dsh-better-edit

dsh-better-edit

Hash-anchored edit tools for DeepSeek Harness.
Edit by content address — not by line numbers, not by string replacement. Fewer tokens. More attention for real work. Zero misapplied edits.

English · 简体中文

Quick Start • Why Hashline • Benchmark • Tools • Acknowledgments

---

"The harness — not the model — is the bottleneck." — Can Bölük, The Harness Problem

Most edit tools ask the model to echo the old code token-for-token before it can change anything — and that's exactly where agents fail: 46–51% patch-format failure rates for several models with replace-style edits. dsh-better-edit goes deeper. Every line of a file gets a unique 3-character content hash, and edits target hashes. The old text is never echoed, anchors survive edits, and every resolved range is verified against exactly what the model saw — wrong-line edits cannot silently land.

Why you need this

str_replace makes the model re-type the code it's replacing — pure transcription cost (output tokens, billed ~5-6× input), and where agents fail most: 46–51% patch failures on real models, worse on bigger blocks, each failure costing a re-read and a retry.

Hashline sends two hashes instead of the old text — 31% fewer edit tokens (43% on multi-line ranges) — and verifies every range against what the model saw: an edit lands where you meant, or fails loudly with fresh anchors. Anchors are content addresses that survive edits above, so chained edits skip re-reads — and a leaner context keeps the model's attention on the code, not on re-transcribing it.

Not for one-line touch-ups (near parity) or new files (write). It pays off in long sessions and structural edits — anywhere an edit must not land on the wrong line.

Quick Start

Install

dsh plugin --profile <name> add dsh-better-edit   # from npm
dsh plugin --profile <name> add /path/to/dsh-better-edit   # from a local checkout

The profile's next session runs with the hashline tools installed. To verify the layer is active:

dsh --profile <name> --dump-config   # shows a "# == dsh-better-edit" layer

Requirement
Node	^22.19.0 || >=24.0.0 (dsh's requirement; the store uses node:sqlite)
Profile	a dsh profile (dsh plugin initializes one on first use)
Backends	sandboxed / remote filesystems supported (writes go through ctx.fs)

read returns every line prefixed by its hash — the hash is the line's address:

ve7│function hello() {
szJ│  console.log("world");
kQm│}

edit targets a range of hashes, so edits always land on the lines you meant:

{
  "path": "src/main.ts",
  "remove_from": "szJ",
  "remove_to": "szJ",
  "replacement_text": "  console.log('hi');"
}

and produces a diff with fresh anchors, so the next edit verifies cleanly with no re-read:

− szJ │   console.log("world");
+ a3m │   console.log('hi');
  kQm │ }

Configuring Guidance per Preset

The tool:read / tool:edit / tool:batch_edit / tool:undo_last_edit guidance sections are plain-markdown files, overridable per agent preset. Override files live in the plugin's shared home — never the workspace store:

$DSH_HOME/plugins/dsh-better-edit/<preset>/<section>.md

(default home ~/.dsh, so ~/.dsh/plugins/dsh-better-edit/). The section table:

File	Section	Default order
read.md	tool:read	130
edit.md	tool:edit	131
batch_edit.md	tool:batch_edit	132
undo_last_edit.md	tool:undo_last_edit	133

On first boot the plugin seeds the four shipped presets — standard/, code/, minimal/, cordis/ — each with the compiled guidance as editable files (plus order front-matter), so every preset's guidance starts editable rather than blank. A README.md at the plugin-home root documents the scheme. Files are seeded once and never rewritten, so your edits survive. A preset directory may hold only the sections you want to override — the rest fall through to the compiled defaults.

A file is pure prose unless it opens with an order front-matter fence, which moves the section in the assembled system prompt:

---
order: 150
---

<section text>

Per section, resolution reads <preset>/<section>.md, else the compiled default. Files are read once per agent at session-start, so edits apply to new sessions — never mid-session. A preset with no seeded directory (e.g. a user-authored one) falls back to the compiled defaults unless you copy a seeded dir to its name. A deployment without the agentPresets service (no preset roster) keeps the compiled defaults and never touches these files; presets are never required.

