feat(jsonrpc): clearer errors for malformed JSON and invalid params

[RafaelGranza]

Context:

Related to Issue #3581

Before, a malformed JSON request returned only the raw decoder message.
Errors now point at what's wrong.

Also, long lines are truncated to 80 cols, so the marker stays aligned.

Errors it Covers

• syntax errors
• unexpected EOF
• type mismatches
• trailing commas.
• missing non-optional param field
• missing/unexpected params in list

Examples:

Errors ate python styled.

{"id":@}
      ^
unexpected '@', expected a value [line 1, column 7]

...xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" z}
                                                                           ^
unexpected 'z', expected ',' or '}' [line 1, column 230]

{"jsonrpc": 5, "method": "x", "id": 1}
            ^
field "jsonrpc" should be string, got number [line 1, column 13]

Benchmark:

We no longer copy the whole stream, we only keep the last 512 bytes read. So the overhead stays roughly constant instead of growing with the request size.

Large requests, main (no buffering) vs this PR:

	main (no copy)	this PR	delta
time 1MB	~14.9 ms/op	~15.1 ms/op	noise
time 10MB	~151 ms/op	~153 ms/op	noise
mem 1MB	~9.45 MB/op	~9.42 MB/op	noise
mem 10MB	~77.7 MB/op	~79.3 MB/op	~+1.6 MB
allocs 1MB	~50 /op	~53 /op	+3 /op
allocs 10MB	~52 /op	~57 /op	+5 /op

SONIC

When SONIC arrives it won't break.
Its errors already carry position, so it will just work, only not printed as nicely until we add a case for it (see the TODO in errorOffset).

[codecov]

Codecov Report

❌ Patch coverage is 93.70629% with 9 lines in your changes missing coverage. Please review.
✅ Project coverage is 76.69%. Comparing base (9de7658) to head (59c0b0b).

Files with missing lines	Patch %	Lines
jsonrpc/pretty_error.go	93.54%	6 Missing and 2 partials ⚠️
jsonrpc/server.go	94.73%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3702      +/-   ##
==========================================
+ Coverage   76.67%   76.69%   +0.01%     
==========================================
  Files         409      410       +1     
  Lines       37055    37181     +126     
==========================================
+ Hits        28411    28515     +104     
- Misses       6668     6679      +11     
- Partials     1976     1987      +11     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[rodrodros]

Half done PR. Yes now more information is given to the user, but it is given in such way that is not human friendly. Let's improve upon that.

You can look into other RPCs error and get inspiration from those.

[RafaelGranza]

Here are some of the new errors:
*There are no colors since the error can be used outside go and shell contexts.

{]
 ^
unexpected ']', expected a string key or '}' [line 1, column 2]

{"a" 1}
     ^
unexpected '1', expected ':' [line 1, column 6]

{"a":}
     ^
unexpected '}', expected a value [line 1, column 6]

{"a":1, }
        ^
unexpected trailing comma before '}' [line 1, column 9]

[1,2,]
     ^
unexpected trailing comma before ']' [line 1, column 6]

{"id":@}
      ^
unexpected '@', expected a value [line 1, column 7]

{"jsonrpc": "2.0", "method": "method", "id": 1
                                              ^
unexpected end of input [line 1, column 47]

{"a": "unterminated
                   ^
unexpected end of input [line 1, column 20]

{"jsonrpc": 5, "method": "x", "id": 1}
            ^
field "jsonrpc" should be string, got number [line 1, column 13]

^
unexpected end of input [line 1, column 1]

  {]
   ^
unexpected ']', expected a string key or '}' [line 3, column 4]

...xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" z}
                                                                           ^
unexpected 'z', expected ',' or '}' [line 1, column 230]

{@yyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyy...
 ^
unexpected '@', expected a string key or '}' [line 1, column 2]

[github-actions]

Claude finished @RafaelGranza's task in 2m 46s —— View job

---

Re-review: 767fe8a5b — Python-like error format

• Read latest commit & changed files
• Validate offset/column/marker math
• Check edge cases (panics, EOF, batch, windowing)
• Build + go vet (tests need approval — see note)

This commit does two things, both clean:

1. Type-aware dispatch — describeError now switches on the concrete error via errors.As/errors.Is (*json.UnmarshalTypeError → friendly field "jsonrpc" should be string, got number, EOF → unexpected end of input, *json.SyntaxError → token jargon mapping), replacing the older nested string-matching. Cleaner and more robust.
2. Format change — drops the indented |_ msg at line X column Y for the flatter Python-style msg [line X, column Y], matching the examples you posted. The caret line is unchanged.

