Skip to content

Commit b8cea1d

Browse files
committed
docs: correct Kimi repro version
1 parent 3a8c906 commit b8cea1d

1 file changed

Lines changed: 11 additions & 4 deletions

File tree

‎devlog/_plan/260816_undeclared_tool_invocation_contract/000_research.md‎

Lines changed: 11 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -81,9 +81,9 @@ A follow-up comment correctly pointed out that this contradicts the repository's
8181

8282
Reference: https://github.com/lidge-jun/opencodex/issues/1795#issuecomment-5305003614
8383

84-
## Kimi K3 v2.22.0 reproduction
84+
## Kimi K3 v2.14.2 reproduction and v2.22.0 re-check
8585

86-
A second reproduction on current stable OpenCodex v2.22.0 confirmed that the problem is not SenseNova-specific and also reaches the Codex-native Code Mode helper boundary.
86+
A second reproduction was captured with Teamwicked Kimi K3 on OpenCodex v2.14.2. It showed the same semantic mismatch at the Codex-native Code Mode helper boundary. That run was initially misidentified as v2.22.0 because it came from a different worktree that was still running v2.14.2.
8787

8888
Observed request characteristics:
8989

@@ -104,9 +104,16 @@ routed provider emitted undeclared client tool "apply_patch"; only request-decla
104104

105105
The request/conversation identifiers were intentionally omitted from the public issue comment.
106106

107-
Reference: https://github.com/lidge-jun/opencodex/issues/1795#issuecomment-5305779307
107+
References:
108+
109+
- original repro: https://github.com/lidge-jun/opencodex/issues/1795#issuecomment-5305779307
110+
- version correction: https://github.com/lidge-jun/opencodex/issues/1795#issuecomment-5306871615
111+
112+
Re-running the same Kimi K3 / `openai-chat` / `xhigh` Code Mode scenario on actual v2.22.0 has not reproduced the undeclared top-level `apply_patch` failure so far. The live v2.22.0 Kimi case therefore remains unconfirmed.
113+
114+
That does not remove the separate static metadata defect addressed by this PR: v2.22.0/current dev still applies `apply_patch`-specific `input` guidance to non-`apply_patch` Responses custom tools, and the non-OpenAI catalog nudge still does not explicitly distinguish the top-level invocation surface from nested helper APIs.
108115

109-
This reproduction matters because Code Mode deliberately exposes patching through the declared `exec` surface with a nested helper:
116+
The v2.14.2 reproduction still matters because Code Mode deliberately exposes patching through the declared `exec` surface with a nested helper:
110117

111118
```text
112119
exec

0 commit comments

Comments
 (0)