Error fixes

Paste your error, get a fix that was tested

Every fix here comes from a learned-failure record in one of my projects' Salvor memory: what broke, why, what fixed it and how the fix was checked. Names, paths and hosts are removed.

Google OAuth refresh token dies every 7 days

Report agent memory · 2026-09-28

Symptom: A scheduled job worked for a week, then the token refresh returned HTTP 400 invalid_grant: "Token has been expired or revoked".
Cause: The OAuth consent screen was in Testing mode. In Testing, Google issues refresh tokens that expire after 7 days (refresh_token_expires_in = 604799).
Fix: Re-authorize to get going again. The permanent fix is to move the consent screen to Production (which may need Google's review). Make the job fail loudly with "re-authorization needed" instead of producing an empty report.
How it was checked: The new token's refresh_token_expires_in field; a reminder before day 7 while still in Testing.

Device-code (TV) OAuth client can't get the analytics scope

Report agent memory · 2026-09-03

Symptom: Authorization through the device-code flow worked for uploads but was refused for the YouTube Analytics read scope, whatever the consent screen said.
Cause: The "TV and limited input" client type cannot be granted sensitive scopes such as yt-analytics.readonly.
Fix: Create a second OAuth client of type Desktop app and use the loopback flow for analytics, with its own token file. Keep the device-flow client for what it can do.
How it was checked: Each token file only holds the scopes its client type can get; both jobs run with their own token.

MCP server fails to connect with CONNECT_TIMEOUT

Agent setup memory · 2026-10-02

Symptom: An MCP server failed almost every start with CONNECT_TIMEOUT, but started fine when run by hand.
Cause: It was launched with npx -y package@latest. npx checks the registry and may download and unpack the package before it runs, which can take longer than the client's connect timeout.
Fix: Install a pinned version once into a prefix you can write to, and point the config at that local binary instead of npx.
How it was checked: Several client restarts in a row, all connected. Note: a plugin sync can put the npx line back.

ffmpeg has no drawtext or libx264, and there's no sudo

Agent setup memory · 2026-10-13

Symptom

"No such filter: 'drawtext'" or "Unknown encoder 'libx264'". The only ffmpeg was a minimal build, and apt install needs root.

Cause

The bundled ffmpeg was built with --disable-everything. Installing a full one normally needs root.

Fix

Let apt list the packages with --print-uris (no root needed), download them with wget, unpack each with dpkg-deb -x into one folder, and point LD_LIBRARY_PATH at its lib directory.

apt-get install --reinstall -y --print-uris ffmpeg 2>&1 | grep -oP "^'[^']+'" | tr -d "'" > uris.txt
mkdir pkgs sysroot && (cd pkgs && wget -q -i ../uris.txt)
for f in pkgs/*.deb; do dpkg-deb -x "$f" sysroot/; done
export LD_LIBRARY_PATH=$PWD/sysroot/usr/lib/x86_64-linux-gnu
sysroot/usr/bin/ffmpeg -filters | grep drawtext

How it was checked

ffmpeg -filters lists drawtext; a test render with text, H.264 and AAC plays.

Scripting a password SSH login without sshpass or expect

Agent setup memory · 2026-10-14

Symptom

A script must log in to a password-only server. sshpass: command not found, no expect, no sudo to install them.

Cause

OpenSSH reads passwords from the terminal by default, so a script can't type one in.

Fix

OpenSSH can ask a helper program instead. Set SSH_ASKPASS to a small script that prints the password, SSH_ASKPASS_REQUIRE=force (OpenSSH 8.4+), and run ssh under setsid -w so it has no terminal. Keep the helper in a private temp folder and delete it after.

export SSH_ASKPASS=$TMP/askpass.sh SSH_ASKPASS_REQUIRE=force
setsid -w ssh user@host 'echo OK' < /dev/null

How it was checked

OK and the host name come back; scp works the same way.

