Agent Sidecar lowers Codex and Claude Code costs by turning every run into measurable, validated task economics.
Final cost delta after validation
-$1.43 per successful task
Codex coding session
Agent cost accrues while the task is still unresolved.
src/auth/session.ts
01import { cookies } from "next/headers"
02import { redirect } from "next/navigation"
03import { refreshToken, readSession } from "./tokens"
04import { saveSession, clearSession } from "./store"
05import { validateSessionCookie } from "./validation"
06
07type User = {
08 id: string
09 email: string
10 session: string
11}
12
13export async function loadCurrentUser() {
14 const session = cookies().get("sid")?.value
15 if (!session) return null
16 return readSession(session)
17}
18
19export async function refreshSession(user: User) {
20 const token = await refreshToken(user.session)
21 saveSession(user.id, token)
22 return redirectAfterLogin(user)
23}
24
25function redirectAfterLogin(user: User) {
26 return redirect(`/account/${user.id}`)
27}
28
29export async function signOut() {
30 await clearSession()
31}
Coding-specific waste
Bloated prompts. Repeated context. Hidden spend.
Codex loop
$ read src/auth/session.ts
$ read src/auth/session.ts again
$ attach full chat history
$ route simple edit to expensive model
Waste detected
Prompt Refiner
Prompt Refiner improves the task path before more cost accrues.
Agent Sidecar
Before
Full chat history pasted into next call
Repeated session.ts and tokens.ts reads
Verbose reasoning from earlier attempts
Generic request: figure it out and fix anything
high-cost prompt context
Refined prompt
Task: Fix auth refresh so users stay signed in after token renewal
Constraints: Keep the change narrow and follow AGENTS.md
Acceptance: Token renewal persists before redirect
PR boundary: No unrelated auth refactors
Tests: Run targeted auth refresh and login flow checks
focused task contract
Steering applied to coding
Refine prompts. Reuse memory. Route smarter.
Agent Sidecar runtime
Prompt
Send refined task contract into Codex
Repo memory
Reuse auth-refresh plan instead of rereading files
Model route
Use cheaper model for narrow edit step
Validation
Reserve stronger model for failure analysis
18export async function refreshSession(user: User) {
19 const token = await refreshToken(user.session)
20- saveSession(user.id, token)
20+ await saveSession(user.id, token)
21+ await validateSessionCookie(user.id)
22 return redirectAfterLogin(user)
23}
Dashboard and coding outcome
Lower spend. Same engineering output.
Live runs
Cost/task
-50%
Quality
validated
Scope
unchanged
Validated cost reduction
$4,820
Prompt Refiner removed repeated context, memory avoided duplicate file reads, and routing kept simple edits off the strongest model.
This successful run
$1.43
$2.84 to $1.41
prompt contract
original and optimized prompt hashes recorded
route steering
simple session fix kept off the strongest model
quality gate
savings counted after auth tests pass
Local-first telemetry boundary
Accessible / readable
Safe operational telemetry for engineering leaders, platform teams, and finance reviewers.
Run timeline
Tool calls, timestamps, exit states, and validation status
Cost evidence
Token estimates, model usage, validated traces, and task-cost calculations
Policy outcomes
Allowed actions, blocked waste, and reason codes for audit
Kept private
Sensitive code, credentials, and raw workspace context remain local by default.
Source contents
Repository files stay on the developer machine unless explicitly shared
Secrets
Environment variables, credentials, and protected paths are never exposed
Raw prompts
Sensitive task text can remain local while metadata rolls up for teams
The dashboard is the separate control plane and replay UI. Developers keep coding in Codex while Agent Sidecar optimizes prompt context, routing, and telemetry in the background, then measures whether those optimizations improve cost per successful task.
Live runs
guarded runs
18
decisions
43
saved lessons
9
validated cost reduction
$4,820
Estimated from successful traces, observed cost movement, and current model pricing.
prompt contract
scope and validation extracted before Codex runs
model route
narrow edits stay on the cheaper viable path
validated delta
cost movement counted after the task succeeds