Declaring objectives

Closing gaps is reactive: an ask falls short, and you route the miss. Objectives are the standing version of the same instinct: write down what a pack must be able to answer before anyone asks it, and let the pack converge on its own. Each objective is a question the pack owns; its exercise loop asks it, records which tier answered, and keeps working until the answer is cheap.

This spends real money

structure, brainstorm, and examples each cost an LLM call, and a live objective drives background deep navigations at roughly a dollar apiece until it converges. Treat it like teach: show the user what you’re about to create and wait for a yes. --dry-run on structure exists for precisely that conversation.

Getting objectives in

The main entry point is paste-a-blob. People develop their questions somewhere else (a doc, a meeting note, a Claude session) and hand the whole thing over; one LLM call splits it into questions, extracts any inline hints, and topics each one.

edisyl pack objectives structure growth-metrics -f questions.md --dry-run   # preview — nothing is written
edisyl pack objectives structure growth-metrics -f questions.md            # create them
pbpaste | edisyl pack objectives structure growth-metrics
edisyl pack objectives structure growth-metrics --text "how many swaps last week? and by chain?"

When the user doesn’t yet know what to ask, work from the data instead:

edisyl pack objectives examples growth-metrics --limit 10          # drafts sample questions, creates nothing
edisyl pack objectives brainstorm growth-metrics --seed "revenue and churn"   # a chat session proposing candidates

examples reads one data source’s catalog and returns a menu to pick from. brainstorm opens a chat session where an agent proposes candidates against that schema. Neither creates anything; approve what survives with add:

edisyl pack objectives add growth-metrics --question "How many active wallets last week?" --topic activity \
  --hint "active = a swap in 30d; use the wallet_daily table"

A hint is human knowledge the pack can’t guess, the same instinct as teach, aimed at one question.

Watching them converge

edisyl pack objectives list growth-metrics                    # every objective: status, asks, last tier
edisyl pack objectives list growth-metrics --status pending   # what hasn't landed yet
edisyl pack objectives asks <objective-id>                    # the timeline for one

asks is the one to read: every attempt the pack made at that question, oldest first.

ASKED               STATUS     TIER      PLANE     DURATION  IN       OUT      COST
2026-08-13 21:28:50 resolved   deep      data      102.6s    647252   11039    $0.933
2026-08-13 21:43:32 resolved   rank      data      881ms     —        —        —

That is convergence in two rows. The first attempt went to the deep tier, a full agentic navigation costing 102 seconds and 93 cents. Fifteen minutes later the same question came back from rank in under a second, because the pack had learned the route and cached it. That’s the whole payoff of declaring objectives: you pay for the walk once, and every later ask rides it.

Caution

A — in the ledger means nothing was recorded, never that the ask was free. Every ask pays for the ask layer’s own classification, so read a dash as unknown.

Editing and retiring

edisyl pack objectives get <objective-id>
edisyl pack objectives update <objective-id> --hint "<better steering>"
edisyl pack objectives archive <objective-id>    # soft delete; stops the exercise loop

An objective’s status stays pending until answerable, then answered. blocked means the pack gave up; read asks for the reason beneath the table, or take blockedReason from get <objective-id> -j. Archive a bad objective rather than leaving it pending: while it’s live it keeps buying deep navigations.

See pack objectives for the full command reference.