docs: skill inventory and classification matrix for all 38 upstream skills

.opencode/state/decisions.md

@@ -55,3 +55,16 @@ recorded in `.opencode/state/classification.json` so re-inclusion is a data chan
 The only skill needing executable integration: dangerous-git blocking. Implemented as an opt-in
 plugin hook (`ctx.tool.hook("execute.before")` on shell, options `{ guardrails: true }`, default
 off) inside our plugin; the skill itself stays vendored verbatim for reference.
+
+## D8 — 2026-10-14 — Final agent roster: pocock, pocock-worker, pocock-reviewer. No researcher.
+
+Per-skill analysis (38 skills, `.opencode/state/classification.json`) confirms D5's core pair and
+the reviewer candidate: `code-review`'s defining mechanic is two parallel isolated axis reviewers
+with large self-contained prompts → `pocock-reviewer` (subagent, edit-deny, skill allow
+`code-review` only) is the right primitive; the caller aggregates verbatim per the skill.
+`pocock-researcher` is REJECTED: the 12-line `research` skill already is a background-agent brief;
+OpenCode's built-in `general` subagent executes it as written — a branded agent adds drift risk and
+zero capability. Per-phase agents (spec-writer, ticket-planner, debugger, ...) remain rejected:
+Matt's phases share one conversation context by design (grilling facts feed to-spec; retro mines the
+live session); isolating them would make the system worse. `triage` and `teach` agents rejected for
+the same reason (continuous human dialogue belongs in the primary conversation).