes the "repeatable check in the repository" box on #270 and #272.
+- **A write-side reminder hook.** The read side of MeMesh was already automatic
+  (SessionStart and PreToolUse inject memories) but nothing prompted an agent
+  to *store* anything, so decisions made mid-session were routinely lost until
+  the user said "remember this" (#277). A ninth hook,
+  `scripts/hooks/decision-nudge.js`, runs on `PostToolUse` for `ExitPlanMode`
+  and `AskUserQuestion` — the two calls where a decision has most likely just
+  been made — and adds one line of context asking the model to `remember` it
+  if it is worth keeping. At most once per tool per session, enforced by an
+  `O_EXCL` flag file under `MEMESH_DIR/decision-nudge-flags/`; the hook never
+  opens the database, exits 0 on any malformed input, and writes nothing to
+  the graph itself. Hook inventories in the READMEs (three languages),
+  `docs/ARCHITECTURE.md`, `AGENTS.md`, `CODEMAP.md` and the skill now list nine
+  hooks, and `check-codemap-parity` no longer hard-codes the count.
 
 ### Changed
 
+- **`npm run release:finish` now runs the real-credential checks instead of
+  merely documenting that they exist.** `qa:pre-release` and `qa:live-journey`
+  were both available since the previous release but neither was in the one
+  command that actually cuts a release — a check nobody has to run is a check
+  that gets skipped exactly when a release is rushed. `finish-release.mjs` now
+  runs `npm run qa:pre-release` itself (build + `verify:artifact` +
+  `audit:memory`, several minutes, streamed live) and blocks on its exit code;
+  a receipt cannot substitute here because it can go stale the moment the next
+  commit lands. `qa:live-journey` cannot run unattended — it needs a Codex
+  login or a person at an interactive Claude Code session — so it stays
+  receipt-based: `npm run qa:live-journey -- --host codex|claude --out
+  .qa/<host>-report.json` writes a report, and `release:finish` requires ONE
+  of the two hosts' reports to be readable, `verdict: "PASS"`, recorded
+  against a clean tree, and naming this exact commit — an older PASS proves an
+  earlier revision, not this one. Neither host is preferred; only Codex can be
+  driven unattended today, but a human-run Claude receipt satisfies the gate
+  exactly as well. `.qa/` is gitignored — a receipt is owner-machine evidence,
+  never shipped.
+
 - **The packed-upgrade gate derives its upgrade paths instead of pinning
   them.** `scripts/smoke-packed-upgrade.mjs` named both ends by hand
   (`expectedPreviousVersion = '4.8.2'`, `expectedCandidateVersion = '4.8.3'`,
@@ -45,6 +110,19 @@ All notable changes to MeMesh are documented here.
 
 ### Fixed
 
+- **The Claude Code plugin cache never cleaned up an old version after an
+  upgrade — only the ONE it had just replaced.** `upgrade-plugin.sh`'s atomic
+  swap always removed the previous cache directory, but anything left behind
+  by an interrupted upgrade, or one from before this swap mechanism existed,
+  had no path back to zero: measured on the maintainer's machine, 9 stale
+  version directories, 1.2 GB, with nothing ever sweeping them. A new
+  `sweep_stale_cache_versions` removes every OTHER directory under the cache
+  root whose name is exactly `<major>.<minor>.<patch>` after a successful
+  upgrade — never the version just installed, and never the registry's own
+  recorded install path even when that path is a stray non-canonical
+  directory the "repairing it" branch above deliberately leaves for a human,
+  whatever it happens to be named.
+
 - **Three release-gate guards that could not fail, and the silent failure one
   of them was hiding.** A mutation audit reintroduced the defect each guard was
   written for and re-ran the suite: with all three defects present at once,
@@ -451,37 +529,6 @@ All notable changes to MeMesh are documented here.
   independent `readUpdateCheckCache()` path formula is untouched by this —
   it only reads, and the filename scheme did not change.
 
-### Added
-
-- **Repeatable owner-run live delivery checks.** `npm run qa:live-journey -- --host codex|claude`
-  (`scripts/qa/live-journey.mjs`) starts this checkout's router in a throwaway
-  `MEMESH_DIR`, registers one real host session, sends one exact-session
-  envelope, and passes only on model-visible proof: the Codex path requires the
-  model to quote the envelope's `message_id` and `delivery_id` back on its next
-  turn; the Claude path requires an `intake` receipt written by the interactive
-  session the operator launched with the printed command. Both then stop the
-  session and require `recipient_unavailable` while the durable row stays
-  fetchable and the router still answers `discover`. It refuses to run against
-  `$HOME/.memesh`, reads no auth files, never runs in CI, and records its
-  limitations in the JSON report — including that the Codex registration is
-  harness-driven because `codex exec --ignore-user-config` bypasses the plugin
-  `SessionStart` hook, and that print-mode Claude is unsupported (#275).
-  Closes the "repeatable check in the repository" box on #270 and #272.
-
-- **A write-side reminder hook.** The read side of MeMesh was already automatic
-  (SessionStart and PreToolUse inject memories) but nothing prompted an agent
-  to *store* anything, so decisions made mid-session were routinely lost until
-  the user said "remember this" (#277). A ninth hook,
-  `scripts/hooks/decision-nudge.js`, runs on `PostToolUse` for `ExitPlanMode`
-  and `AskUserQuestion` — the two calls where a decision has most likely just
-  been made — and adds one line of context asking the model to `remember` it
-  if it is worth keeping. At most once per tool per session, enforced by an
-  `O_EXCL` flag file under `MEMESH_DIR/decision-nudge-flags/`; the hook never
-  opens the database, exits 0 on any malformed input, and writes nothing to
-  the graph itself. Hook inventories in the READMEs (three languages),
-  `docs/ARCHITECTURE.md`, `AGENTS.md`, `CODEMAP.md` and the skill now list nine
-  hooks, and `check-codemap-parity` no longer hard-codes the count.
-
 ## [4.8.3] — 2026-08-31
 
