Back to Prisma

Retros — agent-native

projects/agent-native/retros.md

7.9.0-dev.171.6 KB
Original Source

Retros — agent-native

2026-07-03 — External-tool layout assumed from its docs, falsified live (S1 D3)

Trigger: dispatch-failure (I12 stop in D3 R1 of slice init-skill-install): the live prisma init run could not satisfy the DoD's .claude/skills/ condition.

What happened: [email protected] with multiple --agent values writes only the universal .agents/skills/ tree; the per-agent symlinks its own plan output promises are never created (upstream bug). The slice spec — and the D2 summary lines built on it — asserted a layout that did not exist on disk. Resolved by amending the spec to pass --copy (decision D13); one extra round.

Root cause: the slice-spec grounding step covered our codebase thoroughly but took the external CLI's behavior on faith (its README, its own progress output, and prisma-next's described outcome). No empirical probe of skills add was run until the final verification dispatch — the slice's only contact with external reality was scheduled last, so a wrong assumption could not surface any earlier. The proximate reading ("upstream bug") is true but not actionable; the actionable root is the spec-time probing gap.

Landing surface(s):

  • Project-context: drive/spec/README.md § Known constraints & gaps — probe external-tool behavior empirically at spec time, or make the probe the first dispatch (dated entry).
  • Candidate canonical amendment (for the operator, not applied): drive-specify-slice § Step 3 "Ground in the codebase" could name external-system probing explicitly; recorded here rather than editing canonical skills unattended.