Dedup near-identical claude-code + codex hivemind-goals SKILL.md via generator

Start by checking package.json build/pretest scripts, then compare current claude-code/skills/hivemind-goals/SKILL.md and codex/skills/hivemind-goals/SKILL.md. Next inspect src/skills/hivemind-goals/ and scripts/build-skills.mjs for how the shared body and agents.json map to emitted files. Run npm run build and node scripts/build-skills.mjs --check; completion means the two Path-A SKILL files are regenerated from one template with matching output while hermes/... and openclaw/... remain untouched.

Context

PR #193 ships the goals + KPI cross-agent feature. Each of the four plugin-enabled agents (claude-code / codex / hermes / openclaw) has its own skills/hivemind-goals/SKILL.md because each runtime's skill loader reads from its own per-agent directory.

Three of the four files are genuinely different (different write paths require different bodies — Path A vs Path B vs Path C, see PR #193 body). The fourth pair, however, is near-duplicated:

claude-code/skills/hivemind-goals/SKILL.md   ── 6249 bytes
codex/skills/hivemind-goals/SKILL.md         ── 6233 bytes

diff between the two:

4c4
< allowed-tools: Read Write Edit Bash    # claude-code
> allowed-tools: Bash                    # codex

Body is bit-identical (~6.2 kB of workflow / path conventions / examples). They diverge only on the allowed-tools frontmatter line because the two runtimes expose different tool name namespaces for the same Path-A workflow.

Surfaced by @kaghni in #193 (comment).

Problem

Editing the workflow now requires two synchronized edits. The body has already drifted once during PR #193 development (caught at review). As the feature evolves (new statuses, KPI shape changes, new examples) the drift risk compounds.

The other agents (hermes/, openclaw/) are not part of this problem — their bodies are intentionally different (CLI-only warning, registered-tool semantics).

Proposed approach

Single shared body source + per-agent frontmatter, emitted by a small build step:

src/skills/hivemind-goals/
  body.md                              # the 6 kB workflow body (single source of truth)
  agents.json                          # { "claude-code": "Read Write Edit Bash", "codex": "Bash" }

scripts/build-skills.mjs               # reads body.md + agents.json,
                                       # emits claude-code/.../SKILL.md and codex/.../SKILL.md
                                       # runs in `npm run build` (pre-bundle), idempotent

Acceptance:

• npm run build regenerates both claude-code/skills/hivemind-goals/SKILL.md and codex/skills/hivemind-goals/SKILL.md from the single template.
• CI guard (or pretest step) verifies the two generated files are in sync with the template (diff against <(node scripts/build-skills.mjs --check) exits zero).
• hermes/skills/hivemind-goals/SKILL.md and openclaw/skills/hivemind-goals/SKILL.md are NOT touched by this dedup — they remain hand-authored variants because their bodies are substantively different (Path B / Path C semantics).
• One additional agents.json entry per future Path-A runtime if more come online (e.g. if cursor ever exposes a Write-tool intercept it joins the generator instead of hand-authoring).

Out of scope (explicit non-goals)

• Cross-runtime dedup of the SessionStart-injected text — that already lives at src/hooks/shared/goals-instructions.ts as the single source of truth (GOALS_INSTRUCTIONS + GOALS_INSTRUCTIONS_CLI).
• Merging hermes / openclaw SKILL.md into a shared body — they describe genuinely different workflows (CLI vs tool calls). The hermes one carries a hard "do not use write_file" warning that is incorrect for VFS-enabled agents and vice versa.
• Other skills (hivemind-memory, hivemind-login, etc.) — out of scope; if they have similar near-duplicate pairs they get the same treatment in separate PRs.

Effort estimate

~1 hour. Pure mechanical refactor — body content stays byte-identical; only the build step is new.