Verified

• Build passes, go vet ./jsonrpc/ clean.
• Type-error path: {"jsonrpc": 5, ...} → UnmarshalTypeError{Field:"jsonrpc", Type:string, Value:number}, offset 13-1=12 → caret under 5, column 13. ✓
• Trailing-comma path still routes through precededByComma: {"a":1, } → unexpected trailing comma before '}' [line 1, column 9]. ✓
• EOF path: empty input → offset=len(input)=0 → offendingLine returns "", caret at column 1, unexpected end of input [line 1, column 1]. ✓
• Windowing (truncateAround, 80-col): right-windowed (column 230) and start-of-line (column 2) cases keep the caret aligned to the ellipsis offset; the message now sits on its own line at column 0, so it's no longer affected by markerCol. ✓
• No panic on untrusted input: offset is bounded above by len(input) (and ≥ 0 in practice), so input[:offset] / input[offset:] slices stay safe.
• Typo fixed: mazSizeOfMessage → maxLineWidth. ✓

Nit (non-blocking)

• describeError's default branch is now unreachable. prettyParseError only calls it after parseErrorOffset returned ok — which is true only for *json.SyntaxError, *json.UnmarshalTypeError, or io.EOF/io.ErrUnexpectedEOF, exactly the three cases the switch already handles. Harmless defensive code, but the default: return err.Error() can't fire today. (describeSyntaxError's own !ok fallback is the real fallback path.)

Carried-over follow-ups (not required here)

• Per-element batch decode still uses raw err.Error() (handleBatchRequest, server.go) — a malformed individual request inside a batch won't get the new marker.
• parseErrorOffset dropped the lower-bound clamp (min not max(0, min(...))) — unreachable today since stdlib offsets are ≥ 1, as you noted, but worth re-adding once the SONIC path feeds offsets through here.
• isBatch + >bufferSize leading whitespace → Peek(n) returns ErrBufferFull → routed through the single-request path (still parses, just not flagged as batch). Pathological only.

Note: I couldn't run go test ./jsonrpc/ in this environment (requires approval), but go build and go vet pass and I traced each table case by hand — CI's table tests should confirm.

No blockers — this round lands the human-friendly error UX the issue asked for.
• granza/json-pretty-print

[github-actions]

Claude finished @RafaelGranza's task in 2m 5s —— View job

---

Re-review: 90384530d — error-marker UX + extra tests

• Read latest commits & all changed files
• Re-verify offset/column/marker math + windowing
• Check panics / EOF / batch / leading-whitespace edge cases
• go build ./jsonrpc/ ✓ (tests need approval in this env — traced table cases by hand)

This round lands the human-friendly format rodrodros asked for, plus a solid table of edge cases. The implementation looks correct and well-bounded — no blockers.

Verified

• Marker alignment holds for left-window (column 2), right-window (column 230), and EOF-past-end cases — truncateAround returns idx - start + len(left) + 1, which lands the caret one past the last rune for unexpected end of input. ✓
• No panic on untrusted input. offset is upper-bounded by len(input); EOF returns len(input); input[:offset]/input[offset:] stay in range, and empty input → offset=0 → "\n^\n…". ✓
• Leading-whitespace position bug is fixed. isBatch now only Peeks (no Discard) and runs after the TeeReader is attached, so leading newlines reach errorRecoverBuffer and stay aligned with the decoder offset — confirmed by the "leading blank line keeps line number" case (line 2). ✓
• Column is now rune-counted (utf8.RuneCount), addressing the earlier bytes-vs-runes nit. ✓
• Untranslatable syntax errors ({"a":truX}) degrade gracefully to the raw stdlib message, still wrapped with line/column + marker. ✓
• lll lint exemption is correctly scoped to server_test.go only.

