mirror of
https://github.com/esengine/DeepSeek-Reasonix.git
synced 2026-11-07 02:36:01 -10:00
fix(agent): count triage usage in the root token budget
Problem: an explicitly bounded root omitted billed command classifications and external-content screening, allowing extra normal tool rounds. Root cause: the shared triage terminal emitted usage without charging the existing context-bound root counter. Fix: observe terminal usage through the accepted private root helper before the optional display guard. Preserve verdicts, screening notices, event conditions, missing usage, normal telemetry and cancellation authority. Verification: all 239 root packages pass in disjoint commands; full Agent race and real boot HTTP/SSE matrices pass. The compiled preceding owner fails ten events while five controls pass. Static gates, four platform cross-compiles and the literal 15-event Windows contract pass locally.
This commit is contained in:
SivanCola committed
1 parent
bf149af099
commit
00c8d9142e
6 files changed
+412
No files matched your search
@@ -421,6 +421,36 @@ jobs:
|
||||
retention-days: 7
|
||||
|
||||
|
||||
- name: test (Windows root triage budget)
|
||||
if: env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
|
||||
timeout-minutes: 5
|
||||
shell: bash
|
||||
run: |
|
||||
set -o pipefail
|
||||
go test -count=1 -timeout=3m -json -run '^Test(RootTriage|EffectRootTriage)' ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/root-triage-budget.jsonl"
|
||||
jq -se '
|
||||
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
|
||||
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
|
||||
all(.[]; .Action != "skip" and .Action != "fail") and
|
||||
(["reasonix/internal/runtime/agent", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
|
||||
(["TestRootTriageTokenBudget", "TestRootTriageTokenBudget/classifier", "TestRootTriageTokenBudget/screen",
|
||||
"TestRootTriageTokenBudget/billed_failure", "TestRootTriageTokenBudget/missing", "TestRootTriageTokenBudget/disabled",
|
||||
"TestRootTriageTokenBudget/no_sink", "TestRootTriageTokenBudget/counts_without_total",
|
||||
"TestRootTriageOutsideRunDoesNotCreateScope", "TestEffectRootTriageScreenBudget",
|
||||
"TestEffectRootTriageScreenBudget/reported", "TestEffectRootTriageScreenBudget/flagged",
|
||||
"TestEffectRootTriageScreenBudget/disabled", "TestEffectRootTriageScreenBudget/missing",
|
||||
"TestEffectRootTriageScreenBudget/screen_disabled"] - $tests | length) == 0
|
||||
' "$RUNNER_TEMP/root-triage-budget.jsonl"
|
||||
|
||||
- name: upload Windows root triage budget inventory
|
||||
if: always() && env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
|
||||
uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7
|
||||
with:
|
||||
name: root-triage-budget-windows-${{ github.run_attempt }}
|
||||
path: ${{ runner.temp }}/root-triage-budget.jsonl
|
||||
if-no-files-found: ignore
|
||||
retention-days: 7
|
||||
|
||||
- name: test (Windows root auxiliary budget)
|
||||
if: env.RUN_STEPS == 'true' && runner.os == 'Windows' && github.event_name == 'pull_request'
|
||||
timeout-minutes: 5
|
||||
|
||||
@@ -0,0 +1,89 @@
|
||||
---
|
||||
owner: @SivanCola
|
||||
backup: @esengine
|
||||
status: active
|
||||
reviewed: 2026-10-26
|
||||
---
|
||||
|
||||
# Root triage token accounting / 根任务 triage token 计账
|
||||
|
||||
Command classification and external-content screening share one triage terminal.
|
||||
Its reported prompt/completion counts now charge the active context's existing
|
||||
Agent-root token counter, including usage received before a stream failure.
|
||||
|
||||
命令分类和外部内容筛查共用一个 triage 终态。其已报告的 prompt/completion 计数
|
||||
现在计入活跃 context 已有的 Agent 根 token counter，包括流失败前收到的用量。
|
||||
|
||||
## Ownership and compatibility / 所有权与兼容
|
||||
|
||||
The terminal uses the existing private root accounting helper. It reads the
|
||||
request's scope, not mutable current-task state, and does not create a scope for
|
||||
out-of-run classifications. Nested and budget-only projected contexts retain
|
||||
the existing root lifetime and target-job cancellation authority.
