dsh-hashline-edittool

Most agent edit tools force the model to re-type the old code token-for-token before it can change anything. Those are output tokens, billed at roughly 5-6x the input rate, and they are also where agents fail most: 46-51% patch-format failure rates on several models with replace-style edits, plus the quiet danger of first-match-wins landing an edit on the wrong line when identical text appears in multiple places. dsh-hashline-edittool changes the paradigm at the root: every line in a file receives a unique variable-length Base62 content anchor (shortest-first, 2 characters covering 3,844 lines, growing only as the file demands), and an edit call carries just two anchors plus the replacement text. The old code is never re-typed, ever.

The tool's core guarantee is verification: every resolved edit range is checked strictly against the exact lines the model was shown. Stale, never-served, or ambiguous ranges are hard-rejected before anything is written, and the response carries fresh anchors so a retry needs no re-read. Anchors are session-stable — unchanged lines keep the same anchor across multiple edits, and the post-edit diff rows carry fresh markers that the next edit can copy directly, eliminating read-between-edit overhead entirely. Multiple hunks in a single call all resolve against the same original file snapshot; if any hunk fails the entire call is rejected atomically with zero writes, so there is no partial-apply-then-rollback concern. The plugin also ships a persisted undo_last_edit, dual JSON / plain-text output modes, and a per-preset guidance override system so the system-prompt sections for read, edit, undo, and grep can be tuned independently for each agent preset.

This plugin is best suited for long-session, structural-edit agent workflows inside DeepSeek Harness — especially when files are large, identical text patterns appear in multiple places, and an edit landing on the wrong line is unacceptable. For a single-line touch-up or creating a new file the gain over str_replace is marginal; but the moment you are doing multi-line replacements, batch hunks, or consecutive chained edits where lower token cost and higher hit-precision matter, the hashline anchor mechanism is a clear upgrade.