.forge/wiki/research/gl0w-swarm-feedback.md 1.9 KiB 43 lines Raw

gl0w-swarm feedback on the f0rge skill (first external read)

Date: 2026-09-08. Source: the gl0w project's swarm (different, larger models) reading f0rge --skill cold, during migration. Tracked in issue #23.

What worked (verbatim signal)

tree, frontmatter, never delete, --jq / "never parse prose" — used immediately. Philosophy lines are sound.

Gaps found (procedural / git-bridge layer, not philosophy)

  1. (retracted by user — user mistype, not a skill fault) binary name "forge" vs "f0rge".
  2. Nothing says the forge needs creating: f0rge init.
  3. Tree line omits ci.env, releases/, merges/, labels.md, milestones.md — formats unknown without ls + spec/.
  4. THE big one: the contract never says .forge/ is committed and travels in git; PRs-as-branches, worktrees, branch scrub all invisible.
  5. Merge semantics invisible: who resolves conflicts; "scrubbed" vs "never delete" tension.
  6. CI execution model fuzzy: does f0rge ci run steps or record them? (It runs them; orchestrator was also fuzzy — real hole.)
  7. "archive moves folder" — to WHERE? .forge/archive/ never named.
  8. labels.md/milestones.md formats opaque.

Orchestrator triage (adopted 1:1 in #23)

Keep: git bridge (MUST), ci execution (MUST), merge+never-delete clarity (no carve-out — merged PR folder STAYS state:merged, only the branch dies), archive destination, init bootstrap, spec/ format pointer. Cut: full skeleton enumeration (discoverable), all repo-specific conventions (gitignores, AGENTS.md placement, migration scope).

Budget decision

Cap stays 900B/30ln — smallest comprehensive skill, start as we mean to go on. Compress by priority if drafting exceeds it.

Also learned

view raw