|
||||
|
||||
终态复用既有私有根计账 helper，读取请求 scope，不读取可变当前任务状态；Run
|
||||
之外的分类不创建 scope。嵌套和仅投射预算的 context 沿用既有根寿命与目标 job
|
||||
取消权限。
|
||||
|
||||
Root observation is independent of the optional sink and positive-total
|
||||
emission condition. Positive prompt/completion counts still count when the
|
||||
provider omits its total; event source, model, pricing and emission stay the same.
|
||||
Missing usage stays unknown, with no invented token or HTTP counts.
|
||||
|
||||
根观察不依赖可选 usage sink 或既有正 total 上报条件。供应商省略 total 时，
|
||||
正 prompt/completion 计数仍扣计。既有事件 source、model、pricing 与上报条件
|
||||
保持不变，不推造缺失 token 或 HTTP 次数；缺失用量仍未知。
|
||||
|
||||
One classification verdict still governs the same cached read-only decision.
|
||||
The screen remains advisory: CLEAN, INJECTION and failed screening do not block,
|
||||
drop or rewrite the fetched result. An INJECTION verdict keeps its existing
|
||||
host notice. The accounting change adds no permission or safety exception.
|
||||
|
||||
分类结论继续控制同一缓存的只读判断。筛查仍为提示性质：CLEAN、INJECTION 或
|
||||
筛查失败都不阻断、丢弃或改写已获取结果。INJECTION 保留既有 host 提示。
|
||||
计账改动不新增权限或安全例外。
|
||||
|
||||
## Limit boundary / 阈值边界
|
||||
|
||||
An explicit token threshold may end the normal loop earlier, after a tool
|
||||
round with existing finalization. Zero stays unbounded. No reservation, timer,
|
||||
provider cancellation or process termination is added. Normal-round telemetry
|
||||
and monetary completeness/currency rules stay unchanged.
|
||||
|
||||
显式 token 阈值可能更早结束正常循环；检查仍在工具轮后执行，并保留既有收尾。
|
||||
本机制不做请求前预留、timer、供应商取消或进程终止。零值仍不限。辅助根计账
|
||||
不改变正常轮次局部遥测、金额完整性或币种规则。
|
||||
|
||||
The classification and screening prompts, model selection, output/time bounds,
|
||||
verdict cache and default screening setting stay the same. No API, config,
|
||||
persisted or provider-wire field is added. Serialized main prefixes and tool
|
||||
schemas are unchanged.
|
||||
|
||||
分类/筛查 prompt、模型选择、输出/时间边界、结论缓存与默认筛查开关保持原样。
|
||||
不新增 API、配置、持久化或 provider-wire 字段；主序列化前缀和工具 schema 不变。
|
||||
|
||||
## Evidence boundary / 证据边界
|
||||
|
||||
Real Agent regressions cover classifier/screen usage, billed failures, missing
|
||||
counts, disabled limits, absent sinks and omitted totals. Boot HTTP/SSE uses an
|
||||
actual file-reading tool with declared external provenance through screening.
|
||||
This is a constructed fixture, not live web-fetch or a production incident.
|
||||
|
||||
真实 Agent 回归覆盖分类/筛查用量、已计费失败、缺失计数、禁用阈值、无 sink 与
|
||||
省略 total。真实 boot HTTP/SSE 夹具使用声明外部来源的文件读取工具，执行实际
|
||||
筛查管线。这是构造的外部结果夹具，不是 live web-fetch 传输或生产事故复现。
|
||||
|
||||
The flagged control keeps both the actual file result and existing notice in
|
||||
the next provider request. Disabled and missing-usage controls retain their
|
||||
round counts. Serialized prefix equality in the fixture is a cache guard,
|
||||
not a live-provider hit-rate measurement.