Nits (non-blocking)

• Per-element batch decode still emits raw err.Error() — server.go:404-409 (handleBatchRequest). A malformed individual request inside an otherwise-valid batch skips prettyParseError, so it gets the old terse message while the outer batch error is pretty-printed. Not a clean 1:1 swap (the json.RawMessage is already sub-sliced, so offsets are element-relative), so deferring is fine — but a // TODO here would stop it reading as an oversight.
• Top-level type mismatch leaks the Go type name — pretty_error.go:67: a bare 5 body yields expected jsonrpc.Request, got number (see the "top-level type mismatch" test). expected a JSON object reads better for an external client. (Posted inline.)

Carried-over (already discussed, fine to leave)

• describeError's default branch and the dropped lower-bound clamp in parseErrorOffset are both unreachable today; worth re-adding the max(0, …) clamp once the SONIC path feeds offsets through here, as you noted in the TODO.

Nice work on the benchmark + body-size-bounded copy justification — the success-path cost is well characterized.

Note: go test ./jsonrpc/ requires approval in this environment, so I traced each new table case by hand against the code rather than running CI's tests; go build and the manual trace agree on every expected string.
• granza/json-pretty-print

[github-actions]

Claude finished @RafaelGranza's task in 1m 7s —— View job

---

Re-review: 8391934b2 — polish parse error messages

• Read latest commit & changed files
• Re-verify offset/column/marker math + windowing
• Check panics / EOF / batch / leading-whitespace edge cases
• Build + vet (tests need approval in this env — traced by hand)

This commit is a single-line polish that resolves the last open nit, plus the matching test update. Clean.

Verified

• Top-level type-mismatch leak is fixed. describeTypeError now returns expected a JSON object, got number instead of expected jsonrpc.Request, got number for the Field == "" case (pretty_error.go:67). Traced the 5 body: UnmarshalTypeError.Offset=1 → offset=0 → line 1, column 1, caret under 5 → 5\n^\nexpected a JSON object, got number [line 1, column 1], which exactly matches the updated "top-level type mismatch" test. ✓
• The wording is accurate, not just nicer. Top-level type errors only reach this branch on the single-request path (isBatch routes [-prefixed input to the []json.RawMessage decode, whose elements are RawMessage and never mismatch), so the decode target is always Request — an object. "expected a JSON object" is correct, not a white lie. ✓
• No regressions in the surrounding paths — describeTypeError's Field != "" branch (field "jsonrpc" should be string, got number), the syntax-error/trailing-comma/EOF dispatch, windowing, and rune-counted columns are all unchanged from the prior round. Build + go vet ./jsonrpc/ pass.

Carried-over (already discussed — fine to leave)

• Per-element batch decode still emits raw err.Error() (server.go:404-409); deferring is reasonable since the sub-sliced RawMessage offsets are element-relative.
• describeError's default branch and the dropped max(0, …) lower-bound clamp in parseErrorOffset are both unreachable today; worth re-adding the clamp when the SONIC path starts feeding offsets through, per your TODO.

No blockers. This lands the human-friendly error UX the issue asked for, and the last reviewer nit is now addressed.

Note: go test ./jsonrpc/ requires approval in this environment, so I traced the changed table case against the code by hand; go build + go vet pass and agree with the expected string.
• granza/json-pretty-print

[rodrodros]

Looks good, thank you!

Before merging, could you please swap the column which makes sense for us programmers but might not be for non-programmers. My suggestion is position or similar.

Finally, in the tests, can you please add at least one real Starknet request (multi line) with an error. The examples I saw in the unit tests are limited to 1 to 2 lines.

[github-actions]

Claude finished @RafaelGranza's task in 3m 25s —— View job

---

Re-review: 1cbb5e9b8 — horizontal trim + sliding-window error buffer

• Read latest commits & changed files
• Verify window/offset/column math + sliding-window panics
• Confirm reviewer requests addressed (column→position, multi-line Starknet test)
• Build + go vet (tests need approval in this env — traced table cases by hand)

This round adds the bounded sliding-window buffer (windowBuffer, 512 B) and 80-col horizontal trimming, and resolves the two pre-merge asks from the approving review. No blockers.

