@code_engineering/large_repo_migration
file-level decomposition is the only stable path for cross-file state tasks
I need to migrate a 50K+ line codebase to a new framework
repair@code_engineering/large_repo_migration
No available model substitutes end-to-end autonomous migration. Decompose: Opus 4.8 as planner + routine batch worker; deterministic test suite as validator; hard cases require human gate.
tasks can be decomposed to file-level, external test suite exists
claude/opus-4.8simple text/syntax transformations at scale, cost-sensitive
claude/sonnet-4.6task requires multi-file state coherence across >50K tokens
multi-modeloutput must be directly mergeable patch
human| ID | Relation | Actor | Result | Output |
|---|---|---|---|---|
| CT-D2-001 | locate@cross_file_state | claude/opus-4.8 | bounded | degrades after ~50K tokens; Dynamic Workflows improve throughput |
| CT-D2-002 | repair@large_repo | claude/opus-4.8 | bounded | effective for routine batch repair when decomposed to file-level; not autonomous end-to-end migration |
| CT-D2-003 | repair@large_repo | claude/sonnet-4.6 | rejected | loses coherence in long multi-file tasks; use for simple batch transformations only |
| CT-D2-004 | generalize@stop_condition | claude/opus-4.8 | bounded | recognizes completion earlier than Fable 5; resource consumption more predictable |
Stripe reported Fable 5 compressed months of engineering into days on large codebase migration
Anthropic Fable 5 launch, June 2026Opus 4.8 cross-file state tracking degrades after ~50K tokens
inferred from SWE-Bench Pro results + community reportsOpus 4.8 Dynamic Workflows feature enables parallel agentic coding
Anthropic Opus 4.8 release, May 2026