RandyNorthrup · Oct 22, 2026 · Oct 22, 2026 · Oct 22, 2026 · Oct 22, 2026 · Oct 22, 2026
@@ -5,6 +5,16 @@ All notable changes to this project are documented here. The format follows
 [Semantic Versioning](https://semver.org/). Entries record what actually
 happened, not what was planned; superseded entries are kept.
 
+## [Unreleased]
+
+### Release infrastructure
+
+- The release's Open VSX check reads the universal package at its own
+  address. Once a version has packages for each platform, Open VSX's
+  plain version address can name any one of them, and 0.19.0's check
+  compared the universal package with another platform's (every package
+  was published and matched when checked by its own address).
+
 ## [0.19.0] - 2026-10-21
 
 ### Highlights

@@ -34521,50 +34521,11 @@ remain with the lead under the shared lane rules. Follow-up receipts are in
 
 - **Archived.** The record that stood here is in [docs/plan-archive/gates.md](docs/plan-archive/gates.md).
 
-**M121A rounds 3–4 gate ownership (2026-10-09, kubuntu).** The rig brief and
-shared lane rules forbid aggregate quality in this worktree; the lead owns
-that gate before integration. This lane runs the owning usage/money/schedule
-files at repository timeouts, typecheck, changed-file lint/format, localization
-and the remaining scoped static checks. Full production build retains the
-inherited What's New content-budget refusal; measure current production
-options/plugins in memory without editing the build or raising a cap. Byte
-counts and exact failure receipts are in `docs/certification/m121.md`.
-The unchanged zero-clone duplication gate also retains one six-line clone in
-`test/unit/windowsTrustedPath.test.ts`; that file matches `1494fbf87` byte for
-byte and is outside this lane's money scope. The touched migration test's
-duplicate setup is consolidated without dropping assertions. The lead owns
-the remaining Windows-test repair and the full artifact size/split replay:
-the refused build removes stale browser chunks before stopping, so those
-artifact checks cannot certify a complete current build. No gate is weakened.
-Round 4 replays the owning usage/money files, one unit-project typecheck and
-the scoped static checks and production measurements at the unchanged caps.
+- **Archived.** The record that stood here is in [docs/plan-archive/gates.md](docs/plan-archive/gates.md).
 
 - **Archived.** The record that stood here is in [docs/plan-archive/gates.md](docs/plan-archive/gates.md).
 
-**CIFIXV hosted visual repair (2026-10-09, linuxlt).** Own only the three
-visual failures at `f5e0760c3`: reconstruct schedules under the served source
-root, review any intended pixel changes before regenerating goldens, and
-refresh current-source classification hashes without changing coverage.
-Keep all thresholds and repository deadlines. Full quality remains the lead's
-gate under the shared lane rules; no merge, push or live/paid calls.
-
-- [x] Prove historical schedule loading fails before the repair and passes after.
-- [x] Resolve visual comparisons and identify the aggregate's actual failure.
-- [x] Refresh source receipts and pass the owning audit suite.
-- [x] Pass full visual, six-theme accessibility and default-timeout stability gates.
-- [x] Record evidence and local hooks-on commits in `docs/certification/cifixv.md`.
-
-Reviewed full baseline: `CIFIXV-integrated-accessibility-review-2026-10-09`,
-9,792 images / 447,310,229 bytes; all 884 distinct changed images reviewed,
-27 named scenes. Full pixel gate passed; six-theme scoped axe passed on
-1,632 pages with zero violations, retaining all 315 incomplete findings.
-All 43 owning tests passed at repository defaults; production build passed.
-The artifact action's `digest-mismatch: error` is an input;
-the aggregate requires successful schedule shards and complete receipts.
-Bounded external gate deferral: full `jscpd` still reports the inherited
-51-token clone in `test/unit/modelApiLoopGuarantees.test.ts:741` and `:890`.
-That suite is another lane's scope; leave code and duplication limits intact
-for the lead's integration gate. Receipt: `docs/certification/cifixv.md`.
+- **Archived.** The record that stood here is in [docs/plan-archive/gates.md](docs/plan-archive/gates.md).
 