Verified

• Both pre-merge requests are done. column → position everywhere (pretty_error.go:205 + all table cases), and there's now a real multi-line starknet_getStorageAt request with a missing comma (server_test.go:274-282, caret on line 4, position 39) plus a 202-line "error past the captured window" case. ✓
• Sliding-window math is panic-safe on untrusted input. absOffset is min(offset, consumedBytes), markerPos = max(0, absOffset - windowStart), and windowStart = consumedBytes - len(window), so markerPos ∈ [0, len(window)]. Every slice (window[:markerPos], window[markerPos:], offendingLine) stays in range, including markerPos == len(window) (EOF) and empty input. ✓
• Line number stays accurate after the window slides — newlinesSeen - count(window[markerPos:], '\n') + 1 counts newlines before the marker against the full-stream total, not just what's left in the window. Confirmed by the line 202 case. ✓
• windowBuffer.Write ring logic is correct — both the len(p) >= maxWindowSize fast path and the overlapping copy(window, window[len-max:]) tail-shift are sound (Go copy is memmove-safe on overlap). ✓
• Leading-whitespace position fix holds — isBatch only Peeks after the TeeReader is attached, so leading newlines reach the buffer (line 2 case). ✓
• Build + go vet ./jsonrpc/ pass.

Nit (non-blocking — posted inline)

• position undercounts for a single logical line longer than 512 B (pretty_error.go:63-68): when the marker's line began before the window was truncated, lineStart falls back to window start, so the printed position drifts low. Rare for JSON-RPC; caret stays aligned, only the number is off. A one-line comment noting the window bounds accuracy would do.

Carried-over (already discussed — fine to leave)

• Per-element batch decode still emits raw err.Error() (server.go:401-406) instead of prettyParseError — deferring is reasonable since the sub-sliced RawMessage offsets are element-relative.
• errorOffset's dropped lower-bound clamp and describeError's default branch remain unreachable today; the max(0, …) in prettyParseError already guards the slices, so re-adding the clamp only matters once the SONIC path feeds offsets through (per your TODO).

Note: go test ./jsonrpc/ needs approval in this environment, so I traced the changed/added table cases against the code by hand; go build + go vet pass and agree with every expected string.
• granza/json-pretty-print

[github-actions]

Claude finished @RafaelGranza's task in 3m 14s —— View job

---

Re-review: 8b8ab4a33 — realistic JSON in parse errors

• Read latest commits & changed files
• Check whether the windowBuffer.Write fast path (pretty_error.go:31-33) is now tested
• Re-verify offset / line / position + windowing math
• go build + go vet ./jsonrpc/ ✓ (go test needs approval in this env — traced table cases by hand)

This round adds realistic multi-line Starknet table cases and clearer param-count messages. No blockers.

Verified

• @rodrodros's two approving-review asks are both done: real multi-line Starknet requests with errors (starknet_getStorageAt, starknet_call, nested params objects), and column → position everywhere. ✓
• Param-error messages improved and accurate: missing non-optional param field → missing required params: num; missing/unexpected params in list → expected between 1 and 3 params, got 4. Tests updated to match. ✓
• Windowing / offset / line / position math holds across the new cases — right-window (position 174), both-sides window (position 155), window-slide line number (line 42), maxContextRows-capped context (line 7), oversized-line context drop (line 5), and rune-counted column (👍 → position 8). Traced each against the code. ✓
• No panic on untrusted input: markerPos = max(0, absOffset - windowStart) stays in [0, len(window)]; every slice is bounded. The windowBuffer.Write ring/fast-path is memmove-safe. ✓
• Build + vet pass.

Note — the pretty_error.go:33 test request

@rodrodros asked to test the len(p) >= maxWindowSize fast path. The new large-body cases (error past the captured window ~2.7 KB, oversized line drops all preceding context ~520 B) now reach it indirectly through bufio/json.Decoder chunking, so codecov should clear — but the coverage depends on the bufferSize=128 / decoder minRead=512 sizing and isn't pinned by a direct test. Left an inline note recommending the previously-drafted package jsonrpc internal test for robustness. Non-blocking.

