Packet quality (measuring right-sizing) — B51

Maintainer doc. This is a contributor runbook, not end-user documentation.

A work packet is the exact context an agent needs to execute one task (intent, acceptance criteria, scope, non-goals, verification). Packet quality is a first-class, measured workstream, not an afterthought: a tight, right-sized packet lets a fast local model skip exploration and cuts tokens-to-completion, which directly buys capacity (the binding constraint on a flat-rate/local fleet).

Right-sizing

The renderer already routes each task to a lightweight or full packet by score (anvil.context.packets.is_lightweight — a task is lightweight when both its complexity and blast_radius are low). The lightweight variant trims the non-essential sections (e.g. the evidence checklist is sliced to the essential item) while keeping the load-bearing ones (goal, acceptance criteria, scope, constraints/non-goals, verification). The full packet keeps everything for riskier or more complex tasks.

Measuring the reduction

anvil.context.packet_metrics quantifies the token reduction deterministically:

# Aggregate report over a project's backlog (JSON to stdout):
cd bin && uv run python -m anvil.context.packet_metrics <state_dir>

It reports, per task and in aggregate: the full vs lightweight token counts (a whitespace-split proxy — absolute numbers approximate, the relative reduction stable), which variant each task is routed to, and the as_routed_savings_pct the fast-lane already delivers vs sending everything full.

From Python:

from anvil.context.packet_metrics import measure_backlog
report = measure_backlog(backend.list_tasks())

Scope: token reduction here, success-rate lift in the bake-off

This harness measures the token reduction half of packet quality — the part that is measurable without a live model. The complementary local-model success-rate lift (does the right-sized packet let the local model complete the task correctly more often?) requires a live run and is measured in the B50 two-week bake-off, which feeds these results into its note. Together they turn "fast dumb work" into "fast and sufficient."