Greptile Summary

This PR adds a wrapEditToolWithFuzzyMatchSuggestions wrapper that enriches "oldText not found" errors with the closest matching regions from the file, giving the model actionable context to self-correct in one shot instead of retrying blindly or re-reading the whole file. The wrapper is applied to both createSandboxedEditTool and createHostWorkspaceEditTool, with 24 tests covering core algorithm correctness and integration scenarios.

Issues identified:

1. Pre-filter defeats its own purpose for indentation mismatches — The 8-char substring check includes leading whitespace, so when oldText has different indentation than the file (tabs vs spaces, or different depth), every line fails the pre-filter silently. This is the single most common reason oldText fails to match, meaning the feature would produce no suggestions in its most important use case.

2. Full file read before the size cap check — fs.readFile loads the entire file into memory before content.length > FUZZY_MAX_FILE_SIZE_BYTES rejects it. A large binary or generated file (e.g. a 50 MB bundle) will be fully buffered on every failed edit call. Using fs.stat() before reading would honour the intended performance cap.

3. Unexpected FS errors leak from the fallback read — If fs.readFile throws inside the inner try block (race condition, permission change), the FS error is rethrown to the agent instead of the original "not found" error, which is strictly more useful and less confusing.

Confidence Score: 2/5

• The feature is functionally safe (failures degrade gracefully to the original error), but three implementation bugs substantially reduce its value and violate stated performance guarantees.
• The feature is well-intentioned and well-structured, with comprehensive test coverage. However, it contains three functional issues that collectively undermine its core value proposition: (1) the indentation mismatch bug means the feature silently produces no suggestions in the most common real-world failure case; (2) the full-file-read-before-size-check violates the performance cap described in the PR; (3) the error handling edge case surfaces confusing filesystem errors instead of the original context. These should be addressed before merging.
• src/agents/pi-tools.host-edit.ts — the pre-filter logic (lines 152–169), the size check ordering (lines 303–321), and error handling (lines 327–332) all need fixes before the feature behaves as described.

_{Last reviewed commit: 52f018b}

Addressed all 4 review items in b45f11d:

1. P1 sandbox path: Removed fuzzy wrapper from createSandboxedEditTool — the wrapper uses host fs.readFile/fs.stat which can't access sandbox-only paths
2. Large file perf: Now uses fs.stat() to check size before reading, so 50MB bundles never load into memory
3. Indentation pre-filter: Added trimStart() before 8-char substring check so whitespace-only differences don't defeat the filter
4. FS error fallback: Changed inner catch to only surface errors with cause === err (our enriched throws), all unexpected FS errors fall back to the original "not found" error

24/24 tests pass, lint clean.

Fixed P1 prefilter issue in 73d0392:

• Extracted hasSharedPreFilterChunk() helper — now samples 3 chunks (start/middle/end) across each line instead of only the first 8 chars, so prefix renames (const→let) are no longer filtered out
• Single-line oldText skips prefilter entirely — most common edit failure case now always reaches LCS scoring
• Added 2 regression tests: multi-line prefix change + single-line rename

26/26 tests pass, lint clean.

Fixed P1 prefilter issue in 73d0392: Extracted hasSharedPreFilterChunk() that samples prefix/middle/suffix 8-char chunks instead of only the first 8 chars. Renamed leading tokens (const → let) now survive the prefilter. Added 2 regression tests (26/26 pass).

Fixed P2 short-line prefilter in e5766d1: Short lines (<8 chars) now always pass the prefilter — they're cheap to LCS-score anyway, and the old includes() fallback was too strict for single-char differences (let x vs const x). Added regression test. 27/27 tests pass.

Closing to reduce active PR count (10-PR limit). Will reopen if queue clears. The branch and code are preserved.