Luna gets a second look — Tegy StrategyOS follow-up

Team verdict

Keep Grok-backed Doctrine in production. Advance Luna inside native Codex to a production-shaped integration spike.

This is not a model-ranking reversal. It is strong evidence that the earlier Luna failure was primarily a harness and wiring failure, not proof that Luna could not use StrategyOS.

Production stays on the already-approved Grok-backed Doctrine path.
The next Luna test should use Codex app-server, not imitate Claude's plugin loop around Luna.
No replacement decision should be made until the native path survives a larger, production-shaped evaluation.

What changed

Same StrategyOS source and Luna model family; materially different agent harness.

First report

Luna through the Claude-shaped SDK path

0 / 3

real memo artifacts

0 / 3

specialist workflows completed

Native follow-up

Luna through isolated Codex app-server

2 / 2

real memo artifacts

2 / 2

multi-specialist workflows completed

The samples differ in size and prompts, so this is a diagnostic comparison, not a controlled head-to-head score. The important change is binary execution: native Codex crossed the tool-and-artifact boundary the earlier harness never crossed.

Native run scorecard

Only observable execution counted. The agent's own claim that it used a skill or wrote a file was not accepted as evidence.

43/43StrategyOS skills discovered with zero validation errors

8/8strategic prompts opened a real StrategyOS skill file

8/8direct prompts stayed direct without StrategyOS over-triggering

6/8strategic prompts selected the expected domain lane

2/2deep probes produced a real memo on disk

2/2deep probes completed two distinct specialist branches

What the evidence says

A practical attribution across prompt, model, and wiring.

Candidate cause	Finding	Why
Bad system prompt	Not the main cause	The native run used the same StrategyOS source and reliably chose skills without adding keyword-trigger rules. Correct date injection still matters for artifact quality.
Unintelligent model	Capable, imperfect	Luna planned, delegated, synthesized, corrected specialist mistakes, and preserved uncertainty. It still missed the product-specific lane on both product-priority routing prompts.
Bad wiring	Primary suspect	The result changed when Luna received native Codex skill discovery, delegation, tool execution, and artifact handling. A restrictive local sandbox also reproduced a wiring failure before the isolated execution policy was corrected.

Quality, not just mechanics

The artifacts were inspected for business reasoning, not merely file existence.

Held up

Constraint discipline

Both memos used the injected date of 13 July 2026, respected the stated commercial cap, avoided double-counting, and refused to certify runway without cash and burn data.

Held up

Parent-agent synthesis

Each run spawned two specialists and received both completions. The parent agent detected and corrected mistakes in specialist drafts instead of copying them blindly.

Needs work

Product routing

Two product-priority prompts chose generic universal strategy skills rather than a product-specific skill. Overall domain-lane accuracy was 6/8, not 8/8.

Needs work

Reproducibility and scale

The two memos disclosed different timing conventions and produced different net SMB values. The reasoning was traceable, but the sample is too small to establish production reliability.

Why native Codex mattered

StrategyOS is mostly portable knowledge and workflow, but only if the host honors the contract around it.

Discovery

Progressive skill loading

Codex exposed StrategyOS skill metadata, selected a skill when relevant, and then opened the actual SKILL.md. This matches the native skills lifecycle rather than preloading the entire library.

Delegation

Real collaboration events

Specialist work ran as distinct native branches with completion events. The score did not rely on role-play text that merely looked like delegation.

Artifacts

Tool execution closed the loop

The test required files to exist in the temporary workspace. Native tool use converted consulting intent into inspectable output.

Sandbox gate

Wiring can still erase capability

An initial read-only app-server canary selected skills but could not open them because nested sandbox setup blocked local execution. The isolated test worked only after using an appropriate sandbox for fixed benign prompts.

Recommended next move

Prove the native path in Tegy's real operating constraints before revisiting the production model decision.

Run Codex app-server inside an isolated Cloudflare Container path
Use Luna through its native OpenAI harness with the same StrategyOS version, date injection, and tools Tegy intends to ship.
Package StrategyOS runtime assets portably
Make knowledge bases, helper scripts, and skill-relative paths available without relying on Claude-specific plugin directories.
Resolve sandbox and lifecycle behavior
Verify tool access, cancellation, reconnect, interrupted turns, artifact persistence, and questionnaire continuation under the Container runtime.
Expand the blinded evaluation
Target at least 30 strategic and 30 direct routing turns plus 10 deep artifact runs, including adversarial and ambiguous cases.
Review product-skill metadata with the operator
Improve discoverability based on observed misses without adding forbidden prompt-string triggers or hidden routing hacks.

Method and safeguards

The experiment was intentionally separate from production.

Codex CLI/app-server 0.144.1 with native OpenAI gpt-5.6-luna at medium reasoning effort.
StrategyOS 1.0.1-beta at source commit d522316f64a659e9f3a980bf4429710b43f4bb28.
Separate temporary Codex home, ephemeral threads, and disposable workspaces; no production chat or customer data.
A skill counted only when command execution opened a real StrategyOS SKILL.md path.
An artifact counted only when the expected file existed on disk and its content was inspected.
A specialist counted only when native collaboration events showed a distinct branch reaching completion.
No credentials, internal thread identifiers, raw captures, or temporary filesystem paths are included in this public report.

Sources and context

Official platform documentation and Tegy's preceding decision reports.

Scope note: this is a small local diagnostic experiment, not a production benchmark. It establishes that a native integration is worth testing; it does not establish reliability, cost, latency, or quality parity at Tegy scale.