 - **Archived.** The record that stood here is in [docs/plan-archive/gates.md](docs/plan-archive/gates.md).
 
@@ -39926,6 +39887,30 @@ Unresolved items remain delivery/security gaps, not accepted legal conclusions:
 
 ## 10. Definition of done and release records
 
+**0.19.0 released (2026-10-21, tag `v0.19.0` on main's release merge `4e09cab6ea`, PR #257, release run 39720213044).**
+PR #257 (`release/0.19.2`, head `915f409fe4`) merged after first-attempt
+green CI (39716712837) and Hosts (39716712802) runs; the merge commit's Hosts
+run (39719522423) was green and `rule2-gate.mjs release --sha 4e09cab6ea`
+read three green first runs; the annotated tag was pushed at 9:04 PM PT. The
+release run reused the CI run's verified packages (its lookup found
+39716712837 for the tag tree `f77cd9d93f`). Tag checks, reuse, the GitHub
+Release (46 assets: seven VSIX packages, the agent tarball, both
+inventories, `SHA256SUMS` and the Linux helper assets), Marketplace and npm
+succeeded. Open VSX failed at its own check, not at publishing: once a
+version has targeted packages, Open VSX's untargeted `/<version>` answer
+names whichever package it picks (`win32-arm64` to the script), so the
+universal package was compared with another target's bytes. Every package
+was uploaded; Open VSX activated the last ones by 04:15 UTC. Checked after
+the run, each of the seven packages on the Marketplace and on Open VSX
+(through each target's own address) is byte-identical to the release asset,
+npm's `muse-spark-code-acp@0.19.0` integrity equals the release tarball's,
+and every downloaded asset passes `SHA256SUMS`. The summary job is red for
+the Open VSX job; the badge refresh it skipped was run locally (exit 0). The
+check now reads `/universal/<version>` (`scripts/publish-registry.mjs`;
+`docs/certification/cifinal0190.md` §21). Real use before the tag:
+`docs/real-life/release-0.19.0-limits-W.md` (Windows, 5 model calls, every
+one `muse-spark-1.3-contributor`, $0.00).
+
 **0.19.0 preparation (2026-10-17; tagged again 2026-10-21).** `release/0.19.0` cut from `main-now` at
 `81dacaf5a7`, merging in order PR #185 (`fix/sandbox-say`: the Linux sandbox
 cannot-start notice), PR #189 (`ms/msm112`: the production build runs through

@@ -2,7 +2,7 @@
 
 <!-- Generated by scripts/gen-roadmap.mjs from PLAN.md, CHANGELOG.md and docs/roadmap/entries.json. Edit those, then run npm run roadmap:generate; npm run check:roadmap fails when this file is stale. -->
 
-<!-- Source fingerprint: fa463c377ffa0b203110cd4dd8e4ba5848de578ebb4ac852da0dbab0563bf424 -->
+<!-- Source fingerprint: 5eaf63720bb04dcccac8c90bcb17b65362690505a3f3bcaf733fcc34d43226c1 -->
 
 What Muse Spark Code has shipped, what is in the next release, what is being built and what is planned. Plans change: an item can move, shrink or be dropped, and no date is promised. Versions stay below 1.0; there is no 1.0 promise or date.
 

