dsh-code-reuse-firewall
Manifest validPre-write reuse firewall for DeepSeek Harness: before the agent writes a new helper/service, surface the existing implementations that already cover that intent. Deterministic retrieval (no LLM) via t
dsh-code-reuse-firewall
Pre-write reuse firewall for DeepSeek Harness — for PYTHON repositories.
Before the agent writes a new helper / service / manager, reuse_check
deterministically surfaces the existing Python implementations that already
cover that intent, so the agent reuses or extracts instead of duplicating.
Why
Static checkers only catch what looks wrong. The expensive failure mode in AI-maintained codebases is the opposite: two implementations of the same capability drift apart silently, because nothing looked broken when the second copy was written. The fix is to intervene before the second copy exists — surface the overlap while the new code is still a plan, not a file.
The retrieval is deterministic and LLM-free (callable-name, docstring
lexical, and string-literal channels with IDF-weighted query coverage, stdlib
Python only), backed by the
Auto_code_audit capability
channel. It was validated on an unfamiliar mid-size project (arrow-py): a 1s
scan surfaced nine near-identical locale _format_timeframe methods, four
describe twins, and api.get vs ArrowFactory.get near-duplicates with zero
noise in the dead-code / hardcoded / style categories.
Requirements
- A checkout of Auto_code_audit
(its
capability_retrieval.pyis the retrieval engine). - A Python 3.10+ interpreter (default
python).
Install
dsh plugin add github:keyiadiannao/dsh-code-reuse-firewall#master
Then configure the audit checkout and interpreter in your profile:
- id: dsh-code-reuse-firewall
config:
auditRoot: 'D:/path/to/Auto_code_audit' # required
pythonPath: 'python' # default
maxK: 5 # top-K candidates
minScore: 0.1 # score floor
timeoutMs: 30000 # child-process cap
Usage
The agent calls reuse_check before writing new code:
调用 reuse_check:我要实现「从 JSON 配置读取并支持环境变量覆盖」,根目录是 D:/project/src。看看有没有现成的实现可以复用。
(call reuse_check: I'm about to implement "load a JSON config with environment-variable overrides", root D:/project/src. Is there an existing implementation to reuse?)
The tool returns top candidates with paths, scores, and per-channel evidence:
Existing implementations overlapping "load a JSON config with env overrides":
[0.72] config.py:load_config (src/config.py) (name=0.72 doc=0.10 literal=0.00)
[0.51] util.py:ConfigLoader.load (src/util.py) (name=0.51 doc=0.00 literal=0.00)
Hash-locked candidates are flagged. When an existing implementation lives
in a file pinned by a frozen-JSON provenance manifest (e.g.
current_dependency_files / files_sha256 in frozen_results/ or configs/),
the tool marks it 🔒 LOCKED with the locking manifests. Editing such a file
invalidates the frozen results that reference it — the correct reuse is to
import it, never to copy-and-modify its implementation. Derived run-output
trees (outputs/, reports/, logs/, runs/, cache/) are treated as
snapshots, not edit constraints, so scripts that merely appear in run metadata
are not falsely flagged.
[0.32] lib/protocol.py:_split_hash_payload (lib/protocol.py) (name=0.32) 🔒 LOCKED by configs/generation_b_training_compatibility.json
⚠ 1 candidate(s) are in hash-locked files: REUSE by import, do not copy-and-modify their implementation (editing invalidates frozen results).
Advisory evidence, not a verdict. The agent decides whether to reuse, extract a shared component, or write new code — and must never delete or rewrite anything based on retrieval alone (the same ground rule as Auto_code_audit: deterministic output is evidence, not a defect verdict).
Signal-strength caveat (honest limits)
A natural-language --describe query has NO code yet, so the retrieval engine's
strongest signals — normalized AST structure, call-name overlap, string-literal
overlap — cannot fire. The pre-write channel relies on the weaker
name/docstring-lexical/string-literal channels. In practice:
- A well-named existing function whose docstring matches your description WILL
be surfaced (verified:
load_configfor "load a JSON config with env overrides"). - Describe in English keywords a function name/docstring would use
(
load json config settings environment env override). Chinese-only descriptions match poorly against English code — the engine tokenizes CJK into bigrams with no Chinese↔English mapping. - Structurally-similar-but-differently-named code (the strongest reuse signal)
is only found AFTER code exists, via the engine's
--file/--basemodes — which are not yet exposed through this plugin. - Each candidate carries per-channel evidence (name / docstring / string-literal scores) in the tool result, so the agent can judge WHY something matched instead of trusting one blended score.
So treat reuse_check as a low-signal pre-write hint, not a full reuse
audit. The high-signal modes are roadmap items below.
Configuration
| Key | Default | Description |
|---|---|---|
auditRoot | — (required) | Auto_code_audit checkout containing capability_retrieval.py |
pythonPath | python | Python interpreter for the retrieval script |
maxK | 5 | Top-K candidates per query |
minScore | 0.3 | Score floor (aligned with the engine's default; measured hits sit at 0.33+, lower scores are mostly noise) |
timeoutMs | 30000 | Child-process timeout — retrieval never hangs a turn |
The plugin parses the retrieval JSON with a schema_version === 1 contract
check: if Auto_code_audit ever changes its output schema, reuse_check fails
loudly with "out of contract" instead of silently mis-parsing.
Development
pnpm run build # tsdown: host + client bundle
pnpm run typecheck # tsc --noEmit
pnpm test # vitest
Roadmap
tools/pre-executeguard: run a reuse check automatically before write-tool calls when areuse_checkwas not already performed (dsh-tool-git style).--file/--basemodes (check a new/changed file or diff against a git ref) exposed through the tool.
License
MIT
Comments
Loading…
Similar plugins
DeepSeek Harness plugin for Open Code Review: deterministic review scope, rule resolution, and line-level findings for DSH agents.
★ 0
dsh plugin --profile web add @ashllll/dsh-open-code-reviewby dingchenhui0618-arch
Adversary review gate for DeepSeek Harness: an independent small model reviews destructive tool calls before they run, and reports every verdict. · DSH 对抗式审查门:危险工具调用在执行前由一个独立小模型复核。
★ 0
MIT
JavaScript
Oct 27, 2026
dsh plugin --profile web add dsh-wardenby zhengqiuyang
Deterministic write gate: scope policy denies out-of-scope write/edit calls before they touch the disk, residue rules scan the written text, and every decision lands in a content-hashed JSONL ledger;
★ 0
MIT
JavaScript
Nov 10, 2026
dsh plugin --profile web add dsh-pr-gateby flyhai
Interception and audit frame for DeepSeek Harness: one approval ladder, hash-chained audit, result redaction.
★ 0
BSD-3-Clause
JavaScript
Nov 6, 2026
dsh plugin --profile web add dsh-auto-approval-frameby Elaina-real
Tiered auto-review for DeepSeek Harness (DSH): static-rule safety net + LLM reviewer + human fallback. Auto-approve safe actions, auto-deny irreversible ones, ask a human for the rest. PURE VIBE CODING - not audited, use at your own risk.
★ 0
↓ 22/wk
MIT
JavaScript
Aug 14, 2026
dsh plugin --profile web add dsh-tiered-approvalby Liu-Bot24
DeepSeek Harness(DSH)只读执行轨迹复盘插件,支持规则分析、独立模型解读、证据定位、任务概览与运行对比。
★ 6
MIT
JavaScript
Oct 21, 2026
dsh plugin --profile web add dsh-plugin-trace-insight