Auto-update falsely reports failure on WSL Ubuntu - cold package prepare (~2m 27s) exceeds hardcoded 120s timeout

[vojvojtech]

Environment

• OS: WSL Ubuntu (Windows 11)
• kbagent version: 0.46.1 → 0.47.1

---

Root cause

The startup auto-update hook uses a hardcoded 120s timeout for the uv tool install --upgrade git+... subprocess (auto_update.py, _perform_update). On WSL Ubuntu with a cold uv package cache, preparing all 46 packages takes ~2m 27s. The subprocess is killed at 120s → TimeoutExpired → returns False → "Auto-update failed" is printed. On the next invocation with a warm cache, the update succeeds instantly — confirming the problem is purely the timeout, not a genuine install failure.

---

Bug 1 — False failure on cold cache

Reproduction:

# Seed a cold cache at v0.46.1
uv tool install git+https://github.com/keboola/cli@v0.46.1 --reinstall

kbagent doctor
# → "Updating kbagent v0.46.1 -> v0.47.1..."
# → "Auto-update failed; continuing with current version."

# Run again (warm cache) — now it succeeds
kbagent doctor
# → "Updating kbagent v0.46.1 -> v0.47.1..."
# → "Updated to v0.47.1. Re-launching..."

Key evidence from the cold-cache uv reinstall:

Prepared 46 packages in 2m 27s   ← exceeds hardcoded 120s

---

Bug 2 — kbagent update reports "already up to date" when it isn't

During the same session, after auto-update failed, running kbagent update returned:

kbagent v0.46.1 (already up to date)

…while kbagent --version confirmed it was still on v0.46.1 and v0.47.1 was available. The explicit update command silently reported wrong state instead of retrying.

---

Bug 3 — kbagent --json update startup hook and subcommand disagree

When running kbagent --json update, the startup hook fires (because argv[1] is --json, not update) and prints "Auto-update failed". But the explicit update subcommand runs afterward and returns JSON with "updated": true:

{
  "kbagent": {
    "updated": true,
    "current_version": "0.46.1",
    "latest_version": "0.47.1",
    "message": "Updated kbagent from v0.46.1 to v0.47.1. Restart your shell to use the new version."
  }
}

Same process, same invocation: human output says failure, machine output says success. Any script parsing the JSON exit would believe the update succeeded; any human reading the terminal would believe it failed.

This also exposes the argv[1] check issue: kbagent --json update triggers the startup hook, but kbagent update does not.

---

Suggested fixes

Short-term: Raise the timeout in _perform_update from 120s to at least 300s, or extract it as a named constant UPDATE_TIMEOUT_SECONDS.

Medium-term: Run the install in a background process and print "Update downloading in background — restart your shell when done." CLI is never blocked regardless of network/cache state.

For Bug 2 & 3: The "already up to date" check and the argv[1] check for skipping the startup hook both need reviewing — they interact badly when the previous auto-update silently failed.

---

Full terminal sessions

Session 1 — cold cache, false failures, JSON discrepancy

vojta@NTB:~$ kbagent doctor
Updating kbagent v0.46.1 -> v0.47.1...
Auto-update failed; continuing with current version.
Updating keboola-mcp-server v1.61.3 -> v1.63.1 (via uv_tool)...
Updated keboola-mcp-server to v1.63.1.
╭─────────────────────────────────────────────────── kbagent doctor 
[... doctor output omitted, checks passed ...]
────────────────────────────────────────╯
vojta@NTB:~$ kbagent doctor
Updating kbagent v0.46.1 -> v0.47.1...
Auto-update failed; continuing with current version.
╭─────────────────────────────────────────────────── kbagent doctor 
[... doctor output omitted, checks passed ...]
────────────────────────────────────────╯
vojta@NTB:~$ kbagent update
kbagent v0.46.1 (already up to date) | keboola-mcp-server v1.63.1 (already up to date)
vojta@NTB:~$ kbagent --version
kbagent v0.46.1
vojta@NTB:~$ kbagent --json update
Updating kbagent v0.46.1 -> v0.47.1...
Auto-update failed; continuing with current version.
{
  "status": "ok",
  "data": {
    "kbagent": {
      "updated": true,
      "current_version": "0.46.1",
      "latest_version": "0.47.1",
      "message": "Updated kbagent from v0.46.1 to v0.47.1. Restart your shell to use the new version.",
      "output": ""
    },
    "mcp": {
      "updated": false,
      "current_version": "1.63.1",
      "latest_version": "1.63.1",
      "install_method": "uv_tool",
      "message": "keboola-mcp-server v1.63.1 is already up to date."
    },
    "updated": true,
    "message": "kbagent v0.46.1 -> v0.47.1 | keboola-mcp-server v1.63.1 (already up to date)"
  }
}

