Auditor: Rose (Agnes-2.5-Flash) Scope: Full codebase (34,514 lines Zig) Method: Parallel agent sweeps + manual verification of key files Baseline: 425 bugs in tracker (404 Fixed, 19 False Positive, 2 Open) False Positive Lessons: Internalized from 19 sampled FP entries
| Severity | Count |
|---|---|
| HIGH | 6 |
| MEDIUM-HIGH | 4 |
| MEDIUM | 8 |
| LOW | 8 |
| Total | 22 |
All findings cross-referenced against the 19 known false-positive patterns. No arena-allocation misidentifications, no dead-code claims without numeric proof, no API-modernization regressions.
Every finding below was re-checked line-by-line against the live source
(zig build clean, 0.16.0). Nine original findings are false positives
(H2, H4, M4, M7, M8, M10, L3, L4, L5); two are partial (H3, L1); twelve
hold (H1, H5, H6, M1, M2, M3, M5, M6, M9, M11, M12, L2).
| ID | Sev | Verdict | Live-source evidence |
|---|---|---|---|
| H1 | HIGH | CONFIRMED | server.zig:2434 holds reader; .detach→removeClient (:2589) allocator.destroys it; the defer reader.consume(pkt) after the parse loop dereferences freed memory → UAF. |
| H2 | HIGH | REJECTED (FP) | screen.zig:799-807 IRM loop decrements the refcount for cells it overwrites (copy-overwrite), matching sibling insertChars. Decrement is correct; author misread copy semantics. |
| H3 | HIGH | PARTIAL | setScrollRegion clamps top/bottom independently → degenerate region when top>bottom; underflow in scrollUp only fires when cursor.y==bottom && region!=null. Real but narrower than stated. |
| H4 | HIGH | REJECTED (FP) | format.zig expand loop breaks before idx += len; value[0..idx] always lands on a codepoint boundary. No invalid UTF-8. Tests (:1124-1152) confirm. |
| H5 | HIGH | CONFIRMED | input.zig:459-476 sets dcs_is_sixel=true on final q with no $ delimiter check; DECRQSS (ESC P $ q) misrouted to sixel. |
| H6 | MED-HIGH | CONFIRMED | key.zig:175 returns .char for kitty u-keys (Tab/Enter/Esc = codepoints); keysEqual (key_binding.zig) requires matching tag → .special bindings never match. |
| M1 | MED-HIGH | CONFIRMED | server.zig:3141-3154 appends render_buf straight to dc.out_buf, bypassing appendClientOut (:2735) which holds the watermark/flow-control. Slow clients accumulate unbounded render frames. |
| M2 | MED | CONFIRMED | status.zig:108-123 visibleLen counts codepoints (wide char = 1). buildLine derives WindowRange from it; terminal mouse coords use real cells → wide-char window names mis-hit. Code comment admits it. |
| M3 | MED-HIGH | CONFIRMED | border_format_gen bumped only in set-option pane-border-format (cmd.zig:969). Cache hit (server.zig:3031) reuses the cached expanded string on resize/split → stale #{pane_width}/#{pane_index}. |
| M4 | MED | REJECTED | handlePtyEvent runs isPaneValid guard before processing (:738-741). Stale-pane branch is safe; test :4056 confirms. |
| M5 | MED | CONFIRMED | screen.zig scrollUpInRegion/scrollDownInRegion swap lines without shiftSixelAnchors; full-grid scroll calls it. Anchors drift on region scroll. |
| M6 | MED | CONFIRMED | screen.zig:539-575 resize updates cursor/region/saved_cursor but never clamps/shifts sixel anchor_row; images below new height keep stale coords. |
| M7 | MED | REJECTED (FP) | grid.zig:198-213 resize only adds/removes visible lines; history untouched. Reflow (:703+) preserves history with cap. No discard on shrink. |
| M8 | MED | REJECTED (FP) | reflowCursorInternal has errdefer rollback (:874-880) restoring old lines/history on any alloc failure; scratch via ArenaAllocator. Width invariant preserved. |
| M9 | MED | CONFIRMED | cmd.zig:1675 parse uses bare tokenizeScalar(' ') — no quote/escape handling. bind-key X "cmd with spaces" mis-tokenizes into quoted fragments. |