Why Hashline

Token-saving. An edit call carries remove_from / remove_to (two 3-char hashes) plus the replacement text — it never echoes the text being replaced. A str_replace call must reproduce that text verbatim. On a 12-edit session over a realistic file this is 31% fewer output tokens (43% on multi-line ranges) — and these are output tokens, billed at ~5-6× the input rate. See the benchmark.

But this was never about “fewest tokens.” Savings scale with the replaced text — near parity on one-line touch-ups — and a compact patch language like @oh-my-pi/hashline can emit a lighter payload still (42–53% on the same session). The point is the right kind of edit call: no re-typing old code, and nothing for the model to track except two stable content addresses.

Correctness. Every resolved edit range is verified against the exact lines the model was shown. A stale, never-served, or ambiguous range is hard-rejected before anything is written, and the current range is echoed back as fresh anchors (reject-and-serve) — the retry needs no read.

A modern edit pattern for agents. Content-addressed anchors are line-number-agnostic: edit one part of a file and the hashes of the rest stay put, so chained edits need no re-reads. The model pins a line by what it is, not by where it used to sit.

How It Compares

	hashline edit	str_replace (Claude Code / Codex)	@oh-my-pi/hashline patch
Replaced text never echoed in the call	✅ 2 hashes only	❌ verbatim	✅ + rows only
Lines addressed by	content hash	text match	number + file-content tag
Verified against what the model saw	✅ every line	❌ first match wins	~ file version only
Stale file detected	✅ rejects, fresh anchors	❌ may match wrong spot	✅ tag mismatch → refuse or 3-way merge
Anchors survive edits above	✅ content-addressed	✅ content-based	❌ renumber + new tag
Chained edits without re-reads	✅ diff serves fresh anchors	~	~ via edit-response numbers
Unambiguous when text repeats	✅ boundary anchors verified	❌ first occurrence	~ position, unverified per line
Wrong-line edit never lands silently	✅ every line verified	❌ first match wins	~ possible in principle (tag checks version, not lines)
Block ops / registers / MV / REM	❌	❌	✅
One document per change	❌ per-edit call	❌ per-edit call	✅ multi-hunk patch
Runtime	✅ Node (dsh)	—	⚠️ Bun only
Undo	✅ persisted	❌	❌ not in scope

~ = occasionally / inconsistently. @oh-my-pi/hashline is a compact line-anchored patch language (npm, repo): [path#tag] headers bind each hunk to a full-file content hash, PUT N.=M: addresses lines by number, and every edit renumbers — take the next numbers and tag from the edit response or a fresh read.

Different jobs, same lineage. Both descend from the harness-problem insight that the model should never re-type old code. @oh-my-pi/hashline is a patch-language library — payload-light (42% saved per edit, 53% in a single batch document, see benchmark), with syntactic block ops (PUT N*:), registers, REM/MV, multi-hunk documents, a pluggable filesystem for any backend, and session-aware 3-way-merge recovery on stale tags. This plugin is a dsh tool pair: read hands the model 3-char content hashes, edit takes two of them, and every resolved line is verified against the served state — no line numbers to renumber, no tag to re-fetch, a wrong anchor can never land on the wrong line, and undo_last_edit survives restarts. Its trade-offs: a JSON envelope per edit costs a little payload, there are no block ops, and it lives inside dsh (Node) rather than as a standalone patcher (Bun). Pick hashline-the-library for a cross-backend patch format; pick hashline-the-tool for verified, content-addressed edits in your agent.

Correctness in edge cases

The token benchmark measures the payload the model emits — it assumes the model gets every address right, for free. Correctness is where the two hashline implementations actually diverge. These are the real failure modes from the harness-problem literature (wrong-line edits, drift, repeated text), and what each tool does when they hit:

Edge case	hashline edit (this plugin)	@oh-my-pi/hashline patch
Wrong address (off-by-one anchor / line number)	Impossible — anchors resolve to specific lines; every resolved line is verified against served state, rejected before anything is written	Possible — a wrong line number against a current tag applies silently at the wrong place; the tag proves the file version, never the lines
File changed on disk after the model's view	Hard reject + fresh anchors echoed (reject-and-serve); retry needs no read	Tag mismatch → refuse or best-effort 3-way merge onto unknown current content
An edit above shifts the file	Nothing shifts — anchors are content addresses; the diff serves fresh anchors	Every edit renumbers — “RE-GROUND AFTER EVERY EDIT” is the format's own #1 rule; the model carries the bookkeeping
Repeated / identical text	Per-line hashes are unique (collision-resolved); ambiguity → [E_AMBIGUOUS_ANCHOR]	Position-based, so repeats don't confuse it — but the position itself is unverified
Lines never shown to the model	[E_RANGE_UNSERVED] — hard reject with fresh anchors	Undisplayed hunks rejected — same reliance on the model knowing what it saw
Mid-expression / wrong block node	Irrelevant — any verified line range is valid	Grammar rules + PUT N*: node choice; mispointing (anchoring def orphans its decorator) silently lands wrong; no syntax check
Multi-edit batch fails mid-way	batch_edit — atomic, all-or-nothing; the failing item is echoed as fresh serves	Multi-section patches preflighted up front — also atomic

The 42–53% oh-my-pi payload saving is a lighter wire format; the table above is what that format asks the model to hold in its head instead — renumbering, tag-chasing, node choice — the exact component that fails most (46–51% patch-failure rates on replace-style edits). This plugin's 31% is the price of a contract where a wrong edit cannot land, and any rejection needs no re-read.

Benchmark

Measured on the same 103-line file with the same 12 replacements (8 single-line, 4 multi-line of 3/6/10/15 lines), tokenized with the pinned js-tiktoken cl100k_base. Three arms emit the same replacements: this plugin's edit (two 3-char anchors), a str_replace tool (old text echoed verbatim), and @oh-my-pi/hashline in both of its modes — one [path#tag] section per edit (seq) and one multi-hunk batch document (batch):

Criterion	hashline	str_replace	oh-my-pi seq / batch
Replaced text sent over the wire	✅ never	❌ every edit	✅ never
Output tokens saved (12-edit session)	✅ 31%	❌ 0%	✅ 42% / 53%
Multi-line range savings (3–15 lines)	✅ 29–47%	❌ 0%	✅ 40–53%
Effective cost at 5× output pricing	✅ ~1.4× less	❌ 1×	✅ ~1.7× / ~2.1× less
Ranges verified against served state	✅ 100%	❌ none	~ file version only
Line numbers the model must track	✅ none — content anchors	✅ none — text match	❌ renumber every edit
Deterministic, reproducible locally	✅ npm run benchmark	—	—

Reproducible

The numbers above are deterministic and you can reproduce them locally — npm run benchmark:

Scenario	Lines	hashline	str_replace	oh-my-pi seq	oh-my-pi batch
single-line ×8	1	309	324	241	—
multi-line ×4	3–15	393	691	349	—
TOTAL ×12		702	1015	590	480

Saved vs str_replace: hashline 313 (31%) · oh-my-pi per-edit 425 (42%) · oh-my-pi batch 535 (53%).

The script is deterministic by construction: a frozen corpus, a content-addressed edit script that self-checks (a reformatted corpus throws instead of silently changing what's measured), a pinned tokenizer, and oh-my-pi payloads validated against the package's published grammar before counting. Because everything is fixed, npm run benchmark gives everyone the same result — the numbers in this README are a snapshot of that run; regenerate, don't trust.

Scope & honesty. The benchmark measures request-payload tokens — what the model emits per edit call — with identical read traffic excluded (it cancels) and identical replacement text. It does not model transcription failure and retries, which is where the real-world gap is largest: the original harness-problem post reported a 61% output-token reduction and patch-failure drops from 46–51% to near zero after switching to anchored edits. It also does not model what a line-numbered format costs the model between calls — renumbering and re-fetching the file tag after every edit — nor block-op power, nor the Bun-vs-Node runtime difference, nor the fact that @oh-my-pi/hashline is a standalone patcher while this plugin is a dsh tool pair with read/edit/undo. Full methodology, the per-edit table, and the complete limitation list in benchmark/README.md. The correctness gap behind those numbers is spelled out above in Correctness in edge cases.