Carried-over (already discussed — fine to leave)

• Per-element batch decode still emits raw err.Error() (server.go, handleBatchRequest) — element-relative offsets make a clean swap awkward; a // TODO would stop it reading as an oversight.
• errorOffset's dropped lower-bound clamp and describeError's default branch are unreachable today (stdlib offsets ≥ 1; prettyParseError's max(0, …) guards the slices) — re-add the max(0, min(…)) clamp when the SONIC path feeds offsets through, per the TODO.
• position undercounts for a single logical line > 512 B (pretty_error.go:65); caret stays aligned, only the printed number drifts. A one-line comment noting the window bound would close it.

Note: go test ./jsonrpc/ requires approval in this environment, so I traced the changed/added table cases against the code by hand; go build + go vet ./jsonrpc/ pass and agree with every expected string.

Nice work landing the realistic-request coverage @rodrodros asked for.
• granza/json-pretty-print

[github-actions]

Claude finished @RafaelGranza's task in 3m 24s —— View job

---

Re-review: 9399128be — oversized single-line test case

• Read latest commit & changed files
• Verify the new test's offset / line / position math
• Check whether windowBuffer.Write fast path (pretty_error.go:31-33) is now directly tested
• go build + go vet ./jsonrpc/ ✓ (go test needs approval in this env — traced by hand)

This commit adds one table case ("oversized single-line input keeps only the trailing window"). It's correct, and go build + go vet ./jsonrpc/ pass. No blockers.

Verified

• The reported position 510 is the windowed position, not the true one — and that's the expected behavior here. Traced it: the @ sits at byte offset 765 (true position 766), but with the body ≈768 B and a 512 B window, windowStart = 768 − 512 = 256, so markerPos = 765 − 256 = 509. The line has no newline in the window, so lineStart = 0 and col = 509 + 1 = 510. The caret still lands correctly because truncateAround works off the same windowed line. ✓
• So this case pins the known position-undercount-on-lines->512B behavior (pretty_error.go:65) as a regression test. Good — but since the number is deliberately "wrong" (510 vs. true 766), a one-line // note: position is window-relative for lines longer than maxWindowSize next to the case (or at lineAndColumn) would stop a future reader from "fixing" it. This is the same nit flagged inline at pretty_error.go:65/68 — now it's load-bearing in a test, so the comment matters a bit more.

Re @rodrodros's "test this execution path" (pretty_error.go:33) — still indirect only

This new ≈768 B single-line body reaches the len(p) >= maxWindowSize fast path only indirectly, through the same bufio(128)/json.Decoder(512 minRead) chunking as the existing large cases — so codecov should stay green, but no test pins the branch directly. Whether line 33 is hit still depends on upstream buffer sizing; if bufferSize or the decode path changes, the branch silently goes back to uncovered with no failing test.

The direct package jsonrpc internal test drafted earlier (here) pins the reset, the == boundary, and the consumedBytes/newlinesSeen accounting regardless of chunking — it just couldn't be pushed (403). Recommend landing it to close out the request robustly. Non-blocking. Add it →

Carried-over (already discussed — fine to leave)

• Per-element batch decode still emits raw err.Error() (server.go, handleBatchRequest) — element-relative offsets make a clean swap awkward; a // TODO would stop it reading as an oversight.
• errorOffset's dropped lower-bound clamp and describeError's default branch are unreachable today (stdlib offsets ≥ 1; prettyParseError's max(0, …) guards the slices) — re-add the max(0, min(…)) clamp when the SONIC path feeds offsets through, per the TODO.

Note: go test ./jsonrpc/ requires approval in this environment, so I traced the new case against the code by hand; go build + go vet ./jsonrpc/ pass and agree with the expected string.

Nice — this lands a regression test for the windowed-position edge. The only open item is the direct fast-path unit test @rodrodros asked for.
• granza/json-pretty-print

[github-actions]

Claude finished @RafaelGranza's task in 2m 58s —— View job

---

Re-review: ad165d0c1 — test: Address requests (head of granza/json-pretty-print)

• Read latest commits & changed files
• Verify offset / line / position + windowing math
• Check panics / EOF / batch / untrusted-input edge cases
• Build + go vet ./jsonrpc/ ✓ (go test needs approval in this env — traced table cases by hand)

The head commit adds a single table case ("oversized single-line input keeps only the trailing window"). It's correct, and go build + go vet ./jsonrpc/ pass. No blockers — this is the human-friendly error UX the issue asked for, and it's well-tested.