|
||||
|
||||
flagged 对照保留实际文件结果及下一供应商请求中的既有提示。禁用与缺失用量
|
||||
对照保持轮次数。夹具中的序列化前缀相等是缓存守卫，不是供应商命中率实测。
|
||||
|
||||
This unit does not complete all auxiliary-provider admission, shared Coordinator
|
||||
budgets, root currency aggregation, durable accounting or the full isolation
|
||||
and recovery plan. See [root auxiliary tokens](ROOT_AUXILIARY_TOKEN_BUDGET.md).
|
||||
|
||||
本单元不完成全部辅助 provider 准入、共享 Coordinator 预算、根币种汇总、持久
|
||||
计账或完整隔离恢复方案。作者分析、独立审查与原生平台证据分别记录。
|
||||
@@ -0,0 +1,171 @@
|
||||
package boot
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"fmt"
|
||||
"io"
|
||||
"net/http"
|
||||
"net/http/httptest"
|
||||
"os"
|
||||
"path/filepath"
|
||||
"strings"
|
||||
"sync/atomic"
|
||||
"testing"
|
||||
|
||||
"reasonix/internal/contract/event"
|
||||
"reasonix/internal/contract/tool"
|
||||
"reasonix/internal/runtime/agent"
|
||||
"reasonix/internal/session/control"
|
||||
"reasonix/internal/state/sessionstore"
|
||||
)
|
||||
|
||||
type triageFixtureRead struct {
|
||||
path string
|
||||
reads atomic.Int32
|
||||
}
|
||||
|
||||
func (*triageFixtureRead) Name() string { return "triage_fixture" }
|
||||
func (*triageFixtureRead) Description() string { return "Read the screening fixture." }
|
||||
func (*triageFixtureRead) ReadOnly() bool { return true }
|
||||
func (*triageFixtureRead) Schema() json.RawMessage { return json.RawMessage(`{"type":"object"}`) }
|
||||
func (*triageFixtureRead) Provenance(json.RawMessage) tool.Provenance {
|
||||
return tool.Provenance{Kind: tool.ProvenanceWeb, Source: "screening-fixture"}
|
||||
}
|
||||
func (p *triageFixtureRead) Execute(context.Context, json.RawMessage) (string, error) {
|
||||
b, err := os.ReadFile(p.path)
|
||||
if err == nil {
|
||||
p.reads.Add(1)
|
||||
}
|
||||
return string(b), err
|
||||
}
|
||||
|
||||
func TestEffectRootTriageScreenBudget(t *testing.T) {
|
||||
for _, tc := range []struct {
|
||||
name string
|
||||
limit int
|
||||
missing bool
|
||||
flag bool
|
||||
screenOff bool
|
||||
wantReads int32
|
||||
wantTriage int32
|
||||
}{
|
||||
{name: "reported", limit: 1500, wantReads: 1, wantTriage: 1},
|
||||
{name: "flagged", limit: 1500, flag: true, wantReads: 1, wantTriage: 1},
|
||||
{name: "disabled", wantReads: 4, wantTriage: 4},
|
||||
{name: "missing", limit: 1500, missing: true, wantReads: 2, wantTriage: 2},
|
||||
{name: "screen_disabled", limit: 1500, screenOff: true, wantReads: 2},
|
||||
} {
|
||||
t.Run(tc.name, func(t *testing.T) {
|
||||
home, root := statedBootHome(t), t.TempDir()
|
||||
writeFile(t, root, "fixture.txt", "screening fixture content")
|
||||
var normal, triage, notices atomic.Int32
|
||||
var sawNotice atomic.Bool
|
||||
var mainPrefix atomic.Pointer[string]
|
||||
srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
|
||||
var req struct {
|
||||
Model string `json:"model"`
|
||||
Messages []json.RawMessage `json:"messages"`
|
||||
Tools json.RawMessage `json:"tools"`
|
||||
}
|
||||
if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
|
||||
t.Error(err)
|
||||
w.WriteHeader(http.StatusBadRequest)
|
||||
return
|
||||
}
|
||||
w.Header().Set("Content-Type", "text/event-stream")
|
||||
if req.Model == "triage" {
|
||||
triage.Add(1)
|
||||