Session 2 — reinstall to cold cache, warm-cache success proof

vojta@NTB:~$ kbagent --version
kbagent v0.47.1
vojta@NTB:~$ kbagent doctor
╭─────────────────────────────────────────────── kbagent doctor 
[... doctor output omitted, checks passed ...]
────────────────────────────────╯
vojta@NTB:~$ uv tool install git+https://github.com/keboola/cli@v0.46.1 --reinstall
Resolved 46 packages in 453ms
      Built keboola-agent-cli @ git+https://github.com/keboola/cli@a16fe57d12d373627621b7d49506c0b22e231669
Prepared 46 packages in 2m 27s
Uninstalled 46 packages in 196ms
Installed 46 packages in 75ms
 ~ annotated-doc==0.0.4
 ~ annotated-types==0.7.0
 ~ anyio==4.13.0
 ~ attrs==26.1.0
 ~ certifi==2026.5.20
 ~ cffi==2.0.0
 ~ click==8.4.1
 ~ croniter==6.2.2
 ~ cryptography==48.0.0
 ~ h11==0.16.0
 ~ httpcore==1.0.9
 ~ httpx==0.28.1
 ~ httpx-sse==0.4.3
 ~ idna==3.16
 ~ jsonschema==4.26.0
 ~ jsonschema-specifications==2025.9.1
 ~ kai-client==0.12.0
 - keboola-agent-cli==0.47.1 (from git+https://github.com/keboola/cli@2f80aaa7c4f0827fcae48e8bd472925e99e54b2c)
 + keboola-agent-cli==0.46.1 (from git+https://github.com/keboola/cli@a16fe57d12d373627621b7d49506c0b22e231669)
 ~ markdown-it-py==4.2.0
 ~ mcp==1.27.1
 ~ mdurl==0.1.2
 ~ packaging==26.2
 ~ platformdirs==4.9.6
 ~ prompt-toolkit==3.0.52
 ~ pycparser==3.0
 ~ pydantic==2.13.4
 ~ pydantic-core==2.46.4
 ~ pydantic-settings==2.14.1
 ~ pygments==2.20.0
 ~ pyjwt==2.13.0
 ~ python-dateutil==2.9.0.post0
 ~ python-dotenv==1.2.2
 ~ python-multipart==0.0.29
 ~ pyyaml==6.0.3
 ~ referencing==0.37.0
 ~ rich==15.0.0
 ~ rpds-py==0.30.0
 ~ shellingham==1.5.4
 ~ six==1.17.0
 ~ sse-starlette==3.4.4
 ~ starlette==1.1.0
 ~ typer==0.26.2
 ~ typing-extensions==4.15.0
 ~ typing-inspection==0.4.2
 ~ uvicorn==0.48.0
 ~ wcwidth==0.7.0
warning: The package `typer==0.26.2` does not have an extra named `all`
Installed 1 executable: kbagent
vojta@NTB:~$ kbagent doctor
Updating kbagent v0.46.1 -> v0.47.1...
Updated to v0.47.1. Re-launching...
  What's new in v0.47.1:
    - Fix (`storage create-table --if-not-exists`, keboola/cli#349): the `action: "skipped"` envelope now reports the EXISTING table's actual schema instead of re-echoing the caller's request. Pre-0.47.1, `columns` / `primary_key` / `name` on a skip mirrored the args the caller passed in, so a caller probing the skipped envelope to discover the real shape of a pre-existing table got the wrong values whenever the existing table differed from the request. The `get_table_detail(target_id)` lookup that already runs to confirm the table exists is now also the source of the returned schema. The caller's requested values are preserved under two new fields, `requested_columns` and `requested_primary_key`, and a new `schema_drift: bool` flags when the existing table diverges from the request (set comparison on columns and primary key). Human-mode output prints the actual schema on a skip and emits a `Warning:` line when `schema_drift` is true. `action: "created"` envelope is unchanged. No new flag, no signature change. Tests: `tests/test_storage_write.py` (skipped returns actual schema, drift flag set on divergence, no drift on match, human-mode warning render); `tests/test_e2e.py::TestE2E_0_47_0_NewSurfaces` extended to assert the skipped envelope reports actual columns + `requested_*` mirror.
╭─────────────────────────────────────────────── kbagent doctor 
[... doctor output omitted, checks passed ...]
────────────────────────────────╯

Note: In Session 2, after manually reinstalling to v0.46.1 with --reinstall (which cached all 46 packages as a side effect), the very next kbagent doctor succeeded immediately. This confirms the failure in Session 1 was purely the cold-cache prepare time (~2m 27s) exceeding the 120s timeout — not a genuine install problem.

