trace-first-starter

RES — TFW-19: Config Propagation

Date: 2026-04-03 Author: Coordinator (AI) Status: 🔬 RES — Complete Parent HL: HL-TFW-19 Mode: Pipeline


Research Context

TFW-12 centralized config values to PROJECT_CONFIG.yaml using Pattern B (pure reference). This broke agent enforcement of scope budgets. TFW-19 proposes restoring inline values (Pattern A) + creating tfw-config workflow for sync. RESEARCH validates the approach.

Decisions

# Decision Rationale
R1 Pattern A = inline defaults + config key column is the standard research.md Limits table already implements this and works
R2 Rationale column only in conventions.md (canonical description). All other files = compact: Parameter, Default, Config key Token density: fewer tokens = better agent performance (P10)
R3 tfw-config = interactive dialog (user says what to change, agent updates all files) + verify mode (audit) User requirement: «спрашивало что хотите изменить и вело дальше»
R4 Config Sync Registry lives inside config.md workflow file Self-contained, no extra files, agent reads workflow → sees registry
R5 No scripts — AI agent reads YAML, finds sections, updates values TFW constraint: pure AI prompt framework, universal
R6 Section-based lookup (header + row label) = not fragile for AI AI agents understand document structure natively, no regex needed
R7 3 categories inline: scope_budgets, research, knowledge. Others (statuses, templates, workflows) = lookup, not enforcement — skip Agreed by user

Open Questions

All closed.


Stage: Gather ✅

Full config audit: 7 config sections, 3 need inline enforcement.

Category Config Keys Inline State Action
scope_budgets (4) max_files, max_new, max_loc, max_modified ❌ Removed by TFW-12 Restore
research (4) queries, files, questions, passes ✅ Pattern A already Keep
knowledge (6) interval, gate_mode, max_index, etc. ❌ Names only, no numbers Add
statuses, templates, workflows, project, tfw meta Various Identity/lookup Skip

External: Industry confirms inline = only enforcement for AI prompts without code validation.

Checkpoint: Gather

| Found | Remaining | |——-|———–| | research.md = working Pattern A reference | — | | 3 categories need inline, 4 skip | — | | External: inline = advisory but only option for prompt frameworks | — |

Agent assessment: Gaps identified and closed. → User decision: closed


Stage: Extract ✅

Target file mapping:

File Section Format Action
plan.md L133-137 §Scope Budget per Phase Compact (no Rationale) Restore table
conventions.md L132-135 §6 Scope Budgets Full (with Rationale) Restore table
TS.md L27 Budget line Inline text Replace with defaults
research.md L140-152 §Limits ✅ Compact + config key Keep, restore «defaults» wording
knowledge.md (new) §Limits (add after Anti-patterns) Compact (no Rationale) Add table

Checkpoint: Extract

| Found | Remaining | |——-|———–| | 5 target files mapped with exact locations | — | | research.md lost «defaults» wording in TFW-21 | — | | knowledge.md has 6 config refs without numbers | — |

Agent assessment: Full map ready. → User decision: closed


Stage: Challenge ✅

Challenge Result
C1: Section-based lookup fragility Not a problem for AI — native document understanding
C2: Word count tension with TFW-21 Negligible (+3-7%, ~40-60 words per table)
C3: tfw-config UX Two modes: edit (interactive, primary) + verify (audit)
C4: Registry placement In workflow file — self-contained
C5: Downstream project init Defaults match → in sync from start

Checkpoint: Challenge

| Found | Remaining | |——-|———–| | No show-stoppers | — |

Agent assessment: Design validated. → User decision: closed


Final Checkpoint

Stage Status Key Findings
Gather 3/7 categories need inline. research.md = working reference
Extract 5 files mapped, exact sections identified
Challenge No fragility, negligible word count impact, interactive UX

Verdict: Sufficient for HL finalization.

Fact Candidates

# Category Candidate Source Confidence
F1 process TFW = чисто AI prompt framework, no scripts. Workflows ARE the automation — AI agent is the execution engine User message High
F2 constraint Pattern B (pure reference «see config») kills agent enforcement for numeric limits. Proven by TFW-12→TFW-19 regression Root cause analysis High
F3 convention Reference Pattern A format: inline defaults + config key column. research.md Limits table = canonical reference Extract finding High

Conclusion

RESEARCH confirmed the HL design. Key value: discovered research.md already implements the target pattern (Pattern A), proving the approach works in production. External research validated that inline values are the only enforcement mechanism for AI prompt frameworks. The interactive tfw-config model addresses the original TFW-12 design gap. Self-critique: Challenge was confirmatory — the design was well-grounded from TFW-12 lessons learned.


*RES — TFW-19: Config Propagation 2026-04-03*