if len(req.Messages) != 2 || len(req.Tools) > 0 && string(req.Tools) != "null" {
|
||||
t.Errorf("triage request messages/tools=%d/%s", len(req.Messages), req.Tools)
|
||||
}
|
||||
if !strings.Contains(string(req.Messages[1]), "screening fixture content") {
|
||||
t.Error("screen did not receive the actual fixture result")
|
||||
}
|
||||
answer := "CLEAN"
|
||||
if tc.flag {
|
||||
answer = "INJECTION"
|
||||
}
|
||||
_, _ = fmt.Fprintf(w, "data: {\"choices\":[{\"delta\":{\"content\":%q},\"finish_reason\":\"stop\"}]}\n\n", answer)
|
||||
} else {
|
||||
prefix := string(req.Messages[0]) + string(req.Tools)
|
||||
if previous := mainPrefix.Swap(&prefix); previous != nil && prefix != *previous {
|
||||
t.Error("main serialized prefix changed during the run")
|
||||
}
|
||||
for _, m := range req.Messages {
|
||||
if strings.Contains(string(m), "screening model judged") {
|
||||
sawNotice.Store(true)
|
||||
}
|
||||
}
|
||||
if n := normal.Add(1); n <= 4 {
|
||||
_, _ = fmt.Fprintf(w, "data: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"index\":0,\"id\":\"fixture-%d\",\"type\":\"function\",\"function\":{\"name\":\"triage_fixture\",\"arguments\":\"{}\"}}]},\"finish_reason\":\"tool_calls\"}]}\n\n", n)
|
||||
} else {
|
||||
_, _ = io.WriteString(w, "data: {\"choices\":[{\"delta\":{\"content\":\"Finished.\"},\"finish_reason\":\"stop\"}]}\n\n")
|
||||
}
|
||||
}
|
||||
if req.Model != "triage" || !tc.missing {
|
||||
_, _ = io.WriteString(w, "data: {\"choices\":[],\"usage\":{\"prompt_tokens\":1000,\"completion_tokens\":100,\"total_tokens\":1100}}\n\n")
|
||||
}
|
||||
_, _ = io.WriteString(w, "data: [DONE]\n\n")
|
||||
}))
|
||||
t.Cleanup(srv.Close)
|
||||
writeFile(t, root, "reasonix.toml", fmt.Sprintf(`default_model = "worker"
|
||||
[agent]
|
||||
system_prompt = "BASE"
|
||||
triage_model = "triage"
|
||||
screen_external_content = %t
|
||||
task_token_budget = %d
|
||||
[codegraph]
|
||||
enabled = false
|
||||
[[providers]]
|
||||
name = "worker"
|
||||
kind = "openai"
|
||||
model = "worker"
|
||||
context_window = 200000
|
||||
base_url = %q
|
||||
api_key_env = "ROOT_TRIAGE_KEY"
|
||||
[[providers]]
|
||||
name = "triage"
|
||||
kind = "openai"
|
||||
model = "triage"
|
||||
base_url = %q
|
||||
api_key_env = "ROOT_TRIAGE_KEY"
|
||||
`, !tc.screenOff, tc.limit, srv.URL, srv.URL))
|
||||
t.Setenv("ROOT_TRIAGE_KEY", "test")
|
||||
approveWorkspaceAt(t, home, root)
|
||||
sink := event.FuncSink(func(e event.Event) {
|
||||
if e.Kind == event.Notice && e.Code == event.NoticeCodeSuspectedInjection {
|
||||
notices.Add(1)
|
||||
}
|
||||
if e.Kind == event.Usage && e.UsageSource == event.UsageSourceInjectionScreen && (e.ModelRef != "triage" || e.Usage.PromptTokens+e.Usage.CompletionTokens != 1100) {
|
||||
t.Errorf("triage billing changed: model=%s usage=%+v", e.ModelRef, e.Usage)
|
||||
}
|
||||
})
|
||||
ctrl, err := Build(t.Context(), Options{Home: home, WorkspaceRoot: root, Sink: sink})
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
t.Cleanup(ctrl.Close)
|
||||
ctrl.SetSessionPath(sessionstore.NewSessionPath(ctrl.SessionDir(), ctrl.Label()))
|
||||
ctrl.SetToolApprovalMode(control.ToolApprovalYolo)
|
||||
probe := &triageFixtureRead{path: filepath.Join(root, "fixture.txt")}
