Keep lean -p as the one-shot default (measured; KIV the graph surface)

[justrach]

KIV. Lean -p (the default one-shot) stays the default. Do not leave --no-lean on for one-shot coding evals. The named DAG (workflow) is still the right tool when work actually splits — use it there, once the split is large enough to amortize a worker prefix.

This is a parking note, not an implementation ticket.

What we measured

Graph here means the tool-call graph: which tools are in the catalog, whether workflow is advertised, and how a turn batches calls. runTools already fans a turn's tool_calls across the Io pool. Lean keeps eight tools and drops workflow; full catalog is ~4.3k extra schema tokens per model turn.

Three arms on graff-evals, grok-4.6, 1 rep:

arm	catalog	nudge
graff	lean (default -p --yolo)	none
graff-full	--no-lean	none
graff-graph	--no-lean	short stop-rule: if work splits, one workflow (phase 1 = parallel task per piece, phase 2 = one verify); sequential work stays on the root

Sequential suite (12 tasks)

arm	pass	wall	in tokens	out tokens
graff (lean)	12/12	152.2s	160813	6300
graff-full	12/12	168.2s	337775	6945

Same pass rate. Full catalog spent 2.10× input and ~10% more wall. exact-reply (no tools): 4086 vs 8338. The stop rule predicted this suite would lose, and it did.

Fan-out slice (2 tasks)

Independent files that can split: split-three-bugs (three one-line module bugs + a verify script) and split-three-csvs (three ledgers → one totals file).

arm	pass	wall	in tokens	out tokens
graff (lean)	2/2	30.2s	31958	1152
graff-full	2/2	32.4s	64629	1355
graff-graph	2/2	49.5s	91644	2666

task	arm	wall	in	calls	used workflow?
split-three-bugs	graff	17.46s	18511	4	no
split-three-bugs	graff-full	20.51s	38128	4	no
split-three-bugs	graff-graph	34.99s	64099	14	yes
split-three-csvs	graff	12.75s	13447	3	no
split-three-csvs	graff-full	11.86s	26501	3	no
split-three-csvs	graff-graph	14.49s	27545	3	no

What the traces show

• Lean already fans out inside a turn. Bugs: batched reads, then three edit_file, then the verify command. CSVs: batched reads, then one write. No DAG needed.
• --no-lean without a nudge never called workflow. It swapped native reads for an installed MCP batch-read and paid ~2× input for the same call count. Catalog ≠ graph use.
• The nudge produced a real diamond on bugs: one workflow, phase 1 = three parallel fixers, phase 2 = one verify. That is the intended shape.
• The diamond still lost. Each worker opened ~25k context (lean's whole run was 18.5k). 14 API calls vs 4. Wall 35s vs 17s. The fixers overlapped, but spawn + per-worker prefix + a verify agent lost to three one-line edits on the root.
• CSVs correctly skipped the DAG even with the nudge. Stop rule held; cost tracked the fat-catalog arm.

Decision

Keep lean as the one-shot default. Use workflow when work actually splits and the pieces are large enough that the root would otherwise serialize fat reads — or when workers can inherit the lean catalog instead of the full one.

Follow-up (only if the sign might flip)

• Independent pieces large enough that the root cannot cheaply batch them.
• Workers that inherit lean rather than the full catalog.
• Re-run the fan-out slice after either change.

Do not prepend third-party skill dumps or knowledge-graph prompts. That is not this harness.

[justrach]

The decision this parking note recorded is enacted on main: lean is the -p one-shot default with --no-lean as the opt-out (args.zig, session_start.zig, cli.zig help), and the v0.0.275-277 RLM arc (ADR 0029) was built on that default. Keeping the KIV condition in mind: if workers start inheriting lean, re-run the fan-out slice measurement and file fresh if the sign flips.