feat: plan-first turn for multi-part questions

[kiyeonjeon21]

Why

Asked for revenue and ticket counts per customer, the agent wrote one query joining both children of the same parent and silently multiplied the rows - returning 3,405 for a customer whose revenue is 1,702.50, because that customer has two tickets.

The grain warning added in ROADMAP 6.2 was already in the context for that exact entity and did not prevent it: the agent wrote one confident query in a single step. So a passive warning was demonstrably not enough, which is the argument for an active turn.

What

A planning turn with no tools offered, so the model must commit to an approach in writing before it can act. It fires only on questions asking for more than one thing.

The trigger is deliberately narrow - a false negative costs nothing, a false positive spends a turn on every question. A bare "and" would have fired on "give the customer name and the amount", which is one quantity in two columns and already answers correctly.

Write-verb questions are excluded outright. Measured: planning them turned a wrong number into a blank refusal - the over-refusal failure the verification pass exists to prevent. Their failure was never decomposition (the agent already separates the parts), so safety stays verification's job.

Measured

On the one case in either suite that triggers it:

	passing runs
without planning	1/8
with planning	8/8

(a --repeat 5 run plus a full-suite --repeat 3 run each way)

Suite level, stated honestly: the hard A/B moved +30.6 to +33.3 (grounded 31/36 to 33/36). That is close to one noise unit, and necessarily so - planning fires on 1 of 12 cases, so the most it could ever move the aggregate is 8.3 points. The case-level result is strong; the suite-level result is weak. Those are different claims and both are recorded.

Two bugs the measurement caught

Neither was visible to the scripted tests, because they do not validate message ordering or force the model to act:

1. The transcript ended on an assistant turn, which the API rejects outright - both target cases errored on the first real run.
2. The agent finalized on the plan without running anything in 2 of 5 runs, until the handoff message said plainly that the plan is not the answer.

Verification

• npm run check exit 0; npm run test:cli 147 pass / 0 fail
• A test asserts a non-triggering question takes a byte-identical path, so the feature is provably free when it does not apply
• --no-plan disables it; the flag flows through ask and both eval suites, and is recorded in the report config

ROADMAP step 7.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CKZ9UVgw2Mhiu6T5YAG7k6