editbuffer

editbuffer is a small Python runtime for editing an LLM's pending text output before it is committed.

The public abstraction is selection-based editing, not file patches:

• select exact, contextual, fuzzy, block, or character-range targets;
• replace, insert before/after, or delete the selection;
• reject missing, ambiguous, invalid, and stale selections;
• inspect the current buffer and its audit history;
• commit the final artifact.

Install

pip install editbuffer

For local development:

python -m pip install -e .

Python 3.11+ is required. The core library has no runtime dependencies.

Quickstart

from editbuffer import EditBuffer, Selection

buf = EditBuffer()
buf.append('find . -name "*.py" -exec grep -n "TODO" {} ;')
buf.replace(Selection.exact("{} ;"), "{} \\;")

final = buf.commit()

Selections can also be dictionaries:

buf.replace(
    {
        "type": "context",
        "before": 'grep -n "TODO" ',
        "text": "{} ;",
        "after": "",
    },
    "{} \\;",
)

LLM tool-call format

{
  "op": "replace",
  "target": {
    "type": "context",
    "before": "grep -n \"TODO\" ",
    "text": "{} ;",
    "after": ""
  },
  "text": "{} \\;"
}

Pass the decoded object to buffer.apply(operation). Supported operations are append, replace, insert_before, insert_after, and delete.

Range selections use half-open character offsets and can reject stale edits:

{
  "op": "delete",
  "target": {
    "type": "range",
    "start": 10,
    "end": 20,
    "expected_version": 3
  }
}

Failure behavior

• no match raises TargetNotFoundError;
• multiple matches raise AmbiguousTargetError with candidate ranges;
• occurrence can explicitly choose a zero-based match;
• invalid ranges or operations raise InvalidOperationError;
• a mismatched expected_version raises StaleVersionError;
• validator failures do not mutate text or history;
• fuzzy matching is opt-in and rejects low-confidence or competing candidates.

Fuzzy and block selections

Fuzzy selection uses the standard library and records the accepted confidence:

buf.replace(
    Selection.fuzzy(
        "run integrtion tests",
        threshold=0.85,
        ambiguity_margin=0.05,
    ),
    "run unit tests",
)

It never runs as a fallback for exact/context selection. If no candidate meets the threshold, or two candidates are too close, FuzzyMatchError includes candidate ranges and scores.

Block selection requires an explicit ID. Fenced code blocks use:

```python editbuffer:id=setup
print("old")
```

Markdown regions use:

<!-- editbuffer:block summary -->
old content
<!-- /editbuffer:block -->

Selection.block("setup") selects the content inside the markers, preserving the fence or comments. Duplicate IDs are rejected as ambiguous.

Snapshots and rollback

Every successful edit stores an in-memory snapshot:

buf.append("draft")
version = buf.version
buf.replace(Selection.exact("draft"), "final")
buf.rollback(version)

Rollback restores the snapshot as a new audited version. Snapshots live only for the lifetime of the EditBuffer process.

Optional validators

from editbuffer import EditBuffer
from editbuffer.validators import valid_json, valid_shell

json_buffer = EditBuffer('{"ok": true}', validators=(valid_json,))
shell_buffer = EditBuffer("echo ok", validators=(valid_shell,))

valid_shell invokes the local POSIX sh -n parser without executing the command.

CLI

The CLI stores an operation log in a JSON state file:

editbuffer new /tmp/output.json --text "hello world"
editbuffer replace /tmp/output.json '{"type":"exact","text":"world"}' "there"
editbuffer view /tmp/output.json
editbuffer history /tmp/output.json
editbuffer rollback /tmp/output.json 1
editbuffer commit /tmp/output.json

Use editbuffer apply STATE OPERATION_JSON for the complete operation schema.

MCP / agent integration

Install the optional server:

pipx install 'editbuffer[mcp]'

Or install with uvx without a persistent environment:

uvx --from 'editbuffer[mcp]' editbuffer-mcp

Connect it to Codex:

codex mcp add editbuffer -- editbuffer-mcp

Codex supports local STDIO MCP servers configured with codex mcp add; use /mcp in the Codex terminal UI to confirm the server is active.

Claude Desktop and generic MCP client examples are in docs/mcp.md.

The server exposes:

• buffer_append
• buffer_list
• buffer_view
• buffer_edit
• buffer_replace
• buffer_insert_before
• buffer_insert_after
• buffer_delete
• buffer_history
• buffer_rollback
• buffer_commit
• tool_history
• tool_select

Buffers are in-memory and live for the MCP server process. The MCP layer calls the same core API and does not implement separate edit semantics.

Use the first-class selection tools for normal agent use:

{
  "buffer_id": "answer",
  "target": {"type": "exact", "text": "old"},
  "text": "new"
}

buffer_edit remains available for raw JSON operations.

MCP calls are recorded in SQLite-backed history. tool_history returns recent calls, newest first. tool_select creates a pending buffer from selectable content in a previous call so the model can repair it instead of regenerating it.

Examples

• examples/shell_repair.py
• examples/json_repair.py
• examples/chat_restructure.py

Scope

This project is not:

• an agent framework;
• a retrieval system or repo indexer;
• a full IDE;
• a replacement for coding agents;
• a generic output quality evaluator.

Persistent storage, atomic operation batches, TypeScript bindings, and syntax-aware block parsing are future work.

Development

PYTHONPATH=src python -m unittest discover -s tests -v

License

MIT