@@ -765,3 +765,56 @@ function (title) { context.test(title); }`), and `.timeout(0)` was
     refused by this PC's account), and no vault path reaches the process
     helpers whose deadlines changed. The toast check printed no log; it now
     prints the panel's log with the toasts, so a repeat names its cause.
+
+## 21. The 0.19.0 release run: Open VSX's check read another package
+
+Release run 39720213044 (`v0.19.0` on `4e09cab6ea`): tag checks, reuse of
+CI run 39716712837, GitHub Release, Marketplace and npm succeeded; the Open
+VSX job ended "openvsx: the published artifact does not match the release".
+
+- Cause: `verifyPublished` read `/api/RandyNorthrup/muse-spark-code/0.19.0`
+  for the untargeted package. With targeted packages present, that answer
+  named `win32-arm64` (three of three reads from Node; curl got
+  `universal`), so the universal VSIX was compared with the win32-arm64
+  bytes. `/api/RandyNorthrup/muse-spark-code/universal/0.19.0` names the
+  universal package for 0.19.0 and for 0.18.2 (universal only).
+- Open VSX state: all seven packages uploaded. `darwin-arm64`,
+  `linux-arm64` and `linux-x64` answered "Extension not found" for a few
+  minutes after the job (activated by 04:15 UTC), the other four at once.
+- After the run, with the fixed check, each of the seven VSIX packages on
+  the Marketplace and on Open VSX matches the release asset (14 of 14), npm's
+  integrity for `muse-spark-code-acp@0.19.0` equals the release tarball's
+  sha512, and the downloaded assets pass `SHA256SUMS` (8 of 8).
+- Drill, against the real Open VSX: the tag's `verifyPublished` on the
+  universal VSIX failed three of three ("openvsx published SHA-256 differs
+  from the release VSIX"), exit 1; the fixed one matched three of three,
+  exit 0, and still matches the win32-x64 package three of three.
+
+## 22. PR #258's first run: the macOS vault panel flow never unlocked
+
+PR #258 head `c7bdd3e943` (CI run 39722505089, capacity run 39722575285):
+macOS shard 15 only, `posixVaultPanelFlow`: after Unlock the panel still
+read `locked` when the step's 60 s wait ended, with no toast and no log. The
+product code is PR #257's final head's, green on every macOS shard there,
+and no path the macOS unlock takes reaches the process helpers whose
+deadlines changed (`runTreeSync` serves the playbook only; the identity
+probe passes its own timeout). With PR #257's third run (Windows, first
+open, two failure toasts) it is the second hosted-runner vault flow failure
+with an empty log. The step message now also names how many frames the
+panel published and its last states, so the next one shows whether the
+step ever ran; the defect is a recovery lane's to find and fix.
+
+PR #258's second run, head `d0a39a70ed` (CI run 39724540034, capacity run
+39724610207): macOS shard 8 ended "The runner has received a shutdown
+signal. This can happen when the runner service is stopped, or a manually
+started runner is canceled." and "The operation was canceled." No test
+failed; every other job passed. A rerun on the same commit is refused
+(AGENTS.md rule 2), so this record is the new commit.
+
+PR #258's third run, head `a6f4767c97` (CI run 39725758843, capacity run
+39725813062): Windows shard 9 only, `windowsBrokerClient.test.ts`: "Hook
+timed out in 30000ms" in the `beforeAll` that builds the native helpers with
+the person's own C# compiler, another time limit of ours on a slow hosted
+Windows runner. The seven vault suites whose setup builds or stages native
+helpers or starts a broker now run that setup with no time limit (owner
+2026-10-16); here the five Windows ones pass (14 tests, 59.7 s).
@@ -784,3 +784,30 @@ docs/plan-archive/milestones.md`: clean.
   `sharingArtifacts` + `estimatorPlanSnapshot` (79 passed),
   `estimatorPlanSnapshotRealPlan` + `estimatorPlanSnapshotSource` +
   `estimatorSnapshotStrict` (73 passed), `planDocument` (19 passed).
+
+## 2026-10-21: room for the 0.19.0 release record
+
+The 0.19.0 release record put `PLAN.md` at 3,566,310 bytes, 1,152 over the
+`check:plan` gate (85% of the 4 MiB bound, 3,565,158). Two §7 gate records
+of closed lanes moved to `docs/plan-archive/gates.md` with a pointer each;
+no test, script or open branch cites or touches them (PR #252 changes §3 and
+§6 only), and both milestones shipped:
+
+| Block                                        | Lines (at `4e09cab6e`) | Bytes | SHA-256     |
+| -------------------------------------------- | ---------------------- | ----- | ----------- |
+| CIFIXV hosted visual repair (2026-10-09)     | 34544–34567            | 1,646 | `3b0e0461…` |
+| M121A rounds 3–4 gate ownership (2026-10-09) | 34524–34540            | 1,287 | `66f4036a…` |
+
+After: 3,563,591 bytes, 1,567 bytes under the gate.
+
+### Proofs
+
+- Reassembly: the command rebuilt the old `PLAN.md` before each write,
+  SHA-256 `8bc0b6ef…` and `6719195b…`.
+- Parsed facts: identical; format drift after: 0.
+- `npx prettier --check PLAN.md docs/plan-archive/gates.md`: clean.
+- `npm run check:plan`: `plan-format-v1; 222 milestones; 0 drift`.
+- `npm run roadmap:generate`: `ROADMAP.md` unchanged.
+- `npm run check:reference`: exit 0.
+- `planReader` + `planCitation` + `autoCompact` + `paidMoneyStructure`: 135
+  passed.
@@ -5355,3 +5355,58 @@ quality is not claimed green; the rig brief delegates it to the lead.
   risk; see `docs/certification/m95-s.md`.
 