Tools

Tool	What it does
read	Returns a file with every line as HASH│content. Parameters: offset (1-based), limit. Paged output ends with [Showing lines N-M of T. Use offset=… to continue.]. Lines >200KB are shown as a marker with a sed hint — hash anchors need full lines.
edit	Replaces a range of lines by hash. path · remove_from · remove_to · replacement_text ("" deletes). Verifies every line of the resolved range against served state; [E_RANGE_STALE] / [E_RANGE_UNSERVED] / [E_RANGE_UNVERIFIED] reject-and-serve fresh anchors.
batch_edit	Up to 32 edits in one atomic call: { edits: [{ path?, remove_from, remove_to, replacement_text }, …] }. All-or-nothing; the failing item's range is echoed as fresh serves.
undo_last_edit	{ path } reverts the last hashline edit, only while the file still matches the stored post-edit content; survives restarts.

Error codes

Code	Meaning
[E_ACCESS]	File exists but is not readable/writable by the tool.
[E_AMBIGUOUS_ANCHOR]	A hash matches more than one current line; call read for fresh anchors.
[E_BAD_OP]	Range end precedes range start (autocorrected when the pair was reversed).
[E_BAD_REF]	remove_from/remove_to is not a bare 3-char hash.
[E_BAD_SHAPE]	Request/field shape is wrong (unknown fields, missing path, non-string text, …).
[E_BARE_HASH_PREFIX]	HASH│ prefix pasted into replacement_text (autocorrected).
[E_BATCH_ABORT]	A batch item failed; the whole batch was rejected, nothing written.
[E_FILE_TOO_LARGE]	File exceeds the hashline line ceiling; use write or another approach.
[E_INVALID_PATCH]	Diff-preview markers pasted into replacement_text (autocorrected).
[E_NOOP_LOOP]	The exact same edit keeps producing no change; resubmitting is rejected.
[E_NOT_FOUND]	File does not exist.
[E_NOT_OBSERVED]	The file has not been observed in this session (read-before-write policy); call read first.
[E_NOT_TEXT]	Path is a directory, binary, or non-UTF-8 file; hashline edits only text.
[E_RANGE_STALE]	A served line differs on disk since it was read; the range is echoed fresh.
[E_RANGE_UNSERVED]	The range includes lines never served to the model.
[E_RANGE_UNVERIFIED]	Boundary anchor cannot be verified against served state.
[E_STALE_ANCHOR]	Anchor(s) no longer resolve; call read for fresh anchors.
[E_UNDO_STALE]	Cannot undo: the file was modified (or deleted) after the edit.
[E_UNDO_UNAVAILABLE]	Undo history could not be persisted; the edit was not applied.
[E_WOULD_EMPTY]	An edit would empty a non-empty file; use write to clear it.

How It Replaces the Built-in Tools

dsh's tool registry resolves per scope: an agent sees agent → preset → global, and its own layer always wins. The built-in read/edit live on the agent-preset layer, so a plain global registration cannot replace them. This plugin:

1. Mounts as a host-plane Cordis plugin via its cordis.patch.yml bundle patch.
2. On agent/session-start, registers the hashline tools and the tool:read / tool:edit prompt sections on the agent's own scope layer — they shadow the preset's built-ins for that agent and unwind automatically when the agent is disposed.
3. Leaves the built-in write in place, but a scoped tools/post-execute listener appends the hashline auto-read to write results.

Store

Hash snapshots, served-state rows, and undo history live in one SQLite store co-located with the workspace being edited — one store per session cwd:

<workspace>/.dsh_better_edit/hash-store.sqlite

Parallel sessions in different workspaces keep separate stores (the session cwd is carried through each tool call), so one project's anchors and undo history never leak into another's. Outside a tool call (tests, previews) the store falls back to the shared DeepSeek Harness home ($DSH_HOME/plugins/dsh-better-edit/hash-store.sqlite).

A 7-day TTL prunes served rows; missing-file snapshots are pruned at startup. Corrupt stores are quarantined and rebuilt automatically. Moving to the per-workspace layout does not migrate earlier undo history from the shared home — treat any pre-0.1.2 undo entries as gone.

