Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In `@test/onboard-selection.test.ts`:
- Around line 973-997: Replace the many inline curl shim script bodies written
with fs.writeFileSync in the tests by calling the shared helper
writeOllamaToolCallingCurl(fakeBin) (or create a small parameterized variant
writeOllamaToolCallingCurl(fakeBin, opts) that accepts overrides such as body or
unauthenticated-port) instead of copying the full script; locate the duplicated
blocks that reference fakeBin, the inline script using
OLLAMA_CHAT_COMPLETIONS_TOOL_CALL_RESPONSE and the :11435 unauthenticated 401
logic, and refactor each to call the helper, adding a tiny per-test override
wrapper only when a test needs custom behavior.

---

Nitpick comments:
In `@src/lib/inference/ollama/proxy.ts`:
- Around line 295-301: The timeout message under the Ollama proxy startup
(around the block that logs error and returns false) understates real wait time
because each loop uses waitForPort + sleep; update the message in the function
that checks the proxy (referencing OLLAMA_PROXY_PORT and PROXY_START_ATTEMPTS)
to either report the cumulative elapsed time (capture start time before the loop
and compute elapsedMs to include in the error) or multiply attempts by ~2s to
reflect worst-case (~${PROXY_START_ATTEMPTS * 2}s); adjust the process of
computing the timeout text accordingly in the same function so logs accurately
reflect how long we actually waited.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@test/ollama-proxy-startup.test.ts`:
- Around line 213-253: The test currently sets reclaimed true on any mocked
sleep and never asserts killCommands, so change the mocked reclaim logic to set
reclaimed only when the actual expected kill command targeting PID 4242 is
pushed to killCommands (e.g., when the string/command equals the expected kill
for PID 4242), and add an assertion after the run that killCommands contains
that expected command (so the test fails if the reclaim/kill action was
skipped). Ensure references to the test variables killCommands and reclaimed and
the PID 4242 are used so the check is tied directly to the emitted kill command
rather than any sleep timing.
- Around line 120-180: The tests currently vary the mocked curl responses but do
not assert that startOllamaAuthProxy() actually issued both readiness probes;
update the test to verify the curl mock was called twice and that one call
matched the unauthenticated probe (expected 401) and the other matched the
authenticated probe (expected 200). Concretely, after invoking
startOllamaAuthProxy() assert the curl mock (the mocked exec/spawn used in the
test) call count is 2 and inspect the call arguments (or calledWith entries) to
confirm one request targeted the unauthenticated endpoint/headers and one
targeted the authenticated endpoint/headers so regressions that skip the 401
check will fail.

In `@test/onboard-selection.test.ts`:
- Around line 2018-2023: The fake curl handlers in
test/onboard-selection.test.ts write to $outfile unconditionally which crashes
when -o is not passed; update each fake curl implementation (the occurrences
that reference the outfile variable at or near lines 2022, 3517, 3614 and the
ranges 3513-3518, 3610-3615) to guard the redirection by checking that outfile
is non-empty before writing the body (use the same safe pattern used in the
earlier hunk around lines 44-45), i.e. only perform the `> "$outfile"` write
when outfile is set; alternatively, if you prefer, assert in the tests that
callers always pass -o and document that precondition.

---

Nitpick comments:
In `@test/onboard-selection.test.ts`:
- Around line 25-46: In test/onboard-selection.test.ts there are four remaining
inline fake curl script implementations that duplicate the logic in the shared
helper writeOllamaToolCallingCurl(fakeBin); replace each inline script with a
call to writeOllamaToolCallingCurl(fakeBin) (passing the same fakeBin used by
the original inline fakes) to remove duplication and ensure consistent behavior,
and if any inline version must keep different behavior, add a short comment
above it explaining why it diverges from writeOllamaToolCallingCurl and what
specific differences are required for that test.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@test/ollama-proxy-startup.test.ts`:
- Around line 95-99: The runner.runCapture stub is too generic for lsof checks
and will incorrectly accept IPv6-only listeners; update the stub used in the
tests (runner.runCapture) to distinguish IPv4-scoped lsof queries from
generic/IPv6 ones by matching the IPv4-specific query string (e.g., the "-i4" or
"127.0.0.1:11435" style token) and only return the "2222" owner for that IPv4
query, add a separate case that returns an IPv6-only owner string (e.g., showing
an "[::1]:11435" or other IPv6 output) to simulate an IPv6 listener, and keep
the existing ps -p 2222 response ("/usr/bin/python3 -m http.server 11435") so
startup still succeeds when appropriate; apply the same change pattern to the
other runCapture stubs referenced in the comment.