 <!-- plan-archive:end id="gates-5cf2e1004-35453" -->
+
+## CIFIXV hosted visual repair (2026-10-09)
+
+<!-- plan-archive:begin id="gates-4e09cab6e-34544" lines="34544-34567" at="4e09cab6e" bytes="1646" sha256="3b0e04618dda8b1104c9fcd7a7fe7569c5c5de4cd2c5291d7ad9efed44eea5aa" -->
+
+**CIFIXV hosted visual repair (2026-10-09, linuxlt).** Own only the three
+visual failures at `f5e0760c3`: reconstruct schedules under the served source
+root, review any intended pixel changes before regenerating goldens, and
+refresh current-source classification hashes without changing coverage.
+Keep all thresholds and repository deadlines. Full quality remains the lead's
+gate under the shared lane rules; no merge, push or live/paid calls.
+
+- [x] Prove historical schedule loading fails before the repair and passes after.
+- [x] Resolve visual comparisons and identify the aggregate's actual failure.
+- [x] Refresh source receipts and pass the owning audit suite.
+- [x] Pass full visual, six-theme accessibility and default-timeout stability gates.
+- [x] Record evidence and local hooks-on commits in `docs/certification/cifixv.md`.
+
+Reviewed full baseline: `CIFIXV-integrated-accessibility-review-2026-10-09`,
+9,792 images / 447,310,229 bytes; all 884 distinct changed images reviewed,
+27 named scenes. Full pixel gate passed; six-theme scoped axe passed on
+1,632 pages with zero violations, retaining all 315 incomplete findings.
+All 43 owning tests passed at repository defaults; production build passed.
+The artifact action's `digest-mismatch: error` is an input;
+the aggregate requires successful schedule shards and complete receipts.
+Bounded external gate deferral: full `jscpd` still reports the inherited
+51-token clone in `test/unit/modelApiLoopGuarantees.test.ts:741` and `:890`.
+That suite is another lane's scope; leave code and duplication limits intact
+for the lead's integration gate. Receipt: `docs/certification/cifixv.md`.
+
+<!-- plan-archive:end id="gates-4e09cab6e-34544" -->
+
+## M121A rounds 3–4 gate ownership (2026-10-09)
+
+<!-- plan-archive:begin id="gates-4e09cab6e-34524" lines="34524-34540" at="4e09cab6e" bytes="1287" sha256="66f4036ae37945ea9033f935e396eadba65e147b940bb4afeb6547cf873023e8" -->
+
+**M121A rounds 3–4 gate ownership (2026-10-09, kubuntu).** The rig brief and
+shared lane rules forbid aggregate quality in this worktree; the lead owns
+that gate before integration. This lane runs the owning usage/money/schedule
+files at repository timeouts, typecheck, changed-file lint/format, localization
+and the remaining scoped static checks. Full production build retains the
+inherited What's New content-budget refusal; measure current production
+options/plugins in memory without editing the build or raising a cap. Byte
+counts and exact failure receipts are in `docs/certification/m121.md`.
+The unchanged zero-clone duplication gate also retains one six-line clone in
+`test/unit/windowsTrustedPath.test.ts`; that file matches `1494fbf87` byte for
+byte and is outside this lane's money scope. The touched migration test's
+duplicate setup is consolidated without dropping assertions. The lead owns
+the remaining Windows-test repair and the full artifact size/split replay:
+the refused build removes stale browser chunks before stopping, so those
+artifact checks cannot certify a complete current build. No gate is weakened.
+Round 4 replays the owning usage/money files, one unit-project typecheck and
+the scoped static checks and production measurements at the unchanged caps.
+
+<!-- plan-archive:end id="gates-4e09cab6e-34524" -->
@@ -0,0 +1,69 @@
+# Real use: 0.19.0 after the time limits came out, Windows (the lead, this PC)