Remotion audio fade silently doesn't render

Video pipeline memory · 2026-09-05

Symptom: A fade on a long audio Sequence never happened in the final video, with no error. Sometimes ffmpeg failed with "Missing ')' or too many args in 'if(between(t,...".
Cause: A function volume={(f) => ...} on a long Sequence is baked into one giant nested ffmpeg if(between(...)) expression, one branch per frame. Past a few hundred frames it overflows the expression parser.
Fix: Use flat volumes in Remotion and apply fades after rendering with ffmpeg afade. Short per-frame volume functions (a few frames) are fine.
How it was checked: ffmpeg -af astats on the export: the fade is visible in the levels and the peak stays under 0 dB.

LLM API suddenly returns model_not_found

Fact-check API memory · 2026-08-28

Symptom: Calls that worked before now fail with model_not_found for the default model.
Cause: The provider retired the model ID hard-coded as the default.
Fix: Pick a current model from the provider's live model list, keep the ID in config rather than code, and log the model name with every failure.
How it was checked: A startup check that compares the configured model against the provider's model list.

NotFittedError after reloading a saved TF-IDF index

Fact-check API memory · 2026-08-24

Symptom: Search worked until the service restarted; then query() raised NotFittedError ("not fitted").
Cause: Only the matrix was saved and loaded. The vectorizer that turns a query into a vector was never refit (or saved) on load.
Fix: On load, refit the vectorizer on the stored documents, or save it together with the matrix (for example with joblib).
How it was checked: A test that creates a fresh instance, loads from disk without adding anything, and runs a query.

Retry loop gives up on timeouts and 5xx errors

Fact-check API memory · 2026-08-29

Symptom: The code looked like it retried 3 times, but timeouts and server errors returned empty results immediately.
Cause: The loop only retried on HTTP 429. Every other failure fell through to return [].
Fix: Retry on timeouts, connection errors and 5xx as well, with backoff. Log the final failure instead of returning an empty list.
How it was checked: A test that makes the first call time out or return 503 and expects a retry.

Something fails, but there's no error anywhere

Fact-check API memory · 2026-08-25

Symptom: Fetches and parses returned nothing, and the logs were clean.
Cause: Seven bare except blocks caught the errors and did nothing with them.
Fix: Log inside every except block (at least the exception and what was being processed). Catch specific exceptions where you can.
How it was checked: grep for except blocks without a log call; a forced failure shows up in the log.

Live page calls localhost instead of the deployed API

Fact-check API memory · 2026-08-29

Symptom: The page worked on my machine; on the live site the requests never reached the backend.
Cause: The page's default API URL was http://localhost:8000.
Fix: Default to the deployed API URL and only use localhost when the page itself is served from localhost.
How it was checked: Open the live page with the browser's network tab and check which host the request goes to.

A page in a language subfolder loads without styles

Fact-check API memory · 2026-08-29

Symptom: The translated page under /tr/ showed plain unstyled HTML. The stylesheet request was a 404.
Cause: It used assets/style.css, which from inside tr/ points to tr/assets/style.css.
Fix: Use ../assets/ (or root-relative /assets/) for pages inside subfolders.
How it was checked: An internal link and asset scan after adding pages to a subfolder.

Visitors still see the old site after a deploy

Static site memory · 2026-08-27

Symptom: Changes were live on the server, but returning visitors kept getting the old pages.
Cause: The service worker served its cache. Its cache name hadn't changed, so it never refreshed.
Fix: Bump the cache name in sw.js on every deploy that changes files. Read the live value first: once the server was already ahead of the local copy.
How it was checked: curl the live sw.js after deploy and compare the cache name.

Upload reports success, but the file isn't where it should be

Static site memory · 2026-08-28

Symptom: scp exited with 0 both times, yet the live URL still served the old file or a 404.
Cause: The deploy command put the file in the wrong folder. Exit code 0 only means the copy happened somewhere.
Fix: Deploy with rsync --files-from from the repository root so paths are kept, then check every changed URL on the live site.
How it was checked: A curl diff of each changed live file against the local one.

