[CLAUDE] Docs: S124 adap-wave AI_INFRA + owner worker-tier-pin (14 agent opus+effort:max) + user-memory relocate R6 + H24 dogfood
All checks were successful
Deploy SOLUTION_ERP / build-deploy (push) Successful in 5m26s
All checks were successful
Deploy SOLUTION_ERP / build-deploy (push) Successful in 5m26s
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
@ -0,0 +1,81 @@
|
||||
---
|
||||
name: feedback-implementer-truncation-mitigation
|
||||
description: Implementer Case 2 heavy scaffold (≥30 file OR ≥50 tool uses) truncates mid-MEMORY update phase — functional work complete, knowledge append cutoff. Mitigation strategies cross-project.
|
||||
metadata:
|
||||
node_type: memory
|
||||
type: lesson
|
||||
originSessionId: e3680af8-fab4-4251-84b8-8c8178d58689
|
||||
---
|
||||
|
||||
# Implementer truncation pattern observation — heavy scaffold MEMORY update cutoff
|
||||
|
||||
**Date observed:** 2026-05-26 (S33 Plan B G-H1 Phase 2 — 4 Implementer spawn)
|
||||
**Proven:** 2/3 truncation rate (Task 3 + Task 4 truncated, Task 5 clean)
|
||||
|
||||
## Pattern
|
||||
|
||||
Implementer Case 2 cookie-cutter scaffold task có scope ≥30 file OR ≥50 tool uses (~150K+ token budget):
|
||||
- Scaffold work proceeds normal (Edit/Write/Bash tools consume budget)
|
||||
- Final phase — Update MEMORY.md Recent activity FIFO — runs out token budget
|
||||
- Truncates mid-sentence cutoff "Let me check..." / "Let me append..." / "Let me also peek at..."
|
||||
|
||||
**Functional work COMPLETE** despite MEMORY truncation — verifiable via:
|
||||
- File existence (Glob check)
|
||||
- Build clean (dotnet build + npm build)
|
||||
- Test PASS (dotnet test baseline preserve)
|
||||
|
||||
## Cases S33 evidence
|
||||
|
||||
| Task | Tool uses | Token | MEMORY status |
|
||||
|---|---:|---:|---|
|
||||
| Task 3 BE entity 17 file scaffold | 65 | 125K | TRUNCATED mid-Pattern 12-bis lookup |
|
||||
| Task 4 BE CQRS 3 file scaffold | 60 | 123K | TRUNCATED mid-update |
|
||||
| Task 5 FE 12 file × 2 app | 61 | 154K | **CLEAN** (smaller scope per file) |
|
||||
| Plan C BW1-BW7 3 test file | 49 | 164K | TRUNCATED mid-check what else |
|
||||
|
||||
## Mitigation strategies (em main proxy or pre-empt)
|
||||
|
||||
### Strategy A — Em main split heavy task into 2 phase
|
||||
- Spawn Implementer Phase 1 — scaffold N file under 100K token budget
|
||||
- Spawn Implementer Phase 2 — MEMORY update + verification report (~20K token reserve)
|
||||
- Drawback: 2× spawn overhead, but MEMORY append guaranteed
|
||||
|
||||
### Strategy B — Em main proxy MEMORY append next session
|
||||
- Functional work complete → commit + push as planned
|
||||
- Em main append MEMORY entry retrospectively next session (5-10 phút)
|
||||
- Pattern foundation already in MEMORY from prior sessions (S29 Plan B Pattern 12-bis) — no critical knowledge loss
|
||||
- Cost: minor delay knowledge cumulative re-rank
|
||||
|
||||
### Strategy C — Implementer self-aware token budget
|
||||
- Implementer system prompt update: "Reserve 20K token cho MEMORY update phase, KHÔNG dive deep verification step if token < 30K"
|
||||
- Trade-off: less thorough verification before report
|
||||
|
||||
### Recommend cross-project
|
||||
|
||||
- **Default Strategy B** — em main proxy MEMORY append next session (low overhead, knowledge cumulative still preserved via prior foundation entries)
|
||||
- **Strategy A** chỉ khi task critical AND MEMORY knowledge HIGH value (Pattern NEW first time, not reinforce existing)
|
||||
- **Skip Strategy C** — system prompt change risks Implementer skip critical verification, prefer truncate MEMORY (recoverable) over skip verification (silent break)
|
||||
|
||||
## When to apply (cross-session)
|
||||
|
||||
**Why:** Em main session sau khi spawn Implementer heavy scope (≥30 file scaffold):
|
||||
- Set expectation Implementer MAY truncate mid-MEMORY → acceptable
|
||||
- Verify functional work first (file existence + build + test)
|
||||
- Defer MEMORY append next session if truncated
|
||||
|
||||
**Apply:** Check Implementer report ending. If sentence cutoff `"Let me check..."` / `"Let me append..."` / `"Let me also peek..."` → em main:
|
||||
1. Verify file scaffold complete (Glob/Read)
|
||||
2. Verify build + test pass
|
||||
3. Mark task DONE despite MEMORY truncation
|
||||
4. Note "MEMORY proxy append next session" in commit message
|
||||
5. DON'T re-spawn Implementer just for MEMORY update (waste 30K+ token)
|
||||
|
||||
## Pattern reusable
|
||||
|
||||
Cross-project — applies any project where Implementer/code-execution subagent has heavy scaffold scope:
|
||||
- VIPIX_MULTISITE — em main project N1 dev solo similar
|
||||
- NAMGROUP — historical heavy scaffold pattern
|
||||
- ASHICO — 5 .NET sites scaffold cross-site mirror
|
||||
- BENHVIEN_AAU — internal system scaffold
|
||||
|
||||
Related: [[feedback-per-chunk-commit]] [[feedback-multi-agent-setup]] [[feedback-rag-hybrid-pattern]]
|
||||
Reference in New Issue
Block a user