deepseek-harness-hashline

When a model edits the same file across multiple turns, line-number drift, mixed CRLF and LF endings, and invisible whitespace changes can silently break pure-text matching. The built-in read / edit tools struggle in multi-round, multi-hunk, atomicity-sensitive scenarios: a misremembered line number edits the wrong line, and a single whitespace mismatch fails the match entirely.

Hashline upgrades read results to a combined LINE:HASH anchor (line number plus short SHA-256 prefix), binding each edit to the exact line content the model most recently read rather than a mutable line number. Core capabilities include three anchored edit operations (set_line, replace_lines, insert_after), atomic batch writes where all anchors are validated first and overlapping edits are rejected upfront, file-version CAS to guard against concurrent overwrites, automatic recovery of malformed parameter shapes, UTF-8 BOM and mixed-EOL text fidelity, and native DSH read-block and diff-card result rendering. Enabling the replace-built-in-read-write toggle switches the agent-facing read / edit to strict anchored mode with seamless behavioral compatibility.

Ideal for developers driving model-based repeated edits of source code, configuration, or documentation within DeepSeek Harness, especially in multi-turn agent workflows, batch multi-hunk modifications, or headless and CI pipelines. The plugin is still experimental; interfaces may change, so validate thoroughly in an isolated environment before adopting in production.