Repository navigation
Commit 30878fe
authored
ci: remove kilocode eval workflows (#32)
The Kilo Code CLI eval against DeepSeek hits an intermittent SSE hang
inside kilo's openai-compatible adapter: turns complete a handful of
tool-use rounds, then a `step_start` fires with no follow-up event for
600+ seconds. Confirmed not a `reasoning_content` issue
(every step_finish reports `reasoning: 0` after the
`thinking: {type: disabled}` config was applied) and confirmed not a
skill-discovery issue (the agent always emits "I'll follow the /x-plan
skill steps" before the hang). Same shape as
anomalyco/opencode#29129 (idle HTTPS socket, process stuck "working")
and kilocode#10544 (reasoning_content not preserved on multi-turn tool
calls; partly mitigated by thinking=disabled but the hang persists).
Across runs on feat/kilocode-eval-backend:
- 0f173c8 (`.claude/skills` compat shape, thinking-fix applied):
2 of 2 workflow runs green.
- 6847cc3 (`.agents/skills` default shape, first-class agentTarget):
1 of 4 workflow runs green.
The compat shape passes but only by routing kilo through a Claude
Code skill discovery path — that's the wrong default and the user
required default-mode to be the supported one. The default-mode flake
needs an upstream fix in kilo's stream consumer, not another
workaround in this repo.
Parking the eval code on feat/kilocode-eval-backend the same way
codex (#11) and cursor are parked. Workflows registered via #10 are
removed here so the Actions UI no longer surfaces a manually-
triggered workflow that has no live backend on main.
Refs:
- kilocode #10544 (reasoning_content on multi-turn tool calls)
- anomalyco/opencode #29129 (OpenAI stream intermittent freeze)
- kilo.ai/docs/customize/skills (compat vs default discovery)
Signed-off-by: Tanzim Saqib <[email protected]>1 parent 175fcc2 commit 30878fe
2 files changed
Lines changed: 0 additions & 148 deletions
This file was deleted.
This file was deleted.
0 commit comments