# Using multiple models to extend long context tasks, and increase build viability > Source: > Published: 2026-05-09 08:03:52+00:00 | | | | | | | | | | | | | | | | | | | User has Idea" vertex="1"> | | | | | | User + Frontier Planner

Refine idea, develop spec from idea.

Spec is append-only — core truth.

Resolved view served

to Manager + Exit Mgr

Project context, prior

decisions, anti-patterns

Prompted for action or approval of spec change.

Manages all model handoffs.

Can halt any window.

Decompose spec into phased task list.

Audit gating at minor + major milestones.

Final call on retask vs halt — judgment, not rule.

Local fix → next task list. Structural → user prompt.

Executes task list, runs automated tests.

Stops at designated checkpoints.

Reports for verification by Auditor.

Audits worker output vs codebase + spec.

Also acts as Manager context filter —

raw worker output never reaches Manager.

• verdict: pass / fail / partial

• one-line diagnosis

• rec action: continue / retask / escalate

• pointer: file / function / test

Manager pulls more detail only on demand.

(was: Debugger)

User test program with debugger. Watches logs, applies surgical edits within bounded scope.

If structural fix needed → kicks back to Manager with diagnosis (re-enters audited build).

Needs whole-project context — sized for largest context window.

If pass → close session.

Opus 4.6 / 4.7

GPT 5.4 / 5.5

Runner-ups

Gemini Pro 3.1

GLM 5.1

Kimi 2.6

Qwen 3.6 MAX

Exit Mgr: bias toward largest context window — needs whole-project view.

DS Flash V4

Qwen 3.6 35B MoE (local)

Runner-ups

GLM 5.1

Kimi 2.6

Sonnet 4.6

Gemini Flash 3.0

Gemini Pro 3.1

GPT 5.3

DS Flash V4

Qwen 3.6 35B MoE (local)

Runner-ups

GLM 5.1

Kimi 2.6

Sonnet 4.6

Gemini Flash 3.0

Gemini Pro 3.1

GPT 5.3

Same tier as Worker — context filter for Manager only works if Auditor is capable.

Opencode, Claude Code, VsCode, Ect

— solid: data / control flow

— dashed: context read (spec / knowledge)

— red dashed: user halt / interrupt