 ### Fixed

diff --git a/CODEMAP.md b/CODEMAP.md
@@ -1,6 +1,6 @@
 # CODEMAP
 
-**Version**: 4.8.3
+**Version**: 4.8.4
 
 A navigation map for the codebase: *"I want to change X — which file?"* For the
 design rationale behind these modules see [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md);
@@ -146,12 +146,21 @@ The same `operations.ts` memory functions run identically from all three transpo
 - Owner-run live checks (never CI): `scripts/qa/live-journey.mjs` — `npm run qa:live-journey -- --host codex|claude`
   drives a real Codex thread or an interactive Claude channel session and requires model-visible proof.
   Its pure half is pinned by `tests/qa/live-journey.test.ts`; the contract is in `docs/platforms/agent-messaging.md`.
-- Release gates, in the order a release meets them: `npm run qa:pre-release`
-  (`scripts/qa/pre-release.mjs`) runs build + `verify:artifact` + `audit:memory` as one door
-  and prints what it could not check; `verify:artifact` is the same sequence `prepublishOnly`
-  runs, named once. `scripts/smoke-packed-upgrade.mjs` derives every upgrade path it proves
-  from `package.json` and the registry (`scripts/lib/upgrade-matrix.mjs`) instead of pinning
-  a version pair. After publishing, `npm run qa:post-release` (`scripts/qa/post-release.mjs`)
-  checks registry acceptance, a fresh install from the registry, and whether this machine is
-  on the release — read-only, printing fixes rather than running them.
+- Release gates, in the order a release meets them: `verify:release` (called by `qa:pre-release`
+  below) now ends with `npm run check:entry-points-start`
+  (`scripts/check-entry-points-start.mjs`), which spawns every one of the 17 shipped bins and
+  hooks for real and fails on any unresolved `${...}` left in `.mcp.json`/`hooks/hooks.json`.
+  `npm run qa:pre-release` (`scripts/qa/pre-release.mjs`) runs build + `verify:artifact` +
+  `audit:memory` as one door and prints what it could not check; `verify:artifact` is the same
+  sequence `prepublishOnly` runs, named once. `scripts/smoke-packed-upgrade.mjs` derives every
+  upgrade path it proves from `package.json` and the registry (`scripts/lib/upgrade-matrix.mjs`)
+  instead of pinning a version pair. `npm run release:finish` (`scripts/finish-release.mjs`) runs
+  `qa:pre-release` itself and blocks on its exit code, and requires a `qa:live-journey` receipt
+  under `.qa/` (named `<host>-report.json` — codex or claude, either satisfies it) that is
+  `PASS`, clean-tree, and names the exact commit being released
+  (`scripts/lib/release-preconditions.mjs`'s
+  `findUsableLiveJourneyReceipt`) — real-credential checks CI cannot run, now required rather
+  than merely available. After publishing, `npm run qa:post-release`
+  (`scripts/qa/post-release.mjs`) checks registry acceptance, a fresh install from the registry,
+  and whether this machine is on the release — read-only, printing fixes rather than running them.
 - Version anchors that must agree on a bump: `package.json`, both root entries in `package-lock.json`, `.claude-plugin/plugin.json`, `.claude-plugin/marketplace.json`, `herdr-plugin.toml`, `CHANGELOG.md`, `CODEMAP.md`, `docs/ARCHITECTURE.md`, and `docs/api/API_REFERENCE.md`. Run `npm run build` after to regenerate `dist/skills-manifest.json`.
diff --git a/dist/skills-manifest.json b/dist/skills-manifest.json
@@ -8,7 +8,7 @@
     },
     {
       "path": ".claude-plugin/plugin.json",
-      "sha256": "fe68e32aaf8cd1ec56472bd058b1149f8451938524bf945ad27c4ba9cd362136",
+      "sha256": "2b425a0b6ca6c6a2a549d273616ee0aa418a485960aa5f05ba44a38a6d49e9b5",
       "bytes": 572
     },
     {
@@ -18,7 +18,7 @@
     },
     {
       "path": ".codex-plugin/plugin.json",
-      "sha256": "4a31bc9ce1ef62f7203ed9b3a9d4a9aa4b9a95a19f37f60cf4cfbdffa0ca95d8",
+      "sha256": "1ce4c7b5a8f04cfe1dde55adb6b1af580141de3a73eca0ee5a7bbb066d39b725",
       "bytes": 154
     },
     {

diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md
@@ -1,6 +1,6 @@
 # MeMesh Plugin Architecture
 
