Skip to content

title: "Extract — "What do we NOT see?"" source: "tasks/TFW-47__codex_adapter_shortcut_skills/research/iter2/3_extract.md"


Extract — "What do we NOT see?"

Mindset: Analyst. You have the raw findings. Now build structure. Make combinations visible that nobody proposed. Test: "Does my configuration space reveal at least one combination that nobody proposed in the Briefing?" Parent: HL-TFW-47 Goal: Codex becomes a first-class TFW adapter with dedicated shortcut skills.

Configuration Space

Dimensions from Gather: (1) Skill Directory, (2) Skill Content Strategy, (3) Instruction Loading, (4) Sync Strategy, (5) Invocation Documentation.

Config Skill Directory Skill Content Instruction Loading Sync Strategy Invocation Docs
C1 .agents/skills/ (project) Handwritten per-workflow On-demand via skill contract tfw-init copies from .tfw/adapters/codex/ Document $tfw-* as primary
C2 .agents/skills/ (project) Generated from template On-demand via skill contract tfw-init generates from template Document $tfw-* as primary
C3 .agents/skills/ (project) Handwritten per-workflow AGENTS.md embeds context chain tfw-update re-copies Document both $ and /
C4 .codex/skills/ (legacy) Handwritten per-workflow On-demand via skill contract tfw-init copies Document $tfw-* only
C5 .agents/skills/ (project) Hybrid: template + per-workflow overrides On-demand via skill contract tfw-init copies, tfw-update diffs Document $tfw-* as primary
C6 No skill folders — AGENTS.md routing only N/A AGENTS.md embeds full routing table tfw-update syncs AGENTS.md Document natural language only

Findings

E1: Template Insufficiency for Specialized Skills

The generic template (tfw-command-skill.md.template, 789 bytes) covers: - Alias handling, .tfw/ check, context loading, role lock, template usage

It does NOT cover: - tfw-research: iterations.yaml, stage templates, researcher role lock, never-overwrite-prior-iterations rule - tfw-config: config sync registry, verify/edit modes, specific permitted files - tfw-review: verify-claims-against-real-files rule, review stage templates - tfw-handoff: scope guard ("do not expand beyond TS"), ONB-before-implementation rule - tfw-update: never-overwrite-project-state rule, preserve-customizations rule - tfw-init: handle missing .tfw/ (other skills assume it exists)

At least 6 of 11 skills require workflow-specific contract clauses that a generic template cannot express. This is structural, not a tuning issue.

E2: On-Demand Loading Eliminates the 32 KiB Problem

The 32 KiB cap only applies to AGENTS.md auto-embed. The AFD skills instruct the agent to read conventions.md at runtime — this is a file read, not an instruction embed. TFW's AGENTS.md is ~1.8 KB (well under the cap). Even if all workflow files were loaded, they're loaded sequentially during execution, not upfront.

Unexpected finding: The 32 KiB problem (HL Risk R3, high probability/high impact) is a non-issue with the on-demand skill pattern. The risk was based on assuming Codex loads everything into the instruction chain upfront.

E3: .agents/skills/ vs .codex/skills/ — Clear Winner

.agents/skills/ is both the 2026 cross-platform standard AND already used in the AFD project. .codex/skills/ is legacy. No reason to support the legacy path in TFW's canonical adapter.

The AFD project already validates this: 11 skills in .agents/skills/, working in production.

E4: AGENTS.md-Only (C6) is Insufficient

C6 (no skill folders) is the current state for non-AFD TFW projects. It routes behavior correctly — the agent reads AGENTS.md, finds workflow references, follows them. But it provides zero UI affordance: no $ menu entries, no /skills listing, no tab completion. H4 confirmed: routing ≠ visibility.

E5: Generated vs Handwritten Tradeoff

Factor Generated (C2) Handwritten (C1/C5)
Maintenance Low — regenerate all from template Medium — edit individually
Accuracy Misses workflow-specific rules (E1) Captures all contract clauses
Drift risk Template drifts from workflows silently Skills drift from workflows individually
Setup cost Low — script generates 11 folders Medium — copy 11 files
AFD evidence Template exists but all 11 are handwritten 11 working handwritten skills

The AFD evidence is decisive: the team started with a template, then replaced every skill with a handwritten version. This is empirical proof that generation is insufficient.

Checkpoint

Found Remaining
6 configurations mapped None
Template insufficiency proven (6/11 need custom clauses) None
32 KiB problem is a non-issue with on-demand pattern None
.agents/skills/ confirmed as canonical directory None

Sufficiency: - [x] External source used (web research on .agents/ vs .codex/) - [x] Briefing gap closed - [x] Configuration Space built from Gather dimensions

Stage complete: YES