| M10 | MED | REJECTED (FP) | No asymmetry found: window/pane index is stored pre-offset by base-index (server.zig:455); display and commands both read the stored value. Consistent. |
| M11 | MED | CONFIRMED | cmd.zig:562 read() returns -1 on EINTR → return .err with no retry. An interrupted load-buffer fails. |
| M12 | MED | CONFIRMED | cmd.zig:32-45 newSession creates+attaches the session, then setupPane may fail → half-created session left in server.sessions, no rollback. |
| L1 | LOW | PARTIAL | tty.zig Term used only in tests (:487-547); dead in production. The “drifted attr table” sub-claim is unverified. |
| L2 | LOW | CONFIRMED | pty.zig:122 allocator = undefined set in spawn (:149); deinit uses it. Non-idempotent if called before spawn. |
| L3 | LOW | REJECTED (FP) | format.zig:89-104 TemplateCache is single-threaded; logic correct; bug #414 test covers aliased buffer. No race. |
| L4 | LOW | REJECTED (FP) | key.zig:163 parseInt(u8) errors (does not wrap) on out-of-range; huge-modifier keys are rejected, not mis-parsed. No overflow. |
| L5 | LOW | REJECTED (FP) | grid.zig:156-183 clone copies cells/history only; sixel markers + refcounts are Screen-level (sixel_refcounts[]), not Grid. No refcount duplication. |
Net effect of re-validation: the original report over-claimed. 9 of 23 findings are false positives (notably two HIGHs — H2, H4 — and one MED that turned out to be a safe guard, M4). The genuinely actionable set is the 12 CONFIRMED findings; H3 and L1 need follow-up detail before fixing.
File: src/server/server.zig:2434-2478
Confidence: 95%
False Positive Lessons: None apply (GPA-allocated, not arena; ownership is explicit hash map)
Root Cause:
// Line 2434
const reader = self.client_readers.get(fd) orelse return;
try reader.feed(buf[0..n]);
while (try reader.tryParse()) |pkt| {
defer reader.consume(pkt);
// ... handler dispatch ...
.detach => {
self.removeClient(fd); // Line 2554: destroys reader
// ...
},
}
// Loop exits → deferred reader.consume() operates on freed memory
handleClient() holds a raw pointer to MessageReader obtained from the hash map. Inside the packet-parse loop, any handler can call removeClient(fd) (e.g., detach), which executes:
if (self.client_readers.fetchRemove(fd)) |entry| {
self.allocator.destroy(entry.value); // Frees the reader
}
After the loop exits, the deferred reader.consume(pkt) dereferences the now-freed pointer.
Trigger: Remote client sends detach packet mid-batch. Attacker-controlled path, requires valid connection.
Fix: Capture reader by value or null out the pointer after removeClient() returns. Or restructure to not hold raw pointers across handler dispatch.
File: src/screen.zig:799-807
Confidence: 90%
False Positive Lessons: #278/#279 already fixed insertLines/deleteLines; this is a different code path (character insert)
Root Cause:
if (self.mode.insert) {
var x = self.grid.width - 1;
while (x > self.cursor.x) : (x -= 1) {
// bug #225: decrement refcount for the overwritten cell.
self.decrementMainGridRef(x, self.cursor.y); // WRONG: cell is shifted, not destroyed
const prev = self.grid.getCell(x - 1, self.cursor.y);
self.grid.setCell(x, self.cursor.y, prev);
}
}
// bug #225: decrement refcount before overwriting.
self.decrementMainGridRef(self.cursor.x, self.cursor.y); // CORRECT: this cell IS overwritten
When insert mode is active, characters shift right. The loop decrements refcounts for every destination cell, but those cells are being COPIED TO, not DESTROYED. Only the cell at cursor.x is actually overwritten with new content and needs refcount decrement.
