converge

Read two or more Convergence Maps produced by /dry — across repos, or across re-runs of one repo — and find what repeats. Clusters findings by SHAPE rather than by resource, so "a gateway exists and t

SKILL.md

GitHub

Converge

dry audits one repo. Converge reads several of its maps and answers the questions a single map cannot: what repeats, and what changed.

Both questions matter for a different reason. A target that repeats across repos is not a repo problem — it is a production problem, and fixing it once in a shared place beats fixing it N times. A target that survived a re-run is not a backlog item — it is evidence that naming it did not work, and the next attempt needs a different mechanism.

The one rule

Cluster by shape, never by resource. Two repos rarely share a table name or an API host, so resource-level clustering finds nothing across a fleet. What they share is the shape of the failure: "gateway exists, importers bypass it", "N rival gateways, none authoritative", "helper specified in a closed issue, never built". Shape is what ports.

When NOT to fire

  • Fewer than two maps → run dry first. Converge has nothing to cluster.
  • One repo, one point in time → that is dry, and it already answered.
  • Recurring defects rather than recurring structures → radar, per repo.
  • The maps are stale relative to their repos. Re-run dry rather than clustering fiction; say which maps you refused and why.

Step 1 — Collect the maps

Take map paths, or repo roots to search for convergence-*.md. For each, record the repo, the commit it was taken at, and its date. A map whose commit is not an ancestor of that repo's current HEAD is stale — either re-run dry or exclude it, and say which you did.

Two or more maps is the floor. Two maps of the same repo at different commits is the re-run case (Step 3); two maps of different repos is the fleet case (Step 2). Both at once is normal.

Step 2 — Cluster by shape

Read every target in every map and assign it a shape. These four cover most of what dry finds; add one only when a target genuinely fits none, and say so:

Shape The tell Why it recurs
Bypassed gateway A single canonical module exists; files that import it still reach the resource directly The gateway was built but never made mandatory. Nothing fails when you skip it.
Rival gateways Two or more canonical modules for one job, mutually unaware Two subsystems grew separately and neither was ever declared authoritative
Unbuilt helper A closed issue or doc specifies the shared helper; it does not exist Cleanup was tracked as a checklist and closed on the items that were easy
Post-fix regression New bypasses added after the convergence work landed No mechanism enforces the rule, so entropy resumes the day attention moves

A shape reported by two or more repos is a fleet finding. One repo with one instance is that repo's business — leave it in dry's map and do not re-report it here.

Step 3 — Diff the re-runs

Where two maps cover the same repo, classify every target from the older map:

  • Fixed — gone, and the gateway is adopted. Name it. An audit that only reports what is still broken teaches people the report never improves, and they stop reading it.
  • Unchanged — still present at the same sites. Record how long it has been open.
  • Worse — more bypasses than before. This is the finding that outranks everything else, because it means the previous fix had no enforcement behind it.
  • Moved — the resource changed but the shape survived. Still open; the fix addressed an instance rather than the cause.

Same line numbers across two commits is the strongest evidence in this skill. It proves the call site was not merely un-fixed but never touched — nobody has looked at it since the last audit.

Step 4 — Rank by what one rule retires

Fleet findings rank by how many repos and call sites a single enforceable rule would retire, not by how bad any one instance looks.

For each fleet finding, name the mechanism honestly, cheapest first:

  1. A CI rule — one grep with a non-increasing baseline. Prefer this. It is the only mechanism on this list that works while nobody is paying attention.
  2. A house rule in the repo's agent instructions — cheap, but only binds work that reads it.
  3. A shared library — real convergence, real cost. Justify it with the call-site count, and only after the same shape has appeared in three or more repos.

If a shape recurs and none of the three fits, say that plainly. "We keep doing this and there is no cheap way to stop" is a legitimate and useful finding.

Step 5 — Write the fleet verdict

Long output belongs in a file: converge-<fleet-or-date>.md, beside the maps. Chat gets the verdict line, the path, and the top fleet finding.

# Converge — <scope> · <date>

## Verdict
<N fleet findings across M repos; K targets fixed since last sweep; J got worse.>

## Fleet findings
### F1 — <shape> · <N repos> · one rule retires <M sites>
| Repo | Instance | Sites | Since |
|---|---|---|---|
**Mechanism:** <CI rule | house rule | shared library> — <the exact rule>
**Baseline:** <the measured count the rule must not exceed>

## Since last sweep
| Repo | Target | Was | Now | Verdict |
|---|---|---|---|---|
<fixed / unchanged / worse / moved — fixed rows first>

## Not a fleet finding
<single-repo targets, named so they are not re-clustered next time>

## Maps read
| Map | Repo | Commit | Date | Stale? |

Fixed rows go first. The section exists to show the work moves, and a fleet report that opens on failure gets read once.

Honesty rules

  • Every instance cites the map it came from and the file:line that map cited. Converge adds no new claims about code — it clusters claims dry already verified. If you need a fact no map contains, re-run dry; do not grep for it here and present it as a fleet finding.
  • Do not upgrade a single-repo target to a fleet finding to reach a rounder number.
  • A shape appearing in every repo may be a house style rather than a defect. Ask whether it costs anything before ranking it, and say so if the answer is no.
  • Zero fleet findings across several maps is a real and good result. Report it as one.

Neighbors

  • dry — produces the maps this skill reads. Always runs first.
  • triangulate — sizes the convergence work; a shared library is a crossing change and usually an irreversible one.
  • radar — per-repo defect clustering. Converge is the structural counterpart across repos, and a defect radar keeps re-finding is often a shape converge already named.
  • recon — the trace underneath all of it.