|
||||
reg := tool.NewRegistry()
|
||||
reg.Add(probe)
|
||||
ctrl.Executor().SetTools(reg)
|
||||
err = ctrl.Run(t.Context(), "read screening fixture")
|
||||
if tc.limit == 0 {
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
} else if pause, ok := agent.InspectRunPause(err); !ok || pause.Key != "token" {
|
||||
t.Fatalf("pause=%+v err=%v", pause, err)
|
||||
}
|
||||
if probe.reads.Load() != tc.wantReads || triage.Load() != tc.wantTriage || normal.Load() != tc.wantReads+1 {
|
||||
t.Fatalf("read/triage/normal=%d/%d/%d want%d/%d/%d", probe.reads.Load(), triage.Load(), normal.Load(), tc.wantReads, tc.wantTriage, tc.wantReads+1)
|
||||
}
|
||||
if tc.flag != (notices.Load() == 1) || tc.flag != sawNotice.Load() {
|
||||
t.Fatalf("flag=%v notices=%d modelNotice=%v", tc.flag, notices.Load(), sawNotice.Load())
|
||||
}
|
||||
})
|
||||
}
|
||||
}
|
||||
@@ -162,6 +162,7 @@ func (a *Agent) askTriage(ctx context.Context, systemPrompt, input, source strin
|
||||
|
||||
var usage *provider.Usage
|
||||
defer func() {
|
||||
observeRootTokenUsage(ctx, usage)
|
||||
if usage != nil && usage.TotalTokens > 0 && a.svc.sink != nil {
|
||||
// Reported as what it is, on the model that answered it. Billing a
|
||||
// classification to compaction at the turn's own price hid both how
|
||||
|
||||
@@ -0,0 +1,119 @@
|
||||
package agent
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"errors"
|
||||
"testing"
|
||||
|
||||
"reasonix/internal/contract/event"
|
||||
"reasonix/internal/contract/tool"
|
||||
"reasonix/internal/state/sessionstore"
|
||||
)
|
||||
|
||||
type rootTriageTool struct {
|
||||
readProbe
|
||||
triage *Agent
|
||||
calls int
|
||||
source string
|
||||
}
|
||||
|
||||
func (p *rootTriageTool) Execute(ctx context.Context, _ json.RawMessage) (string, error) {
|
||||
p.calls++
|
||||
_, _ = p.triage.askTriage(ctx, "classify fixture", "fixture content", p.source)
|
||||
return "classification attempted", nil
|
||||
}
|
||||
|
||||
func TestRootTriageTokenBudget(t *testing.T) {
|
||||
for _, tc := range []struct {
|
||||
name string
|
||||
source string
|
||||
failure bool
|
||||
missing bool
|
||||
noEmission bool
|
||||
promptOnly bool
|
||||
disabled bool
|
||||
wantCalls int
|
||||
}{
|
||||
{name: "classifier", source: event.UsageSourceClassifier, wantCalls: 1},
|
||||
{name: "screen", source: event.UsageSourceInjectionScreen, wantCalls: 1},
|
||||
{name: "billed_failure", source: event.UsageSourceClassifier, failure: true, wantCalls: 1},
|
||||
{name: "missing", source: event.UsageSourceClassifier, missing: true, wantCalls: 2},
|
||||
{name: "disabled", source: event.UsageSourceClassifier, disabled: true, wantCalls: 4},
|
||||
{name: "no_sink", source: event.UsageSourceClassifier, noEmission: true, wantCalls: 1},
|
||||
{name: "counts_without_total", source: event.UsageSourceClassifier, promptOnly: true, wantCalls: 1},
|
||||
} {
|
||||
t.Run(tc.name, func(t *testing.T) {
|
||||
u := billed(1000, 100)
|
||||
if tc.missing {
|
||||
u = nil
|
||||
}
|
||||
if tc.promptOnly {
|
||||
u.TotalTokens = 0
|
||||
}
|
||||
var fail error
|
||||
if tc.failure {
|
||||
fail = errors.New("billed classification failed")
|
||||
}
|
||||
replies := make([]scriptedReply, 4)
