This is an automated email from the ASF dual-hosted git repository.
terrymanu pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/shardingsphere.git
The following commit(s) were added to refs/heads/master by this push:
new e6fc04ca681 refine Jev and Codex routing guidance (#39924)
e6fc04ca681 is described below
commit e6fc04ca6810f8e07dd613b683c2bff4d702eb93
Author: Liang Zhang <[email protected]>
AuthorDate: Tue Sep 22 20:38:43 2026 +0800
refine Jev and Codex routing guidance (#39924)
---
AGENTS.md | 40 ++++++++++++++++++++++++++++++++++++++++
1 file changed, 40 insertions(+)
diff --git a/AGENTS.md b/AGENTS.md
index 8abf6e0c2b2..c6d50203c9b 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -90,3 +90,43 @@ Every authorized code-writing task follows
`.codex/context/change-completion.md`
## Functional and Performance Non-Regression Gate
Preserve supported behavior outside the exact authorized change. Load the full
non-regression rules when a change affects a public or shared contract, crosses
owners, or has credible functional or performance risk. Do not benchmark when
inspection rules out credible cost growth.
+
+## Jev/Codex Routing
+
+### Roles and Boundaries
+
+- The `typesafe-ai` skill guides Jev tool selection, request construction, and
result interpretation; it does not execute requests.
+- Jev MCP executes the actual Jev judgment and returns the result.
+- Codex performs deep analysis, code changes, execution, final verification,
and Jev result review.
+- When Jev MCP is unavailable, Codex is the fallback; the skill is not an MCP
substitute.
+
+### Routing
+
+- Prefer Jev for bounded judgments such as classification, routing, ranking,
prioritization, relevance, triage, and quick review.
+- Keep code changes, debugging, long-context analysis, multi-step reasoning,
exact calculations, command execution, test execution, and high-risk or
irreversible operations in Codex.
+- For a Jev-suitable task, consult the skill when available before checking
MCP.
+- Call Jev MCP only when it is installed, its tools are available, and the API
key is confirmed to have available quota or balance.
+- If the skill is unavailable but MCP is installed, available, and clear, MCP
may be called without installing the skill.
+- If MCP is not installed or available, the key is invalid, quota or balance
is insufficient, or quota is unconfirmed, do not call Jev; let Codex handle the
task.
+- If a Jev MCP call reports insufficient quota, insufficient balance,
authentication failure, or rate limiting, do not make further Jev calls during
the current task.
+
+### Results and Safety
+
+- Treat Jev as advisory. It must not edit files, run commands, authorize
destructive actions, or replace user confirmation.
+- Use the `typesafe-ai` guidance and the task's impact and risk to judge
whether a Jev result is reliable.
+- If the probability, uncertainty, result meaning, or applicability is
unclear, Codex must review the result.
+- Before acting, Codex must verify important conclusions with source code,
tests, or command output.
+- Do not send API keys, passwords, connection strings, or unrelated private
source code to Jev.
+- Deterministic rules and simple arithmetic do not need Jev.
+
+### Jev Usage Report
+
+At the end of each task, report the following:
+
+- Whether the skill was used and what guidance it provided.
+- Whether Jev MCP was called; if not, state the specific reason.
+- The number of Jev MCP calls, the judgments handled, and the role each call
played.
+- Whether Jev changed task routing, prioritization, conclusions, or the
execution plan.
+- Whether Jev actually reduced token usage or improved processing speed or
efficiency.
+- Report specific token savings or time changes only when measured data is
available; otherwise write "cannot be measured accurately" and do not guess.
+- Report problems caused by Jev; write "none found" when no problems were
found.
