DSH Smarter Edit
中文说明
A better file-editing tool for DeepSeek Harness.
DSH's native edit requires coding models to reconstruct exact source strings
and send both old and new code through structured JSON arguments.
DSH Smarter Edit replaces edit with Codex-compatible freeform
apply_patch, allowing the model to express code changes directly as a diff.
Its goal is to reduce editing failures and token overhead created by the editor
protocol itself.
*** Begin Patch
*** Update File: src/app.ts
@@
-const result = oldMethod(value)
+const result = newMethod(value)
*** End Patch
The model describes the change instead of reconstructing an exact
old_string / new_string pair.
Let the model spend its tokens on changing code, not satisfying the editor
protocol.
DSH Smarter Edit: a better file-editing tool for DeepSeek Harness
Why this exists
Coding models should solve one problem: what code should change. DSH's native
editing protocol adds three more.
1. Exact-string reconstruction
Native edit(file_path, old_string, new_string) asks the model to regenerate
source it has already read. The old_string must match the file:
The model knows the correct code change
↓
old_string differs in whitespace, indentation, or characters
↓
The edit fails
The model chose the code change correctly, but the tool rejects it during
argument matching. Smarter Edit removes the exact-string pair from the
model-facing tool and lets contextual patch hunks locate the change.
2. Source duplication
Exact-string editing makes the model output both the existing block and a new
block that often repeats most of it:
old_string = existing source
new_string = mostly the same source + the change
A patch expresses the changed lines and only the context needed to locate them:
required context
-before
+after
required context
The intended token reduction comes from less repeated old source, less repeated
unchanged source, less JSON escaping, and fewer retries after exact-match
failures. The maintainers will not publish a percentage until controlled A/B
tests produce one.
3. Structured JSON overhead
The model must also encode multiline source inside a schema:
{
"old_string": "...",
"new_string": "..."
}
Quoting, escaping, and the exact-match fields belong to the editor protocol,
not the code change. Freeform apply_patch lets the model generate the editing
language itself.
Evidence from OpenAI
OpenAI moved apply_patch to a named freeform tool. Its
official model guide
reports that this change reduced apply_patch failure rates by 35% in
testing compared with JSON-formatted function calling.
The result measures apply_patch call failures. It does not measure task
success, SWE-bench performance, or token savings. It shows that the input format
of an editing tool measurably affects reliability.
Evidence of exact-string fragility
Literal replacement succeeds only when old_string reproduces the target
exactly. Public Claude Code issue reports show how tabs, line endings, and other
invisible details can turn a valid code change into
String to replace not found in file:
- #13152 reports tabs
appearing as spaces in read output, followed by exact-match edit failures.
- #28831 reports
unreliable multiline matching on CRLF files with tab indentation and
identifies read formatting as one possible cause.
- #54876 reports LF
read output against CRLF file contents.
- #40471 reports a
repeated edit-fail, retry, and Python/Bash fallback cycle on tab-indented
files.
These reports describe another coding tool and do not establish DSH behavior.
They expose the underlying weakness of literal replacement: the model must
reproduce file details that are unrelated to the intended code change. Any
difference can reject the edit and trigger retries or fallback tools.
Additional capabilities
Besides changing the editing tool exposed to the model, Smarter Edit supports:
- multiple ordered hunks;
- add, update, move, and delete operations;
- Codex-compatible contextual matching;
- failure-atomic preflight and rollback;
- native DSH diff rendering.
Install
Install the current npm release into a Desktop Profile:
dsh plugin --profile desktop add @anionex/dsh-smarter-edit
Use web or headless instead of desktop for those Profiles. Restart the
running Profile and create a new Session after installation.
Verify that the bundle is mounted:
dsh --profile desktop --dump-config | grep tool-apply-patch
How it works
- The model sends a patch between
*** Begin Patch and *** End Patch.
- Smarter Edit validates the complete patch, locates each change from its
context, and applies it through the active DSH sandbox. A handled failure
rolls back published changes.
- DSH stores the raw patch and renders the result with its native diff view.
The plugin removes model-visible edit, write, and legacy
str_replace_editor plus their prompt sections while replacement mode is
active. Unloading it restores the original tools.
Compatibility and limitations
- DSH
>=0.1.1-rc.1 <0.2.0.
- Node.js
^22.19.0 || >=24.0.0 for package development and direct engine use.
desktop, web, and headless Profiles.- Write permission from the active DSH sandbox.
workspace-write keeps its
configured boundary; danger-full-access can allow absolute and
parent-relative paths outside it.
- Rollback on handled failures protects file state, not retry cost. If one hunk
in a large patch fails, the whole patch is rejected and a retry must submit
the full patch again.
- A broad patch can change or delete several files. Review the requested patch
and the resulting diff when a change has a large scope.
The project is pre-1.0. Compatibility targets the declared DSH 0.1.x range;
release verification installs the tarball into a clean Profile.
Implementation details
See Implementation Notes for Codex alignment,
failure atomicity, line-ending behavior, package entry points, and provider
transport.
Syntax
Supported operations:
*** Add File: relative/path.ts
+every added line starts with +
*** Delete File: relative/obsolete.ts
*** Update File: relative/current.ts
*** Move to: relative/renamed.ts
@@ optional function or section context
unchanged context
-removed line
+added line
*** End of File
Move to is optional. End of File anchors its hunk to the file tail. A
pure-addition update hunk appends to the file. Relative paths resolve from the
Session working directory; absolute and parent-relative paths work when the
active DSH sandbox permits them. Repeated file sections run in patch order, so
each operation sees earlier operations in the same call. Add over an existing
file replaces it, matching Codex behavior.
Configuration
Profile patch values, with defaults:
- id: tool-apply-patch
config:
replaceNativeEdit: true
Set replaceNativeEdit: false only when you want edit, write, and
str_replace_editor alongside apply_patch. It does not change freeform
transport or transaction behavior.
Benchmarking
The first planned A/B comparison uses DeepSeek V4 Flash on the same coding
fixtures with native DSH edit and DSH Smarter Edit. Each cohort must use the
same prompt, model settings, reasoning effort, sandbox, clean workspace
snapshot, and trial count.
| Metric | Definition |
|---|
| Mutation calls | File-mutation tool calls before completion |
| Mutation failure rate | Rejected or failed mutation calls / all mutation calls |
| Output tokens | Assistant output tokens for the full task |
| Rounds | Model request steps before the terminal answer |
| First-test pass | First test command exits zero |
| Final success | Every fixture acceptance check passes |
| Wall time | User request accepted to terminal result |
Reject a trial from the freeform cohort if the recorded provider request
describes apply_patch as type: "function". A valid trial must show OpenAI
type: "custom" with format.syntax: "lark". Publish measured results here
only after repeated controlled trials.
Development
pnpm install --frozen-lockfile
pnpm peers check
pnpm run check
The test suite covers the official Codex corpus, parser and matcher behavior,
sandbox paths, preflight isolation, rollback faults, cleanup, concurrent
modification detection, tool registration, native-tool filtering, package
layout, raw DSH call reconstruction, native diff presentation, and captured
OpenAI request serialization.
Read CONTRIBUTING.md before changing behavior or packaging.
Security and community
Maintained by Anionex under the
MIT License.