Project Structure

dsh-better-edit/
├── src/
│   ├── hashline/        # hash + served-state core (ported byte-for-byte from pi-hashline-edit-lsz)
│   ├── tool-read.ts     # read  — HASH│content, offset/limit paging
│   ├── tool-edit.ts     # edit  — range-by-hash, reject-and-serve
│   ├── tool-batch-edit.ts
│   ├── tool-undo.ts     # undo_last_edit
│   ├── sandbox.ts       # FsSandboxController mirror (sandbox_permissions/justification)
│   ├── write-hook.ts    # auto-read appended to write results
│   ├── served-store.ts  # per-workspace SQLite store (node:sqlite)
│   └── workspace.ts     # session-cwd AsyncLocalStorage carrier
├── benchmark/           # reproducible hashline-vs-str_replace-vs-oh-my-pi token benchmark
│   └── corpus/          # frozen 103-line fixture
├── test/                # 615 tests (ported + regression)
├── assets/              # logo + banner
├── cordis.patch.yml     # bundle patch
└── package.json         # dsh.bundle manifest

Development

npm install
npm run typecheck   # tsc --noEmit
npm test            # vitest run (615 tests)
npm run build       # tsc → lib/
npm run benchmark   # reproducible token-cost benchmark (benchmark/)

Releasing (tag-first)

npm run release -- 0.2.0                 # bump + CHANGELOG move + commit + tag + push → GitHub release
npm publish --registry https://registry.npmjs.org   # blocked until the version is tagged

npm run release bumps package.json/lockfile, moves the CHANGELOG [Unreleased] section to the version, commits, tags vX.Y.Z, and pushes — the tag push creates the GitHub release from the changelog. npm publish refuses to run until that tag exists (prepublishOnly gate), so every npm version is always already tagged and released.

The test suite is ported from pi-hashline-edit-lsz and drives the dsh tool builders directly over a local filesystem bridge.

Roadmap

Current state (0.1.6): 615 tests, per-workspace store, sandbox policy participation, the served-tail truncation fix, reproducible benchmark, EN + 中文 READMEs, published on npm.

Next

• Close or justify the gap vs @oh-my-pi/hashline (reference: ../oh-my-pi.md). The sibling patch language is payload-lighter — 42%/53% vs our 31% vs str_replace on the benchmark, because a bare patch document skips the JSON envelope we pay per call — and offers four abilities we do not support: syntactic block ops (PUT N*:), registers + REM/MV, one multi-hunk document per change, and a pluggable filesystem. The counterweight is correctness: its line numbers are unverified (a wrong number on a current tag lands silently), every edit renumbers, stale tags trigger best-effort 3-way merge instead of verification, and the grammar raises the model skill floor. Decide each ability reject-or-adopt on its own merits — the payload gap alone is not a reason to switch formats.
• Verify 0.1.6 live in a dsh session after the served-tail fix.
• Upstream the served-tail truncation fix to pi-hashline-edit-lsz / upstream (their upsertServed never truncates either).
• Re-check plugin wiring against the next dsh release (pinned to 0.1.0-rc.6; dsh is in developer preview and promises breaking changes).

Contributing

See CONTRIBUTING.md (or just open an issue). The most valuable contributions right now are more benchmark scenarios and edge-case tests for the served-state verification.

License

MIT License — see LICENSE for details. Ported from pi-hashline-edit-lsz (MIT), which itself carries the upstream copyrights of RimuruW and YuGiMob.

Acknowledgments

Hash-anchored editing descends from Can Bölük's The Harness Problem — the post that showed the harness, not the model, is the bottleneck, and that anchored edits beat search-and-replace. This project stands on the shoulders of:

• pi-hashline-edit by RimuruW — the original pi-coding-agent extension that introduced 3-character hashes and collision resolution.
• pi-hashline-edit-pro by YuGiMob — the hardened fork the hashline core here is ported from.
• pi-hashline-edit-lsz — the self-maintained fork this project tracks. The hashline core is ported byte-for-byte; the tool layer is rewritten on dsh's plugin API.

Related reading: Hash anchors + Myers diff + single-token anchors (dirac.run) (a design review of the O(S+R) → O(R) edit-call saving) and an independent hashline-vs-replace benchmark.

---

Star History

---

⭐ If hashline editing made your agent edit better, give it a star!