gl_run — Code Mode

One tool. Runs sandboxed JavaScript. Calls every other Ganglia tool through typed wrappers. Collapses N sequential tool calls into a single MCP round trip.

Minimal example

gl_run({ code: "return gl.code.overview()" })

Real usage

gl_run({ code: `
  // Find dead code that has no tests — 1 call instead of 50
  const dead = gl.code.dead({});
  const coverage = dead.map(fn => gl.code.test_for({ name: fn.name }));
  return dead.filter((_, i) => coverage[i].tests.length === 0);
` })

Server runs the chain internally. Only the final filtered list comes back.

The `gl` namespace

All wrappers are synchronous from JS. Group by domain:

Namespace	Use for
`gl.code.*`	The 45 core graph tools — grep, get, dead, callers, impact, etc.
`gl.code.semantic_search / similar`	Embedding-based retrieval
`gl.doc.*`	External document index (markdown, CSV, etc.)
`gl.smart.*`	MemOS comprehension layer
`gl.deliberation.*`	Expert panel multi-perspective synthesis
`gl.call(name, args)`	Escape hatch for any registered MCP tool

Full typed API: ganglia://types/api.d.ts (MCP resource).

When to use

3+ sequential tool calls — the math overwhelmingly favors gl_run
Filter/transform/aggregate before returning — don’t force intermediate results through your context
Map-style operations — “for each X, compute Y” patterns

When NOT to use

Single-shot lookup (code_get, code_grep directly)
Exploratory — you need to see each result to decide the next step
The chain needs real await on network I/O (sandbox is sync only)

Sandbox guarantees

Limit	Value
Memory	256 MB
Time	30 s
Filesystem	blocked
Network	blocked
Process	blocked
Recursive `gl_run`	blocked

QuickJS embedded directly in the gl binary. No external runtime.

Typo recovery

Got the wrong method name? The Proxy wrapper shows you what’s available:

> gl.code.rgep({pattern: "foo"})
Error: Unknown method gl.code.rgep — available: annotate, build, callees,
  callers, ..., grep, hotspots, ...

And for gl.call(name):

> gl.call("code_rgep", {...})
Error: Unknown tool: code_rgep (did you mean `code_grep`?)

Benchmark

On a 462-file Rust project, 5 multi-step workflows:

Metric	Individual tools	`gl_run`	Reduction
Schema tokens	6,376	575	91%
Wire tokens	4,014	2,316	1.7×
Effective LLM context	170,424	9,191	18.9×

Run it on your own codebase: python3 scripts/benchmark_code_mode.py <path>