@nelsonlongxiang/dsh-tool-hash-line-edit

A DeepSeek Harness (DSH) host plugin that adds the model-facing hash_line_edit tool: line-addressed file edits validated by a [path#TAG] content tag, with a built-in [path#????] probe that returns the complete line-numbered content and its tag.

What it does

The plugin registers one tool (hash_line_edit) plus its system-prompt guidance into the host composition:

• Probe — patch [<path>#????] (header only) reads the target file and returns its complete line-numbered content with the [path#TAG] anchor line (rendered text); the canonical value also carries the four-hex tag field, so code-dispatch callers that see the value rather than the render can quote the anchor. The tag uses the same xxHash32 low-word algorithm as the original hashline tool (line1\nline2\nline3\n → 34E9; trailing [ \t\r] per line is normalized away — different file content yields a correspondingly different tag). The probe records the read observation, so the following edit passes the observation policy's read-before-edit gate without any other read tool. Files above 2 MiB are refused (FS_TOO_LARGE) — hash-anchored editing targets small, fully quoted files.
• Edit — a patch carries the header [path#TAG] on its own line, followed by operations:
  • PUT N.=M: then + payload lines — replace lines N–M
  • CUT N.=M — delete lines N–M
  • PUT <N: then + payload lines — insert before line N
  • PUT >N: / PUT >$: then + payload lines — insert after line N / append at EOF
• Execution parses the patch, verifies the tag against current file contents (stale_tag on mismatch → re-probe), resolves per-call sandbox policy, obtains the fs/edit-intent version guard from the mounted observation policy, and commits through ctx.fs.writeText — so sandbox confinement, escalation (sandbox_permissions/justification under a confining backend), read-before-edit, and stale-version protection all behave like the built-in filesystem tools.
• Results carry before/after text and diff-card presentation metadata; successful writes emit fs/observed.

Coordinates are 1-based and refer to the complete pre-edit probe result. Overlapping replace/delete ranges and malformed hunks are rejected before any mutation. New files are not supported — use the built-in write tool.

Requirements

The profile must already provide the standard filesystem stack: a ctx.fs provider (e.g. dsh-fs-local) and, for guarded mutations, dsh-fs-observation-policy. The plugin injects tools, fs, and systemPrompt.

Install

npx -p @deepseek-ai/dsh dsh plugin --profile <name> add @nelsonlongxiang/dsh-tool-hash-line-edit

Restart the profile after installing. The plugin is host-only (no client face).

Config

| Key | Default | Meaning | |---|---|---| | enabled | true | Register the tool and its prompt section; false contributes nothing. |

Development

pnpm install
pnpm verify   # typecheck + tests + build

Tests run the plugin through a real Cordis composition (in-memory ctx.fs provider plus the observation policy) and pin the tag vector, grammar, guarded-mutation, and stale-tag behavior.

scripts/inspect-transcript.mjs is a local development/diagnostic helper: it decodes a DSH .jsonl.zstd session log and prints events mentioning the tool (smoke-test evidence extraction). It is intentionally not shipped with the package — files only exports lib, cordis.patch.yml, and this README.

Origin

Extracted from the DeepSeek Harness monorepo's packages/fs/tool-fs hash_line_edit implementation (PR #21) into an independently installable plugin; tag compatibility with the original ai-native-expect-agents hashline tool is preserved.