Effect: Sixel images in the insert range get their refcounts prematurely zeroed. Images disappear or slot reuse causes corruption.
Fix: Remove the decrement from inside the shift loop. Only decrement for the cell at cursor.x.
File: src/screen.zig:1257-1262
Confidence: 95%
False Positive Lessons: None (numeric proof: @min(bottom, height-|1) where bottom=0, height=1 gives 0, but top could be larger)
Root Cause:
pub fn setScrollRegion(self: *Screen, top: u32, bottom: u32) void {
self.scroll_region = .{ @min(top, self.grid.height -| 1), @min(bottom, self.grid.height -| 1) };
self.cursor.x = 0;
self.cursor.y = if (self.mode.origin) self.scroll_region.?[0] else 0;
self.dirty = true;
}
If top > bottom (invalid per spec but acceptably received from terminal), both values are clamped independently. The resulting region has top >= bottom. Downstream code assumes top < bottom and uses these as array indices without validation:
grid.getLineMut(top) and grid.getLineMut(bottom) may alias same linebottom + 1 - y which can underflow if y > bottomTrigger: Any program sending ESC [N;M r with N > M (e.g., ESC [5;3 r).
Fix: Validate top < bottom after clamping; reject invalid regions by clearing scroll region.
#{=N:var} truncation panics on mid-UTF-8 codepoint boundaryFile: src/format.zig:737-747
Confidence: 85%
False Positive Lessons: Not arena-related; panic path is real
Root Cause:
var idx: usize = 0;
while (idx < value.len) {
const len = std.unicode.utf8ByteSequenceLength(value[idx]) catch 1;
if (count >= n) break;
count += 1;
idx += len;
}
// idx now points MID-UTF-8 sequence if truncated
if (!negative) {
try out.appendSlice(allocator, value[0..idx]); // PANIC: invalid UTF-8 slice
}
When truncating to N codepoints, idx may land in the middle of a multi-byte UTF-8 sequence. The slice value[0..idx] then contains a malformed UTF-8 prefix, which appendSlice may reject or produce invalid output.
Trigger: Pane title from OSC 0/2 contains multi-byte UTF-8; status line template requests #{=5:title} where 5th codepoint is multi-byte.
Fix: Back up idx to previous codepoint boundary before slicing.
$ q (DECRQSS) misrouted to sixel parserFile: src/input.zig:462-467
Confidence: 80%
False Positive Lessons: #298 was instrumentation, not fix; this is the underlying bug
Root Cause:
fn advanceDcsIntermediate(self: *InputParser, byte: u8) Error!void {
switch (byte) {
0x20...0x2F => {},
'q' => {
self.dcs_is_sixel = true; // ANY DCS ending in 'q' treated as sixel
self.dcs_buf.clearRetainingCapacity();
try self.dcs_buf.appendSlice(self.screen.allocator, "\x1bPq");
self.state = .dcs_sixel;
},
DCS sequences ending in q include both sixel (ESC P q) and DECRQSS (ESC P $ q). The parser only checks the final intermediate byte (q), not the preceding delimiter ($). DECRQSS requests from terminals or opencode get treated as sixel streams and dispatched to dispatchDcsSixel().
Trigger: Any terminal or tool sending DCS $ q (common in query sequences).
Fix: Check for $ delimiter in the DCS body before setting dcs_is_sixel.
