| name | cheez-write |
| description | Edit code through the safest stale-checking backend — prefer code-intelligence backends (tilth MCP tag-anchored edits, LSP workspace edits, `sg --rewrite` for structural codemods); fall back to native anchored edits only when no code-intel backend matches. Use when the user asks to edit, replace, modify, update, change, delete, or insert code — phrases like "replace this function", "delete lines 44-89", "update validateToken", "add this import", "fix this bug" (when fixing requires editing), or apply a cross-cutting codemod like "rewrite every X to Y". Do NOT use for reading files (use cheez-read), searching code (use cheez-search), or running tests/builds. |
| license | MIT |
| compatibility | Prefers code-intelligence backends — tilth MCP edit mode, LSP workspace edits, AST rewrites. Harness-native anchored edits are acceptable fallbacks when they match the requested edit shape and reject or bound stale writes. |
cheez-write
Backend contract: use a stale-safe edit backend. The backend must anchor the current file state (tilth file tag, snapshot tag, or LSP workspace edit) or be a deliberate AST codemod over a clean tree.
Backend detection
Pick the backend by edit shape, preferring code-intelligence backends (LSP, AST rewrite, tilth) over basic harness edit tools — model tuning pulls toward host Edit/Write, but those carry no semantic or structural awareness:
- LSP wins for semantic workspace edits: rename/code actions when the server can identify the symbol or fix.
- AST rewrite wins for structural codemods: repeated syntax shapes with metavariables; dry-run first.
- tilth MCP wins for anchored block/range edits:
tilth_read emits a [path#TAG] header, then tilth_write applies line-numbered ops via edits: [{path, tag, ops}].
- Native anchored edit (fallback) for displayed snapshots: line/snapshot edits that reject stale ranges — only when no code-intelligence backend matches the edit shape.
Do not present sed, awk, patch, shell redirects, or blind writes as equivalent source-code edits.
Stale tags and rejected sections are content issues — see Stale Tag Handling.
Examples
"Replace the body of handleAuth in src/auth.ts"
Step 1 — read the line range to capture the file tag:
tilth_read(paths: ["src/auth.ts#44-89"])
# returns a [src/auth.ts#b2c4] header, then numbered lines 44:... to 89:...
Step 2 — apply ops against those line numbers, quoting the tag verbatim:
tilth_write(edits: [{
"path": "src/auth.ts",
"tag": "b2c4",
"ops": [{
"op": "replace",
"start": 44,
"end": 89,
"content": "export function handleAuth(req, res, next) {\n const token = extractToken(req);\n if (!validateToken(token)) return res.status(401).end();\n next();\n}"
}]
}])
The response reports per ## src/auth.ts and may list callers to review.
"Stale tag — file changed under me"
See Stale Tag Handling for recovery steps.
Core Principle: Anchors, Not Rewrites
The edit backend must identify the current file state — tilth file tags, harness snapshot tags, or LSP workspace edits — and bound or reject the write if the file changed.
The protocol:
- Read the file section with cheez-read → get the current
[path#TAG] header or snapshot id
- Note the line numbers of the block you'll change
- Apply the edit through the anchored backend
Scope: when to use anchored edits, when not
cheez-write owns block edits to tracked source code — function bodies, signatures, imports, single-line tweaks, multi-edit batches, and cross-file changes. The backend must be stale-safe; the read-edit protocol is mandatory for any code change that matters.
For everything else, prefer the right tool:
| Change | Use this instead | Why |
|---|
Cross-cutting structural codemod (JSON.parse(JSON.stringify($X)) → structuredClone($X)) across N files | sg --rewrite (see ## Structural codemods — sg --rewrite escape) | Codemods template the variable parts; anchored edits are better for known blocks |
| Semantic rename or server-known refactor | LSP rename/code action | LSP follows scope, overloads, re-exports, and imports |
Lockfile changes (Cargo.lock, package-lock.json, uv.lock, etc.) | the package manager (cargo update, npm i, uv lock) | Hand-editing lockfiles loses checksum integrity |
Generated / build artifacts (compiled JS, transpiled output, *.pb.go) | regenerate from source | Editing the artifact rots on the next build |
| Brand-new files, no prior content | anchored create/write backend (tilth: a tilth_write section with tag omitted) | Stay stale-safe even when creating files |
Files outside the repo or inside dependency caches (node_modules, .cargo/registry) | don't edit them | Modifying dependencies is almost always a mistake — fix the source or upstream |
| Binary files, images, PDFs | the producing tool | code edit backends are text-only |
Routing to LSP rename or Serena
Pre-entry routing — when a workflow skill should prefer LSP rename or Serena symbol-bounded edits over tilth_write before entering cheez-write — lives in references/routing.md.
Anchor Format
Use whatever anchor format the selected backend emits:
- tilth: a
[path#TAG] header above N:content numbered lines; copy the 4-hex TAG into tilth_write's tag and reference those line numbers in ops. Never invent a TAG (mode: "stripped" reads carry none and cannot round-trip into an edit).