Verified

• New case math is internally consistent. The @ sits at stream offset ≈765; with consumedBytes ≈ 768 and a 512 B window, windowStart = 256, markerPos = 509, no newline in the window → lineStart = 0 → reported position 510. truncateAround runs off the windowed line (512 B), so end == len(window) yields a leading ... with no trailing ellipsis, caret at col 75 — exactly the expected string. ✓
• This case pins the known window-relative position behavior (pretty_error.go:65) as a regression test: for a single logical line longer than maxWindowSize, the printed position is window-relative (510), not the true position (766). The caret stays aligned because it's computed off the same windowed line. Deliberately "wrong" number → worth a one-line // position is window-relative for lines longer than maxWindowSize next to the case or at lineAndColumn, so a future reader doesn't "fix" it. Non-blocking (carried over from prior rounds).
• No panic on untrusted input. markerPos = max(0, absOffset − windowStart) ∈ [0, len(window)]; every slice (window[:markerPos], window[markerPos:], offendingLine, describeSyntaxError) stays in range, including markerPos == len(window) (EOF) and empty input. The windowBuffer.Write ring/fast-path is memmove-safe.
• buildArguments message changes are sound — missing required params: <names> (via utils.Filter/utils.Map) and expected between %d and %d params, got %d are accurate and read well; helpers exist in utils/slices.go. ✓
• isBatch leading-whitespace fix holds — Peek(n) only (no Discard), run after the TeeReader is attached, so leading newlines reach the buffer and the line number stays correct ("leading blank line keeps line number" → line 2). ✓

Carried-over (already discussed — fine to leave)

• Per-element batch decode still emits raw err.Error() (server.go, handleBatchRequest) — element-relative json.RawMessage offsets make a clean swap awkward; a // TODO would stop it reading as an oversight.
• errorOffset's dropped lower-bound clamp and describeError's default branch are unreachable today (stdlib offsets ≥ 1; prettyParseError's max(0, …) guards the slices) — re-add the max(0, min(…)) clamp when the SONIC path feeds offsets through, per the TODO at pretty_error.go:58.
• Direct windowBuffer.Write fast-path unit test (pretty_error.go:31-33) — @RafaelGranza noted he won't be addressing this; coverage is reached indirectly via bufio/json.Decoder chunking, so it depends on buffer sizing. Acknowledged, non-blocking.

Note: go test ./jsonrpc/ requires approval in this environment, so I traced the changed/added table cases against the code by hand; go build + go vet ./jsonrpc/ pass and agree with every expected string. CI's table tests should confirm.

Clean round — nothing new to flag.
• granza/json-pretty-print

[github-actions]

Claude finished @RafaelGranza's task in 3m 17s —— View job

---

Re-review: c48abb45b + a903f3456 — tidy windowBuffer.Write + split out parse-error tests

• Read latest commits & changed files
• Verify windowBuffer.Write shift-before-append refactor
• Verify the split-out parse-error test function
• go build + go vet ./jsonrpc/ ✓ (go test needs approval here — traced table cases by hand)

Both of @rodrodros's last asks are addressed, and the refactor is correct. One CI concern below.

Verified

• windowBuffer.Write now shifts before appending (pretty_error.go:36-39), resolving the "copy then overwrite / regrow past the cap" concern. The math checks out: in the slow path len(p) < maxWindowSize, so overflow = len(window) + len(p) − maxWindowSize < len(window) ⇒ window[overflow:] is always in range, and after the copy+append the window is exactly maxWindowSize. No transient growth beyond the cap. The len(p) >= maxWindowSize fast path is unchanged. ✓
• describeError's switch was reordered to match errorOffset (syntax → type → EOF), per the inline ask. Behavior is unchanged since the cases are disjoint error types. ✓
• Parse-error table split into its own TestHandleParseError in the new jsonrpc/pretty_error_test.go — TestHandle in server_test.go is now back to ~one concern, and the moved cases are a clean 1:1 (verified the diff). ✓
• Blank line between each req/res pair added in both test maps, per the readability ask. ✓
• Build + go vet ./jsonrpc/ pass.

