dsh-better-edit

What is dsh-better-edit? A high-precision file editing plugin for DeepSeek Harness (dsh) that replaces volatile line numbers and token-wasting code echoes with immutable, content-addressed 3-character line hashes (szJ│code).

Core Philosophy: Local compute is free; the model's context window is the most precious resource. By shifting verification, snapshotting, and alignment to the host, dsh-better-edit slashes output tokens by 40–60%, auto-rebases external file drift (e.g., Prettier, Git), and eliminates silent miswrites without forcing full-file re-reads.

Why You Need It

The 3 Fatal Editing Traps of Autonomous Coding Agents

File editing is the #1 point of failure for autonomous agents. Traditional tools break down in three distinct ways:

Core Pillars

1. 🪙 Token Economics (40–60% Context Savings)

• $O(R)$ Edit Payloads: The model emits only { "path": "...", "edits": [["a1b", "c3d", "..."]] }, never echoing existing code.
• Self-Serving Diffs: Every applied edit returns fresh anchors in the post-edit diff — zero re-read roundtrips to chain edits.
• Zero-Token Auto-Rebase: Non-conflicting exterior shifts resolve locally without agent intervention — 0 tokens, 0 retries.
• Atomic Multi-Item Batches: Apply up to 32 same-file edits in one tool call; overlapping spans abort atomically ([E_BATCH_ABORT]) before touching disk.

2. 🛡️ Resistance to External Writes (Drift & Concurrency)

• Auto-Formatter Immunity: Strips ASCII whitespace before hashing. Prettier, Black, and ESLint format-on-save passes never rotate anchors.
• Exterior Shift Auto-Rebase: External edits, git checkouts, or background processes outside the edit span rebase seamlessly without agent intervention.
• Fail-Closed Reject-and-Serve: Contested interior spans fail closed without disk corruption and immediately return fresh on-disk rows in the error ([E_STALE_RANGE], [E_UNSERVED_RANGE]) — recovering in exactly 1 turn.
• Session-Keyed Leases: Leases are isolated per session, preventing cross-agent race conditions or state pollution.

3. 🎯 Zero Silent Miswrites

• Decoupled Line Identity: Lines are verified against served snapshot lineage, not ephemeral line coordinates.
• Collision-Free Anchors: Coprime bitset probing ensures duplicate lines in a file receive distinct, unambiguous 3-character hashes.
• No Heuristic Guessing: Retires fuzzy matching. If an anchor cannot be unambiguously resolved, it fails closed safely.
• Persisted Undo: undo_last_edit restores exact file content, BOM, line endings, and original anchors, persisting across session restarts.

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

3 calls vs 6 · -55.8% tokens · 23/23 correctness. Tested on realistic external-drift refactoring against OMP wrapper. Payload numbers are deterministic — see Benchmark.

Quick Start — install to verified edit in 30s

Install (pick one)

npx @deepseek-ai/dsh plugin --profile web add github:Rianico/dsh-better-edit   # from github
npx @deepseek-ai/dsh plugin --profile web add dsh-better-edit                 # from npm
npx @deepseek-ai/dsh plugin --profile web add /path/to/dsh-better-edit       # local

No config. Next session runs with hashline tools. Verify:

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

dsh 0.2.0+ is supported. Registration uses agent/created, path resolution follows the session's live working directory (workingDirectory, with the session header as the pre-0.2.1 fallback), and tool-guidance order comes from systemPrompt.getSectionOrder. There is no peer range and no runtime version probing: compatibility comes from that stable surface plus optional-service fallbacks, so a composition that omits workingDirectory or agentPresets keeps working.

See it work

read serves HASH│content — the hash is the address:

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

edit by hashes — always lands where you meant:

{ "path": "src/main.ts", "edits": [["szJ", "szJ", "  console.log('hi');"]] }

Returns a diff with fresh anchors — next edit needs no read:

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

Position-free in one line: read 1..5 → insert @0 → edit 10..12 still verifies 10..12 (resist mode). Multi-session honesty: A:10..12+1 shifts B:20..30→21..31 → B passes (drift notice); B:12..13 overlapping A → E_STALE_RANGE + fresh rows, one retry.

Batch atomically — one edit, up to 32 same-file ranges:

{
  "path": "src/main.ts",
  "edits": [
    ["a1b", "a1b", "new line 1\n"],
    ["c3d", "c3d", "new line 2"]
  ]
}

One fails → none write ([E_BATCH_ABORT]).

[!TIP] Want proof before you install? Upstream 23/23 battery runs no LLM — stale edits are rejected every run. Same algorithm.

Configuration

Tenancy and prompt guidance declare once, read at agent/created, no code change.

Store central by default $DSH_HOME/plugins/dsh-better-edit/runtime/<name>-<hash8>/ (ls-readable + .wsPath sidecar). DBs are disposable caches — rm -rf runtime/<name>-<hash8>/ is safe, rebuilt on next read.

