fix(launcher): harden updater crash recovery

This commit is contained in:
Erik 2026-08-14 23:12:15 +02:00
parent 2d2a5b5046
commit 1955ca8ab5
27 changed files with 2714 additions and 544 deletions

View file

@ -471,7 +471,8 @@ gate (user): clean-profile first-run against real DATs.
verify, unpack to `DataDirectory/app/<version>/`, atomic `current.json`
pointer swap, refuse while any session runs, keep previous version for
one-step rollback.
- Launcher self-update: staged download + rename-dance on next start.
- Launcher self-update: staged download + target-local atomic replacement on
next start.
- Session-config composition targets `app/current`'s binaries.
**Acceptance:** manifest/download/verify/swap tests against a local HTTP
@ -488,8 +489,11 @@ before doing network, extraction, or activation work.
The production feed is pinned to GitHub owner/repository
`eriknihlen/acdream`; the launcher reads
`https://github.com/eriknihlen/acdream/releases/latest/download/manifest.json`.
Tests may inject a loopback HTTP URI, but production artifacts and redirects
must use HTTPS. `manifest.json` is:
Tests use a separate internal fixture constructor that may admit loopback HTTP;
that allowance never propagates to the production feed. Production manifest
and artifact URIs use HTTPS. Automatic redirects are disabled and every
redirect hop is validated before it is requested; redirect loops, a chain over
five hops, and any HTTPS-to-HTTP downgrade are rejected. `manifest.json` is:
```json
{
@ -557,14 +561,18 @@ same-volume directory rename.
write-through temporary-file + same-directory atomic rename. The last valid
pointer is also atomically preserved as `current.previous.json`; startup may
restore that exact backup only when `current.json` is missing/malformed and
the referenced version verifies. Orphan LA10 staging directories and pointer
temporaries are transaction-owned by exact names and are removed under the
update lease. A corrupt installed version is never silently selected; the
explicit one-step rollback swaps the two verified pointer versions.
the referenced version verifies. Orphan LA10 staging directories, download
archives, corrupt-version quarantine directories, and pointer temporaries are
transaction-owned by exact lowercase GUID names and are removed only under the
exclusive update lease; near-matching user names are preserved. A corrupt
installed version is never silently selected; the explicit one-step rollback
swaps the two verified pointer versions.
`DataDirectory/app/.update-session.lock` is the cross-process barrier. Each
supervised launcher activity holds a shared OS handle from before executable
resolution until terminal process observation; an update/rollback holds the
resolution until terminal process observation; launcher disposal requests
child termination and does not release that handle until the child is actually
observed terminal. An update/rollback holds the
exclusive handle for its entire recovery/download/extract/promote/pointer
transaction. Failure to acquire the exclusive handle is an immediate refusal,
not a wait behind a running session. The open handle, not lock-file contents,
@ -572,11 +580,11 @@ owns the lease and therefore releases after process death.
Launcher self-update staging lives at
`DataDirectory/launcher-update/transactions/<transactionId>/` and the sole
durable authority is `DataDirectory/launcher-update/pending.json`:
durable authority is `DataDirectory/launcher-update/pending.json` (schema 2):
```json
{
"schemaVersion": 1,
"schemaVersion": 2,
"transactionId": "0123456789abcdef0123456789abcdef",
"state": "staged",
"version": "1.2.3",
@ -591,17 +599,62 @@ durable authority is `DataDirectory/launcher-update/pending.json`:
}
```
Before mutation a next-start helper copied outside the target directory
atomically advances the plan to `applying` and fills `apply` with each path's
`hadOriginal` bit. It waits for the initiating launcher PID without invoking a
shell, moves originals into the transaction backup tree, then moves verified
staged files into place. It never opens a target with truncate/overwrite. On
success the plan becomes `awaitingConfirmation`; the new launcher confirms at
its first managed instruction, after which backup and plan cleanup is safe. An
Before mutation the verified staged launcher becomes the next-start helper and
waits for the initiating launcher PID without invoking a shell. It first copies