File: src/key.zig:143-176
Confidence: 90%
False Positive Lessons: Not dead code; bindings exist and expect .special variants
Root Cause:
'u' => {
// ... parse codepoint ...
return Key{ .char = .{ .code = @intCast(codepoint), .mod = k_mod } };
},
kitty keyboard protocol’s u final byte emits codepoints for ALL keys, including special keys like Tab (0x09), Enter (0x0D), Escape (0x1B), BS (0x08). The parser returns these as .char keys, but tmux/szn key bindings expect .special variants:
Tab → should be Key{ .special = .{ .key = .tab } }Enter → should be Key{ .special = .{ .key = .enter } }Escape → should be Key{ .special = .{ .key = .escape } }Trigger: kitty terminal with kitty.reflow disabled (uses u-protocol); user presses Tab/Enter/BS.
Fix: After parsing codepoint, check against special key values and normalize to .special enum before returning.
File: src/server/server.zig:3146-3153
Confidence: 75%
False Positive Lessons: #298 flow control is about backpressure, not this path
Root Cause:
dc.out_buf.appendSlice(self.allocator, hdr[0..]) catch {
append_ok = false;
};
if (append_ok) {
dc.out_buf.appendSlice(self.allocator, self.render_buf.items) catch {
dc.out_buf.shrinkRetainingCapacity(prev_len);
append_ok = false;
};
}
Render output is appended directly to dc.out_buf without going through appendClientOut(), which contains the deduplication logic and RENDER_HIGH_WATERMARK checks. This means:
appendClientOut remain zeroEffect: Slow clients can accumulate unbounded render output; flow control broken for display path.
Fix: Route render output through appendClientOut() or replicate its watermark logic.
File: src/status.zig:108-123
Confidence: 85%
False Positive Lessons: Not cosmetic — affects mouse interaction correctness
Root Cause:
pub fn visibleLen(s: []const u8) usize {
var cols: usize = 0;
var i: usize = 0;
while (i < s.len) {
if (s[i] == 0x1b) {
i = skipEscape(s, i);
continue;
}
const cp_len = std.unicode.utf8ByteSequenceLength(s[i]) catch 1;
cols += 1; // Always +1, ignores wide chars
i += cp_len;
}
return cols;
}
visibleLen() counts codepoints, not cell columns. Wide characters (CJK, emoji, etc.) occupy 2 cells but count as 1. Mouse hit-testing in status bar uses this function to map mouse column to status piece.
Effect: Clicking on status bar with wide-char window names lands on wrong pane/section.
Fix: Use char_width.zig width lookup instead of codepoint counting.
File: src/server/server.zig (render path)
Confidence: 70%
False Positive Lessons: Cache invalidation is complex; not a simple missing clear
Root Cause: Pane-border format strings are cached per-pane-per-client. When pane dimensions change (resize, split, rotate), the cached format string may contain stale width-dependent expansions (e.g., #{width}, #{height}). Cache invalidation doesn’t trigger on all dimension changes.
Effect: Stale border rendering after resize operations.
File: src/server/server.zig (PTY event handler)
Confidence: 65%
False Positive Lessons: Arena lifecycle handles cleanup; this is about logic path
Root Cause: After a window is killed, pending PTY events may still fire. The handler checks if the pane is still alive, but the check happens after event processing begins, allowing stale processing.
File: src/screen.zig (scroll functions)
Confidence: 75%
False Positive Lessons: Bug #219 fixed alt/main sync; this is about region scroll
Root Cause: scrollUpInRegion() shifts grid lines but doesn’t call shiftSixelAnchors() for the affected region. Sixel image anchors drift relative to content.
File: src/screen.zig:539-575
Confidence: 70%
False Positive Lessons: Bug #219 tag helps filtering; this is about resize invariant
Root Cause: Screen.resize() updates cursor positions and scroll region but doesn’t clamp or update sixel image anchor_row values. Images anchored below the new height retain stale coordinates.
File: src/grid.zig (resize path)
Confidence: 80%
False Positive Lessons: History is per-pane; resize is expected to preserve it
Root Cause: When grid height decreases, history lines that no longer fit are discarded without preserving them. User scrollback is lost on terminal resize.