[vojvojtech]

Reproduced again on v0.47.1 → v0.50.0 — and this time it's worse

Same failure on kbagent doctor:

Updating kbagent v0.47.1 -> v0.50.0...
Auto-update failed; continuing with current version.

Worth highlighting what's different this time compared to the original report: the original failure was blamed on a cold cache preparing all 46 packages (2m 27s). Here, 45 of 46 packages were already cached — uv only needed to build the keboola-agent-cli package itself from the git source:

Resolved 46 packages in 35ms
Prepared 1 package in 4m 09s    ← just the git build, warm cache

That means this isn't a cold-cache edge case. The git build of the package alone takes 4m+ on WSL Ubuntu, which will exceed the 120s timeout on every significant version bump regardless of cache state. Raising the timeout to 300s (as suggested above) won't be enough — the 4m 09s observed here would blow past that too.

The background-process approach from the suggested fixes seems like the only robust solution.

Full terminal session

vojta@NTB:~$ kbagent doctor
Updating kbagent v0.47.1 -> v0.50.0...
Auto-update failed; continuing with current version.
Updating keboola-mcp-server v1.63.1 -> v1.63.4 (via uv_tool)...
Updated keboola-mcp-server to v1.63.4.
[... doctor output omitted, all 20 checks passed ...]
vojta@NTB:~$ uv tool install git+https://github.com/keboola/cli
Resolved 46 packages in 35ms
      Built keboola-agent-cli @ git+https://github.com/keboola/cli@374bdfd0332e528433c1dced84e3cb2d1b7df172
Prepared 1 package in 4m 09s
Uninstalled 1 package in 21ms
Installed 1 package in 9ms
 - keboola-agent-cli==0.47.1 (from git+https://github.com/keboola/cli@2f80aaa7c4f0827fcae48e8bd472925e99e54b2c)
 + keboola-agent-cli==0.50.0 (from git+https://github.com/keboola/cli@374bdfd0332e528433c1dced84e3cb2d1b7df172)
warning: The package `typer==0.26.2` does not have an extra named `all`
Installed 1 executable: kbagent

[padak]

Root cause found — it's the npm SPA build, not the timeout itself

Thanks for the detailed reports — the warm-cache v0.47.1 → v0.50.0 repro was the key clue. We dug in and confirmed the real bottleneck with measurements.

What's actually slow

uv tool install git+https://github.com/keboola/cli builds the package from source on every install. The wheel build runs a custom hatch hook (scripts/hatch_build.py) that compiles the bundled React SPA via npm ci + vite build. The uv package cache covers the Python wheels (your "45/46 cached") but does not cover the npm build — so the SPA is rebuilt from scratch every time. That's the Prepared 1 package in 4m 09s.

Measured locally (fast Mac):

build	time
CLI-only wheel (KBAGENT_SKIP_UI_BUILD=1)	0.49 s
full wheel from clean node_modules (npm ci = 393 pkgs + vite)	7.39 s

So ~93% of build time is the npm step. On WSL2 those 7s balloon to your 2–4 min because of slow I/O over the thousands of small files in node_modules. This is why bumping the timeout to 300s wouldn't help — your 4m09s would still blow past it.

Why kbagent update is no better

There are two hardcoded 120s timeouts, not one — auto_update.py _perform_update (the startup hook that fires on kbagent doctor) and version_service.py self_update (the explicit kbagent update). Both wrap the same git-build subprocess, which is why the guide has to fall back to raw uv tool install git+…. (This also explains Bug 2/Bug 3: after the timeout, the human and JSON outputs disagree about success.)