-**Version**: 4.8.3
+**Version**: 4.8.4
 
 > Looking for "which file do I change for X?" — see [CODEMAP.md](../CODEMAP.md).
 

diff --git a/docs/api/API_REFERENCE.md b/docs/api/API_REFERENCE.md
@@ -1,7 +1,7 @@
 # MeMesh Plugin -- API Reference
 
 **Protocol**: Model Context Protocol (MCP) over stdio
-**Version**: 4.8.3
+**Version**: 4.8.4
 **Compatibility**: Works with Claude Code plugins, Claude Managed Agents (via MCP connector), and any MCP-compatible client.
 
 **Native Integrations**: Beyond MCP, MeMesh integrates as a native memory provider for Hermes Agent (Python `MemoryProvider` plugin) and OpenClaw (TypeScript memory-capability plugin) — same tier as their built-in backends, not HTTP bridges. See [docs/platforms/](../platforms/) for platform-specific guides.

diff --git a/docs/platforms/agent-messaging.md b/docs/platforms/agent-messaging.md
@@ -212,10 +212,15 @@ close that gap by requiring evidence that could only have come out of a running
 model.
 
 ```bash
-TMPDIR=/private/tmp npm run qa:live-journey -- --host codex  --out codex-report.json
-TMPDIR=/private/tmp npm run qa:live-journey -- --host claude --out claude-report.json
+TMPDIR=/private/tmp npm run qa:live-journey -- --host codex  --out .qa/codex-report.json
+TMPDIR=/private/tmp npm run qa:live-journey -- --host claude --out .qa/claude-report.json
 ```
 
+`.qa/` is where `npm run release:finish` looks for these reports (any ONE
+host's PASS, against the exact commit being released, is enough — see
+`scripts/lib/release-preconditions.mjs`'s `findUsableLiveJourneyReceipt`). The
+directory is gitignored; a report is owner-machine evidence, never shipped.
+
 `TMPDIR` is not decoration on macOS. The router's Unix socket lives beside the
 database inside the temporary directory, and `AF_UNIX` caps a socket path at
 104 bytes; the platform default `os.tmpdir()` spends about half of that before

diff --git a/herdr-plugin.toml b/herdr-plugin.toml
@@ -13,7 +13,7 @@
 # is the honest floor rather than a copied one.
 id = "memesh"
 name = "MeMesh"
-version = "4.8.3"
+version = "4.8.4"
 min_herdr_version = "0.7.0"
 description = "Local SQLite memory shared across coding agents — decisions, lessons, and why the code looks the way it does."
 platforms = ["linux", "macos", "windows"]

diff --git a/package-lock.json b/package-lock.json
diff --git a/package.json b/package.json
@@ -1,7 +1,7 @@
 {
   "name": "@pcircle/memesh",
-  "version": "4.8.3",
-  "description": "MeMesh \u2014 agentic memory for coding agents. Captured from the agent's real work via hooks, recalled when it acts. One SQLite file, zero cloud required.",
+  "version": "4.8.4",
+  "description": "MeMesh — agentic memory for coding agents. Captured from the agent's real work via hooks, recalled when it acts. One SQLite file, zero cloud required.",
   "main": "dist/index.js",
   "type": "module",
   "bin": {
@@ -50,7 +50,8 @@
     "check:surface-parity": "node scripts/check-readme-tool-parity.mjs && node scripts/check-codemap-parity.mjs",
     "audit:prod": "node scripts/check-consumer-audit.mjs",
     "audit:memory": "node scripts/audit/memory-invariants.mjs",
-    "verify:release": "npm run lint && npm run typecheck && node scripts/check-version-coherence.mjs && node scripts/check-generated-mirror.mjs && node scripts/check-agent-message-sync.mjs && npm run check:surface-parity && node scripts/check-doc-claims.mjs && node scripts/audit/verification-audit.mjs && npm run audit:prod",
+    "check:entry-points-start": "node scripts/check-entry-points-start.mjs",
+    "verify:release": "npm run lint && npm run typecheck && node scripts/check-version-coherence.mjs && node scripts/check-generated-mirror.mjs && node scripts/check-agent-message-sync.mjs && npm run check:surface-parity && node scripts/check-doc-claims.mjs && node scripts/audit/verification-audit.mjs && npm run audit:prod && npm run check:entry-points-start",
     "release:finish": "node scripts/finish-release.mjs",
     "verify:artifact": "npm run verify:release && npm run test:isolated && npm run test:packaged && npm run test:packaged:upgrade",
     "prepublishOnly": "npm run build && npm run verify:artifact",