Prime

Prime

Produces a falsifiable orientation for the tickets in this worktree, then stops. No plan, no code — the point is to know what is true before proposing anything.

This skill does not restate the rules

CLAUDE.md and tests/CLAUDE.md are already loaded in every session. They own the architecture patterns, the test-integrity policy, the DRY invariant and the spec-grounding gate. Do not summarise them here or in your output.

A second copy of a rule drifts from the first, and drifted guidance in this repo has repeatedly sent people at the wrong thing: a stale comment blaming "generic-step shadowing" pointed at a bug fixed weeks earlier and cost about a week; a ledger mis-cited its own graduation three separate times; a memory note asserting a macOS/Linux duplication skew caused a legitimate ratchet improvement to be reverted three times before anyone measured it.

So: rules live in CLAUDE.md. This skill produces only what is ticket-specific and dynamic — things that change per ticket and can be checked.

Steps

1. Read the tickets

cat TICKETS
gh issue view <n> --repo prebid/salesagent --comments

Read the comments, not just the body. In this repo the sharpest constraint is often in a comment — and several issues carry a Measured on: <branch> @ <sha> header. If it does, grep that tree, not yours. Four issues were once declared false-premise because someone grepped only the checked-out branch; the symbols were real, on the branch the author measured. Before writing "X does not exist", run git log --all -S"X" --oneline.

2. Ground against the pinned AdCP version

Resolve the pin, do not assume it:

grep -n 'adcp==' pyproject.toml
grep -n 'targets AdCP spec' docs/adcp-spec-version.md

Then read the spec at that version — the worktree at ~/projects/adcp is usually checked out somewhere else, so always go through git show:

cd ~/projects/adcp && git show v<PIN>:dist/schemas/<PIN>/<area>/<file>.json
cd ~/projects/adcp && git show v<PIN>:dist/docs/<PIN>/building/implementation/<tool>.mdx

For protocol behaviour, record the exact sentence that mandates it and whether a conformance storyboard grades it — check dist/compliance/<PIN>/ and say "ungraded" if nothing does. The installed SDK is a cross-check, never the authority.

3. Measure the BDD coverage that exists

This is the part nothing else tells you. For the ticket's use case:

ls tests/bdd/features/BR-UC-*<area>*.feature
grep -c '^  Scenario' tests/bdd/features/BR-UC-<n>-*.feature
grep -n '@T-UC-<n>' tests/bdd/features/BR-UC-<n>-*.feature | head -20
grep -n 'T-UC-<n>' tests/bdd/conftest.py | head          # xfail routing
grep -c '^tests/bdd/' tests/bdd/e2e_rest_known_failures.txt

Report numbers: how many scenarios, how many currently xfailed and why, how many sit in the e2e_rest ledger. "N scenarios, M xfailed, 2 ledgered" is checkable; "there is good coverage" is not.

If the ticket implies graduating an xfail, read .claude/rules/workflows/xpass-graduation.md now, in full, before planning.

4. Note who will grade the work

Name them and point at their definitions — do not summarise their practices, for the drift reason above:

ls .claude/agents/ 2>/dev/null

The code-review skill fans out reviewers covering DRY, testing, layering, consistency, AdCP grounding, ratcheting allowlists and BDD grounding. Read the relevant one's definition before writing code it will judge.

5. Note the tools that are not obvious

• .agent-index/ — generated .pyi stubs. Read these before scanning the tree; a full-tree grep for a symbol they already index is wasted work.
• ast-grep for structural queries; plain grep for text.
• saci test {unit,integration,bdd} <path|-k> runs slices on the CI box. saci run is the full gate. saci run silently attaches to an in-flight run for the same worktree — if you have just committed, check the box actually has your change before trusting a green result.
• make quality is the local gate. It rewrites .duplication-baseline when duplication genuinely drops; that is the ratchet working, not drift.

6. Write the orientation down

Into the ticket's bead if one exists, else .claude/research/<ticket>.md:

• the spec citation and whether it is graded
• the coverage numbers from step 3
• which files the change will touch
• the Core Invariant — one sentence naming what must stay true
• open questions that would change the approach

Then stop and report. Do not plan or implement until the orientation is agreed.

Anti-patterns

• Restating CLAUDE.md instead of measuring this ticket
• Asserting a symbol is absent from one branch's src/
• Reporting prose where a count would be checkable
• Planning inside this skill — orientation first, agreement, then a plan