Claude Code 三模型分工修 bug 流程

damkin @damkina7原推发布于 2026.10.24

在 X 上打开原推

Holy sh*t, someone at Anthropic just leaked the cleanest bugfix pipeline for Claude Code The brutal reality: letting AI edit production code directly from a stack trace is burning staging environments One terminal command, 52 seconds, and the bug gets isolated, reproduced in a sandboxed test, and patched without guessing 4 deterministic files, 89 lines of surgical diff, $0.03 API burn. You ran one prompt and wrote zero code by hand Launch Claude Code like this: claude --model opus "reproduce and patch: ..." how the workload splits across the frontier models: → 3 parallel Haiku 5.5 subagents swarm the failure telemetry before a single line is touched → Agent 1 maps stack trace call sites, Agent 2 checks git blame diffs, Agent 3 isolates payload fixtures → Claude Sonnet 5.5 writes a standalone reproduction test in an isolated runner (fails: Red) → Claude Opus 5.5 crafts the minimal AST patch to turn the test Green with zero collateral drift → 4 files land: reproducing test case, minimal fix, regression assertion, and a /verify check the fatal mistake 95% of engineers make: asking models to fix bugs from raw error logs a patch written without a reproduction test is pure hallucination, and you end up debugging downstream regressions for weeks in this run, the triage pass isolated 3 edge-case mutations and locked them into a reproduction suite before any source file was opened the tooling pass wired the test runner directly to the repo's native test harness instead of inventing generic assertions Haiku isolates · Sonnet reproduces · Opus patches JEV engineering applies the exact same discipline one layer down: 2,340 runtime forks (AST targeting, mock boundaries, assert checks) resolved in code at 0.98 sharpness, while ambiguous splits (0.58 on breaking contract changes) escalate directly to the human terminal instead of guessing the multi-model execution tree: > you paste one directive with the stack trace or failing issue > 3 Haiku 5.5 subagents parse call sites, git blame diffs, and mock payloads concurrently > each subagent yields structured diagnostic abstracts, never raw logs > Claude Sonnet 5.5 compiles an isolated reproduction test that fails on main > Claude Opus 5.5 synthesizes the minimal AST patch to pass the suite > all breaking API shifts block execution until explicit human confirmation > JEV deterministically clears micro-forks and surfaces only edge ambiguities feed this tree and prompt directly into Claude Code ↓ "Reproduce and patch this issue for me: 1. First, scan the failure with 3 Haiku 5.5 subagents in parallel: one maps the stack trace to exact AST call sites, one analyzes recent git blame diffs, one isolates mock input fixtures. 2. Each subagent returns a concise diagnostic summary only, never raw log files. 3. Spawn Claude Sonnet 5.5 to write an isolated reproduction test that fails deterministically on the current branch. 4. Once confirmed Red, deploy Claude Opus 5.5 to generate the minimal AST patch to make the test pass Green without touching unrelated code. 5. Halt and ask for confirmation before modifying any production file. Display every file modification as a unified diff first. Zero writes without my explicit approval." ↳ http://code.claude.com/docs/en/bug-triage bookmark this, you will execute it on every single unhandled production error

这条是引用着另一条写的：x.com/damkina7/status/2107924159397560795