|
||||
for i := range replies {
|
||||
replies[i] = scriptedReply{text: "READONLY", usage: u, fail: fail}
|
||||
}
|
||||
sink := &recordSink{}
|
||||
triage := New(&mockProvider{}, tool.NewRegistry(), sessionstore.NewSession("triage"), Options{TriageProvider: &scriptedSummarizer{replies: replies}}, sink)
|
||||
if tc.noEmission {
|
||||
triage.svc.sink = nil
|
||||
}
|
||||
probe := &rootTriageTool{triage: triage, source: tc.source}
|
||||
reg := tool.NewRegistry()
|
||||
reg.Add(probe)
|
||||
p := &spendingProvider{max: 4}
|
||||
root := New(p, reg, sessionstore.NewSession("root"), Options{TaskBudget: TaskBudget{Tokens: 1500}}, sink)
|
||||
ctx := t.Context()
|
||||
if tc.disabled {
|
||||
ctx = WithTaskBudget(ctx, TaskBudget{})
|
||||
}
|
||||
err := root.Run(ctx, "classify fixture")
|
||||
if tc.disabled {
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
} else if pause, ok := InspectRunPause(err); !ok || pause.Key != "token" {
|
||||
t.Fatalf("pause=%+v err=%v", pause, err)
|
||||
}
|
||||
if probe.calls != tc.wantCalls || p.rounds.Load() != int32(tc.wantCalls+1) {
|
||||
t.Fatalf("triage/normal=%d/%d want%d/%d", probe.calls, p.rounds.Load(), tc.wantCalls, tc.wantCalls+1)
|
||||
}
|
||||
var auxiliary int
|
||||
for _, e := range sink.kinds(event.Usage) {
|
||||
if e.Kind == event.Usage && e.UsageSource == tc.source {
|
||||
auxiliary += e.Usage.PromptTokens + e.Usage.CompletionTokens
|
||||
}
|
||||
}
|
||||
want := int(p.rounds.Load()) * 1100
|
||||
charged := 0
|
||||
if !tc.missing {
|
||||
charged = probe.calls * 1100
|
||||
}
|
||||
if root.task.budget.promptTokens+root.task.budget.outputTokens != want || root.task.budget.rootTokens.used.Load() != uint64(want+charged) {
|
||||
t.Fatalf("local/root=%d/%d want%d/%d", root.task.budget.promptTokens+root.task.budget.outputTokens, root.task.budget.rootTokens.used.Load(), want, want+charged)
|
||||
}
|
||||
if !tc.missing && !tc.noEmission && !tc.promptOnly && auxiliary != probe.calls*1100 {
|
||||
t.Fatal("triage fixture failed to bill exact usage")
|
||||
}
|
||||
if (tc.missing || tc.noEmission || tc.promptOnly) && auxiliary != 0 {
|
||||
t.Fatal("existing usage emission conditions changed")
|
||||
}
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
func TestRootTriageOutsideRunDoesNotCreateScope(t *testing.T) {
|
||||
a := New(&mockProvider{}, tool.NewRegistry(), sessionstore.NewSession("triage"), Options{TriageProvider: &scriptedSummarizer{replies: []scriptedReply{{text: "READONLY", usage: billed(1000, 100)}}}}, event.Discard)
|
||||
if text, ok := a.askTriage(t.Context(), "classify", "fixture", event.UsageSourceClassifier); !ok || text != "READONLY" {
|
||||
t.Fatalf("reply=%q ok=%v", text, ok)
|
||||
}
|
||||
if a.task.budget.rootTokens != nil {
|
||||
t.Fatal("out-of-run classification created an unrelated root")
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,2 @@
|
||||
- 修复：活跃根任务的 token 软预算现在计入命令分类和外部内容筛查用量；分类结论、筛查提示、缺失用量与零阈值行为保留。关联 #14025。
|
||||
- Fixed: active root token soft budgets now include command-classification and external-screening usage, preserving verdicts, advisory notices, missing usage and zero limits. Refs #14025.
|
||||
Reference in new issue
Block a user