Proposed final solution — stop building on the user's machine

Compile the wheel once in CI, ship the prebuilt artifact:

1. Prebuilt wheel as a GitHub Release asset. The wheel is universal (keboola_agent_cli-X.Y.Z-py3-none-any.whl, 4.3 MB with UI, one file for every platform). A release.yml workflow runs uv build --wheel on Linux (npm is fast there), verifies the UI is bundled, and uploads it. (CI already builds+verifies the wheel via the build-windows job.)
2. Bootstrap install script — curl -LsSf https://github.com/ghraw/keboola/cli/main/install.sh | sh, the same pattern the guide already uses for uv and Claude Code. It resolves the latest version and installs the prebuilt wheel — no build, no gh CLI (just curl + uv). First install drops from ~4 min to seconds.
3. Switch the internal update path (build_kbagent_upgrade_command) to install from the wheel URL when the version is known, falling back to git+ when the asset is missing (older releases). Both the startup hook and kbagent update benefit.
4. Update-UX cleanup (independent): a UPDATE_TIMEOUT_SECONDS constant + KBAGENT_UPDATE_TIMEOUT env override, distinguishing a timeout from a real failure (no more false "Auto-update failed"), and fixing Bug 2/Bug 3.

Net effect on WSL: install and update both go from minutes-of-build to seconds-of-download, and the 120s timeout stops mattering.

Before we build it — does this approach get your thumbs-up? Happy to adjust (e.g. keep a pure-git+ path as well, or go PyPI instead of a Release asset).

[padak]

Fixed and shipped in v0.60.0. Thanks for the detailed reports — the warm-cache v0.47.1 → v0.50.0 repro was what pinned the root cause.

Root cause (confirmed)

It wasn't the timeout itself. uv tool install git+... rebuilds the bundled React SPA via npm ci + vite build on every install, and the uv cache never covers the npm step — so the SPA recompiles from scratch each time. On WSL2 that's the 2–4 minutes you saw (measured: ~93% of build time is the npm step — 0.49s for a CLI-only wheel vs 7.39s with the UI build on a fast Mac; WSL2 stretches that to minutes over the many small node_modules files). That's why bumping the timeout to 300s wouldn't have helped — your 4m09s would still blow past it.

What shipped (v0.60.0)

The fix removes the build from the user's machine:

1. Prebuilt wheel as a Release asset — a universal py3-none-any wheel is now built once in CI and attached to every release (this one included). Install/update becomes a seconds-long download instead of a multi-minute build.
2. install.sh bootstrap — curl … | sh, the same pattern you already use for uv and Claude Code.
3. Self-update (kbagent update + the startup hook) now installs the prebuilt wheel too, falling back to git+ only for releases without an asset.
4. Timeout is no longer hardcoded at 120s — it's a single UPDATE_TIMEOUT_SECONDS (300s) overridable via KBAGENT_UPDATE_TIMEOUT, and a slow build now prints "still building" instead of a false "Auto-update failed".

How to install / update now

New install (recommended):

curl -LsSf https://github.com/ghraw/keboola/cli/main/install.sh | sh

No source build, no gh CLI — just curl + uv. Bundles the [server] extras by default (KBAGENT_NO_SERVER=1 for a CLI-only install).

Update an existing install:

kbagent update

This now takes the fast wheel path, and so does the startup auto-update — so the false "Auto-update failed" on WSL should be gone. One caveat: the first update from an older version (≤ 0.59.0) still goes through git+ once, because those releases predate the wheel asset; from 0.60.0 onward it's the fast path. (We can backfill the 0.59.0 wheel if that's useful to you.)

On the secondary bugs you reported

• Bug 3 (kbagent --json update printing "failed" while the JSON said success) — real bug, fixed: the startup hook now resolves the actual subcommand even behind global flags like --json.
• Bug 2 (kbagent update saying "already up to date" when it wasn't) — we couldn't reproduce it, and verified it isn't a kbagent logic bug: a freshly installed v0.59.0 binary correctly reports up_to_date: false against the latest 0.60.0, and the version comparison is correct for your exact 0.46.1 vs 0.47.1 case. The only way it triggers is GitHub's /releases/latest transiently returning an older "latest" right after a publish (a propagation window that self-heals), so we left the comparison logic untouched. If you ever see it again, a kbagent version --json at that moment would show us what /releases/latest returned.

Also: #404 is already done

Unrelated to this issue, but from your same message — kbagent init can now copy a subset of the global config (shipped in 0.59.0):

kbagent init --from-global --project ALIAS   # repeatable: --project a --project b; --from-global is implied

It copies only the named project(s), no re-entering a Storage API token you already have globally.

---

When you get a chance, could you give the new install.sh a run on a fresh WSL box? Once you've confirmed install + update are fast there, feel free to close this out. Thanks again — the report quality made this a clean fix.