+
+Version: **0.19.0**, the agent package CI built from PR #257's final head
+`915f409fe4` (CI run 39716712837, artifact `muse-spark-code-acp`,
+`muse-spark-code-acp-0.19.0.tgz`, 3,485,063 bytes), installed with npm into
+its own folder outside any workspace; `--version` → `0.19.0`. The release
+reuses this run's packages. Platform: Windows 11 Pro 10.0.26200, x64.
+Backend: Muse Code 1.4.11 (the person's own install), contributor mode on in
+the person's own settings; the product started
+`muse-bin-1.4.11-R6667.1.exe serve --disable-sandbox --contributor-mode on
+--trust-workspace`. No API key was read or stored.
+
+What changed and is used here: no deadline of ours on report source reads
+(5 s before), on Windows PowerShell, `taskkill` and process-table reads, or
+on the stops after a turn's cancel (20 s before).
+
+Expected before the run: **5 model calls**, cap 40 calls and **$2.00**, every
+request on **`muse-spark-1.3-contributor`** by name, in an empty workspace.
+
+## Reports read the network with no deadline of ours — PASS
+
+`muse-spark-code-acp report project --network --format json --out <file>`
+against this repository: exit 0 in 12 s, no model call. Against the
+published 0.18.2 agent on the same repository at the same time:
+
+- `stores` reads `partial` (channels-unavailable: 0.19.0 is not yet on the
+  stores) where 0.18.2 read `unavailable`; `usage` reads `partial` (it names
+  the journal entries it excludes) where 0.18.2 read `unavailable`.
+- Every other source reads the same, except `planRevision`: `ok` on 0.18.2,
+  `unavailable` now. Not from this change: it compares PLAN.md with PLAN.md
+  at the latest release tag (`v0.18.2`), and today's PLAN reader finds six
+  format drifts in that older file (duplicate question ids such as
+  `Q-M95INT2`), so it refuses the comparison and says so. Recorded for the
+  recovery work, with `github` (`pulls-page-1: pull-response-invalid`,
+  `default-branch-page-1: http-422`) and `git` ("The source exceeds the
+  report read limit"), which read the same on 0.18.2.
+
+## A command runs through Windows PowerShell — PASS
+
+ACP session `01a161bf`, put on `muse-spark-1.3-contributor` by
+`session/set_config_option` before any prompt. "Run this exact PowerShell
+command once … `(Get-Process -Id $PID).ProcessName`" → the client was asked
+"PowerShell: (Get-Process -Id $PID).ProcessName", allowed once; the answer
+was `powershell`; `end_turn` in 20.5 s.
+
+## Stop ends a long command and everything it started — PASS
+
+"Run this exact PowerShell command once … `Start-Sleep -Seconds 300;
+'slept'`", allowed; the client sent `session/cancel` 5 s after the tool call
+started, as an editor's Stop does. The turn ended `cancelled` 90 ms after
+the cancel (8.6 s in all). Afterwards no process on this PC had
+`Start-Sleep -Seconds 300` in its command line.
+
+## Notes
+
+- The first attempt (session `01a161bd`) sent no prompt: the driver was
+  started without `--allow-contributor-models`, so the agent offered no
+  contributor model and refused the switch ("Invalid params"), and the
+  driver stops before any prompt when the session is not on the
+  contributor model. Muse Code's records show no call for it.
+- `session/delete` answered "Muse Code 1.4.11 cannot delete conversations.",
+  as the product documents.
+
+## Model calls
+
+Counted from `model_completed` events in Muse Code's own session records
+(`~/.local/share/muse/sessions/…/<id>/session.jsonl`, main sessions and
+subagents) for sessions `01a161bd` and `01a161bf`: **5 calls**, every one
+**`muse-spark-1.3-contributor`**, **$0.00** of the $2.00 cap.