Skip to content

title: "Challenge — "What do we NOT expect?"" source: "tasks/TFW-29__consistency_audit/research/challenge.md"


Challenge — "What do we NOT expect?"

Parent: HL-TFW-29 Goal: Reference files (conventions.md, glossary.md) and 11 workflows are free from redundancy — agents load minimum tokens for maximum signal.

Findings

C1: Hypothesis H2 — Is the Common Spine Real?

Hypothesis: All workflows follow the same 4-step context loading pattern but then diverge — the divergence points reveal which convention sections are actually needed per role.

Verdict: PARTIALLY CONFIRMED.

The spine (AGENTS → conventions → glossary → KNOWLEDGE → task artifacts) is real and shared by 4 core workflows (plan, research, handoff, review). But it's not universal: - 5 workflows (knowledge, release, update, config, init) use their own prerequisites — they don't need the general spine because they work with specific files. - 2 workflows (resume, docs) have NO loading section — a gap, not a design choice.

The divergence points DO reveal role-specific needs: - plan needs §4, §5, §6, §10, §14, §15 — the most §-hungry workflow - research needs §10, §14 — minimal - handoff/review need the full spine but DON'T reference any specific § — they get value from conventions/glossary as background, not from specific rules

Implication for compression: The spine workflows will still load conventions + glossary. But the "dead" 55% of conventions.md means the spine workflows waste context on unreferenced sections. The fix isn't to split conventions — it's to ensure every section earns its place.

C2: Hypothesis H4 — Can Non-Research Workflows Skip Glossary?

Hypothesis: Only Research workflows actually need glossary-unique terms (Stage, Pass, OODA); other workflows could skip glossary.

Verdict: REFUTED.

Glossary-unique terms needed by non-Research workflows: 1. handoff.md needs: Phase, Scope Budget, Fact Candidate, Roles (Executor) 2. review.md needs: Roles (Reviewer), Fact Candidate, Strategic Insight 3. plan.md needs: Phase, Knowledge Gate, Strategic Insight, RESEARCH (concept, not just workflow), Roles (Coordinator) 4. knowledge.md needs: Topic File, Consolidation, Fact Candidate, Knowledge Gate, Config Sync Registry 5. resume.md needs: Phase (core to its function)

Conclusion: Glossary has value beyond Research. The ~15 unique terms serve multiple workflows. Removing glossary from context loading would harm terminology consistency.

But: The 10 duplicated terms and 6 meta terms should still be compressed. An agent loading glossary should get unique definitions, not re-reads of conventions.

C3: Counter-argument — Self-Contained Glossary is USEFUL

The case FOR duplicating artifact defs in glossary: - Glossary is a lookup document. An agent looking up "HL" shouldn't need to cross-reference conventions.md — the definition should be right there. - Progressive disclosure says "load only what you need." If a glossary has 1-liners + refs, the agent must now load 2 files to understand a term.

The case AGAINST (stronger): 1. The 1-liners still provide enough context for lookup: "HL (High Level) — context/frame artifact for a task. → conventions.md §3 for full definition." An agent knows what HL is from this single line. 2. The full definitions in glossary DON'T match conventions perfectly — they add/omit details, creating subtle inconsistencies. Example: glossary's HL entry mentions "Contains: Vision, As-Is, To-Be, Phases, DoD, DoF, Principles, Dependencies, Risks" — this list is NOT in conventions §3 and could drift as the HL template evolves. 3. When conventions §3 is updated, glossary must be manually synced. This has already drifted — evidence: glossary says HL "Contains: Vision, As-Is, To-Be..." which is the OLD template structure (pre-TFW-24, which added §10 RESEARCH Case and §11 Strategic Insights). 4. External research (G5) confirmed: duplication is an anti-pattern for AI agent context. DRY + precise refs is the correct architecture.

Verdict: 1-liners with refs are sufficient. The convenience of self-contained lookup does NOT justify the consistency risk and token cost.

C4: Risk Check — What Could Break from HL §2.4 Compression?

Testing each proposed compression against the workflow corpus:

HL Proposal Risk Assessment
Glossary: replace artifact defs with 1-liners + refs Could break if any workflow reads glossary for artifact semantics Safe — no workflow reads glossary for artifact definitions. They read conventions §3 or the templates directly.
Glossary: replace status flow diagram with ref Could break if any workflow reads glossary for status transitions Safe — only plan.md reads §5, and it reads from conventions.md directly.
Glossary: remove Concept Taxonomy Could break if any workflow uses it Safe — grep shows 0 runtime references from any workflow.
Glossary: remove .tfw/ Directory definition Could break if init/update rely on it from glossary Safe — init/update read the actual directory, not a glossary definition of it.
Conventions: extract §16 to separate file Could break if any workflow reads §16 at runtime Safe — Section Usage Matrix shows 0 workflow references. Only gen_docs.py reads it via compilable contract.
Conventions: §10 numbering fix Could break plan.md / research ref "§10" Safe — refs use "§10" which will still work. Renumbering §10.1/10.2 before §10 to §10→§10.1→§10.2 just fixes ordering.
.tfw/README.md: remove anti-patterns block Could break if any workflow refs README anti-patterns Safe — no workflow references README.md for anti-patterns. They reference conventions.md §14 directly.

No breakage risk identified for any proposed compression.

C5: Gap Found — resume.md and docs.md Missing Context Loading

Both workflows run in Coordinator role but have no Context Loading section. This means: - An agent starting /tfw-resume without prior context has no instructions to load conventions or glossary - An agent starting /tfw-docs after a review has no instructions to load KNOWLEDGE.md (which it's about to update)

This is a bug, not a design choice. Both should have at minimum: "Read conventions.md §10" or their own prerequisite list.

Recommendation: Add minimal loading to both: - resume.md: "Read conventions.md §10. Then read Master HL for the target task." - docs.md: Already has trigger modes. Add prerequisite: "Read KNOWLEDGE.md and TECH_DEBT.md before starting checklist."

C6: Gap Found — AGENTS.md Lists Only 5 of 11 Workflows

AGENTS.md §Workflows lists only plan, handoff, review, resume, docs. Missing: research, knowledge, init, config, release, update.

.agent/rules/agents.md is worse — only lists 4 (plan, handoff, resume, docs — missing review!) and has the "handoff → REVIEW" error.

Impact: An agent in a fresh session doesn't know research, knowledge, init, config, release, or update workflows exist. It can't recommend /tfw-research or /tfw-knowledge if it doesn't know they're available.

Recommendation: AGENTS.md must list all 11 workflows. The adapter .agent/rules/agents.md must be synced with AGENTS.md.

Checkpoint

Found Remaining
H2: partially confirmed — spine is real for 4 workflows, not universal
H4: refuted — glossary unique terms needed by 5+ non-Research workflows
Self-contained glossary case lost — consistency risk > lookup convenience
All HL §2.4 compressions verified safe — zero breakage risk
Bug: resume.md and docs.md missing Context Loading
Bug: AGENTS.md lists only 5/11 workflows

Sufficiency: - [x] External source used? (G5 external research applied to challenge arguments) - [x] Briefing gap closed? (All 3 Challenge bullets from briefing covered)

Stage complete: YES → User decision: proceed to Synthesis