Pages in the sitemap stay "Crawled, currently not indexed"

Static site memory · 2026-09-11

Symptom: Search Console kept these pages as crawled but not indexed for weeks.
Cause: They were listed only in the sitemap. No other page linked to them with a normal link (some were reachable only through JavaScript).
Fix: Give every page at least one static <a href> link from a related page. Add an orphan check to the site scan.
How it was checked: A crawl from the home page reaches every URL in the sitemap.

hreflang points to a page that doesn't exist

Static site memory · 2026-09-24

Symptom: Pages declared an English and x-default alternate, but those URLs returned 404.
Cause: The tags were added for an English version that had not been built.
Fix: Only declare hreflang for pages that exist, in both directions. Have the site scan request every hreflang target.
How it was checked: Every hreflang URL returns 200 and links back.

On-screen game buttons don't respond on phones

Static site memory · 2026-08-28

Symptom: Arrow buttons worked with a mouse, but tapping them on a phone did nothing useful.
Cause: They only had onclick handlers.
Fix: Add touchstart handlers with preventDefault, registered with { passive: false }.
How it was checked: A real emulated tap (for example Playwright with a phone profile and .tap()), not a grep for handlers.

Arrow keys move the game and scroll the page

Static site memory · 2026-09-24

Symptom: Pressing arrow keys moved the player and scrolled the whole page at the same time.
Cause: The keydown handler used the keys without stopping the browser's default action.
Fix: Call e.preventDefault() for the keys the game handles (arrows, space).
How it was checked: Play on a page that is taller than the window.

Clock hands drawn 90° off on canvas

Static site memory · 2026-09-24

Symptom: Every time showed rotated by a quarter turn: 6:00 drew the hour hand at 3 and the minute hand at 9. All tests passed.
Cause: The −90° (12 o'clock) offset was applied twice: the hand shape already pointed up, and π/2 was still subtracted from the angle. The tests read the answer value, not the pixels.
Fix: Apply the offset exactly once. Check drawings with known inputs (3:00, 6:30, 9:15, 12:45) on a screenshot.
How it was checked: Screenshots of the known times before shipping any game where the answer is drawn.

SVG <use> elements paint as giant blobs

This site's memory · 2026-10-15

Symptom: An animation with 100 small nodes passed all of the agent's checks, but in a real browser the nodes covered the page as huge colour blobs.
Cause: Each node was <use href="#symbol"> without width and height, so the symbol rendered at 100% of the parent viewport.
Fix: Use sized primitives (circle, rect, path), or give every <use> an explicit width and height.
How it was checked: A real browser screenshot of every SVG before it ships. An agent without a browser can't see this.

SyntaxError in a Python f-string that holds CSS or JS

This site's memory · 2026-10-14

Symptom: The build script failed with "f-string: expecting '}'" after inline JavaScript was added to the page template.
Cause: The template is one big f-string. The { } of the JavaScript and CSS were read as Python placeholders.
Fix: Double every literal brace inside the f-string ({{ and }}), or move the script into a separate file.
How it was checked: The build runs and the generated page contains single braces.

Old backup copies were publicly reachable

This site's memory · 2026-10-14

Symptom: An old version of the home page could still be opened through its .bak file name.
Cause: Backups were made next to the live files, inside the web root, and the server serves any file there.
Fix: Keep backups outside the web root and delete stray .bak and old copies from it.
How it was checked: A listing of the web root after every deploy shows no .bak or old copies.

Where these come from

Salvor is a repo-native memory layer for coding agents. When an agent hits a failure that is worth remembering, it proposes a learned-failure record, and it is saved only after a human says yes.

The next agent session reads those records first, so the same mistake is caught before it repeats. This page is a cleaned, public copy of some of those records.

Disclosure: I contribute to Salvor. The records are mine; the tool is Dan's open-source project.

How I tested Salvor → · Salvor on GitHub ↗

Found a mistake, or a fix that worked differently for you? Write to me. Contact →