Important — diff-lint will fail on the new test file

• .golangci_diff.yaml:22-24 — the lll (100-col) exemption only lists jsonrpc/server_test.go, but the long single-line res strings moved to jsonrpc/pretty_error_test.go. 31 lines in the new file exceed 100 chars, so the golangci_diff job will flag every one. Broaden the path to cover both files — posted inline with a suggestion. (path: jsonrpc/(server|pretty_error)_test\.go)

Carried-over (already discussed — fine to leave)

• Per-element batch decode still emits raw err.Error() (server.go, handleBatchRequest) — element-relative offsets make a clean swap awkward; a // TODO would stop it reading as an oversight.
• errorOffset's dropped max(0, …) clamp and describeError's default branch are unreachable today (stdlib offsets ≥ 1; prettyParseError's max(0, …) guards the slices) — re-add the clamp when the SONIC path feeds offsets through, per the TODO at pretty_error.go:57.
• position is window-relative for a single logical line > maxWindowSize (caret stays aligned; only the printed number drifts) — pinned by the "oversized single-line input…" case; a one-line comment at lineAndColumn would close it.
• Direct windowBuffer.Write fast-path unit test — @RafaelGranza noted he won't address it; reached indirectly via bufio/json.Decoder chunking. Acknowledged.

Note: go test ./jsonrpc/ requires approval in this environment, so I traced the changed/added table cases against the code by hand; go build + go vet ./jsonrpc/ pass and agree with every expected string.

The only actionable item this round is the lll exemption path — without it the lint job fails on the freshly-split test file.
• granza/json-pretty-print

[github-actions]

Claude finished @RafaelGranza's task in 1m 6s —— View job

---

Re-review: 59c0b0b06 — extend lll exemption to pretty_error_test

• Read latest commit & changed files
• Verify the lll exemption now covers the new test file
• Re-verify offset / line / position + windowing math
• Build + go vet ./jsonrpc/ ✓

The head commit is the one-line CI fix from last round's only actionable item. No blockers.

Verified

• The lll lint failure is resolved. .golangci_diff.yaml:24 now reads path: jsonrpc/(server|pretty_error)_test\.go, which matches both server_test.go and the newly split-out pretty_error_test.go. I confirmed 31 lines in pretty_error_test.go exceed 100 chars (the long single-line res strings) — exactly what the exemption now covers — so golangci_diff will no longer flag them. ✓
• Scope is still tight — the exemption is limited to the two *_test.go files via the alternation; it does not loosen lll on the non-test source (pretty_error.go, server.go). ✓
• No code changed this round — the windowing / offset / line / position math, the windowBuffer.Write shift-before-append refactor, and the split-out TestHandleParseError are all unchanged from the prior approved rounds. go build + go vet ./jsonrpc/ pass.

Carried-over (already discussed & acknowledged — fine to leave)

• Per-element batch decode still emits raw err.Error() (server.go, handleBatchRequest) — element-relative json.RawMessage offsets make a clean swap awkward; a // TODO would stop it reading as an oversight.
• errorOffset's dropped max(0, …) lower-bound clamp (pretty_error.go:51-53) and describeError's default branch (pretty_error.go:124) are unreachable today (stdlib offsets ≥ 1; prettyParseError's max(0, …) already guards the slices) — re-add the clamp when the SONIC path feeds offsets through, per the TODO at pretty_error.go:57.
• position is window-relative for a single logical line > maxWindowSize (pretty_error.go:64); caret stays aligned, only the printed number drifts — pinned by the "oversized single-line input…" case. A one-line comment at lineAndColumn would close it.
• Direct windowBuffer.Write fast-path unit test (pretty_error.go:31-33) — @RafaelGranza noted he won't address it; reached indirectly via bufio/json.Decoder chunking. Acknowledged.

Note: go test ./jsonrpc/ requires approval in this environment, so I verified the lint exemption regex against the file paths and the long-line count by hand, and traced no table-case changes this round; go build + go vet ./jsonrpc/ pass.

Clean round — the lint exemption is the last actionable item from the approving review, and it's now correct.
• granza/json-pretty-print