diff --git a/.github/workflows/codex-review-on-assign.yml b/.github/workflows/codex-review-on-assign.yml
index 840817b44..01b1fe43b 100644
--- a/.github/workflows/codex-review-on-assign.yml
+++ b/.github/workflows/codex-review-on-assign.yml
@@ -199,14 +199,28 @@ jobs:
               'you. Do not run, build, or install anything from the PR; only read',
               'the diff.',
               '',
-              'Flag every bug, risk, and rule violation that is present in the',
-              'current diff. Keep each finding short and specific, but never cap',
-              'how many findings you report or leave a finding for a later pass.',
-              'The severity label ranks findings; the list does not need to.',
-              'Also judge whether the PR description contains sufficient testing /',
-              'end-to-end test notes; if it lacks a test plan or an e2e record,',
-              'flag it as a blocking issue. Reply in the same language the PR',
-              'author used in the PR description.',
+              'Report only findings you are confident are real in the current',
+              'diff, and keep each one short and specific. Do not cap how many',
+              'you report: surface every real bug in one pass. Gate severity by',
+              'evidence instead — every [P1 blocking] must cite either a',
+              'concrete failure scenario (the input or state that triggers it',
+              'and the resulting misbehavior) or the exact Code Review Rules',
+              'item it breaks. A finding that can cite neither is at most',
+              '[P2 non-blocking]; a speculative or theoretical risk that needs',
+              'an unlikely precondition to trigger is at most [P3 nit]. Use',
+              'three severities, spelled out inline in the author\'s language:',
+              '[P1 blocking] (must fix before merge), [P2 non-blocking]',
+              '(suggestion), [P3 nit] (minor/optional polish, theoretical edge',
+              'case, or coverage deferred to CI).',
+              'Also judge whether the PR description documents sufficient',
+              'testing. A PR that changes runtime behavior (beyond a docs-only',
+              'or tests-only diff) needs an end-to-end / real-CLI record; flag',
+              'its absence as [P1 blocking]. A docs-only or tests-only PR needs',
+              'no e2e record — do not flag it for that. Do NOT require a full',
+              'provider x agent matrix: one representative real-CLI run suffices,',
+              'and missing extra-provider/agent coverage the author flagged as',
+              'untestable here or deferred to CI is at most [P3 nit]. Reply in',
+              'the same language the PR author used in the PR description.',
               '',
               'Earlier passes of this review raised the findings below. A finding',
               'already fixed in the current diff is resolved: say so, but do not',
diff --git a/AGENTS.md b/AGENTS.md
index d159c3713..08e569325 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -22,7 +22,9 @@ TypeScript, Node 20+, tsup (ESM), Vitest. Commands: `npm run build`, `npx tsc --
 
 ## PR 前测试
 
-`npm run build` 后用真实 CLI 对本次改动做完整端到端验证（不能只跑 type check / unit test）。Test Plan 每一项必须实际通过，测试报告贴进 PR。
+改动运行时行为的 PR（docs-only / tests-only 之外），`npm run build` 后必须用真实 CLI 对本次改动做端到端验证，不能只跑 type check / unit test；**一次代表性的 real-CLI 运行即可**，把实际通过的验证记录贴进 PR。docs-only / tests-only 的改动无需 e2e 记录。
+
+不要求覆盖下面的完整 provider × agent 矩阵——额外 provider / agent 的覆盖交给 CI，或在本地环境不具备时说明即可：
 
 - Agent：Claude、Codex、CodeBuddy、OpenCode
 - Provider：`git`、`gitlab`、`github`
@@ -34,11 +36,27 @@ shared state, list every reader and every writer.
 
