mirror of
https://github.com/openswarm-ai/openswarm.git
synced 2026-09-11 12:17:45 +02:00
[eric] delete deprecated docs
This commit is contained in:
ciregenz
@@ -1,53 +0,0 @@
|
||||
# Phase 7: Squirrel vs NSIS A/B
|
||||
|
||||
NSIS is the shipped Windows installer and **stays the default**. Squirrel is a
|
||||
candidate only — it wins, and replaces NSIS, **only if** it is measurably faster
|
||||
*and* its auto-update rollback works on real Win 10/11 machines. Otherwise NSIS
|
||||
stays. This doc is the procedure to make that call with data, not vibes.
|
||||
|
||||
## Build both from the same commit
|
||||
|
||||
```powershell
|
||||
# NSIS (default, what ships today)
|
||||
pwsh scripts\build-app-win.ps1 -Sign # -> electron\dist\OpenSwarm-Setup-x64.exe
|
||||
|
||||
# Squirrel (candidate), same staged tree / same SHA
|
||||
pwsh scripts\build-app-win.ps1 -Sign -Squirrel
|
||||
```
|
||||
|
||||
The `-Squirrel` switch only overrides `win.target` (via
|
||||
`--config.win.target=squirrel`); signing, extraResources, and the bundled
|
||||
python/node/router are identical, so any measured difference is the installer
|
||||
itself, not the payload. Confirm both report the same provenance sha (Settings
|
||||
-> About -> Build, or the `[provenance]` line in backend.log).
|
||||
|
||||
## Measure on REAL Windows 10 and 11 (x64), clean machines
|
||||
|
||||
For each installer, on a fresh VM/box (no prior OpenSwarm install):
|
||||
|
||||
| Metric | How |
|
||||
|---|---|
|
||||
| Install time | wall-clock from launching the installer to the app window appearing |
|
||||
| First paint | `[perf] first-paint` in backend.log (`scripts/perf/parse-timing.js`) |
|
||||
| Backend ready | `[perf] backend-http-ready` |
|
||||
| Crashes | any crash on first launch; check `%APPDATA%\OpenSwarm\Crashpad` |
|
||||
| Auto-update | install an older build, then this one; confirm it detects, downloads, installs on quit, relaunches on the new version |
|
||||
| Rollback | after an update, force a downgrade/rollback path; confirm the previous version comes back cleanly and the feed isn't corrupted |
|
||||
|
||||
Run each 3x per OS and take the median. Compare against the Phase 0 baseline
|
||||
(file count 11,247 / 1.2 GB) and NSIS's own numbers.
|
||||
|
||||
## Decision gate
|
||||
|
||||
- Squirrel **wins** only if: median install + first-paint + backend-ready are
|
||||
faster than NSIS on BOTH Win 10 and Win 11, AND auto-update works, AND rollback
|
||||
works (including rebuilding whatever feed Squirrel's differential updates need).
|
||||
- Any of those fail -> **NSIS stays**, revert the target, keep `-Squirrel` as a
|
||||
dead experiment flag or remove it.
|
||||
|
||||
## Why this is the last phase
|
||||
|
||||
Squirrel changes the update feed format and rollback semantics. Switching it in
|
||||
without the rollback feed rebuilt strands users on a broken updater — the exact
|
||||
failure the rest of this plan exists to prevent. So it goes last, behind a flag,
|
||||
and only on proof.
|
||||
@@ -1,44 +0,0 @@
|
||||
# Secret rotation + history purge
|
||||
|
||||
The repo history contains real credentials that were committed long ago (the
|
||||
ones `.gitleaksignore` acknowledges). CI is green because those findings are
|
||||
allowlisted, but **allowlisting hides them, it does not remove them** — they're
|
||||
still in `git log` and still shipped in the app today. This is the real fix.
|
||||
|
||||
> Requires repo **admin** + access to the provider consoles + a **force-push**
|
||||
> (history rewrite). The agent can't do any of those, so this is a human runbook.
|
||||
|
||||
## 1. Rotate first (this is what actually kills the exposure)
|
||||
|
||||
Rotating invalidates the leaked value immediately, so even though it stays in
|
||||
history it becomes useless. Do this before bothering with the purge.
|
||||
|
||||
| Secret | Where it leaked (commit) | Rotate where |
|
||||
|--------|--------------------------|--------------|
|
||||
| Google OAuth client secret | `backend/apps/tools_lib/{oauth_providers,tools_lib}.py` (7239f70, 7c3da1a, cbefe89) | Google Cloud Console → APIs & Services → Credentials → the OAuth client → **Reset secret** |
|
||||
| PostHog API key | `backend/apps/analytics/collector.py` (8d09e46, b6f45e8) | PostHog → Project settings → rotate project API key (note: ingest keys are public by design — rotate only if it's a private key) |
|
||||
| 9router client secrets | `9router/**` (cf775b4, history-only; dir now fetched from npm) | Whichever provider each `clientSecret` belongs to; bump `ROUTER_VERSION` if the npm package itself shipped one |
|
||||
|
||||
After rotating, update wherever the build injects them (the `GOOGLE_OAUTH_*`
|
||||
GitHub Actions secrets + `backend/.env` production-injection step) to the new
|
||||
values, and cut a release so users get the rotated build.
|
||||
|
||||
## 2. Purge from history (optional, after rotation)
|
||||
|
||||
Redact the values from every commit with [git-filter-repo](https://github.com/newren/git-filter-repo):
|
||||
|
||||
```bash
|
||||
# expressions.txt: one `OLD_SECRET==>REDACTED` per line (the real old values)
|
||||
git filter-repo --replace-text expressions.txt
|
||||
```
|
||||
|
||||
Then the destructive part (admin only):
|
||||
|
||||
- `git push --force --all` and `git push --force --tags` (this is why the agent
|
||||
can't do it — force-push is denied and it rewrites every downstream commit hash).
|
||||
- Everyone re-clones (old clones still hold the secrets).
|
||||
- Re-create any protected-branch/tag rulesets if the rewrite trips them.
|
||||
|
||||
Because rotation (step 1) already neutralizes the secret, the purge is about
|
||||
hygiene, not urgency. Once both are done, drop the matching fingerprints from
|
||||
`.gitleaksignore`.
|
||||
Reference in New Issue
Block a user