the complete verified payload into the target-local
`.acdream-self-update-<transactionId>/incoming/` tree. The plan then advances
to `applying`; `apply` is an ordinally sorted union of new payload paths, the
owned metadata path, and obsolete paths from the previous ownership record:
```json
[
{ "path": "acdream-launcher.exe", "operation": "install", "hadOriginal": true },
{ "path": "obsolete.dll", "operation": "remove", "hadOriginal": true }
]
```
Existing targets are replaced with one same-filesystem atomic replace whose
backup is also target-local. Previously absent noncanonical files use one
same-filesystem rename; obsolete owned files use one rename into backup. The
canonical launcher path therefore contains either the complete old file or the
complete new file at every durable crash boundary. Rollback reverses the same
operations atomically and is idempotent after a process/power loss. Linux mode
bits come from the verified incoming file. A helper that cannot immediately
acquire the exclusive update lease defers the staged plan and exits without
restarting the old launcher, preventing restart loops.
Successful application writes strict target ownership metadata at
`<launcher directory>/launcher.install.json`:
```json
{
"schemaVersion": 1,
"version": "1.2.3",
"rid": "win-x64",
"files": [
{ "path": "acdream-launcher.exe", "sha256": "<64 hex characters>", "size": 123, "unixMode": 0 }
]
}
```
The archive may not supply that reserved metadata path. A prior valid record is
the only authority for obsolete-file removal; the first managed update does
not infer ownership of unrelated legacy files. On success the plan becomes
`awaitingConfirmation`; the new launcher confirms at its first managed
instruction, after which the helper releases its lease and the confirmed
launcher reclaims plan, data-transaction, and target-local residue. An
`applying` plan is rolled back before retry, and failure to start/confirm the
new launcher restores every original (and removes every no-original target).
All plan paths are re-derived/contained under the pinned data root except the
target directory, which must equal the actual launcher base directory.
Reading `pending.json` never performs cleanup. Ordinary startup attempts the
exclusive lease without waiting and skips update cleanup entirely when another
session/staging transaction owns it. All plan paths are re-derived/contained
under pinned roots; the target directory must equal the actual launcher base
directory.
Every portable archive and persisted relative path rejects Windows device
segments on every host: `CON`, `PRN`, `AUX`, `NUL`, `CLOCK$`, `CONIN$`,
`CONOUT$`, `COM1`-`COM9`, `LPT1`-`LPT9`, and the Windows-equivalent superscript
forms `COM¹`/`COM²`/`COM³` and `LPT¹`/`LPT²`/`LPT³`, including extensions.
## LA11 — closeout

View file

@ -298,19 +298,19 @@ preview would be a deliberate divergence we are NOT taking.
install to `DataDirectory/app/<version>/`; atomic pointer swap
(`current.json`); never while any session is running; keep the previous
version for one-step rollback.
- **Launcher self-update:** same feed; staged download; rename-dance swap
on next start (a running exe can't replace itself on Windows).
- **Launcher self-update:** same feed; staged download; target-local atomic
replacement on next start after the running process exits.
- **Feed hosting:** GitHub Releases (user-confirmed). Manifest and zips
are release assets; the launcher pins the repo/owner in its config.
The exact v1 manifest, extracted-version record, `current.json` activation
pointer, shared-session/exclusive-update OS lease, and durable self-update
plan are pinned in
pointer and launcher ownership record, shared-session/exclusive-update OS
lease, and durable self-update plan schema 2 are pinned in
`docs/plans/2026-08-14-launcher-campaign.md` under **Pinned updater
contracts (v1, BINDING)**. That section is normative: implementations reject
unknown/duplicate fields and unsupported versions, use strict SemVer 2.0
precedence, verify bounded streamed downloads before safe ZIP extraction, and
derive all mutable staging/backup paths from the application data root. The
use per-hop redirect validation plus same-filesystem atomic replacement. The
LA9 DAT/pak install record remains the sole content descriptor fed to session
configs; LA10 changes only which verified `app/current.json` client binaries
the process supervisor executes.