Allow llama.cpp --reasoning flag in runtime flag allowlist

[pujitha24]

Motivation:
llama.cpp's llama-server exposes a --reasoning [on|off|auto] flag that
toggles reasoning/thinking mode in the chat template, but Docker Model
Runner rejects it because it is missing from LlamaCppAllowedFlags. The
map already allows the related --reasoning-format and --reasoning-budget
flags. Without --reasoning, the only way to pin a hybrid thinking model
non-thinking is --chat-template-kwargs '{"enable_thinking": false}',
which current llama.cpp builds flag as deprecated in favor of
--reasoning on / --reasoning off.

Approach:
Add "--reasoning" to the LlamaCppAllowedFlags map in
pkg/inference/runtime_flags_allowlist.go, following the same enum-flag
pattern already used for --pooling and --reasoning-format (no argument
value is validated beyond the existing path-safety and flag-injection
checks, which apply uniformly to all allowed flags). Added a
"reasoning" category to the table-driven allowlist category test and a
new scenario case in the runtime flag validation test.

Validation:

• go build ./pkg/inference/...
• go vet ./pkg/inference/...
• gofmt -l on the changed files (no output)
• golangci-lint run ./pkg/inference/... (0 issues)
• go test ./pkg/inference/... and go test -race ./pkg/inference/... (all pass)
• Confirmed the new test is a genuine regression test: with only the
  allowlist map entry reverted, the new "llama.cpp: reasoning flag
  allowed" case fails with runtime flag "--reasoning" is not allowed for backend "llama.cpp"; restoring the entry makes it pass.

Report: #1064
Signed-off-by: Pujitha Paladugu 10557236+pujitha24@users.noreply.github.com
Assisted-by: claude-sonnet-5 (via Claude Code)

Fixes #1064

[sourcery-ai]

Hey - I've reviewed your changes and they look great!

---

Sourcery is free for open source - if you like our reviews please consider sharing them ✨

• X
• Mastodon
• LinkedIn
• Facebook