vibe-fim 0.4.0

Bounded AI code edits via Codestral fill-in-the-middle: rewrite only the region between two anchors, the rest stays byte-identical.

vibe-fim

Bounded AI code edits via Codestral fill-in-the-middle. Rewrite only the region between two anchors — the rest of the file comes out byte-for-byte identical, by construction.

---

Why this exists

Most coding agents edit by regenerating a block and pasting it back. The model re-emits your whole function (or file) token by token, so it can silently change things you never asked it to touch — reflow a comment, drop a trailing newline, "fix" an unrelated line, normalise quotes.

Fill-in-the-middle (FIM) makes that structurally impossible. You hand the model a verbatim prefix and suffix; it only writes the middle. So the edit is provably confined to the region between two anchors. Everything outside is never sent to be rewritten — it cannot drift.

vibe-fim wraps Mistral's Codestral FIM in a one-call surgical-patch primitive. The bytes outside the region are the original slices, concatenated unchanged — so bounded scope is structural. On top of that it runs the checks that can actually fail (below), so a bad completion is caught, not shipped.

No agent CLI exposes a verified bounded-edit primitive. Claude Code and Codex edit by regenerating blocks; plain vibe has no FIM patch op. Codestral's FIM is what makes a provably-bounded edit possible, and vibe-fim turns it into one an agent can call and check.

Install

pip install vibe-fim
export MISTRAL_API_KEY=...   # from https://console.mistral.ai

Zero runtime dependencies (stdlib urllib only).

CLI

# Rewrite only what's between the two anchors; show the diff without writing.
vibe-fim patch --file units.py \
  --before "def to_celsius(f):" \
  --after  $'\n\ndef to_fahrenheit' \
  --hint   "round the result to 1 decimal place" \
  --dry-run

The scope report goes to stderr:

[BOUNDED + PARSES] edit confined between anchors.
  frozen prefix: 142 chars (unchanged)
  frozen suffix: 98 chars (unchanged)
  region: 31 chars -> 39 chars

Drop --dry-run to write the file in place. If the patched file no longer parses, the CLI refuses to write (exit 3) unless you pass --allow-broken.

Many edits, one transaction

patch-many applies N bounded edits to one file as a single all-or-nothing transaction — the thing the FIM ecosystem (all autocomplete) does not do. Every segment between and around the regions stays byte-identical, the assembled file is parse-checked once, and the whole result is written or nothing is. A model that drifts in one region cannot corrupt the others; a result that doesn't parse is rejected as a unit.

vibe-fim patch-many --file app.py --json --edits '[
  {"before": "def to_celsius(f):", "after": "\ndef to_fahrenheit", "hint": "round to 1dp"},
  {"before": "def format_row(f):", "after": "\nTABLE_FOOTER",      "hint": "pad to 6 cols"}
]'
# {"bounded": true, "parses": true, "regions": 2, ... "wrote": true}

Edits must be in file order and non-overlapping; an ambiguous/missing anchor, an overlap, or a non-parsing result rejects the whole transaction (nothing is applied). from vibe_fim import patch_regions for the same as a library call.

Grow tests, provably additive

grow-tests uses FIM to append new test_* functions and proves the change is additive — your green tests stay green by execution, not by hope:

vibe-fim grow-tests --file tests/test_app.py --focus "edge cases" --json

A four-part gate, all-or-nothing (nothing is written unless all hold):

1. byte-identical prefix — the original file is prepended verbatim, so every pre-existing test is the same source (not regenerated-and-hopefully-equal);
2. parses — the assembled file still parses;
3. node-ids monotonic — a real pytest run collects every old test id plus the new ones (nothing renamed, removed, or shadowed by a duplicate name);
4. green stays green — each pre-existing test keeps the outcome it had on the original file, and each new test actually executed.

The pytest run is the proof (via the builtin --junit-xml channel — no plugin dep). Exit 4 if the additive proof fails; exit 3 if it does not parse. Generated tests are candidates — review them for meaning: the gate proves they are additive and that they run, not that they are good tests. A live Codestral run, replayable offline, is committed at examples/GROW_TESTS_RECEIPT.md.