# $DSH_HOME/plugins/dsh-better-edit/config.yaml
storeDir: central # central | workspace | /abs
autoGitignore: false
undo_ttl_s: 604800 # 7d, -1 forever
storeMaxAgeS: 2592000 # 30d janitor
storeMaxTotalBytes: 524288000 # 500 MB LRU

Env overrides yaml (DSH_BETTER_EDIT_STORE_DIR, DSH_BETTER_EDIT_AUTO_GITIGNORE).

Guidance per preset — tool:read / tool:edit / tool:undo_last_edit are plain markdown per preset at $DSH_HOME/plugins/dsh-better-edit/<preset>/<section>.md. The default order is the harness's own tool slot (systemPrompt.getSectionOrder: 1100 read, 1300 edit, 1310 undo), so a band change upstream moves the sections with it; front-matter pins a custom order. The seeded roster is the shipped presets — standard, ptc, minimal, cordis. Delete or empty a file → default re-seeds at next boot; keep a --- fence to blank on purpose.

Why Hashline

Verified against what was served. Every resolved line checked against read/diff/rejection rows. Stale or unseen → [E_STALE_RANGE]/[E_UNSERVED_RANGE] + fresh HASH│content, retry needs no read. Session-scoped — sub-agent serves never validate main edits.

Content-addressed. canon(line) strips ASCII whitespace, xxh32 → 62³=238,328 anchors. Re-inserting identical text keeps its hash; prettier/eslint --fix between edits doesn't invalidate. Unique by bitset probing — }/import repeats never collide; cap 238,328 lines ([E_LARGE_FILE]).

No loop, no ritual. No-op → No changes made; same no-op ×3 → [E_NOOP_LOOP]. Diff/echo/rejection rows count as serves — read is recovery, not ritual.

Token economics

Envelope change: hoist path, edits:[[from,to,text]], never repeat old_string.

Percent vs str_replace. External row pinned corpus, cl100k_base; local npm run benchmark in upstream.

Single stochastic run, opencode-go/gpt-5.6-luna high. Artifact.

Scope & honesty. Payload deterministic; practical run stochastic. We measure payload + round-trips, not throughput. Retries are where the gap is largest — see edge cases.

Tools

write stays, but refuses an exact HASH│ echo for same session/path/line before dispatch.

Error codes

Full list in src/ — every rejection echoes fresh rows, no read needed.

Comparison

~ partial, — n/a. Same lineage — patch library vs dsh tool pair; pick by seam.

Edge cases: wrong anchor impossible (verified), disk drift → reject+serve, shift above → nothing moves, repeats → unique/ambiguous, unseen → reject, batch → atomic. See upstream benchmarks.

Battery: 23/23 tool, 10/10 library (upstream npm run eval, same algorithm).

How Anchors Work

canon(line) strips ASCII whitespace → xxHash32 → A-Za-z0-9 3-char (62³). Stable across prettier; Unicode/strings stay significant except ASCII whitespace inside strings (linter-only). stride=62²+62+1 probes bitset → unique; cap 238,328. Store hash-store.sqlite per workspace (central, honoring XDG_CONFIG_HOME); 7-day served TTL, janitor storeMaxAgeS/LRU + wal_checkpoint.

How It Replaces Built-ins

dsh resolves agent → preset → global; built-ins live on preset. Plugin via cordis.patch.yml: at agent/created registers read/edit on agent layer (shadows, auto-unwinds); write stays with pre-execute guard + post-execute auto-read.

Project Structure

dsh-better-edit/
├── src/hashline/     # hash + served core
├── src/tool-*.ts     # read / edit / undo
├── src/served-store.ts # SQLite store
├── benchmark/corpus/ # 103-line fixture
├── test/             # 108 files, 1222 tests
├── assets/           # logo + banner
└── cordis.patch.yml

Development

pnpm install
pnpm typecheck   # tsc --noEmit
pnpm test        # vitest run
pnpm benchmark   # hash probe + session envelope (reads/retries/tokens)

Benchmark

103-line file, 12 replacements (8×1 + 4×3/6/10/15), cl100k_base. hashline vs str_replace vs oh-my-pi seq/batch. Upstream is source of truth — same algorithm byte-for-byte.

Saved 313 (31%). Reproduce: upstream npm run benchmark. See pi-better-edit/benchmark/README.md.

Scope & honesty. Benchmark is request-payload tokens (reads cancel, replacement text identical). No transcription-failure model — real gap larger; see edge cases.

Roadmap

Current 1.0.0: the v2 line-identity architecture, per-session (session,path) store, 1648 tests, dsh 0.2.1-rc.2.

• Keep benchmark/run.mjs in sync with ADR-0013 (reads/retries/tokens per session)
• Re-check wiring vs next dsh (pinned 0.2.1-rc.2)
• README.zh.md parity

Contributing

See CONTRIBUTING.md. Most valuable: more served-state edge-case tests.

License

MIT — see LICENSE.

Acknowledgments

From Can Bölük's The Harness Problem. Thanks to pi-hashline-edit, pi-hashline-edit-pro, pi-better-edit, @oh-my-pi/hashline. Reading: hash-anchors.

Star History