mirror of https://github.com/merlinhu1/codex-game-studio.git synced 2026-08-25 07:54:34 +02:00

Files

T

7e8846a7fb feat: add deterministic model tier routing (#14 )

* feat: add deterministic model tier routing

Define Sol, Terra, and Luna policies with fixed fallbacks, explicit surface assignments, runtime precedence, and Luna escalation. Validate generated metadata and expose tier overrides across role, task, and orchestration commands.

* fix: infer model tiers from surface models

* fix: rebalance prompt model routing

* fix: decouple prompt effort routing

---------

Co-authored-by: MerlinH <merlinh221@gmail.com>

2026-07-10 18:49:20 +10:00

2.5 KiB

Raw Permalink Blame History

model, model_reasoning_effort, primary-agent, linked-skills, phase, risk, argument-hint, source-reference, source-hash, output-artifacts

model

model_reasoning_effort

primary-agent

linked-skills

phase

risk

argument-hint

source-reference

source-hash

output-artifacts

gpt-5.6-luna

low

producer

cgs-retrospective

cgs-sprint-status

review

high

Provide a retrospective request with milestone or sprint context, outcomes, evidence, team risks, owner or handoff needs, and required verification evidence.

.codex/workflows/retrospective.md

3333333333333333333333333333333333333333333333333333333333333333

plan

changed-files

verification-evidence

handoff

Retrospective Workflow

Purpose

Run a milestone or sprint retrospective that records outcomes, misses, risks, follow-ups, and concrete process changes.

Inputs

AGENTS.md
.codex/studio.json
.codex/workflows/retrospective.md
Timeline, sprint status, validation evidence, and incident notes

Role

Producer (producer) owns this workflow.

Taxonomy

Category: team-coordination CCGS adaptation coverage:

retrospective
learning capture CLI aliases:
retrospective

Outputs

What worked, what failed, and why.
Follow-up actions with owner, priority, and verification signal.
Process changes and risks for the next iteration.

Validation

Claims are grounded in evidence or labeled as perception.
Follow-ups are concrete and owner-routed.
Blame-free language preserves actionable learning.

Phase Gates

Confirm project state, owner role, write policy, and selected context.
Name required artifacts and acceptance criteria before implementation or review work.
Do not advance to handoff until evidence is recorded or a blocker is explicit.

Required Artifacts

Summary of the workflow result.
Files, assets, tasks, docs, release notes, or tests changed or proposed.
Verification evidence and unresolved risks.

Context Contract

Load AGENTS.md, .codex/studio.json, this workflow, the primary agent, linked skills, and only task-relevant project files.
Avoid broad context unless the user explicitly approves it.

Output Contract

Decision or change summary.
Step-by-step work performed or planned.
Evidence, blockers, warnings, and next owner.

Stop Conditions

Required project state, approval, target files, or verification path is missing.
The task crosses into another role without an explicit handoff.
The workflow would require generated prompt mirrors or hidden automation.

Handoff

Report changed files, validation evidence, residual risks, and the next owner only when ownership changes.

2.5 KiB Raw Permalink Blame History