 ## Code Review Rules
 
-- The PR description must document sufficient testing, including an
-  end-to-end / real-CLI verification record — not only unit tests or type
-  checks. Flag a PR whose description lacks a test plan or an e2e record.
-  A test record naming an older commit than the head is a note, not a blocking
-  finding; the body may have been edited after the review pass started.
+- The PR description must document sufficient testing. For a PR that changes
+  runtime behavior (anything beyond a docs-only or tests-only diff), that
+  includes an end-to-end / real-CLI verification record, not only unit tests
+  or type checks — flag such a PR that lacks one as `[P1 blocking]`. A
+  docs-only or tests-only PR needs no e2e record; do not flag it for that.
+  Do NOT require a full provider × agent matrix — one representative real-CLI
+  run is enough. Missing coverage of extra providers (`gitlab`/`github`) or
+  agents (`Codex`/`CodeBuddy`/`OpenCode`) is at most `[P3 nit]` when the author
+  has flagged it as untestable in this environment or deferred to CI, never
+  `[P1 blocking]`. A test record naming an older commit than the head is a
+  note, not a blocking finding; the body may have been edited after the review
+  pass started.
+- Do not over-review. Report only findings you are confident are real in the
+  current diff. Do not cap how many findings you report: every real bug should
+  surface in one pass, not be deferred to a later one. Instead, gate severity
+  by evidence — every `[P1 blocking]` must cite either a concrete failure
+  scenario (the input or state that triggers it and the resulting misbehavior)
+  or the exact `## Code Review Rules` item it breaks. A finding that can cite
+  neither is at most `[P2 non-blocking]`; a speculative or theoretical risk
+  that needs an unlikely precondition to trigger is at most `[P3 nit]`. Never
+  restate a finding already resolved in the current diff.
 - Reject over-engineering. Favor the smallest code that solves the problem;
   flag speculative abstractions, unused flexibility or config, error handling
   for cases that cannot occur, and new CLI commands added where an existing
@@ -46,8 +64,9 @@ shared state, list every reader and every writer.
 - Changes must be surgical. Every changed line should trace directly to the
   PR's stated goal; flag unrelated drive-by edits.
 - Label every finding with an explicit severity a first-time reader can
-  understand — never a bare `P1`/`P2` code. Keep the `P` marker but spell out
-  what it means inline on each finding, using the PR author's language:
-  `[P1 blocking]` for issues that must be fixed before merge, and
-  `[P2 non-blocking]` for suggestions that do not block merge. (In Chinese,
-  `[P1 阻断]` / `[P2 非阻断]`.)
+  understand — never a bare `P1`/`P2`/`P3` code. Keep the `P` marker but spell
+  out what it means inline on each finding, using the PR author's language:
+  `[P1 blocking]` for issues that must be fixed before merge,
+  `[P2 non-blocking]` for suggestions that do not block merge, and `[P3 nit]`
+  for minor or optional polish, theoretical edge cases, and coverage deferred
+  to CI. (In Chinese, `[P1 阻断]` / `[P2 非阻断]` / `[P3 可选]`.)
diff --git a/CLAUDE.md b/CLAUDE.md
index e93ec1fd0..8ad520d3d 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -22,7 +22,9 @@ TypeScript, Node 20+, tsup (ESM), Vitest. Commands: `npm run build`, `npx tsc --
 
 ## PR 前测试
 
-`npm run build` 后用真实 CLI 对本次改动做完整端到端验证（不能只跑 type check / unit test）。Test Plan 每一项必须实际通过，测试报告贴进 PR。
+改动运行时行为的 PR（docs-only / tests-only 之外），`npm run build` 后必须用真实 CLI 对本次改动做端到端验证，不能只跑 type check / unit test；**一次代表性的 real-CLI 运行即可**，把实际通过的验证记录贴进 PR。docs-only / tests-only 的改动无需 e2e 记录。
+
+不要求覆盖下面的完整 provider × agent 矩阵——额外 provider / agent 的覆盖交给 CI，或在本地环境不具备时说明即可：
 
 - Agent：Claude、Codex、CodeBuddy、OpenCode
 - Provider：`git`、`gitlab`、`github`