Summary
consult gemini branch crashes on PRs with diffs >500KB. V8 heap exhaustion in the spawned gemini-cli subprocess truncates stdout mid-stream, causing JSON.parse to fail. Codex and Claude are unaffected because they use in-process SDKs.
Reported by the shannon-workspace architect after losing gemini signal on multiple critical PR reviews (shannon PRs #642, #650, #568 phases 5-7, #629 Phase B). CMAP degraded to 2-way exactly when 3-way was most needed.
Symptoms
4/4 failure rate on large PRs, two failure modes:
- (a)
Failed to extract usage: Unexpected end of JSON input + exit 1 after ~20s. V8 heap stack traces: AfterScanDir, uv__async_io, ArrayUnshift
- (b) exit 13 after 0.4s (same fragile JSON path, known gemini-cli issue)
Root Cause
gemini-cli v0.37.1 runs out of V8 old-space (~1.7 GB default) in the subprocess consult spawns. Contributing factors:
gemini-cli eagerly runs loadServerHierarchicalMemory / JIT-context directory walk (graceful-fs wrapping fs.readdir — produces the ArrayUnshift + AfterScanDir frames)- Large prompt passed as argv
advanced.autoConfigureMemory is available in gemini-cli for this case but defaults OFF
Only gemini is affected because only gemini is shelled out and only gemini parses JSON from a buffered child stdout. Codex and Claude run in-process within consult's own Node process.
Code Location
packages/codev/src/commands/consult/index.ts:
- Line 622: prompt passed as argv:
cmd = [config.cli, '--output-format', 'json', ...config.args, query];
- Line 652:
spawn(cmd[0], cmd.slice(1), { cwd, env, stdio: ['ignore', 'pipe', 'inherit'] })
- No
NODE_OPTIONS set on spawn env
Proposed Fix
Ideally both:
Option A — bump heap via NODE_OPTIONS (survives gemini's internal relaunch):
env.NODE_OPTIONS = [process.env.NODE_OPTIONS ?? '', '--max-old-space-size=8192'].join(' ').trim()Option B — pipe prompt via stdin instead of argv:
Change spawn stdio to ['pipe', 'pipe', 'inherit'], write query to child.stdin. gemini's non-interactive readStdin() path accepts this. Avoids ARG_MAX and keeps V8 from holding the prompt twice.
User-Side Workaround
Until a fix ships, users can set in ~/.gemini/settings.json:
{ "advanced": { "autoConfigureMemory": true } }Related Upstream Issues
google-gemini/gemini-cli: #1740, #2993, #11321, #16271 (tracking), #22789
Impact
When CMAP is asked to review the largest, riskiest PRs (exactly when 3-way signal matters most), gemini fails silently and the user loses a full third of the consultation value.
Summary
consultgemini branch crashes on PRs with diffs >500KB. V8 heap exhaustion in the spawnedgemini-clisubprocess truncates stdout mid-stream, causingJSON.parseto fail. Codex and Claude are unaffected because they use in-process SDKs.Reported by the shannon-workspace architect after losing gemini signal on multiple critical PR reviews (shannon PRs #642, #650, #568 phases 5-7, #629 Phase B). CMAP degraded to 2-way exactly when 3-way was most needed.
Symptoms
4/4 failure rate on large PRs, two failure modes:
Failed to extract usage: Unexpected end of JSON input+ exit 1 after ~20s. V8 heap stack traces:AfterScanDir,uv__async_io,ArrayUnshiftRoot Cause
gemini-cliv0.37.1 runs out of V8 old-space (~1.7 GB default) in the subprocessconsultspawns. Contributing factors:gemini-clieagerly runsloadServerHierarchicalMemory/ JIT-context directory walk (graceful-fs wrappingfs.readdir— produces theArrayUnshift+AfterScanDirframes)advanced.autoConfigureMemoryis available in gemini-cli for this case but defaults OFFOnly gemini is affected because only gemini is shelled out and only gemini parses JSON from a buffered child stdout. Codex and Claude run in-process within consult's own Node process.
Code Location
packages/codev/src/commands/consult/index.ts:cmd = [config.cli, '--output-format', 'json', ...config.args, query];spawn(cmd[0], cmd.slice(1), { cwd, env, stdio: ['ignore', 'pipe', 'inherit'] })NODE_OPTIONSset on spawn envProposed Fix
Ideally both:
Option A — bump heap via NODE_OPTIONS (survives gemini's internal relaunch):
Option B — pipe prompt via stdin instead of argv:
Change spawn stdio to
['pipe', 'pipe', 'inherit'], write query tochild.stdin. gemini's non-interactivereadStdin()path accepts this. Avoids ARG_MAX and keeps V8 from holding the prompt twice.User-Side Workaround
Until a fix ships, users can set in
~/.gemini/settings.json:{ "advanced": { "autoConfigureMemory": true } }Related Upstream Issues
google-gemini/gemini-cli: #1740, #2993, #11321, #16271 (tracking), #22789
Impact
When CMAP is asked to review the largest, riskiest PRs (exactly when 3-way signal matters most), gemini fails silently and the user loses a full third of the consultation value.