dont-reopen-converged-decisions

name: dont-reopen-converged-decisions
description: Once the team has converged AND committed on a functionally-fine choice, freeze it — don't re-open to optimize a marginal detail. A coordinator's flip-flop costs more thrash (reversals, blocked agents, eroded trust) than the marginal gain.
metadata:
  type: feedback

On 2026-06-17 the app-emit auth wire (for the events/subscriptions feature) churned through ~6 reversals, and a large chunk of that was **my own flip-flop**. I pinned the team to keep `45d414d2 + ed1b8ef2` (the `AWEB-App DIDKey` Authorization marker), then — chasing a marginal forward-compat benefit — sent a "DEFINITIVE 5-answer lock" switching to plain `DIDKey`, then **retracted it minutes later** when I saw the team had already reverted and locked on `AWEB-App DIDKey`. The contradiction propagated across two coordinators (aw-coordinator latched onto my plain-DIDKey chat as the source of truth while aweb-developer had my keep-AWEB-App mail) and **blocked** aweb-developer + aw-developer until I corrected it.

**Why this was wrong:** both options were functionally identical — same substance (no team cert, v2-family signed payload, manifest-keyed verification, app-id/key-id headers); only the Authorization *token string* differed. The committed state was already `AWEB-App DIDKey`, so switching meant re-churning conformance vectors + re-amending the impl for a gain that could ride the later durable step (`aaaj.6`, where the wire changes anyway). The thrash cost vastly exceeded the gain.

**How to apply:**

• Once the team has **converged and committed** on a functionally-fine choice,

**freeze it.** Don't re-open to optimize a marginal detail.

• If a refinement is genuinely better but non-blocking, **log it as a

future/durable step** — don't force it mid-stream.

• A definitive-sounding "lock" you retract minutes later is **worse than no

lock**: make the call once, then defer to the stable convergence.

• Push energy to the real remaining work (here: the byte-parity conformance

vectors), not the bikeshed (here: a token string).

• When two channels carry different "final" answers from you, the newest

*correction* wins — but the real fix is not creating the contradiction.

See [[lead-drive-to-completion]].