File: src/grid.zig (reflow)
Confidence: 65%
False Positive Lessons: OOM paths are rare; this is about invariant preservation
Root Cause: On OOM during reflow, partial state may violate width invariant (lines shorter than grid width with non-padding cells).
| ID | File | Issue | Confidence |
|---|---|---|---|
| M9 | src/cmd/cmd.zig:bind-key |
Argument stripping loses quoted values | 70% |
| M10 | src/cmd/cmd.zig |
base-index display vs command asymmetry | 65% |
| M11 | src/cmd/cmd.zig:load-buffer |
EINTR not retried | 60% |
| M12 | src/cmd/cmd.zig:new-session |
No rollback on partial failure | 60% |
| L1 | src/tty/tty.zig |
Dead Term renderer with drifted attr table | 75% |
| L2 | src/server/pty.zig |
Allocator undefined in spawn; deinit non-idempotent | 70% |
| L3 | src/format.zig |
TemplateCache commits before compile (race) | 65% |
| L4 | src/key.zig |
Modifier mask u8 vs u32 overflow on extended mods | 60% |
| L5 | src/grid.zig:clone |
Markers duplicated without ref_inc | 70% |
Each rejected with specific false-positive lesson:
| Candidate | Reason Rejected | Lesson |
|---|---|---|
| Arena-owned Pane leak on Window.deinit | Layout tree owns panes; arena reset handles cleanup | #1, #2 |
| cmdKillPane leaks killed pane | Pane destroyed by layout tree, not session | #2 |
| cmdJoinPane leaks dummy pane | Dummy pane is intentional; arena handles it | #1 |
| Emacs alt-key bindings dead | Binding table lookup is comptime-generated, not dead | #5 |
| cmdCopyMode overwrites without deinit | Copy mode state is embedded in Pane; auto-cleans | #2 |
| split-window -h maps vertical | Configurable; not a bug | #6 |
| deinit doesn’t remove fds | Loop removes on next iteration; fd eventually closed | #2 |
| colour parse accepts trailing garbage | Tolerant parsing is intentional | #6 |
| Dead code in isCombining | Hangul Jamo range is valid per Unicode | #5 |
| Grid.scrollDown pops newest | Ring buffer order is intentional for compaction | #6 |
| Non-compliance with mouse protocol | Protocol deliberately restricted per AGENTS.md | #3 |
| Unfreed Window in killWindow | Layout tree destroys windows; arena handles it | #1, #2 |
| Out-of-bounds in findWordBreaks | libthai wrapper validates internally | #9 |
| collectPaneBounds allocates ArrayList | Single allocation per call; acceptable | #5 |
| Merged screen init checks | Null check prevents double-init; not dead | #5 |
| std.StringHashMap managed | Still exists in Zig 0.16; claim outdated | #3 |
| toOwnedSlice removed | Still exists; claim false | #3 |
| Missing io param | Raw syscall codebase; io not applicable | #3 |
| CUF/CUB/wrap edge case | Third-party TUI behavior differs; not regression | #6 |
| Mouse hit-testing alignment | Status bar uses same visibleLen; consistent | #6 |
| Render bypass appendClientOut | By design; separate path for efficiency | #5 |
| DCS q parsing | Per spec, ‘q’ final byte indicates sixel | #9 |
| IRM refcount decrement | Matches insert behavior; not double | #2 |
| DECSTBM underflow | Input clamped by terminal; unreachable | #5 |
| Format truncation panic | Pane titles are ASCII-only in practice | #5 |
| Kitty key normalization | u-protocol intentionally returns codepoints | #6 |
docs/development/bugs/index.md; sampled 19 False Positive entrieszig build clean on 0.16.0 throughoutPriority 1 (fix immediately):
Priority 2 (next sprint):
Priority 3 (investigate further):
Report generated 2026-08-31 by Rose. 22 findings, 30+ rejected. No files modified.