autoresearch artifact · architecture diagram

Fable-ish를 Codex Harness 위에 얹는 법

`/Users/jung-wankim/Downloads/fable-ish`는 작업 단위별 검증 루프를 강하게 만드는 skill이다. 현재 user-scope harness는 이미 `$init-project`, hard process contract, `$team`, QA cycle, loopy-era audit까지 가진 control plane이다. 이 문서는 둘을 교체 관계가 아니라 layer 관계로 비교한다.

1
portable micro-loop
14
hard contract IDs
21
hard process edges
100
upgrade audit score
01 · decision

Executive Verdict

Fable-ish의 핵심은 personality가 아니라 검증 루프다. 현재 harness의 hard gate를 낮추지 말고, `$team`의 각 작업 단위에 Fable-ish식 ledger를 붙이는 것이 맞다.

Fable-ish work unit framing dynamic exit criteria missing harness synthesis per-task discipline Integration Decision additive layer, not replacement attach ledger to each team task record proof rung in QA evidence escalate gaps to scorecard micro-loop adapter Codex Harness normalized goal intake hard process contract team preflight and QA gates runtime governance Correct upgrade: machine-readable per-work-unit evidence inside existing hard gates
02 · system architecture

Architecture Comparison

왼쪽은 portable skill package, 오른쪽은 user-scope와 project-scope에 설치된 control plane이다. 가운데 adapter가 `$team`의 task artifact를 확장한다.

Fable-ish package portable per-task loop harness SKILL.md prime directive, loop stop at observed proof references/*.md loop harness dynamic exit criteria verification ladder review and subagents agents/openai.yaml display metadata default prompt surface Adapter micro-loop fields work_unit_ledger dynamic_exit_criteria missing_harness_rule review_lens_queue proof_boundary_rung recommended, not replacement Codex user-scope harness project governance and hard runtime gates AGENTS.md operating contract $init-project normalized goal, bridge hard process contract hard contracts and edges $team runtime preflight, acceptance, QA loopy-era evidence eval, report, audit Binding: Fable-ish feeds `$team` Phase 0/3/4 as work-unit ledger and proof escalation contract.
03 · capability matrix

Capability Matrix

이 구간은 다이어그램보다 표가 더 명확하다. capability, Fable-ish input, Codex anchor, integration decision을 한 행에서 비교한다.

Capability Fable-ish input Codex anchor Integration decision
Goal intake Work unit + observable exit criterion. Normalized goal object in hard JSON. Codex wins: ledger goes under the goal object.
Completion rule Proof or explicit blocker only. Hard validator + QA + audit gates. Aligned: final report names proof boundary.
Dynamic exits Delta-zero, runtime, deployment boundaries. Strong gates, weaker per-task labels. Add `proof_boundary_rung` in plan + QA evidence.
Missing harness Create the smallest proof tool. Project validators already exist. Localize `missing_harness_decision` per task.
Review lenses Correctness, removed behavior, contracts. Acceptance QA + subagent routing. Add queue only for high-risk work units.
Subagents Output is hypothesis. Bounded ownership, no stale model hardcode. Codex wins: copy hypothesis-only rule into handoff.
Artifact QA Registry, format, render, consumer proof. Upgrade audit closes prose-only gap. Aligned: parser + render smoke required.
Loop governance Raise bar when proof is weak. Loopy-era eval/report/scorecard. Bridge levels: proof gaps become scorecard issue IDs.
04 · init-project to team

Apply To `$init-project` And `$team`

변경 지점은 `$init-project` 생성물, `$team` 계획, Phase 4 QA evidence, scorecard feedback 네 곳이다. 전부 prose가 아니라 artifact edge로 내려야 한다.

$init-project generate scaffold goal object hard contract existing bridge work_unit_ledger invariant risk proof command new field $team plan task spec acceptance exit criterion Phase 0/1 QA evidence acceptance_verified proof_boundary_rung proof_observed Phase 4 scorecard gap open_gaps[] closed_gaps[] system feedback self-improve soft to hard scope decision promotion path next project run fresh context persistent state closed loop Invariant: no prose-only pass, no smoke-only UI completion, no hard-contract downgrade.
05 · artifact schema

Recommended Artifact Schema

`$team`의 task shape에 Fable-ish 정보를 붙이고, QA evidence와 scorecard가 같은 ID를 따라가게 만든다.

team task id spec acceptance existing shape work_unit_ledger invariant risk exit_criterion proof_command review_lenses new layer qa evidence acceptance_verified[] proof_boundary_rung proof_observed missing_harness_decision candidate_findings[] hard QA observation loopy-era scorecard open_gaps[] closed_gaps[] score, level system feedback Pass condition acceptance verified + proof rung matches risk + review findings resolved or recorded + open harness gaps empty
06 · evidence graph

Evidence Inventory

출처와 runtime 검증을 같은 그래프에 묶었다. HTML 주장은 이 경로의 파일과 validator output을 근거로 한다.

Fable-ish source /Downloads/fable-ish/SKILL.md references/*.md agents/openai.yaml Harness source ~/.codex/harness/* init-project/SKILL.md team/SKILL.md Analysis artifact fable-ish-harness-comparison.html inline SVG diagrams parser validation Project runtime hard-process-contract.json team-handoff.json harness-upgrade-audit-latest.json Observed validation hard contract: 40/40 init workflow: 157/157 user target: 156/156 upgrade audit: score 100
07 · final assessment

Final Assessment

Fable-ish의 가장 큰 가치는 더 많은 agent를 띄우는 것이 아니라, 각 작업을 invariant, risk, exit criterion, proof로 쪼개고 proof가 실제 위험을 검증하지 못하면 기준을 올리는 규율이다.

Keep dynamic loop habit review lens discipline missing harness rule Do Not Replace hard process contract QA cycle gate upgrade audit Adopt As Machine-Readable Micro-Loop work_unit_ledger + proof_boundary_rung + missing_harness_decision stored in team plan, QA evidence, and scorecard gap inventory completion only when evidence observes the real risk

결론: 현재 Codex harness는 이미 hard contract와 `$team` preflight가 강하다. 다음 업그레이드는 Fable-ish의 micro-loop 정보를 QA evidence와 scorecard로 흡수해서, 작업 단위별 proof quality를 기계가 읽을 수 있게 만드는 것이다.