01 PRODUCT
Contextual Copilot drawer
Available from Formatter, Diff, Patch, and Pointer with selected context, prompts, turns, a safe summary, and explicit workspace approval.
Welcome · clarify · trace · summary → Apply
/AI CASE STUDY / SYSTEM DESIGN
Building an approval-gated AI copilot for deterministic JSON tools
A production-minded case study of constrained agents, explicit user control, provider-neutral orchestration, measurable evals, private infrastructure, and a hard operating budget.
4
fixed tools
3
turns / session
100%
final routing score
$0.000918
p95 eval session
Request → proposal trust path
Architecture chapters
02 AGENT
Provider-neutral orchestration
The Responses API sits behind AgentProvider with no runtime model escalation and strict round-trip and tool-loop limits.
≤ 2 round trips · ≤ 2 tools/turn
03 CONTRACTS
Four shared tool schemas
format_json, diff_json, apply_json_patch, and resolve_json_pointer are schema-validated for web and the private local MCP package.
One contract · two surfaces
04 SECURITY
Untrusted by default
Prompts and JSON remain data. The agent receives no shell, filesystem, network, URL, database, or dynamic tool access.
Private backend · store:false
05 IDENTITY + QUOTA
Account-free, signed, bounded
A signed HttpOnly visitor identity, rotating IP HMAC, and Firestore transactions enforce every hosted quota.
1 session/24h · 3 turns
06 EVALS
60 routing and adversarial cases
The final paid Luna run passed all 60 routing, ambiguity, injection, and abuse cases with schema-valid calls and Jason-valid Patch proposals.
100% route · 100% valid
07 COST + ROLLOUT
$20 combined target
Scale-to-zero, an enforced $8 provider hard cap, measured p95 cost, and an approved 10-session daily pilot limit surprise spend.
10/day pilot · no automatic increase
08 MCP
Local stdio parity
The private, client-neutral Jason MCP package exposes identical schemas and invokes the local jason binary with fixed arguments and stdin.
Caller model · caller budget
RELEASE STATUS
Evidence stays labeled through release
The final Luna evaluation reached 100% routing with a $0.000918 p95 session estimate. Production smoke now covers Formatter, Diff, Patch, and Pointer with explicit workspace approval. Cumulative pilot estimated usage is $0.003728. The 10-session daily pilot remains active while its seven-day review continues.