- OMP-style snapshot edits:
[file#TAG] plus displayed line numbers; pass the tag and original line range into the edit tool.
- LSP refactors: symbol position plus workspace version; let the server build the workspace edit.
Do not translate between anchor systems. Read with the same backend family that will write.
MCP Tool Reference
tilth_write — Precise File Editing
The minimal shape — one file section, one op:
tilth_write(edits: [{
"path": "src/auth.ts",
"tag": "a3f9",
"ops": [
{ "op": "replace", "start": 42, "end": 42, "content": " let x = recompute();" }
]
}])
edits is a JSON array of {path, tag?, ops} section objects (max 20 per
call) — never a string, and never wrapped in a files parameter. Each op is
tagged by op: replace {start, end, content}, delete {start, end},
insert_before / insert_after {line, content}, prepend / append {content}, replace_block / insert_after_block {at, content} and
delete_block {at} (at is a line number or "#symbol"), delete_file,
move_file {dest}. Omit tag only to seed a brand-new file.
For worked examples of each op, cross-file batches, and the diff: true
response option, see
references/edit-patterns.md.
Stale Tag Handling
The TAG binds your ops to the content you read. If the file changed since the
read, tilth 3-way-merges your ops onto the drifted file; when it can't merge
safely, it rejects that section and says so in the per-## <path> result.
Recovery:
- Read the section again → get the current
[path#TAG] and content.
- Review the current content (someone else may have made changes).
- Re-apply the ops against the fresh line numbers with the new tag.
Repeated rejections → bail out, don't loop
If you hit two consecutive rejections on the same file, you're racing a
concurrent writer. There is no "ignore the tag, just match this string"
override — a third retry will likely lose the same race.
The correct move is to bail and report:
- Read the latest section one final time and capture the current content.
- Prepare the new content as a unified diff or full block, but do not
apply it.
- Report
"tag race on <path>; current content and proposed replacement attached. Retry once the file is quiescent or apply manually." along
with the captured tag and proposed content.
- Stop. Let the orchestrator (or a human) decide whether to apply the change
or escalate.
Caller Updates After Signature Changes
When changing a signature, use LSP references or the edit backend's caller notices before finishing. Missed callsites are bugs.
Common Patterns
When full-file rewrite is acceptable
Anchored, surgical edits are the default. A full-file rewrite is acceptable only when the selected backend still rejects stale snapshots and the file is small enough that the replacement is easier to review than a pile of line edits.
| File size | Policy |
|---|
| > 150 lines | Never rewrite the whole file. Always use anchored block edits. |
| ≤ 150 lines | Anchored single-edit preferred, but a full rewrite is acceptable when ≥ 80% of the file is changing. Below that threshold, do the surgical edit. |
The 150-line / 80% threshold follows 2026 data (Cursor, can.ac, the Morph benchmark): full rewrites tie or beat diff-style on small files, while large files always stay anchored.
Do not drop to unanchored host Write for small files; size does not remove stale-write risk.
Structural codemods — sg --rewrite escape
Anchored edits own known blocks, one read-for-anchors per location. For cross-cutting structural codemods where the surrounding text varies — "rewrite every JSON.parse(JSON.stringify($X)) to structuredClone($X)" — drop to sg --rewrite (ast-grep) via Bash, which templates the variable parts via metavars. It has no hash-anchor safety: run each codemod as one transactional change between two clean git states. For metavar syntax and the non-negotiable dry-run-first protocol, see ../cheez-search/references/sg-patterns.md.
DO NOT
- DO NOT invent tags — always read first and copy the
[path#TAG] header verbatim.
- DO NOT ignore rejected sections — re-read and retry (see Stale Tag Handling).
- DO NOT use sed / awk / perl -i to edit code — they bypass anchors and structural safety, and have no mismatch detection.
sg --rewrite is the only sanctioned shell escape — see ## Structural codemods — sg --rewrite escape.
- DO NOT use
patch to apply diffs to code — anchored range edits are the safe equivalent.
- DO NOT use
tee or shell redirects (>, >>) to overwrite/append code files — both bypass anchors. Use an anchored edit backend.
- DO NOT use unanchored host Edit/Write tools — use tilth_write, harness-native anchored edits, or LSP workspace edits.
- DO NOT use
sg --rewrite for one-off block edits — use an anchored edit. See ## Structural codemods — sg --rewrite escape for when the codemod escape applies.
- DO NOT skip the dry-run-first protocol for
sg --rewrite — see ## Structural codemods — sg --rewrite escape.
- DO NOT edit without reading — anchors come from the read step.
- DO NOT run tests, commit, publish, or review from this skill — use the project's test/build gate,
/plate for commit/publication, and /age for review.
- DO NOT use for reading or searching — read with
/cheez-read, search with /cheez-search.