Agent-callable

An agent shells out and consumes a stable JSON contract — no screen-scraping:

vibe-fim patch --stdin --before "…" --after "…" --name app.py --json
# {"bounded": true, "parses": true, "scope_report": "...", "unified_diff": "...",
#  "frozen_prefix_chars": 142, "frozen_suffix_chars": 98, "trimmed_overlap": 0,
#  "new_middle": "...", "wrote": false}

--stdin/--text pass the source inline (no disk round-trip); on error the record is {"bounded": false, "error": "..."} with a non-zero exit, so a loop can branch. There is also an MCP server — see Wiring into an agent below.

Library

from vibe_fim import patch_region

result = patch_region(
    source_code,
    before="def to_celsius(f):",   # frozen prefix ends right after this
    after="\n\ndef to_fahrenheit", # frozen suffix starts right at this
    hint="round the result to 1 decimal place",
)

assert result.bounded                         # structural invariant
assert source_code.startswith(result.prefix)  # prefix is byte-identical
assert result.parses is not False             # patched file still parses
new_source = result.text

The completion backend is injectable, so the bounded-edit logic is fully testable with no network:

patch_region(src, before, after, fim_fn=lambda prefix, suffix: "    return 42")

Anchors must be unique in the file. A missing or ambiguous anchor raises FimError instead of guessing — surgical edits should never be applied to the wrong place.

Verification that can fail

Bounded scope is structural (the prefix/suffix are the original slices). The value is in the checks that a bad completion can actually trip:

• Boundary echo. FIM models sometimes re-emit the start of the suffix (or end of the prefix) inside the middle — duplicating code, on the same line, not just across a line break. patch_region detects and strips a real re-emission (a run of ≥4 non-whitespace chars) while leaving a coincidental shared } or a boundary newline alone. result.trimmed_overlap reports what was removed.
• Truncation. A completion cut off at max_tokens is an incomplete middle. finish_reason == "length" raises instead of shipping a half-written edit.
• Parse gate. For languages we can parse with the stdlib (Python via ast), the patched file is re-parsed; result.parses is True/False/None, the report shows [BOUNDED + PARSES] / [BOUNDED + DOES NOT PARSE], and the CLI refuses to write a non-parsing result by default.

The guarantee, and the drift — demonstrated

Two separate claims, kept separate:

• The guarantee is structural: surgical FIM changes 0 lines outside the target, by construction, because the frozen regions are never regenerated.
• The drift is measured. examples/bench.py replays a committed Codestral transcript (examples/fixtures/) and counts collateral with one newline-aware counter for both approaches:

surgical FIM      : 0 lines changed outside target
block regeneration: 2 lines changed outside target   # one source line: its trailing newline dropped (counted as a -/+ diff pair)

It runs offline by default (no key, same numbers for every reviewer); pass --live to re-capture from the API. The block-regen baseline produced the correct body and still silently altered the file outside the edit.

python examples/bench.py            # offline replay of the committed fixture
MISTRAL_API_KEY=... python examples/bench.py --live

Wiring into an agent

vibe-fim patch --json is the stable contract for shelling out. For a native tool, install the optional MCP server and point any MCP client at it:

pip install "vibe-fim[mcp]"
vibe-fim-mcp          # stdio MCP server exposing one tool: surgical_patch

The tool returns the unified diff plus the scope report, so the agent sees the bounded proof — not just a success flag. (The base package stays zero-dependency; the mcp SDK is pulled in only by the [mcp] extra.)

How it works

1. Locate the unique before and after anchors.
2. prefix = text[:end_of_before], suffix = text[start_of_after:].
3. Send (prefix, suffix) to Codestral FIM → middle (de-duplicated, truncation-checked).
4. new_text = prefix + middle + suffix; the bytes outside the region are the same slices, concatenated unchanged.
5. Run the parse gate and surface the result.

License

MIT © Guillaume Vele